xref: /titanic_50/usr/src/uts/sun4u/cpu/opl_olympus.c (revision 4e5b757fbcf21077677360be274461dcd9064106)
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 2007 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 <vm/seg_kmem.h>
44 #include <sys/cpuvar.h>
45 #include <sys/opl_olympus_regs.h>
46 #include <sys/opl_module.h>
47 #include <sys/async.h>
48 #include <sys/cmn_err.h>
49 #include <sys/debug.h>
50 #include <sys/dditypes.h>
51 #include <sys/cpu_module.h>
52 #include <sys/sysmacros.h>
53 #include <sys/intreg.h>
54 #include <sys/clock.h>
55 #include <sys/platform_module.h>
56 #include <sys/ontrap.h>
57 #include <sys/panic.h>
58 #include <sys/memlist.h>
59 #include <sys/ndifm.h>
60 #include <sys/ddifm.h>
61 #include <sys/fm/protocol.h>
62 #include <sys/fm/util.h>
63 #include <sys/fm/cpu/SPARC64-VI.h>
64 #include <sys/dtrace.h>
65 #include <sys/watchpoint.h>
66 #include <sys/promif.h>
67 
68 /*
69  * Internal functions.
70  */
71 static int cpu_sync_log_err(void *flt);
72 static void cpu_payload_add_aflt(struct async_flt *, nvlist_t *, nvlist_t *);
73 static void opl_cpu_sync_error(struct regs *, ulong_t, ulong_t, uint_t, uint_t);
74 static int  cpu_flt_in_memory(opl_async_flt_t *, uint64_t);
75 
76 /*
77  * Error counters resetting interval.
78  */
79 static int opl_async_check_interval = 60;		/* 1 min */
80 
81 uint_t cpu_impl_dual_pgsz = 1;
82 
83 /*
84  * PA[22:0] represent Displacement in Jupiter
85  * configuration space.
86  */
87 uint_t	root_phys_addr_lo_mask = 0x7fffffu;
88 
89 /*
90  * set in /etc/system to control logging of user BERR/TO's
91  */
92 int cpu_berr_to_verbose = 0;
93 
94 static int min_ecache_size;
95 static uint_t priv_hcl_1;
96 static uint_t priv_hcl_2;
97 static uint_t priv_hcl_4;
98 static uint_t priv_hcl_8;
99 
100 /*
101  * Olympus error log
102  */
103 static opl_errlog_t	*opl_err_log;
104 
105 /*
106  * UE is classified into four classes (MEM, CHANNEL, CPU, PATH).
107  * No any other ecc_type_info insertion is allowed in between the following
108  * four UE classess.
109  */
110 ecc_type_to_info_t ecc_type_to_info[] = {
111 	SFSR_UE,	"UE ",	(OPL_ECC_SYNC_TRAP), OPL_CPU_SYNC_UE,
112 	"Uncorrectable ECC",  FM_EREPORT_PAYLOAD_SYNC,
113 	FM_EREPORT_CPU_UE_MEM,
114 	SFSR_UE,	"UE ",	(OPL_ECC_SYNC_TRAP), OPL_CPU_SYNC_UE,
115 	"Uncorrectable ECC",  FM_EREPORT_PAYLOAD_SYNC,
116 	FM_EREPORT_CPU_UE_CHANNEL,
117 	SFSR_UE,	"UE ",	(OPL_ECC_SYNC_TRAP), OPL_CPU_SYNC_UE,
118 	"Uncorrectable ECC",  FM_EREPORT_PAYLOAD_SYNC,
119 	FM_EREPORT_CPU_UE_CPU,
120 	SFSR_UE,	"UE ",	(OPL_ECC_SYNC_TRAP), OPL_CPU_SYNC_UE,
121 	"Uncorrectable ECC",  FM_EREPORT_PAYLOAD_SYNC,
122 	FM_EREPORT_CPU_UE_PATH,
123 	SFSR_BERR, "BERR ", (OPL_ECC_SYNC_TRAP), OPL_CPU_SYNC_OTHERS,
124 	"Bus Error",  FM_EREPORT_PAYLOAD_SYNC,
125 	FM_EREPORT_CPU_BERR,
126 	SFSR_TO, "TO ", (OPL_ECC_SYNC_TRAP), OPL_CPU_SYNC_OTHERS,
127 	"Bus Timeout",  FM_EREPORT_PAYLOAD_SYNC,
128 	FM_EREPORT_CPU_BTO,
129 	SFSR_TLB_MUL, "TLB_MUL ", (OPL_ECC_SYNC_TRAP), OPL_CPU_SYNC_OTHERS,
130 	"TLB MultiHit",  FM_EREPORT_PAYLOAD_SYNC,
131 	FM_EREPORT_CPU_MTLB,
132 	SFSR_TLB_PRT, "TLB_PRT ", (OPL_ECC_SYNC_TRAP), OPL_CPU_SYNC_OTHERS,
133 	"TLB Parity",  FM_EREPORT_PAYLOAD_SYNC,
134 	FM_EREPORT_CPU_TLBP,
135 
136 	UGESR_IAUG_CRE, "IAUG_CRE", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
137 	"IAUG CRE",  FM_EREPORT_PAYLOAD_URGENT,
138 	FM_EREPORT_CPU_CRE,
139 	UGESR_IAUG_TSBCTXT, "IAUG_TSBCTXT",
140 	OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
141 	"IAUG TSBCTXT",  FM_EREPORT_PAYLOAD_URGENT,
142 	FM_EREPORT_CPU_TSBCTX,
143 	UGESR_IUG_TSBP, "IUG_TSBP", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
144 	"IUG TSBP",  FM_EREPORT_PAYLOAD_URGENT,
145 	FM_EREPORT_CPU_TSBP,
146 	UGESR_IUG_PSTATE, "IUG_PSTATE", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
147 	"IUG PSTATE",  FM_EREPORT_PAYLOAD_URGENT,
148 	FM_EREPORT_CPU_PSTATE,
149 	UGESR_IUG_TSTATE, "IUG_TSTATE", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
150 	"IUG TSTATE",  FM_EREPORT_PAYLOAD_URGENT,
151 	FM_EREPORT_CPU_TSTATE,
152 	UGESR_IUG_F, "IUG_F", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
153 	"IUG FREG",  FM_EREPORT_PAYLOAD_URGENT,
154 	FM_EREPORT_CPU_IUG_F,
155 	UGESR_IUG_R, "IUG_R", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
156 	"IUG RREG",  FM_EREPORT_PAYLOAD_URGENT,
157 	FM_EREPORT_CPU_IUG_R,
158 	UGESR_AUG_SDC, "AUG_SDC", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
159 	"AUG SDC",  FM_EREPORT_PAYLOAD_URGENT,
160 	FM_EREPORT_CPU_SDC,
161 	UGESR_IUG_WDT, "IUG_WDT", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
162 	"IUG WDT",  FM_EREPORT_PAYLOAD_URGENT,
163 	FM_EREPORT_CPU_WDT,
164 	UGESR_IUG_DTLB, "IUG_DTLB", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
165 	"IUG DTLB",  FM_EREPORT_PAYLOAD_URGENT,
166 	FM_EREPORT_CPU_DTLB,
167 	UGESR_IUG_ITLB, "IUG_ITLB", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
168 	"IUG ITLB",  FM_EREPORT_PAYLOAD_URGENT,
169 	FM_EREPORT_CPU_ITLB,
170 	UGESR_IUG_COREERR, "IUG_COREERR",
171 	OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
172 	"IUG COREERR",  FM_EREPORT_PAYLOAD_URGENT,
173 	FM_EREPORT_CPU_CORE,
174 	UGESR_MULTI_DAE, "MULTI_DAE", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
175 	"MULTI DAE",  FM_EREPORT_PAYLOAD_URGENT,
176 	FM_EREPORT_CPU_DAE,
177 	UGESR_MULTI_IAE, "MULTI_IAE", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
178 	"MULTI IAE",  FM_EREPORT_PAYLOAD_URGENT,
179 	FM_EREPORT_CPU_IAE,
180 	UGESR_MULTI_UGE, "MULTI_UGE", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
181 	"MULTI UGE",  FM_EREPORT_PAYLOAD_URGENT,
182 	FM_EREPORT_CPU_UGE,
183 	0,		NULL,		0,		0,
184 	NULL,  0,	   0,
185 };
186 
187 int (*p2get_mem_info)(int synd_code, uint64_t paddr,
188 		uint64_t *mem_sizep, uint64_t *seg_sizep, uint64_t *bank_sizep,
189 		int *segsp, int *banksp, int *mcidp);
190 
191 
192 /*
193  * Setup trap handlers for 0xA, 0x32, 0x40 trap types.
194  */
195 void
196 cpu_init_trap(void)
197 {
198 	OPL_SET_TRAP(tt0_iae, opl_serr_instr);
199 	OPL_SET_TRAP(tt1_iae, opl_serr_instr);
200 	OPL_SET_TRAP(tt0_dae, opl_serr_instr);
201 	OPL_SET_TRAP(tt1_dae, opl_serr_instr);
202 	OPL_SET_TRAP(tt0_asdat, opl_ugerr_instr);
203 	OPL_SET_TRAP(tt1_asdat, opl_ugerr_instr);
204 }
205 
206 static int
207 getintprop(pnode_t node, char *name, int deflt)
208 {
209 	int	value;
210 
211 	switch (prom_getproplen(node, name)) {
212 	case sizeof (int):
213 		(void) prom_getprop(node, name, (caddr_t)&value);
214 		break;
215 
216 	default:
217 		value = deflt;
218 		break;
219 	}
220 
221 	return (value);
222 }
223 
224 /*
225  * Set the magic constants of the implementation.
226  */
227 /*ARGSUSED*/
228 void
229 cpu_fiximp(pnode_t dnode)
230 {
231 	int i, a;
232 	extern int vac_size, vac_shift;
233 	extern uint_t vac_mask;
234 
235 	static struct {
236 		char	*name;
237 		int	*var;
238 		int	defval;
239 	} prop[] = {
240 		"l1-dcache-size", &dcache_size, OPL_DCACHE_SIZE,
241 		"l1-dcache-line-size", &dcache_linesize, OPL_DCACHE_LSIZE,
242 		"l1-icache-size", &icache_size, OPL_ICACHE_SIZE,
243 		"l1-icache-line-size", &icache_linesize, OPL_ICACHE_LSIZE,
244 		"l2-cache-size", &ecache_size, OPL_ECACHE_SIZE,
245 		"l2-cache-line-size", &ecache_alignsize, OPL_ECACHE_LSIZE,
246 		"l2-cache-associativity", &ecache_associativity, OPL_ECACHE_NWAY
247 	};
248 
249 	for (i = 0; i < sizeof (prop) / sizeof (prop[0]); i++)
250 		*prop[i].var = getintprop(dnode, prop[i].name, prop[i].defval);
251 
252 	ecache_setsize = ecache_size / ecache_associativity;
253 
254 	vac_size = OPL_VAC_SIZE;
255 	vac_mask = MMU_PAGEMASK & (vac_size - 1);
256 	i = 0; a = vac_size;
257 	while (a >>= 1)
258 		++i;
259 	vac_shift = i;
260 	shm_alignment = vac_size;
261 	vac = 1;
262 }
263 
264 #ifdef	OLYMPUS_C_REV_B_ERRATA_XCALL
265 /*
266  * Quick and dirty way to redefine locally in
267  * OPL the value of IDSR_BN_SETS to 31 instead
268  * of the standard 32 value. This is to workaround
269  * REV_B of Olympus_c processor's problem in handling
270  * more than 31 xcall broadcast.
