xref: /titanic_44/usr/src/uts/sun4u/cpu/us3_cheetah.c (revision aad98a6d8e89f8f5a62a1793da807d4bc5e5b159)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 #include <sys/types.h>
30 #include <sys/systm.h>
31 #include <sys/ddi.h>
32 #include <sys/sysmacros.h>
33 #include <sys/archsystm.h>
34 #include <sys/vmsystm.h>
35 #include <sys/machparam.h>
36 #include <sys/machsystm.h>
37 #include <sys/machthread.h>
38 #include <sys/cpu.h>
39 #include <sys/cmp.h>
40 #include <sys/elf_SPARC.h>
41 #include <vm/hat_sfmmu.h>
42 #include <vm/seg_kmem.h>
43 #include <sys/cpuvar.h>
44 #include <sys/cheetahregs.h>
45 #include <sys/us3_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/prom_debug.h>
51 #include <sys/prom_plat.h>
52 #include <sys/cpu_module.h>
53 #include <sys/sysmacros.h>
54 #include <sys/intreg.h>
55 #include <sys/clock.h>
56 #include <sys/platform_module.h>
57 #include <sys/machtrap.h>
58 #include <sys/ontrap.h>
59 #include <sys/panic.h>
60 #include <sys/memlist.h>
61 #include <sys/bootconf.h>
62 #include <sys/ivintr.h>
63 #include <sys/atomic.h>
64 #include <sys/fm/protocol.h>
65 #include <sys/fm/cpu/UltraSPARC-III.h>
66 #include <vm/vm_dep.h>
67 
68 #ifdef	CHEETAHPLUS_ERRATUM_25
69 #include <sys/cyclic.h>
70 #endif	/* CHEETAHPLUS_ERRATUM_25 */
71 
72 /*
73  * Setup trap handlers.
74  */
75 void
76 cpu_init_trap(void)
77 {
78 	CH_SET_TRAP(tt_pil15, ch_pil15_interrupt_instr);
79 
80 	CH_SET_TRAP(tt0_fecc, fecc_err_instr);
81 	CH_SET_TRAP(tt1_fecc, fecc_err_tl1_instr);
82 	CH_SET_TRAP(tt1_swtrap0, fecc_err_tl1_cont_instr);
83 }
84 
85 static int
86 getintprop(pnode_t node, char *name, int deflt)
87 {
88 	int	value;
89 
90 	switch (prom_getproplen(node, name)) {
91 	case sizeof (int):
92 		(void) prom_getprop(node, name, (caddr_t)&value);
93 		break;
94 
95 	default:
96 		value = deflt;
97 		break;
98 	}
99 
100 	return (value);
101 }
102 
103 /*
104  * Set the magic constants of the implementation.
105  */
106 /*ARGSUSED*/
107 void
108 cpu_fiximp(pnode_t dnode)
109 {
110 	int i, a;
111 
112 	static struct {
113 		char	*name;
114 		int	*var;
115 		int	defval;
116 	} prop[] = {
117 		"dcache-size", &dcache_size, CH_DCACHE_SIZE,
118 		"dcache-line-size", &dcache_linesize, CH_DCACHE_LSIZE,
119 		"icache-size", &icache_size, CH_ICACHE_SIZE,
120 		"icache-line-size", &icache_linesize, CH_ICACHE_LSIZE,
121 		"ecache-size", &ecache_size, CH_ECACHE_MAX_SIZE,
122 		"ecache-line-size", &ecache_alignsize, CH_ECACHE_MAX_LSIZE,
123 		"ecache-associativity", &ecache_associativity, CH_ECACHE_NWAY
124 	};
125 
126 	extern int exec_lpg_disable, use_brk_lpg, use_stk_lpg, use_zmap_lpg;
127 
128 
129 	for (i = 0; i < sizeof (prop) / sizeof (prop[0]); i++)
130 		*prop[i].var = getintprop(dnode, prop[i].name, prop[i].defval);
131 
132 	ecache_setsize = ecache_size / ecache_associativity;
133 
134 	vac_size = CH_VAC_SIZE;
135 	vac_mask = MMU_PAGEMASK & (vac_size - 1);
136 	i = 0; a = vac_size;
137 	while (a >>= 1)
138 		++i;
139 	vac_shift = i;
140 	shm_alignment = vac_size;
141 	vac = 1;
142 
143 	/*
144 	 * Cheetah's large page support has problems with large numbers of
145 	 * large pages, so just disable large pages out-of-the-box.