271  */
272 #undef	IDSR_BN_SETS
273 #define	IDSR_BN_SETS    31
274 #endif	/* OLYMPUS_C_REV_B_ERRATA_XCALL */
275 
276 void
277 send_mondo_set(cpuset_t set)
278 {
279 	int lo, busy, nack, shipped = 0;
280 	uint16_t i, cpuids[IDSR_BN_SETS];
281 	uint64_t idsr, nackmask = 0, busymask, curnack, curbusy;
282 	uint64_t starttick, endtick, tick, lasttick;
283 #if (NCPU > IDSR_BN_SETS)
284 	int index = 0;
285 	int ncpuids = 0;
286 #endif
287 #ifdef	OLYMPUS_C_REV_A_ERRATA_XCALL
288 	int bn_sets = IDSR_BN_SETS;
289 	uint64_t ver;
290 
291 	ASSERT(NCPU > bn_sets);
292 #endif
293 
294 	ASSERT(!CPUSET_ISNULL(set));
295 	starttick = lasttick = gettick();
296 
297 #ifdef	OLYMPUS_C_REV_A_ERRATA_XCALL
298 	ver = ultra_getver();
299 	if (((ULTRA_VER_IMPL(ver)) == OLYMPUS_C_IMPL) &&
300 		((OLYMPUS_REV_MASK(ver)) == OLYMPUS_C_A))
301 		bn_sets = 1;
302 #endif
303 
304 #if (NCPU <= IDSR_BN_SETS)
305 	for (i = 0; i < NCPU; i++)
306 		if (CPU_IN_SET(set, i)) {
307 			shipit(i, shipped);
308 			nackmask |= IDSR_NACK_BIT(shipped);
309 			cpuids[shipped++] = i;
310 			CPUSET_DEL(set, i);
311 			if (CPUSET_ISNULL(set))
312 				break;
313 		}
314 	CPU_STATS_ADDQ(CPU, sys, xcalls, shipped);
315 #else
316 	for (i = 0; i < NCPU; i++)
317 		if (CPU_IN_SET(set, i)) {
318 			ncpuids++;
319 
320 			/*
321 			 * Ship only to the first (IDSR_BN_SETS) CPUs.  If we
322 			 * find we have shipped to more than (IDSR_BN_SETS)
323 			 * CPUs, set "index" to the highest numbered CPU in
324 			 * the set so we can ship to other CPUs a bit later on.
325 			 */
326 #ifdef	OLYMPUS_C_REV_A_ERRATA_XCALL
327 			if (shipped < bn_sets) {
328 #else
329 			if (shipped < IDSR_BN_SETS) {
330 #endif
331 				shipit(i, shipped);
332 				nackmask |= IDSR_NACK_BIT(shipped);
333 				cpuids[shipped++] = i;
334 				CPUSET_DEL(set, i);
335 				if (CPUSET_ISNULL(set))
336 					break;
337 			} else
338 				index = (int)i;
339 		}
340 
341 	CPU_STATS_ADDQ(CPU, sys, xcalls, ncpuids);
342 #endif
343 
344 	busymask = IDSR_NACK_TO_BUSY(nackmask);
345 	busy = nack = 0;
346 	endtick = starttick + xc_tick_limit;
347 	for (;;) {
348 		idsr = getidsr();
349 #if (NCPU <= IDSR_BN_SETS)
350 		if (idsr == 0)
351 			break;
352 #else
353 		if (idsr == 0 && shipped == ncpuids)
354 			break;
355 #endif
356 		tick = gettick();
357 		/*
358 		 * If there is a big jump between the current tick
359 		 * count and lasttick, we have probably hit a break
360 		 * point.  Adjust endtick accordingly to avoid panic.
361 		 */
362 		if (tick > (lasttick + xc_tick_jump_limit))
363 			endtick += (tick - lasttick);
364 		lasttick = tick;
365 		if (tick > endtick) {
366 			if (panic_quiesce)
367 				return;
368 			cmn_err(CE_CONT, "send mondo timeout "
369 				"[%d NACK %d BUSY]\nIDSR 0x%"
370 				"" PRIx64 "  cpuids:", nack, busy, idsr);
371 #ifdef	OLYMPUS_C_REV_A_ERRATA_XCALL
372 			for (i = 0; i < bn_sets; i++) {
373 #else
374 			for (i = 0; i < IDSR_BN_SETS; i++) {
375 #endif
376 				if (idsr & (IDSR_NACK_BIT(i) |
377 				    IDSR_BUSY_BIT(i))) {
378 					cmn_err(CE_CONT, " 0x%x",
379 						cpuids[i]);
380 				}
381 			}
382 			cmn_err(CE_CONT, "\n");
383 			cmn_err(CE_PANIC, "send_mondo_set: timeout");
384 		}
385 		curnack = idsr & nackmask;
386 		curbusy = idsr & busymask;
387 
388 #ifdef OLYMPUS_C_REV_B_ERRATA_XCALL
389 		/*
390 		 * Only proceed to send more xcalls if all the
391 		 * cpus in the previous IDSR_BN_SETS were completed.
392 		 */
393 		if (curbusy) {
394 			busy++;
395 			continue;
396 		}
397 #endif /* OLYMPUS_C_REV_B_ERRATA_XCALL */
398 
399 #if (NCPU > IDSR_BN_SETS)
400 		if (shipped < ncpuids) {
401 			uint64_t cpus_left;
402 			uint16_t next = (uint16_t)index;
403 
404 			cpus_left = ~(IDSR_NACK_TO_BUSY(curnack) | curbusy) &
405 			    busymask;
406 
407 			if (cpus_left) {
408 				do {
409 					/*
410 					 * Sequence through and ship to the
411 					 * remainder of the CPUs in the system
412 					 * (e.g. other than the first
413 					 * (IDSR_BN_SETS)) in reverse order.
414 					 */
415 					lo = lowbit(cpus_left) - 1;
416 					i = IDSR_BUSY_IDX(lo);
417 					shipit(next, i);
418 					shipped++;
419 					cpuids[i] = next;
420 
421 					/*
422 					 * If we've processed all the CPUs,
423 					 * exit the loop now and save
424 					 * instructions.
425 					 */
426 					if (shipped == ncpuids)
427 						break;
428 
429 					for ((index = ((int)next - 1));
430 						index >= 0; index--)
431 						if (CPU_IN_SET(set, index)) {
432 							next = (uint16_t)index;
433 							break;
434 						}
435 
436 					cpus_left &= ~(1ull << lo);
437 				} while (cpus_left);
438 				continue;
439 			}
440 		}
441 #endif
442 #ifndef	OLYMPUS_C_REV_B_ERRATA_XCALL
443 		if (curbusy) {
444 			busy++;
445 			continue;
446 		}
447 #endif	/* OLYMPUS_C_REV_B_ERRATA_XCALL */
448 #ifdef SEND_MONDO_STATS
449 		{
450 			int n = gettick() - starttick;
451 			if (n < 8192)
452 				x_nack_stimes[n >> 7]++;
453 		}
454 #endif
455 		while (gettick() < (tick + sys_clock_mhz))
456 			;
457 		do {
458 			lo = lowbit(curnack) - 1;
459 			i = IDSR_NACK_IDX(lo);
460 			shipit(cpuids[i], i);
461 			curnack &= ~(1ull << lo);
462 		} while (curnack);
463 		nack++;
464 		busy = 0;
465 	}
466 #ifdef SEND_MONDO_STATS
467 	{
468 		int n = gettick() - starttick;
469 		if (n < 8192)
470 			x_set_stimes[n >> 7]++;
471 		else
472 			x_set_ltimes[(n >> 13) & 0xf]++;
473 	}
474 	x_set_cpus[shipped]++;
475 #endif
476 }
477 
478 /*
479  * Cpu private initialization.
480  */
481 void
482 cpu_init_private(struct cpu *cp)
483 {
484 	if (!(IS_OLYMPUS_C(cpunodes[cp->cpu_id].implementation))) {
485 		cmn_err(CE_PANIC, "CPU%d Impl %d: Only SPARC64-VI is supported",
486 			cp->cpu_id, cpunodes[cp->cpu_id].implementation);
487 	}
488 
489 	adjust_hw_copy_limits(cpunodes[cp->cpu_id].ecache_size);
490 }
491 
492 void
493 cpu_setup(void)
494 {
495 	extern int at_flags;
496 	extern int disable_delay_tlb_flush, delay_tlb_flush;
497 	extern int cpc_has_overflow_intr;
498 	uint64_t cpu0_log;
499 	extern	 uint64_t opl_cpu0_err_log;
500 
501 	/*
502 	 * Initialize Error log Scratch register for error handling.
503 	 */
504 
505 	cpu0_log = va_to_pa(&opl_cpu0_err_log);
506 	opl_error_setup(cpu0_log);
507 
508 	/*
509 	 * Enable MMU translating multiple page sizes for
510 	 * sITLB and sDTLB.
511 	 */
512 	opl_mpg_enable();
513 
514 	/*
515 	 * Setup chip-specific trap handlers.
516 	 */
517 	cpu_init_trap();
518 
519 	cache |= (CACHE_VAC | CACHE_PTAG | CACHE_IOCOHERENT);
520 
521 	at_flags = EF_SPARC_32PLUS | EF_SPARC_SUN_US1 | EF_SPARC_SUN_US3;
522 
523 	/*
524 	 * Due to the number of entries in the fully-associative tlb
525 	 * this may have to be tuned lower than in spitfire.
526 	 */
527 	pp_slots = MIN(8, MAXPP_SLOTS);
528 
529 	/*
530 	 * Block stores do not invalidate all pages of the d$, pagecopy
531 	 * et. al. need virtual translations with virtual coloring taken
532 	 * into consideration.  prefetch/ldd will pollute the d$ on the
533 	 * load side.
534 	 */
535 	pp_consistent_coloring = PPAGE_STORE_VCOLORING | PPAGE_LOADS_POLLUTE;
536 
537 	if (use_page_coloring) {
538 		do_pg_coloring = 1;
539 		if (use_virtual_coloring)
540 			do_virtual_coloring = 1;
541 	}
542 
543 	isa_list =
544 	    "sparcv9+vis2 sparcv9+vis sparcv9 "
545 	    "sparcv8plus+vis2 sparcv8plus+vis sparcv8plus "
546 	    "sparcv8 sparcv8-fsmuld sparcv7 sparc";
547 
548 	cpu_hwcap_flags = AV_SPARC_VIS | AV_SPARC_VIS2 |
549 	    AV_SPARC_POPC | AV_SPARC_FMAF;
550 
551 	/*
552 	 * On SPARC64-VI, there's no hole in the virtual address space
553 	 */
554 	hole_start = hole_end = 0;
555 
556 	/*
557 	 * The kpm mapping window.
558 	 * kpm_size:
559 	 *	The size of a single kpm range.
560 	 *	The overall size will be: kpm_size * vac_colors.
561 	 * kpm_vbase:
562 	 *	The virtual start address of the kpm range within the kernel
563 	 *	virtual address space. kpm_vbase has to be kpm_size aligned.
564 	 */
565 	kpm_size = (size_t)(128ull * 1024 * 1024 * 1024 * 1024); /* 128TB */
566 	kpm_size_shift = 47;
567 	kpm_vbase = (caddr_t)0x8000000000000000ull; /* 8EB */
568 	kpm_smallpages = 1;
569 
570 	/*
571 	 * The traptrace code uses either %tick or %stick for
572 	 * timestamping.  We have %stick so we can use it.
573 	 */
574 	traptrace_use_stick = 1;
575 
576 	/*
577 	 * SPARC64-VI has a performance counter overflow interrupt
578 	 */
579 	cpc_has_overflow_intr = 1;
580 
581 	/*
582 	 * Use SPARC64-VI flush-all support
583 	 */
584 	if (!disable_delay_tlb_flush)
585 		delay_tlb_flush = 1;
586 
587 	/*
588 	 * Declare that this architecture/cpu combination does not support
589 	 * fpRAS.