146 	 */
147 	exec_lpg_disable = 1;
148 	use_brk_lpg = 0;
149 	use_stk_lpg = 0;
150 	use_zmap_lpg = 0;
151 }
152 
153 void
154 send_mondo_set(cpuset_t set)
155 {
156 	int lo, busy, nack, shipped = 0;
157 	uint16_t i, cpuids[IDSR_BN_SETS];
158 	uint64_t idsr, nackmask = 0, busymask, curnack, curbusy;
159 	uint64_t starttick, endtick, tick, lasttick;
160 #if (NCPU > IDSR_BN_SETS)
161 	int index = 0;
162 	int ncpuids = 0;
163 #endif
164 #ifdef	CHEETAHPLUS_ERRATUM_25
165 	int recovered = 0;
166 	int cpuid;
167 #endif
168 
169 	ASSERT(!CPUSET_ISNULL(set));
170 	starttick = lasttick = gettick();
171 
172 #if (NCPU <= IDSR_BN_SETS)
173 	for (i = 0; i < NCPU; i++)
174 		if (CPU_IN_SET(set, i)) {
175 			shipit(i, shipped);
176 			nackmask |= IDSR_NACK_BIT(shipped);
177 			cpuids[shipped++] = i;
178 			CPUSET_DEL(set, i);
179 			if (CPUSET_ISNULL(set))
180 				break;
181 		}
182 	CPU_STATS_ADDQ(CPU, sys, xcalls, shipped);
183 #else
184 	for (i = 0; i < NCPU; i++)
185 		if (CPU_IN_SET(set, i)) {
186 			ncpuids++;
187 
188 			/*
189 			 * Ship only to the first (IDSR_BN_SETS) CPUs.  If we
190 			 * find we have shipped to more than (IDSR_BN_SETS)
191 			 * CPUs, set "index" to the highest numbered CPU in
192 			 * the set so we can ship to other CPUs a bit later on.
193 			 */
194 			if (shipped < IDSR_BN_SETS) {
195 				shipit(i, shipped);
196 				nackmask |= IDSR_NACK_BIT(shipped);
197 				cpuids[shipped++] = i;
198 				CPUSET_DEL(set, i);
199 				if (CPUSET_ISNULL(set))
200 					break;
201 			} else
202 				index = (int)i;
203 		}
204 
205 	CPU_STATS_ADDQ(CPU, sys, xcalls, ncpuids);
206 #endif
207 
208 	busymask = IDSR_NACK_TO_BUSY(nackmask);
209 	busy = nack = 0;
210 	endtick = starttick + xc_tick_limit;
211 	for (;;) {
212 		idsr = getidsr();
213 #if (NCPU <= IDSR_BN_SETS)
214 		if (idsr == 0)
215 			break;
216 #else
217 		if (idsr == 0 && shipped == ncpuids)
218 			break;
219 #endif
220 		tick = gettick();
221 		/*
222 		 * If there is a big jump between the current tick
223 		 * count and lasttick, we have probably hit a break
224 		 * point.  Adjust endtick accordingly to avoid panic.