590 	 */
591 	fpras_implemented = 0;
592 }
593 
594 /*
595  * Called by setcpudelay
596  */
597 void
598 cpu_init_tick_freq(void)
599 {
600 	/*
601 	 * For SPARC64-VI we want to use the system clock rate as
602 	 * the basis for low level timing, due to support of mixed
603 	 * speed CPUs and power managment.
604 	 */
605 	if (system_clock_freq == 0)
606 		cmn_err(CE_PANIC, "setcpudelay: invalid system_clock_freq");
607 
608 	sys_tick_freq = system_clock_freq;
609 }
610 
611 #ifdef SEND_MONDO_STATS
612 uint32_t x_one_stimes[64];
613 uint32_t x_one_ltimes[16];
614 uint32_t x_set_stimes[64];
615 uint32_t x_set_ltimes[16];
616 uint32_t x_set_cpus[NCPU];
617 uint32_t x_nack_stimes[64];
618 #endif
619 
620 /*
621  * Note: A version of this function is used by the debugger via the KDI,
622  * and must be kept in sync with this version.  Any changes made to this
623  * function to support new chips or to accomodate errata must also be included
624  * in the KDI-specific version.  See us3_kdi.c.
625  */
626 void
627 send_one_mondo(int cpuid)
628 {
629 	int busy, nack;
630 	uint64_t idsr, starttick, endtick, tick, lasttick;
631 	uint64_t busymask;
632 
633 	CPU_STATS_ADDQ(CPU, sys, xcalls, 1);
634 	starttick = lasttick = gettick();
635 	shipit(cpuid, 0);
636 	endtick = starttick + xc_tick_limit;
637 	busy = nack = 0;
638 	busymask = IDSR_BUSY;
639 	for (;;) {
640 		idsr = getidsr();
641 		if (idsr == 0)
642 			break;
643 
644 		tick = gettick();
645 		/*
646 		 * If there is a big jump between the current tick
647 		 * count and lasttick, we have probably hit a break
648 		 * point.  Adjust endtick accordingly to avoid panic.
649 		 */
650 		if (tick > (lasttick + xc_tick_jump_limit))
651 			endtick += (tick - lasttick);
652 		lasttick = tick;
653 		if (tick > endtick) {
654 			if (panic_quiesce)
655 				return;
656 			cmn_err(CE_PANIC, "send mondo timeout "
657 				"(target 0x%x) [%d NACK %d BUSY]",
658 					cpuid, nack, busy);
659 		}
660 
661 		if (idsr & busymask) {
662 			busy++;
663 			continue;
664 		}
665 		drv_usecwait(1);
666 		shipit(cpuid, 0);
667 		nack++;
668 		busy = 0;
669 	}
670 #ifdef SEND_MONDO_STATS
671 	{
672 		int n = gettick() - starttick;
673 		if (n < 8192)
674 			x_one_stimes[n >> 7]++;
675 		else
676 			x_one_ltimes[(n >> 13) & 0xf]++;
677 	}
678 #endif
679 }
680 
681 /*
682  * init_mmu_page_sizes is set to one after the bootup time initialization
683  * via mmu_init_mmu_page_sizes, to indicate that mmu_page_sizes has a
684  * valid value.
685  *
686  * mmu_disable_ism_large_pages and mmu_disable_large_pages are the mmu-specific
687  * versions of disable_ism_large_pages and disable_large_pages, and feed back
688  * into those two hat variables at hat initialization time.
689  *
690  */
691 int init_mmu_page_sizes = 0;
692 
693 static uint_t mmu_disable_large_pages = 0;
694 static uint_t mmu_disable_ism_large_pages = ((1 << TTE64K) |
695 	(1 << TTE512K) | (1 << TTE32M) | (1 << TTE256M));
696 static uint_t mmu_disable_auto_data_large_pages = ((1 << TTE64K) |
697 	(1 << TTE512K) | (1 << TTE32M) | (1 << TTE256M));
698 static uint_t mmu_disable_auto_text_large_pages = ((1 << TTE64K) |
699 	(1 << TTE512K));
700 
701 /*
702  * Re-initialize mmu_page_sizes and friends, for SPARC64-VI mmu support.
703  * Called during very early bootup from check_cpus_set().
704  * Can be called to verify that mmu_page_sizes are set up correctly.
705  *
706  * Set Olympus defaults. We do not use the function parameter.
707  */
708 /*ARGSUSED*/
709 int
710 mmu_init_mmu_page_sizes(int32_t not_used)
711 {
712 	if (!init_mmu_page_sizes) {
713 		mmu_page_sizes = MMU_PAGE_SIZES;
714 		mmu_hashcnt = MAX_HASHCNT;
715 		mmu_ism_pagesize = DEFAULT_ISM_PAGESIZE;
716 		mmu_exported_pagesize_mask = (1 << TTE8K) |
717 		    (1 << TTE64K) | (1 << TTE512K) | (1 << TTE4M) |
718 		    (1 << TTE32M) | (1 << TTE256M);
719 		init_mmu_page_sizes = 1;
720 		return (0);
721 	}
722 	return (1);
723 }
724 
725 /* SPARC64-VI worst case DTLB parameters */
726 #ifndef	LOCKED_DTLB_ENTRIES
727 #define	LOCKED_DTLB_ENTRIES	5	/* 2 user TSBs, 2 nucleus, + OBP */
728 #endif
729 #define	TOTAL_DTLB_ENTRIES	32
730 #define	AVAIL_32M_ENTRIES	0
731 #define	AVAIL_256M_ENTRIES	0
732 #define	AVAIL_DTLB_ENTRIES	(TOTAL_DTLB_ENTRIES - LOCKED_DTLB_ENTRIES)
733 static uint64_t ttecnt_threshold[MMU_PAGE_SIZES] = {
734 	AVAIL_DTLB_ENTRIES, AVAIL_DTLB_ENTRIES,
735 	AVAIL_DTLB_ENTRIES, AVAIL_DTLB_ENTRIES,
736 	AVAIL_DTLB_ENTRIES, AVAIL_DTLB_ENTRIES};
737 
738 /*
739  * The function returns the mmu-specific values for the
740  * hat's disable_large_pages, disable_ism_large_pages, and
741  * disable_auto_data_large_pages and
742  * disable_text_data_large_pages variables.
743  */
744 uint_t
745 mmu_large_pages_disabled(uint_t flag)
746 {
747 	uint_t pages_disable = 0;
748 	extern int use_text_pgsz64K;
749 	extern int use_text_pgsz512K;
750 
751 	if (flag == HAT_LOAD) {
752 		pages_disable =  mmu_disable_large_pages;
753 	} else if (flag == HAT_LOAD_SHARE) {
754 		pages_disable = mmu_disable_ism_large_pages;
755 	} else if (flag == HAT_AUTO_DATA) {
756 		pages_disable = mmu_disable_auto_data_large_pages;
757 	} else if (flag == HAT_AUTO_TEXT) {
758 		pages_disable = mmu_disable_auto_text_large_pages;
759 		if (use_text_pgsz512K) {
760 			pages_disable &= ~(1 << TTE512K);
761 		}
762 		if (use_text_pgsz64K) {
763 			pages_disable &= ~(1 << TTE64K);
764 		}
765 	}
766 	return (pages_disable);
767 }
768 
769 /*
770  * mmu_init_large_pages is called with the desired ism_pagesize parameter.
771  * It may be called from set_platform_defaults, if some value other than 32M
772  * is desired.  mmu_ism_pagesize is the tunable.  If it has a bad value,
773  * then only warn, since it would be bad form to panic due to a user typo.
774  *
775  * The function re-initializes the mmu_disable_ism_large_pages variable.
776  */
777 void
778 mmu_init_large_pages(size_t ism_pagesize)
779 {
780 	switch (ism_pagesize) {
781 	case MMU_PAGESIZE4M:
782 		mmu_disable_ism_large_pages = ((1 << TTE64K) |
783 		    (1 << TTE512K) | (1 << TTE32M) | (1 << TTE256M));
784 		mmu_disable_auto_data_large_pages = ((1 << TTE64K) |
785 		    (1 << TTE512K) | (1 << TTE32M) | (1 << TTE256M));
786 		break;
787 	case MMU_PAGESIZE32M:
788 		mmu_disable_ism_large_pages = ((1 << TTE64K) |
789 		    (1 << TTE512K) | (1 << TTE256M));
790 		mmu_disable_auto_data_large_pages = ((1 << TTE64K) |
791 		    (1 << TTE512K) | (1 << TTE4M) | (1 << TTE256M));
792 		adjust_data_maxlpsize(ism_pagesize);
793 		break;
794 	case MMU_PAGESIZE256M:
795 		mmu_disable_ism_large_pages = ((1 << TTE64K) |
796 		    (1 << TTE512K) | (1 << TTE32M));
797 		mmu_disable_auto_data_large_pages = ((1 << TTE64K) |
798 		    (1 << TTE512K) | (1 << TTE4M) | (1 << TTE32M));
799 		adjust_data_maxlpsize(ism_pagesize);
800 		break;
801 	default:
802 		cmn_err(CE_WARN, "Unrecognized mmu_ism_pagesize value 0x%lx",
803 		    ism_pagesize);
804 		break;
805 	}
806 }
807 
808 /*
809  * Function to reprogram the TLBs when page sizes used
810  * by a process change significantly.
811  */
812 void
813 mmu_setup_page_sizes(struct hat *hat, uint64_t *ttecnt, uint8_t *tmp_pgsz)
814 {
815 	uint8_t pgsz0, pgsz1;
816 
817 	/*
818 	 * Don't program 2nd dtlb for kernel and ism hat
819 	 */
820 	ASSERT(hat->sfmmu_ismhat == NULL);
821 	ASSERT(hat != ksfmmup);
822 
823 	/*
824 	 * hat->sfmmu_pgsz[] is an array whose elements
825 	 * contain a sorted order of page sizes.  Element
826 	 * 0 is the most commonly used page size, followed
827 	 * by element 1, and so on.
828 	 *
829 	 * ttecnt[] is an array of per-page-size page counts
830 	 * mapped into the process.
831 	 *
832 	 * If the HAT's choice for page sizes is unsuitable,
833 	 * we can override it here.  The new values written
834 	 * to the array will be handed back to us later to
835 	 * do the actual programming of the TLB hardware.
836 	 *
837 	 */
838 	pgsz0 = (uint8_t)MIN(tmp_pgsz[0], tmp_pgsz[1]);
839 	pgsz1 = (uint8_t)MAX(tmp_pgsz[0], tmp_pgsz[1]);
840 
841 	/*
842 	 * This implements PAGESIZE programming of the sTLB
843 	 * if large TTE counts don't exceed the thresholds.
844 	 */
845 	if (ttecnt[pgsz0] < ttecnt_threshold[pgsz0])
846 		pgsz0 = page_szc(MMU_PAGESIZE);
847 	if (ttecnt[pgsz1] < ttecnt_threshold[pgsz1])
848 		pgsz1 = page_szc(MMU_PAGESIZE);
849 	tmp_pgsz[0] = pgsz0;
850 	tmp_pgsz[1] = pgsz1;
851 	/* otherwise, accept what the HAT chose for us */
852 }
853 
854 /*
855  * The HAT calls this function when an MMU context is allocated so that we
856  * can reprogram the large TLBs appropriately for the new process using
857  * the context.
858  *
859  * The caller must hold the HAT lock.
860  */
861 void
862 mmu_set_ctx_page_sizes(struct hat *hat)
863 {
864 	uint8_t pgsz0, pgsz1;
865 	uint8_t new_cext;
866 
867 	ASSERT(sfmmu_hat_lock_held(hat));
868 	/*
869 	 * Don't program 2nd dtlb for kernel and ism hat
870 	 */
871 	if (hat->sfmmu_ismhat || hat == ksfmmup)
872 		return;
873 
874 	/*
875 	 * If supported, reprogram the TLBs to a larger pagesize.