225 		 */
226 		if (tick > (lasttick + xc_tick_jump_limit))
227 			endtick += (tick - lasttick);
228 		lasttick = tick;
229 		if (tick > endtick) {
230 			if (panic_quiesce)
231 				return;
232 #ifdef	CHEETAHPLUS_ERRATUM_25
233 			cpuid = -1;
234 			for (i = 0; i < IDSR_BN_SETS; i++) {
235 				if (idsr & (IDSR_NACK_BIT(i) |
236 				    IDSR_BUSY_BIT(i))) {
237 					cpuid = cpuids[i];
238 					break;
239 				}
240 			}
241 			if (cheetah_sendmondo_recover && cpuid != -1 &&
242 			    recovered == 0) {
243 				if (mondo_recover(cpuid, i)) {
244 					/*
245 					 * We claimed the whole memory or
246 					 * full scan is disabled.
247 					 */
248 					recovered++;
249 				}
250 				tick = gettick();
251 				endtick = tick + xc_tick_limit;
252 				lasttick = tick;
253 				/*
254 				 * Recheck idsr
255 				 */
256 				continue;
257 			} else
258 #endif	/* CHEETAHPLUS_ERRATUM_25 */
259 			{
260 				cmn_err(CE_CONT, "send mondo timeout "
261 				    "[%d NACK %d BUSY]\nIDSR 0x%"
262 				    "" PRIx64 "  cpuids:", nack, busy, idsr);
263 				for (i = 0; i < IDSR_BN_SETS; i++) {
264 					if (idsr & (IDSR_NACK_BIT(i) |
265 					    IDSR_BUSY_BIT(i))) {
266 						cmn_err(CE_CONT, " 0x%x",
267 						    cpuids[i]);
268 					}
269 				}
270 				cmn_err(CE_CONT, "\n");
271 				cmn_err(CE_PANIC, "send_mondo_set: timeout");
272 			}
273 		}
274 		curnack = idsr & nackmask;
275 		curbusy = idsr & busymask;
276 #if (NCPU > IDSR_BN_SETS)
277 		if (shipped < ncpuids) {
278 			uint64_t cpus_left;
279 			uint16_t next = (uint16_t)index;
280 
281 			cpus_left = ~(IDSR_NACK_TO_BUSY(curnack) | curbusy) &
282 				busymask;
283 
284 			if (cpus_left) {
285 				do {
286 					/*
287 					 * Sequence through and ship to the
288 					 * remainder of the CPUs in the system
289 					 * (e.g. other than the first
290 					 * (IDSR_BN_SETS)) in reverse order.
291 					 */
292 					lo = lowbit(cpus_left) - 1;
293 					i = IDSR_BUSY_IDX(lo);
294 					shipit(next, i);
295 					shipped++;
296 					cpuids[i] = next;
297 
298 					/*
299 					 * If we've processed all the CPUs,
300 					 * exit the loop now and save
301 					 * instructions.
302 					 */
303 					if (shipped == ncpuids)
304 						break;
305 
306 					for ((index = ((int)next - 1));
307 					    index >= 0; index--)
308 						if (CPU_IN_SET(set, index)) {
309 							next = (uint16_t)index;
310 							break;
311 						}
312 
313 					cpus_left &= ~(1ull << lo);
314 				} while (cpus_left);
315 #ifdef	CHEETAHPLUS_ERRATUM_25
316 				/*
317 				 * Clear recovered because we are sending to
318 				 * a new set of targets.
319 				 */
320 				recovered = 0;
321 #endif
322 				continue;
323 			}
324 		}
325 #endif
326 		if (curbusy) {
327 			busy++;
328 			continue;
329 		}
330 
331 #ifdef SEND_MONDO_STATS
332 		{
333 			int n = gettick() - starttick;
334 			if (n < 8192)
335 				x_nack_stimes[n >> 7]++;
336 		}
337 #endif
338 		while (gettick() < (tick + sys_clock_mhz))
339 			;
340 		do {
341 			lo = lowbit(curnack) - 1;
342 			i = IDSR_NACK_IDX(lo);
343 			shipit(cpuids[i], i);
344 			curnack &= ~(1ull << lo);
345 		} while (curnack);
346 		nack++;
347 		busy = 0;
348 	}
349 #ifdef SEND_MONDO_STATS
350 	{
351 		int n = gettick() - starttick;
352 		if (n < 8192)
353 			x_set_stimes[n >> 7]++;
354 		else
355 			x_set_ltimes[(n >> 13) & 0xf]++;
356 	}
357 	x_set_cpus[shipped]++;
358 #endif
359 }
360 
361 /*
362  * Handles error logging for implementation specific error types.