876 	 */
877 	pgsz0 = hat->sfmmu_pgsz[0];
878 	pgsz1 = hat->sfmmu_pgsz[1];
879 	ASSERT(pgsz0 < mmu_page_sizes);
880 	ASSERT(pgsz1 < mmu_page_sizes);
881 	new_cext = TAGACCEXT_MKSZPAIR(pgsz1, pgsz0);
882 	if (hat->sfmmu_cext != new_cext) {
883 #ifdef DEBUG
884 		int i;
885 		/*
886 		 * assert cnum should be invalid, this is because pagesize
887 		 * can only be changed after a proc's ctxs are invalidated.
888 		 */
889 		for (i = 0; i < max_mmu_ctxdoms; i++) {
890 			ASSERT(hat->sfmmu_ctxs[i].cnum == INVALID_CONTEXT);
891 		}
892 #endif /* DEBUG */
893 		hat->sfmmu_cext = new_cext;
894 	}
895 	/*
896 	 * sfmmu_setctx_sec() will take care of the
897 	 * rest of the dirty work for us.
898 	 */
899 }
900 
901 /*
902  * This function assumes that there are either four or six supported page
903  * sizes and at most two programmable TLBs, so we need to decide which
904  * page sizes are most important and then adjust the TLB page sizes
905  * accordingly (if supported).
906  *
907  * If these assumptions change, this function will need to be
908  * updated to support whatever the new limits are.
909  */
910 void
911 mmu_check_page_sizes(sfmmu_t *sfmmup, uint64_t *ttecnt)
912 {
913 	uint64_t sortcnt[MMU_PAGE_SIZES];
914 	uint8_t tmp_pgsz[MMU_PAGE_SIZES];
915 	uint8_t i, j, max;
916 	uint16_t oldval, newval;
917 
918 	/*
919 	 * We only consider reprogramming the TLBs if one or more of
920 	 * the two most used page sizes changes and we're using
921 	 * large pages in this process.
922 	 */
923 	if (sfmmup->sfmmu_flags & HAT_LGPG_FLAGS) {
924 		/* Sort page sizes. */
925 		for (i = 0; i < mmu_page_sizes; i++) {
926 			sortcnt[i] = ttecnt[i];
927 		}
928 		for (j = 0; j < mmu_page_sizes; j++) {
929 			for (i = mmu_page_sizes - 1, max = 0; i > 0; i--) {
930 				if (sortcnt[i] > sortcnt[max])
931 					max = i;
932 			}
933 			tmp_pgsz[j] = max;
934 			sortcnt[max] = 0;
935 		}
936 
937 		oldval = sfmmup->sfmmu_pgsz[0] << 8 | sfmmup->sfmmu_pgsz[1];
938 
939 		mmu_setup_page_sizes(sfmmup, ttecnt, tmp_pgsz);
940 
941 		/* Check 2 largest values after the sort. */
942 		newval = tmp_pgsz[0] << 8 | tmp_pgsz[1];
943 		if (newval != oldval) {
944 			sfmmu_reprog_pgsz_arr(sfmmup, tmp_pgsz);
945 		}
946 	}
947 }
948 
949 /*
950  * Return processor specific async error structure
951  * size used.
952  */
953 int
954 cpu_aflt_size(void)
955 {
956 	return (sizeof (opl_async_flt_t));
957 }
958 
959 /*
960  * The cpu_sync_log_err() function is called via the [uc]e_drain() function to
961  * post-process CPU events that are dequeued.  As such, it can be invoked
962  * from softint context, from AST processing in the trap() flow, or from the
963  * panic flow.  We decode the CPU-specific data, and take appropriate actions.
964  * Historically this entry point was used to log the actual cmn_err(9F) text;
965  * now with FMA it is used to prepare 'flt' to be converted into an ereport.
966  * With FMA this function now also returns a flag which indicates to the
967  * caller whether the ereport should be posted (1) or suppressed (0).
968  */
969 /*ARGSUSED*/
970 static int
971 cpu_sync_log_err(void *flt)
972 {
973 	opl_async_flt_t *opl_flt = (opl_async_flt_t *)flt;
974 	struct async_flt *aflt = (struct async_flt *)flt;
975 
976 	/*
977 	 * No extra processing of urgent error events.
978 	 * Always generate ereports for these events.
979 	 */
980 	if (aflt->flt_status == OPL_ECC_URGENT_TRAP)
981 		return (1);
982 
983 	/*
984 	 * Additional processing for synchronous errors.
985 	 */
986 	switch (opl_flt->flt_type) {
987 	case OPL_CPU_INV_SFSR:
988 		return (1);
989 
990 	case OPL_CPU_SYNC_UE:
991 		/*
992 		 * The validity: SFSR_MK_UE bit has been checked
993 		 * in opl_cpu_sync_error()
994 		 * No more check is required.
995 		 *
996 		 * opl_flt->flt_eid_mod and flt_eid_sid have been set by H/W,
997 		 * and they have been retrieved in cpu_queue_events()
998 		 */
999 
1000 		if (opl_flt->flt_eid_mod == OPL_ERRID_MEM) {
1001 			ASSERT(aflt->flt_in_memory);
1002 			/*
1003 			 * We want to skip logging only if ALL the following
1004 			 * conditions are true:
1005 			 *
1006 			 *	1. We are not panicing already.
1007 			 *	2. The error is a memory error.
1008 			 *	3. There is only one error.
1009 			 *	4. The error is on a retired page.
1010 			 *	5. The error occurred under on_trap
1011 			 *	protection AFLT_PROT_EC
1012 			 */
1013 			if (!panicstr && aflt->flt_prot == AFLT_PROT_EC &&
1014 			    page_retire_check(aflt->flt_addr, NULL) == 0) {
1015 				/*
1016 				 * Do not log an error from
1017 				 * the retired page
1018 				 */
1019 				softcall(ecc_page_zero, (void *)aflt->flt_addr);
1020 				return (0);
1021 			}
1022 			if (!panicstr)
1023 				cpu_page_retire(opl_flt);
1024 		}
1025 		return (1);
1026 
1027 	case OPL_CPU_SYNC_OTHERS:
1028 		/*
1029 		 * For the following error cases, the processor HW does
1030 		 * not set the flt_eid_mod/flt_eid_sid. Instead, SW will attempt
1031 		 * to assign appropriate values here to reflect what we
1032 		 * think is the most likely cause of the problem w.r.t to
1033 		 * the particular error event.  For Buserr and timeout
1034 		 * error event, we will assign OPL_ERRID_CHANNEL as the
1035 		 * most likely reason.  For TLB parity or multiple hit
1036 		 * error events, we will assign the reason as
1037 		 * OPL_ERRID_CPU (cpu related problem) and set the
1038 		 * flt_eid_sid to point to the cpuid.
1039 		 */
1040 
1041 		if (opl_flt->flt_bit & (SFSR_BERR|SFSR_TO)) {
1042 			/*
1043 			 * flt_eid_sid will not be used for this case.
1044 			 */
1045 			opl_flt->flt_eid_mod = OPL_ERRID_CHANNEL;
1046 		}
1047 		if (opl_flt->flt_bit & (SFSR_TLB_MUL|SFSR_TLB_PRT)) {
1048 			    opl_flt->flt_eid_mod = OPL_ERRID_CPU;
1049 			    opl_flt->flt_eid_sid = aflt->flt_inst;
1050 		}
1051 
1052 		/*
1053 		 * In case of no effective error bit
1054 		 */
1055 		if ((opl_flt->flt_bit & SFSR_ERRS) == 0) {
1056 			    opl_flt->flt_eid_mod = OPL_ERRID_CPU;
1057 			    opl_flt->flt_eid_sid = aflt->flt_inst;
1058 		}
1059 		break;
1060 
1061 		default:
1062 			return (1);
1063 	}
1064 	return (1);
1065 }
1066 
1067 /*
1068  * Retire the bad page that may contain the flushed error.
1069  */
1070 void
1071 cpu_page_retire(opl_async_flt_t *opl_flt)
1072 {
1073 	struct async_flt *aflt = (struct async_flt *)opl_flt;
1074 	(void) page_retire(aflt->flt_addr, PR_UE);
1075 }
1076 
1077 /*
1078  * Invoked by error_init() early in startup and therefore before
1079  * startup_errorq() is called to drain any error Q -
1080  *
1081  * startup()
1082  *   startup_end()
1083  *     error_init()
1084  *       cpu_error_init()
1085  * errorq_init()
1086  *   errorq_drain()
1087  * start_other_cpus()
1088  *
1089  * The purpose of this routine is to create error-related taskqs.  Taskqs
1090  * are used for this purpose because cpu_lock can't be grabbed from interrupt
1091  * context.
1092  *
1093  */
1094 /*ARGSUSED*/
1095 void
1096 cpu_error_init(int items)
1097 {
1098 	opl_err_log = (opl_errlog_t *)
1099 	    kmem_alloc(ERRLOG_ALLOC_SZ, KM_SLEEP);
1100 	if ((uint64_t)opl_err_log & MMU_PAGEOFFSET)
1101 		cmn_err(CE_PANIC, "The base address of the error log "
1102 		    "is not page aligned");
1103 }
1104 
1105 /*
1106  * We route all errors through a single switch statement.
1107  */
1108 void
1109 cpu_ue_log_err(struct async_flt *aflt)
1110 {
1111 	switch (aflt->flt_class) {
1112 	case CPU_FAULT:
1113 		if (cpu_sync_log_err(aflt))
1114 			cpu_ereport_post(aflt);
1115 		break;
1116 
1117 	case BUS_FAULT:
1118 		bus_async_log_err(aflt);
1119 		break;
1120 
1121 	default:
1122 		cmn_err(CE_WARN, "discarding async error %p with invalid "
1123 		    "fault class (0x%x)", (void *)aflt, aflt->flt_class);
1124 		return;
1125 	}
1126 }
1127 
1128 /*
1129  * Routine for panic hook callback from panic_idle().
1130  *
1131  * Nothing to do here.
1132  */
1133 void
1134 cpu_async_panic_callb(void)
1135 {
1136 }
1137 
1138 /*
1139  * Routine to return a string identifying the physical name
1140  * associated with a memory/cache error.
1141  */
1142 /*ARGSUSED*/
1143 int
1144 cpu_get_mem_unum(int synd_status, ushort_t flt_synd, uint64_t flt_stat,
1145     uint64_t flt_addr, int flt_bus_id, int flt_in_memory,
1146     ushort_t flt_status, char *buf, int buflen, int *lenp)
1147 {
1148 	int synd_code;
1149 	int ret;
1150 
1151 	/*
1152 	 * An AFSR of -1 defaults to a memory syndrome.
1153 	 */
1154 	synd_code = (int)flt_synd;
1155 
1156 	if (&plat_get_mem_unum) {
1157 		if ((ret = plat_get_mem_unum(synd_code, flt_addr, flt_bus_id,
1158 			flt_in_memory, flt_status, buf, buflen, lenp)) != 0) {
1159 			buf[0] = '\0';
1160 			*lenp = 0;
1161 		}
1162 		return (ret);
1163 	}
1164 	buf[0] = '\0';
1165 	*lenp = 0;
1166 	return (ENOTSUP);
1167 }
1168 
1169 /*
1170  * Wrapper for cpu_get_mem_unum() routine that takes an
1171  * async_flt struct rather than explicit arguments.
1172  */
1173 int
1174 cpu_get_mem_unum_aflt(int synd_status, struct async_flt *aflt,
1175     char *buf, int buflen, int *lenp)
1176 {
1177 	/*
1178 	 * We always pass -1 so that cpu_get_mem_unum will interpret this as a
1179 	 * memory error.
1180 	 */
1181 	return (cpu_get_mem_unum(synd_status, aflt->flt_synd,
1182 	    (uint64_t)-1,
1183 	    aflt->flt_addr, aflt->flt_bus_id, aflt->flt_in_memory,
1184 	    aflt->flt_status, buf, buflen, lenp));
1185 }
1186 
1187 /*
1188  * This routine is a more generic interface to cpu_get_mem_unum()
1189  * that may be used by other modules (e.g. mm).
1190  */
1191 /*ARGSUSED*/
1192 int
1193 cpu_get_mem_name(uint64_t synd, uint64_t *afsr, uint64_t afar,
1194     char *buf, int buflen, int *lenp)
1195 {
1196 	int synd_status, flt_in_memory, ret;
1197 	ushort_t flt_status = 0;
1198 	char unum[UNUM_NAMLEN];
1199 
1200 	/*
1201 	 * Check for an invalid address.
1202 	 */
1203 	if (afar == (uint64_t)-1)
1204 		return (ENXIO);
1205 
1206 	if (synd == (uint64_t)-1)
1207 		synd_status = AFLT_STAT_INVALID;
1208 	else
1209 		synd_status = AFLT_STAT_VALID;
1210 
1211 	flt_in_memory = (*afsr & SFSR_MEMORY) &&
1212 		pf_is_memory(afar >> MMU_PAGESHIFT);
1213 
1214 	ret = cpu_get_mem_unum(synd_status, (ushort_t)synd, *afsr, afar,
1215 		CPU->cpu_id, flt_in_memory, flt_status, unum,
1216 		UNUM_NAMLEN, lenp);
1217 	if (ret != 0)
1218 		return (ret);
1219 
1220 	if (*lenp >= buflen)
1221 		return (ENAMETOOLONG);
1222 
1223 	(void) strncpy(buf, unum, buflen);
1224 
1225 	return (0);
1226 }
1227 
1228 /*
1229  * Routine to return memory information associated
1230  * with a physical address and syndrome.
1231  */
1232 /*ARGSUSED*/
1233 int
1234 cpu_get_mem_info(uint64_t synd, uint64_t afar,
1235     uint64_t *mem_sizep, uint64_t *seg_sizep, uint64_t *bank_sizep,
1236     int *segsp, int *banksp, int *mcidp)
1237 {
1238 	int synd_code = (int)synd;
1239 
1240 	if (afar == (uint64_t)-1)
1241 		return (ENXIO);
1242 
1243 	if (p2get_mem_info != NULL)
1244 		return ((p2get_mem_info)(synd_code, afar,
1245 			mem_sizep, seg_sizep, bank_sizep,
1246 			segsp, banksp, mcidp));
1247 	else
1248 		return (ENOTSUP);
1249 }
1250 
1251 /*
1252  * Routine to return a string identifying the physical
1253  * name associated with a cpuid.
1254  */
1255 int
1256 cpu_get_cpu_unum(int cpuid, char *buf, int buflen, int *lenp)
1257 {
1258 	int ret;
1259 	char unum[UNUM_NAMLEN];
1260 
1261 	if (&plat_get_cpu_unum) {
1262 		if ((ret = plat_get_cpu_unum(cpuid, unum, UNUM_NAMLEN, lenp))
1263 			!= 0)
1264 			return (ret);
1265 	} else {
1266 		return (ENOTSUP);
1267 	}
1268 
1269 	if (*lenp >= buflen)
1270 		return (ENAMETOOLONG);
1271 
1272 	(void) strncpy(buf, unum, *lenp);
1273 
1274 	return (0);
1275 }
1276 
1277 /*
1278  * This routine exports the name buffer size.
1279  */
1280 size_t
1281 cpu_get_name_bufsize()
1282 {
1283 	return (UNUM_NAMLEN);
1284 }
1285 
1286 /*
1287  * Flush the entire ecache by ASI_L2_CNTL.U2_FLUSH
1288  */
1289 void
1290 cpu_flush_ecache(void)
1291 {
1292 	flush_ecache(ecache_flushaddr, cpunodes[CPU->cpu_id].ecache_size,
1293 	    cpunodes[CPU->cpu_id].ecache_linesize);
1294 }
1295 
1296 static uint8_t
1297 flt_to_trap_type(struct async_flt *aflt)
1298 {
1299 	if (aflt->flt_status & OPL_ECC_ISYNC_TRAP)
1300 		return (TRAP_TYPE_ECC_I);
1301 	if (aflt->flt_status & OPL_ECC_DSYNC_TRAP)
1302 		return (TRAP_TYPE_ECC_D);
1303 	if (aflt->flt_status & OPL_ECC_URGENT_TRAP)
1304 		return (TRAP_TYPE_URGENT);
1305 	return (TRAP_TYPE_UNKNOWN);
1306 }
1307 
1308 /*
1309  * Encode the data saved in the opl_async_flt_t struct into
1310  * the FM ereport payload.
1311  */
1312 /* ARGSUSED */
1313 static void
1314 cpu_payload_add_aflt(struct async_flt *aflt, nvlist_t *payload,
1315 		nvlist_t *resource)
1316 {
1317 	opl_async_flt_t *opl_flt = (opl_async_flt_t *)aflt;
1318 	char unum[UNUM_NAMLEN];
1319 	char sbuf[21]; /* sizeof (UINT64_MAX) + '\0' */
1320 	int len;
1321 
1322 
1323 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_SFSR) {
1324 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_SFSR,
1325 			DATA_TYPE_UINT64, aflt->flt_stat, NULL);
1326 	}
1327 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_SFAR) {
1328 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_SFAR,
1329 			DATA_TYPE_UINT64, aflt->flt_addr, NULL);
1330 	}
1331 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_UGESR) {
1332 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_UGESR,
1333 			DATA_TYPE_UINT64, aflt->flt_stat, NULL);
1334 	}
1335 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_PC) {
1336 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_PC,
1337 		    DATA_TYPE_UINT64, (uint64_t)aflt->flt_pc, NULL);
1338 	}
1339 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_TL) {
1340 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_TL,
1341 		    DATA_TYPE_UINT8, (uint8_t)aflt->flt_tl, NULL);
1342 	}
1343 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_TT) {
1344 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_TT,
1345 		    DATA_TYPE_UINT8, flt_to_trap_type(aflt), NULL);
1346 	}
1347 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_PRIV) {
1348 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_PRIV,
1349 		    DATA_TYPE_BOOLEAN_VALUE,
1350 		    (aflt->flt_priv ? B_TRUE : B_FALSE), NULL);
1351 	}
1352 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_FLT_STATUS) {
1353 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_FLT_STATUS,
1354 			DATA_TYPE_UINT64, (uint64_t)aflt->flt_status, NULL);
1355 	}
1356 
1357 	switch (opl_flt->flt_eid_mod) {
1358 	case OPL_ERRID_CPU:
1359 		(void) snprintf(sbuf, sizeof (sbuf), "%llX",
1360 			(u_longlong_t)cpunodes[opl_flt->flt_eid_sid].device_id);
1361 		(void) fm_fmri_cpu_set(resource, FM_CPU_SCHEME_VERSION,
1362 			NULL, opl_flt->flt_eid_sid,
1363 			(uint8_t *)&cpunodes[opl_flt->flt_eid_sid].version,
1364 			sbuf);
1365 		fm_payload_set(payload,
1366 			FM_EREPORT_PAYLOAD_NAME_RESOURCE,
1367 			DATA_TYPE_NVLIST, resource, NULL);
1368 		break;
1369 
1370 	case OPL_ERRID_CHANNEL:
1371 		/*
1372 		 * No resource is created but the cpumem DE will find
1373 		 * the defective path by retreiving EID from SFSR which is
1374 		 * included in the payload.
1375 		 */
1376 		break;
1377 
1378 	case OPL_ERRID_MEM:
1379 		(void) cpu_get_mem_unum_aflt(0, aflt, unum, UNUM_NAMLEN, &len);
1380 		(void) fm_fmri_mem_set(resource, FM_MEM_SCHEME_VERSION,
1381 			NULL, unum, NULL, (uint64_t)-1);
1382 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_RESOURCE,
1383 			DATA_TYPE_NVLIST, resource, NULL);
1384 		break;
1385 
1386 	case OPL_ERRID_PATH:
1387 		/*
1388 		 * No resource is created but the cpumem DE will find
1389 		 * the defective path by retreiving EID from SFSR which is
1390 		 * included in the payload.
1391 		 */
1392 		break;
1393 	}
1394 }
1395 
1396 /*
1397  * Returns whether fault address is valid for this error bit and
1398  * whether the address is "in memory" (i.e. pf_is_memory returns 1).
1399  */
1400 /*ARGSUSED*/
1401 static int
1402 cpu_flt_in_memory(opl_async_flt_t *opl_flt, uint64_t t_afsr_bit)
1403 {
1404 	struct async_flt *aflt = (struct async_flt *)opl_flt;
1405 
1406 	if (aflt->flt_status & (OPL_ECC_SYNC_TRAP)) {
1407 		return ((t_afsr_bit & SFSR_MEMORY) &&
1408 		    pf_is_memory(aflt->flt_addr >> MMU_PAGESHIFT));
1409 	}
1410 	return (0);
1411 }
1412 
1413 /*
1414  * In OPL SCF does the stick synchronization.
1415  */
1416 void
1417 sticksync_slave(void)
1418 {
1419 }
1420 
1421 /*
1422  * In OPL SCF does the stick synchronization.
1423  */
1424 void
1425 sticksync_master(void)
1426 {
1427 }
1428 
1429 /*
1430  * Cpu private unitialization.  OPL cpus do not use the private area.
1431  */
1432 void
1433 cpu_uninit_private(struct cpu *cp)
1434 {
1435 	cmp_delete_cpu(cp->cpu_id);
1436 }
1437 
1438 /*
1439  * Always flush an entire cache.
1440  */
1441 void
1442 cpu_error_ecache_flush(void)
1443 {
1444 	cpu_flush_ecache();
1445 }
1446 
1447 void
1448 cpu_ereport_post(struct async_flt *aflt)
1449 {
1450 	char *cpu_type, buf[FM_MAX_CLASS];
1451 	nv_alloc_t *nva = NULL;
1452 	nvlist_t *ereport, *detector, *resource;
1453 	errorq_elem_t *eqep;
1454 	char sbuf[21]; /* sizeof (UINT64_MAX) + '\0' */
1455 
1456 	if (aflt->flt_panic || panicstr) {
1457 		eqep = errorq_reserve(ereport_errorq);
1458 		if (eqep == NULL)
1459 			return;
1460 		ereport = errorq_elem_nvl(ereport_errorq, eqep);
1461 		nva = errorq_elem_nva(ereport_errorq, eqep);
1462 	} else {
1463 		ereport = fm_nvlist_create(nva);
1464 	}
1465 
1466 	/*
1467 	 * Create the scheme "cpu" FMRI.
1468 	 */
1469 	detector = fm_nvlist_create(nva);
1470 	resource = fm_nvlist_create(nva);
1471 	switch (cpunodes[aflt->flt_inst].implementation) {
1472 	case OLYMPUS_C_IMPL:
1473 		cpu_type = FM_EREPORT_CPU_SPARC64_VI;
1474 		break;
1475 	default:
1476 		cpu_type = FM_EREPORT_CPU_UNSUPPORTED;
1477 		break;
1478 	}
1479 	(void) snprintf(sbuf, sizeof (sbuf), "%llX",
1480 	    (u_longlong_t)cpunodes[aflt->flt_inst].device_id);
1481 	(void) fm_fmri_cpu_set(detector, FM_CPU_SCHEME_VERSION, NULL,
1482 	    aflt->flt_inst, (uint8_t *)&cpunodes[aflt->flt_inst].version,
1483 	    sbuf);
1484 
1485 	/*
1486 	 * Encode all the common data into the ereport.
1487 	 */
1488 	(void) snprintf(buf, FM_MAX_CLASS, "%s.%s.%s",
1489 	    FM_ERROR_CPU, cpu_type, aflt->flt_erpt_class);
1490 
1491 	fm_ereport_set(ereport, FM_EREPORT_VERSION, buf,
1492 	    fm_ena_generate(aflt->flt_id, FM_ENA_FMT1), detector, NULL);
1493 
1494 	/*
1495 	 * Encode the error specific data that was saved in
1496 	 * the async_flt structure into the ereport.
1497 	 */
1498 	cpu_payload_add_aflt(aflt, ereport, resource);
1499 
1500 	if (aflt->flt_panic || panicstr) {
1501 		errorq_commit(ereport_errorq, eqep, ERRORQ_SYNC);
1502 	} else {
1503 		(void) fm_ereport_post(ereport, EVCH_TRYHARD);
1504 		fm_nvlist_destroy(ereport, FM_NVA_FREE);
1505 		fm_nvlist_destroy(detector, FM_NVA_FREE);
1506 		fm_nvlist_destroy(resource, FM_NVA_FREE);
1507 	}
1508 }
1509 
1510 void
1511 cpu_run_bus_error_handlers(struct async_flt *aflt, int expected)
1512 {
1513 	int status;
1514 	ddi_fm_error_t de;
1515 
1516 	bzero(&de, sizeof (ddi_fm_error_t));
1517 
1518 	de.fme_version = DDI_FME_VERSION;
1519 	de.fme_ena = fm_ena_generate(aflt->flt_id, FM_ENA_FMT1);
1520 	de.fme_flag = expected;
1521 	de.fme_bus_specific = (void *)aflt->flt_addr;
1522 	status = ndi_fm_handler_dispatch(ddi_root_node(), NULL, &de);
1523 	if ((aflt->flt_prot == AFLT_PROT_NONE) && (status == DDI_FM_FATAL))
1524 		aflt->flt_panic = 1;
1525 }
1526 
1527 void
1528 cpu_errorq_dispatch(char *error_class, void *payload, size_t payload_sz,
1529     errorq_t *eqp, uint_t flag)
1530 {
1531 	struct async_flt *aflt = (struct async_flt *)payload;
1532 
1533 	aflt->flt_erpt_class = error_class;
1534 	errorq_dispatch(eqp, payload, payload_sz, flag);
1535 }
1536 
1537 void
1538 adjust_hw_copy_limits(int ecache_size)
1539 {
1540 	/*
1541 	 * Set hw copy limits.
1542 	 *
1543 	 * /etc/system will be parsed later and can override one or more
1544 	 * of these settings.
1545 	 *
1546 	 * At this time, ecache size seems only mildly relevant.
1547 	 * We seem to run into issues with the d-cache and stalls
1548 	 * we see on misses.
1549 	 *
1550 	 * Cycle measurement indicates that 2 byte aligned copies fare
1551 	 * little better than doing things with VIS at around 512 bytes.
1552 	 * 4 byte aligned shows promise until around 1024 bytes. 8 Byte
1553 	 * aligned is faster whenever the source and destination data
1554 	 * in cache and the total size is less than 2 Kbytes.  The 2K
1555 	 * limit seems to be driven by the 2K write cache.
1556 	 * When more than 2K of copies are done in non-VIS mode, stores
1557 	 * backup in the write cache.  In VIS mode, the write cache is
1558 	 * bypassed, allowing faster cache-line writes aligned on cache
1559 	 * boundaries.
1560 	 *
1561 	 * In addition, in non-VIS mode, there is no prefetching, so
1562 	 * for larger copies, the advantage of prefetching to avoid even
1563 	 * occasional cache misses is enough to justify using the VIS code.
1564 	 *
1565 	 * During testing, it was discovered that netbench ran 3% slower
1566 	 * when hw_copy_limit_8 was 2K or larger.  Apparently for server
1567 	 * applications, data is only used once (copied to the output
1568 	 * buffer, then copied by the network device off the system).  Using
1569 	 * the VIS copy saves more L2 cache state.  Network copies are
1570 	 * around 1.3K to 1.5K in size for historical reasons.
1571 	 *
1572 	 * Therefore, a limit of 1K bytes will be used for the 8 byte
1573 	 * aligned copy even for large caches and 8 MB ecache.  The
1574 	 * infrastructure to allow different limits for different sized
1575 	 * caches is kept to allow further tuning in later releases.
1576 	 */
1577 
1578 	if (min_ecache_size == 0 && use_hw_bcopy) {
1579 		/*
1580 		 * First time through - should be before /etc/system
1581 		 * is read.
1582 		 * Could skip the checks for zero but this lets us
1583 		 * preserve any debugger rewrites.
1584 		 */
1585 		if (hw_copy_limit_1 == 0) {
1586 			hw_copy_limit_1 = VIS_COPY_THRESHOLD;
1587 			priv_hcl_1 = hw_copy_limit_1;
1588 		}
1589 		if (hw_copy_limit_2 == 0) {
1590 			hw_copy_limit_2 = 2 * VIS_COPY_THRESHOLD;
1591 			priv_hcl_2 = hw_copy_limit_2;
1592 		}
1593 		if (hw_copy_limit_4 == 0) {
1594 			hw_copy_limit_4 = 4 * VIS_COPY_THRESHOLD;
1595 			priv_hcl_4 = hw_copy_limit_4;
1596 		}
1597 		if (hw_copy_limit_8 == 0) {
1598 			hw_copy_limit_8 = 4 * VIS_COPY_THRESHOLD;
1599 			priv_hcl_8 = hw_copy_limit_8;
1600 		}
1601 		min_ecache_size = ecache_size;
1602 	} else {
1603 		/*
1604 		 * MP initialization. Called *after* /etc/system has
1605 		 * been parsed. One CPU has already been initialized.
1606 		 * Need to cater for /etc/system having scragged one
1607 		 * of our values.
1608 		 */
1609 		if (ecache_size == min_ecache_size) {
1610 			/*
1611 			 * Same size ecache. We do nothing unless we
1612 			 * have a pessimistic ecache setting. In that
1613 			 * case we become more optimistic (if the cache is
1614 			 * large enough).
1615 			 */
1616 			if (hw_copy_limit_8 == 4 * VIS_COPY_THRESHOLD) {
1617 				/*
1618 				 * Need to adjust hw_copy_limit* from our
1619 				 * pessimistic uniprocessor value to a more
1620 				 * optimistic UP value *iff* it hasn't been
1621 				 * reset.
1622 				 */
1623 				if ((ecache_size > 1048576) &&
1624 				    (priv_hcl_8 == hw_copy_limit_8)) {
1625 					if (ecache_size <= 2097152)
1626 						hw_copy_limit_8 = 4 *
1627 						    VIS_COPY_THRESHOLD;
1628 					else if (ecache_size <= 4194304)
1629 						hw_copy_limit_8 = 4 *
1630 						    VIS_COPY_THRESHOLD;
1631 					else
1632 						hw_copy_limit_8 = 4 *
1633 						    VIS_COPY_THRESHOLD;
1634 					priv_hcl_8 = hw_copy_limit_8;
1635 				}
1636 			}
1637 		} else if (ecache_size < min_ecache_size) {
1638 			/*
1639 			 * A different ecache size. Can this even happen?
1640 			 */
1641 			if (priv_hcl_8 == hw_copy_limit_8) {
1642 				/*
1643 				 * The previous value that we set
1644 				 * is unchanged (i.e., it hasn't been
1645 				 * scragged by /etc/system). Rewrite it.
1646 				 */
1647 				if (ecache_size <= 1048576)
1648 					hw_copy_limit_8 = 8 *
1649 					    VIS_COPY_THRESHOLD;
1650 				else if (ecache_size <= 2097152)
1651 					hw_copy_limit_8 = 8 *
1652 					    VIS_COPY_THRESHOLD;
1653 				else if (ecache_size <= 4194304)
1654 					hw_copy_limit_8 = 8 *
1655 					    VIS_COPY_THRESHOLD;
1656 				else
1657 					hw_copy_limit_8 = 10 *
1658 					    VIS_COPY_THRESHOLD;
1659 				priv_hcl_8 = hw_copy_limit_8;
1660 				min_ecache_size = ecache_size;
1661 			}
1662 		}
1663 	}
1664 }
1665 
1666 #define	VIS_BLOCKSIZE		64
1667 
1668 int
1669 dtrace_blksuword32_err(uintptr_t addr, uint32_t *data)
1670 {
1671 	int ret, watched;
1672 
1673 	watched = watch_disable_addr((void *)addr, VIS_BLOCKSIZE, S_WRITE);
1674 	ret = dtrace_blksuword32(addr, data, 0);
1675 	if (watched)
1676 		watch_enable_addr((void *)addr, VIS_BLOCKSIZE, S_WRITE);
1677 
1678 	return (ret);
1679 }
1680 
1681 void
1682 opl_cpu_reg_init()
1683 {
1684 	uint64_t	this_cpu_log;
1685 
1686 	/*
1687 	 * We do not need to re-initialize cpu0 registers.
1688 	 */
1689 	if (cpu[getprocessorid()] == &cpu0)
1690 		return;
1691 
1692 	/*
1693 	 * Initialize Error log Scratch register for error handling.
1694 	 */
1695 
1696 	this_cpu_log = va_to_pa((void*)(((uint64_t)opl_err_log) +
1697 		ERRLOG_BUFSZ * (getprocessorid())));
1698 	opl_error_setup(this_cpu_log);
1699 
1700 	/*
1701 	 * Enable MMU translating multiple page sizes for
1702 	 * sITLB and sDTLB.
1703 	 */
1704 	opl_mpg_enable();
1705 }
1706 
1707 /*
1708  * Queue one event in ue_queue based on ecc_type_to_info entry.
1709  */
1710 static void
1711 cpu_queue_one_event(opl_async_flt_t *opl_flt, char *reason,
1712     ecc_type_to_info_t *eccp)
1713 {
1714 	struct async_flt *aflt = (struct async_flt *)opl_flt;
1715 
1716 	if (reason &&
1717 	    strlen(reason) + strlen(eccp->ec_reason) < MAX_REASON_STRING) {
1718 		(void) strcat(reason, eccp->ec_reason);
1719 	}
1720 
1721 	opl_flt->flt_bit = eccp->ec_afsr_bit;
1722 	opl_flt->flt_type = eccp->ec_flt_type;
1723 	aflt->flt_in_memory = cpu_flt_in_memory(opl_flt, opl_flt->flt_bit);
1724 	aflt->flt_payload = eccp->ec_err_payload;
1725 
1726 	ASSERT(aflt->flt_status & (OPL_ECC_SYNC_TRAP|OPL_ECC_URGENT_TRAP));
1727 	cpu_errorq_dispatch(eccp->ec_err_class,
1728 		(void *)opl_flt, sizeof (opl_async_flt_t),
1729 		ue_queue,
1730 		aflt->flt_panic);
1731 }
1732 
1733 /*
1734  * Queue events on async event queue one event per error bit.
1735  * Return number of events queued.
1736  */
1737 int
1738 cpu_queue_events(opl_async_flt_t *opl_flt, char *reason, uint64_t t_afsr_errs)
1739 {
1740 	struct async_flt *aflt = (struct async_flt *)opl_flt;
1741 	ecc_type_to_info_t *eccp;
1742 	int nevents = 0;
1743 
1744 	/*
1745 	 * Queue expected errors, error bit and fault type must must match
1746 	 * in the ecc_type_to_info table.
1747 	 */
1748 	for (eccp = ecc_type_to_info; t_afsr_errs != 0 && eccp->ec_desc != NULL;
1749 		eccp++) {
1750 		if ((eccp->ec_afsr_bit & t_afsr_errs) != 0 &&
1751 		    (eccp->ec_flags & aflt->flt_status) != 0) {
1752 			/*
1753 			 * UE error event can be further
1754 			 * classified/breakdown into finer granularity
1755 			 * based on the flt_eid_mod value set by HW.  We do
1756 			 * special handling here so that we can report UE
1757 			 * error in finer granularity as ue_mem,
1758 			 * ue_channel, ue_cpu or ue_path.
1759 			 */
1760 			if (eccp->ec_flt_type == OPL_CPU_SYNC_UE) {
1761 				opl_flt->flt_eid_mod =
1762 					(aflt->flt_stat & SFSR_EID_MOD)
1763 					>> SFSR_EID_MOD_SHIFT;
1764 				opl_flt->flt_eid_sid =
1765 					(aflt->flt_stat & SFSR_EID_SID)
1766 					>> SFSR_EID_SID_SHIFT;
1767 				/*
1768 				 * Need to advance eccp pointer by flt_eid_mod
1769 				 * so that we get an appropriate ecc pointer
1770 				 *
1771 				 * EID			# of advances
1772 				 * ----------------------------------
1773 				 * OPL_ERRID_MEM	0
1774 				 * OPL_ERRID_CHANNEL	1
1775 				 * OPL_ERRID_CPU	2
1776 				 * OPL_ERRID_PATH	3
1777 				 */
1778 				eccp += opl_flt->flt_eid_mod;
1779 			}
1780 			cpu_queue_one_event(opl_flt, reason, eccp);
1781 			t_afsr_errs &= ~eccp->ec_afsr_bit;
1782 			nevents++;
1783 		}
1784 	}
1785 
1786 	return (nevents);
1787 }
1788 
1789 /*
1790  * Sync. error wrapper functions.
1791  * We use these functions in order to transfer here from the
1792  * nucleus trap handler information about trap type (data or
1793  * instruction) and trap level (0 or above 0). This way we
1794  * get rid of using SFSR's reserved bits.
1795  */
1796 
1797 #define	OPL_SYNC_TL0	0
1798 #define	OPL_SYNC_TL1	1
1799 #define	OPL_ISYNC_ERR	0
1800 #define	OPL_DSYNC_ERR	1
1801 
1802 void
1803 opl_cpu_isync_tl0_error(struct regs *rp, ulong_t p_sfar, ulong_t p_sfsr)
1804 {
1805 	uint64_t t_sfar = p_sfar;
1806 	uint64_t t_sfsr = p_sfsr;
1807 
1808 	opl_cpu_sync_error(rp, t_sfar, t_sfsr,
1809 	    OPL_SYNC_TL0, OPL_ISYNC_ERR);
1810 }
1811 
1812 void
1813 opl_cpu_isync_tl1_error(struct regs *rp, ulong_t p_sfar, ulong_t p_sfsr)
1814 {
1815 	uint64_t t_sfar = p_sfar;
1816 	uint64_t t_sfsr = p_sfsr;
1817 
1818 	opl_cpu_sync_error(rp, t_sfar, t_sfsr,
1819 	    OPL_SYNC_TL1, OPL_ISYNC_ERR);
1820 }
1821 
1822 void
1823 opl_cpu_dsync_tl0_error(struct regs *rp, ulong_t p_sfar, ulong_t p_sfsr)
1824 {
1825 	uint64_t t_sfar = p_sfar;
1826 	uint64_t t_sfsr = p_sfsr;
1827 
1828 	opl_cpu_sync_error(rp, t_sfar, t_sfsr,
1829 	    OPL_SYNC_TL0, OPL_DSYNC_ERR);
1830 }
1831 
1832 void
1833 opl_cpu_dsync_tl1_error(struct regs *rp, ulong_t p_sfar, ulong_t p_sfsr)
1834 {
1835 	uint64_t t_sfar = p_sfar;
1836 	uint64_t t_sfsr = p_sfsr;
1837 
1838 	opl_cpu_sync_error(rp, t_sfar, t_sfsr,
1839 	    OPL_SYNC_TL1, OPL_DSYNC_ERR);
1840 }
1841 
1842 /*
1843  * The fj sync err handler transfers control here for UE, BERR, TO, TLB_MUL
1844  * and TLB_PRT.
1845  * This function is designed based on cpu_deferred_error().
1846  */
1847 
1848 static void
1849 opl_cpu_sync_error(struct regs *rp, ulong_t t_sfar, ulong_t t_sfsr,
1850     uint_t tl, uint_t derr)
1851 {
1852 	opl_async_flt_t opl_flt;
1853 	struct async_flt *aflt;
1854 	int trampolined = 0;
1855 	char pr_reason[MAX_REASON_STRING];
1856 	uint64_t log_sfsr;
1857 	int expected = DDI_FM_ERR_UNEXPECTED;
1858 	ddi_acc_hdl_t *hp;
1859 
1860 	/*
1861 	 * We need to look at p_flag to determine if the thread detected an
1862 	 * error while dumping core.  We can't grab p_lock here, but it's ok
1863 	 * because we just need a consistent snapshot and we know that everyone
1864 	 * else will store a consistent set of bits while holding p_lock.  We
1865 	 * don't have to worry about a race because SDOCORE is set once prior
1866 	 * to doing i/o from the process's address space and is never cleared.
1867 	 */
1868 	uint_t pflag = ttoproc(curthread)->p_flag;
1869 
1870 	pr_reason[0] = '\0';
1871 
1872 	/*
1873 	 * handle the specific error
1874 	 */
1875 	bzero(&opl_flt, sizeof (opl_async_flt_t));
1876 	aflt = (struct async_flt *)&opl_flt;
1877 	aflt->flt_id = gethrtime_waitfree();
1878 	aflt->flt_bus_id = getprocessorid();
1879 	aflt->flt_inst = CPU->cpu_id;
1880 	aflt->flt_stat = t_sfsr;
1881 	aflt->flt_addr = t_sfar;
1882 	aflt->flt_pc = (caddr_t)rp->r_pc;
1883 	aflt->flt_prot = (uchar_t)AFLT_PROT_NONE;
1884 	aflt->flt_class = (uchar_t)CPU_FAULT;
1885 	aflt->flt_priv = (uchar_t)
1886 		(tl == 1 ? 1 : ((rp->r_tstate & TSTATE_PRIV) ?  1 : 0));
1887 	aflt->flt_tl = (uchar_t)tl;
1888 	aflt->flt_panic = (uchar_t)(tl != 0 || aft_testfatal != 0 ||
1889 	    (t_sfsr & (SFSR_TLB_MUL|SFSR_TLB_PRT)) != 0);
1890 	aflt->flt_core = (pflag & SDOCORE) ? 1 : 0;
1891 	aflt->flt_status = (derr) ? OPL_ECC_DSYNC_TRAP : OPL_ECC_ISYNC_TRAP;
1892 
1893 	/*
1894 	 * If SFSR.FV is not set, both SFSR and SFAR/SFPAR values are uncertain.
1895 	 * So, clear all error bits to avoid mis-handling and force the system
1896 	 * panicked.
1897 	 * We skip all the procedures below down to the panic message call.
1898 	 */
1899 	if (!(t_sfsr & SFSR_FV)) {
1900 		opl_flt.flt_type = OPL_CPU_INV_SFSR;
1901 		aflt->flt_panic = 1;
1902 		aflt->flt_payload = FM_EREPORT_PAYLOAD_SYNC;
1903 		cpu_errorq_dispatch(FM_EREPORT_CPU_INV_SFSR,
1904 			(void *)&opl_flt, sizeof (opl_async_flt_t), ue_queue,
1905 			aflt->flt_panic);
1906 		fm_panic("%sErrors(s)", "invalid SFSR");
1907 	}
1908 
1909 	/*
1910 	 * If either UE and MK bit is off, this is not valid UE error.
1911 	 * If it is not valid UE error, clear UE & MK_UE bits to prevent
1912 	 * mis-handling below.
1913 	 * aflt->flt_stat keeps the original bits as a reference.
1914 	 */
1915 	if ((t_sfsr & (SFSR_MK_UE|SFSR_UE)) !=
1916 	    (SFSR_MK_UE|SFSR_UE)) {
1917 		t_sfsr &= ~(SFSR_MK_UE|SFSR_UE);
1918 	}
1919 
1920 	/*
1921 	 * If the trap occurred in privileged mode at TL=0, we need to check to
1922 	 * see if we were executing in the kernel under on_trap() or t_lofault
1923 	 * protection.  If so, modify the saved registers so that we return
1924 	 * from the trap to the appropriate trampoline routine.
1925 	 */
1926 	if (!aflt->flt_panic && aflt->flt_priv && tl == 0) {
1927 		if (curthread->t_ontrap != NULL) {
1928 			on_trap_data_t *otp = curthread->t_ontrap;
1929 
1930 			if (otp->ot_prot & OT_DATA_EC) {
1931 				aflt->flt_prot = (uchar_t)AFLT_PROT_EC;
1932 				otp->ot_trap |= (ushort_t)OT_DATA_EC;
1933 				rp->r_pc = otp->ot_trampoline;
1934 				rp->r_npc = rp->r_pc + 4;
1935 				trampolined = 1;
1936 			}
1937 
1938 			if ((t_sfsr & (SFSR_TO | SFSR_BERR)) &&
1939 			    (otp->ot_prot & OT_DATA_ACCESS)) {
1940 				aflt->flt_prot = (uchar_t)AFLT_PROT_ACCESS;
1941 				otp->ot_trap |= (ushort_t)OT_DATA_ACCESS;
1942 				rp->r_pc = otp->ot_trampoline;
1943 				rp->r_npc = rp->r_pc + 4;
1944 				trampolined = 1;
1945 				/*
1946 				 * for peeks and caut_gets errors are expected
1947 				 */
1948 				hp = (ddi_acc_hdl_t *)otp->ot_handle;
1949 				if (!hp)
1950 					expected = DDI_FM_ERR_PEEK;
1951 				else if (hp->ah_acc.devacc_attr_access ==
1952 				    DDI_CAUTIOUS_ACC)
1953 					expected = DDI_FM_ERR_EXPECTED;
1954 			}
1955 
1956 		} else if (curthread->t_lofault) {
1957 			aflt->flt_prot = AFLT_PROT_COPY;
1958 			rp->r_g1 = EFAULT;
1959 			rp->r_pc = curthread->t_lofault;
1960 			rp->r_npc = rp->r_pc + 4;
1961 			trampolined = 1;
1962 		}
1963 	}
1964 
1965 	/*
1966 	 * If we're in user mode or we're doing a protected copy, we either
1967 	 * want the ASTON code below to send a signal to the user process
1968 	 * or we want to panic if aft_panic is set.
1969 	 *
1970 	 * If we're in privileged mode and we're not doing a copy, then we
1971 	 * need to check if we've trampolined.  If we haven't trampolined,
1972 	 * we should panic.
1973 	 */
1974 	if (!aflt->flt_priv || aflt->flt_prot == AFLT_PROT_COPY) {
1975 		if (t_sfsr & (SFSR_ERRS & ~(SFSR_BERR | SFSR_TO)))
1976 			aflt->flt_panic |= aft_panic;
1977 	} else if (!trampolined) {
1978 		aflt->flt_panic = 1;
1979 	}
1980 
1981 	/*
1982 	 * If we've trampolined due to a privileged TO or BERR, or if an
1983 	 * unprivileged TO or BERR occurred, we don't want to enqueue an
1984 	 * event for that TO or BERR.  Queue all other events (if any) besides
1985 	 * the TO/BERR.
1986 	 */
1987 	log_sfsr = t_sfsr;
1988 	if (trampolined) {
1989 		log_sfsr &= ~(SFSR_TO | SFSR_BERR);
1990 	} else if (!aflt->flt_priv) {
1991 		/*
1992 		 * User mode, suppress messages if
1993 		 * cpu_berr_to_verbose is not set.
1994 		 */
1995 		if (!cpu_berr_to_verbose)
1996 			log_sfsr &= ~(SFSR_TO | SFSR_BERR);
1997 	}
1998 
1999 	if (((log_sfsr & SFSR_ERRS) &&
2000 		(cpu_queue_events(&opl_flt, pr_reason, t_sfsr) == 0)) ||
2001 	    ((t_sfsr & SFSR_ERRS) == 0)) {
2002 		opl_flt.flt_type = OPL_CPU_INV_SFSR;
2003 		aflt->flt_payload = FM_EREPORT_PAYLOAD_SYNC;
2004 		cpu_errorq_dispatch(FM_EREPORT_CPU_INV_SFSR,
2005 			(void *)&opl_flt, sizeof (opl_async_flt_t), ue_queue,
2006 			aflt->flt_panic);
2007 	}
2008 
2009 	if (t_sfsr & (SFSR_UE|SFSR_TO|SFSR_BERR)) {
2010 		cpu_run_bus_error_handlers(aflt, expected);
2011 	}
2012 
2013 	/*
2014 	 * Panic here if aflt->flt_panic has been set.  Enqueued errors will
2015 	 * be logged as part of the panic flow.
2016 	 */
2017 	if (aflt->flt_panic) {
2018 		if (pr_reason[0] == 0)
2019 			strcpy(pr_reason, "invalid SFSR ");
2020 
2021 		fm_panic("%sErrors(s)", pr_reason);
2022 	}
2023 
2024 	/*
2025 	 * If we queued an error and we are going to return from the trap and
2026 	 * the error was in user mode or inside of a copy routine, set AST flag
2027 	 * so the queue will be drained before returning to user mode.  The
2028 	 * AST processing will also act on our failure policy.
2029 	 */
2030 	if (!aflt->flt_priv || aflt->flt_prot == AFLT_PROT_COPY) {
2031 		int pcb_flag = 0;
2032 
2033 		if (t_sfsr & (SFSR_ERRS &
2034 			~(SFSR_BERR | SFSR_TO)))
2035 			pcb_flag |= ASYNC_HWERR;
2036 
2037 		if (t_sfsr & SFSR_BERR)
2038 			pcb_flag |= ASYNC_BERR;
2039 
2040 		if (t_sfsr & SFSR_TO)
2041 			pcb_flag |= ASYNC_BTO;
2042 
2043 		ttolwp(curthread)->lwp_pcb.pcb_flags |= pcb_flag;
2044 		aston(curthread);
2045 	}
2046 }
2047 
2048 /*ARGSUSED*/
2049 void
2050 opl_cpu_urgent_error(struct regs *rp, ulong_t p_ugesr, ulong_t tl)
2051 {
2052 	opl_async_flt_t opl_flt;
2053 	struct async_flt *aflt;
2054 	char pr_reason[MAX_REASON_STRING];
2055 
2056 	/* normalize tl */
2057 	tl = (tl >= 2 ? 1 : 0);
2058 	pr_reason[0] = '\0';
2059 
2060 	bzero(&opl_flt, sizeof (opl_async_flt_t));
2061 	aflt = (struct async_flt *)&opl_flt;
2062 	aflt->flt_id = gethrtime_waitfree();
2063 	aflt->flt_bus_id = getprocessorid();
2064 	aflt->flt_inst = CPU->cpu_id;
2065 	aflt->flt_stat = p_ugesr;
2066 	aflt->flt_pc = (caddr_t)rp->r_pc;
2067 	aflt->flt_class = (uchar_t)CPU_FAULT;
2068 	aflt->flt_tl = tl;
2069 	aflt->flt_priv = (uchar_t)
2070 		(tl == 1 ? 1 : ((rp->r_tstate & TSTATE_PRIV) ?  1 : 0));
2071 	aflt->flt_status = OPL_ECC_URGENT_TRAP;
2072 	aflt->flt_panic = 1;
2073 	/*
2074 	 * HW does not set mod/sid in case of urgent error.
2075 	 * So we have to set it here.
2076 	 */
2077 	opl_flt.flt_eid_mod = OPL_ERRID_CPU;
2078 	opl_flt.flt_eid_sid = aflt->flt_inst;
2079 
2080 	if (cpu_queue_events(&opl_flt, pr_reason, p_ugesr) == 0) {
2081 		opl_flt.flt_type = OPL_CPU_INV_UGESR;
2082 		aflt->flt_payload = FM_EREPORT_PAYLOAD_URGENT;
2083 		cpu_errorq_dispatch(FM_EREPORT_CPU_INV_URG,
2084 			(void *)&opl_flt, sizeof (opl_async_flt_t),
2085 			ue_queue, aflt->flt_panic);
2086 	}
2087 
2088 	fm_panic("Urgent Error");
2089 }
2090 
2091 /*
2092  * Initialization error counters resetting.
2093  */
2094 /* ARGSUSED */
2095 static void
2096 opl_ras_online(void *arg, cpu_t *cp, cyc_handler_t *hdlr, cyc_time_t *when)
2097 {
2098 	hdlr->cyh_func = (cyc_func_t)ras_cntr_reset;
2099 	hdlr->cyh_level = CY_LOW_LEVEL;
2100 	hdlr->cyh_arg = (void *)(uintptr_t)cp->cpu_id;
2101 
2102 	when->cyt_when = cp->cpu_id * (((hrtime_t)NANOSEC * 10)/ NCPU);
2103 	when->cyt_interval = (hrtime_t)NANOSEC * opl_async_check_interval;
2104 }
2105 
2106 void
2107 cpu_mp_init(void)
2108 {
2109 	cyc_omni_handler_t hdlr;
2110 
2111 	hdlr.cyo_online = opl_ras_online;
2112 	hdlr.cyo_offline = NULL;
2113 	hdlr.cyo_arg = NULL;
2114 	mutex_enter(&cpu_lock);
2115 	(void) cyclic_add_omni(&hdlr);
2116 	mutex_exit(&cpu_lock);
2117 }
2118 
2119 int heaplp_use_stlb = 0;
2120 
2121 void
2122 mmu_init_kernel_pgsz(struct hat *hat)
2123 {
2124 	uint_t tte = page_szc(segkmem_lpsize);
2125 	uchar_t new_cext_primary, new_cext_nucleus;
2126 
2127 	if (heaplp_use_stlb == 0) {
2128 		/* do not reprogram stlb */
2129 		tte = TTE8K;
2130 	}
2131 
2132 	new_cext_nucleus = TAGACCEXT_MKSZPAIR(tte, TTE8K);
2133 	new_cext_primary = TAGACCEXT_MKSZPAIR(TTE8K, tte);
2134 
2135 	hat->sfmmu_cext = new_cext_primary;
2136 	kcontextreg = ((uint64_t)new_cext_nucleus << CTXREG_NEXT_SHIFT) |
2137 		((uint64_t)new_cext_primary << CTXREG_EXT_SHIFT);
2138 }
2139 
2140 size_t
2141 mmu_get_kernel_lpsize(size_t lpsize)
2142 {
2143 	uint_t tte;
2144 
2145 	if (lpsize == 0) {
2146 		/* no setting for segkmem_lpsize in /etc/system: use default */
2147 		return (MMU_PAGESIZE4M);
2148 	}
2149 
2150 	for (tte = TTE8K; tte <= TTE4M; tte++) {
2151 		if (lpsize == TTEBYTES(tte))
2152 			return (lpsize);
2153 	}
2154 
2155 	return (TTEBYTES(TTE8K));
2156 }
2157 
2158 /*
2159  * The following are functions that are unused in
2160  * OPL cpu module. They are defined here to resolve
2161  * dependencies in the "unix" module.
2162  * Unused functions that should never be called in
2163  * OPL are coded with ASSERT(0).
2164  */
2165 
2166 void
2167 cpu_disable_errors(void)
2168 {}
2169 
2170 void
2171 cpu_enable_errors(void)
2172 { ASSERT(0); }
2173 
2174 /*ARGSUSED*/
2175 void
2176 cpu_ce_scrub_mem_err(struct async_flt *ecc, boolean_t t)
2177 { ASSERT(0); }
2178 
2179 /*ARGSUSED*/
2180 void
2181 cpu_faulted_enter(struct cpu *cp)
2182 {}
2183 
2184 /*ARGSUSED*/
2185 void
2186 cpu_faulted_exit(struct cpu *cp)
2187 {}
2188 
2189 /*ARGSUSED*/
2190 void
2191 cpu_check_allcpus(struct async_flt *aflt)
2192 {}
2193 
2194 /*ARGSUSED*/
2195 void
2196 cpu_ce_log_err(struct async_flt *aflt, errorq_elem_t *t)
2197 { ASSERT(0); }
2198 
2199 /*ARGSUSED*/
2200 void
2201 cpu_check_ce(int flag, uint64_t pa, caddr_t va, uint_t psz)
2202 { ASSERT(0); }
2203 
2204 /*ARGSUSED*/
2205 void
2206 cpu_ce_count_unum(struct async_flt *ecc, int len, char *unum)
2207 { ASSERT(0); }
2208 
2209 /*ARGSUSED*/
2210 void
2211 cpu_busy_ecache_scrub(struct cpu *cp)
2212 {}
2213 
2214 /*ARGSUSED*/
2215 void
2216 cpu_idle_ecache_scrub(struct cpu *cp)
2217 {}
2218 
2219 /* ARGSUSED */
2220 void
2221 cpu_change_speed(uint64_t divisor, uint64_t arg2)
2222 { ASSERT(0); }
2223 
2224 void
2225 cpu_init_cache_scrub(void)
2226 {}
2227 
2228 /* ARGSUSED */
2229 int
2230 cpu_get_mem_sid(char *unum, char *buf, int buflen, int *lenp)
2231 {
2232 	if (&plat_get_mem_sid) {
2233 		return (plat_get_mem_sid(unum, buf, buflen, lenp));
2234 	} else {
2235 		return (ENOTSUP);
2236 	}
2237 }
2238 
2239 /* ARGSUSED */
2240 int
2241 cpu_get_mem_addr(char *unum, char *sid, uint64_t offset, uint64_t *addrp)
2242 {
2243 	if (&plat_get_mem_addr) {
2244 		return (plat_get_mem_addr(unum, sid, offset, addrp));
2245 	} else {
2246 		return (ENOTSUP);
2247 	}
2248 }
2249 
2250 /* ARGSUSED */
2251 int
2252 cpu_get_mem_offset(uint64_t flt_addr, uint64_t *offp)
2253 {
2254 	if (&plat_get_mem_offset) {
2255 		return (plat_get_mem_offset(flt_addr, offp));
2256 	} else {
2257 		return (ENOTSUP);
2258 	}
2259 }
2260 
2261 /*ARGSUSED*/
2262 void
2263 itlb_rd_entry(uint_t entry, tte_t *tte, uint64_t *va_tag)
2264 { ASSERT(0); }
2265 
2266 /*ARGSUSED*/
2267 void
2268 dtlb_rd_entry(uint_t entry, tte_t *tte, uint64_t *va_tag)
2269 { ASSERT(0); }
2270 
2271 /*ARGSUSED*/
2272 void
2273 read_ecc_data(struct async_flt *aflt, short verbose, short ce_err)
2274 { ASSERT(0); }
2275 
2276 /*ARGSUSED*/
2277 int
2278 ce_scrub_xdiag_recirc(struct async_flt *aflt, errorq_t *eqp,
2279     errorq_elem_t *eqep, size_t afltoffset)
2280 {
2281 	ASSERT(0);
2282 	return (0);
2283 }
2284 
2285 /*ARGSUSED*/
2286 char *
2287 flt_to_error_type(struct async_flt *aflt)
2288 {
2289 	ASSERT(0);
2290 	return (NULL);
2291 }
2292