363  */
364 /*ARGSUSED*/
365 int
366 cpu_impl_async_log_err(void *flt, errorq_elem_t *eqep)
367 {
368 	/* There aren't any error types which are specific to cheetah only */
369 	return (CH_ASYNC_LOG_UNKNOWN);
370 }
371 
372 /*
373  * Figure out if Ecache is direct-mapped (Cheetah or Cheetah+ with Ecache
374  * control ECCR_ASSOC bit off or 2-way (Cheetah+ with ECCR_ASSOC on).
375  * We need to do this on the fly because we may have mixed Cheetah+'s with
376  * both direct and 2-way Ecaches.
377  */
378 int
379 cpu_ecache_nway(void)
380 {
381 	return (CH_ECACHE_NWAY);
382 }
383 
384 /*
385  * Note that these are entered into the table: Fatal Errors (PERR, IERR,
386  * ISAP, EMU) first, orphaned UCU/UCC, AFAR Overwrite policy, finally IVU, IVC.
387  * Afar overwrite policy is:
388  *   UCU,UCC > UE,EDU,WDU,CPU > CE,EDC,EMC,WDC,CPC > TO,BERR
389  */
390 ecc_type_to_info_t ecc_type_to_info[] = {
391 
392 	/* Fatal Errors */
393 	C_AFSR_PERR,	"PERR ",	ECC_ALL_TRAPS,	CPU_FATAL,
394 		"PERR Fatal",
395 		FM_EREPORT_PAYLOAD_SYSTEM2,
396 		FM_EREPORT_CPU_USIII_PERR,
397 	C_AFSR_IERR,	"IERR ", 	ECC_ALL_TRAPS,	CPU_FATAL,
398 		"IERR Fatal",
399 		FM_EREPORT_PAYLOAD_SYSTEM2,
400 		FM_EREPORT_CPU_USIII_IERR,
401 	C_AFSR_ISAP,	"ISAP ",	ECC_ALL_TRAPS,	CPU_FATAL,
402 		"ISAP Fatal",
403 		FM_EREPORT_PAYLOAD_SYSTEM1,
404 		FM_EREPORT_CPU_USIII_ISAP,
405 	C_AFSR_EMU,	"EMU ",		ECC_ASYNC_TRAPS, CPU_FATAL,
406 		"EMU Fatal",
407 		FM_EREPORT_PAYLOAD_MEMORY,
408 		FM_EREPORT_CPU_USIII_EMU,
409 
410 	/* Orphaned UCC/UCU Errors */
411 	C_AFSR_UCU,	"OUCU ",	ECC_ORPH_TRAPS, CPU_ORPH,
412 		"Orphaned UCU",
413 		FM_EREPORT_PAYLOAD_L2_DATA,
414 		FM_EREPORT_CPU_USIII_UCU,
415 	C_AFSR_UCC,	"OUCC ",	ECC_ORPH_TRAPS, CPU_ORPH,
416 		"Orphaned UCC",
417 		FM_EREPORT_PAYLOAD_L2_DATA,
418 		FM_EREPORT_CPU_USIII_UCC,
419 
420 	/* UCU, UCC */
421 	C_AFSR_UCU,	"UCU ",		ECC_F_TRAP,	CPU_UE_ECACHE,
422 		"UCU",
423 		FM_EREPORT_PAYLOAD_L2_DATA,
424 		FM_EREPORT_CPU_USIII_UCU,
425 	C_AFSR_UCC,	"UCC ",		ECC_F_TRAP,	CPU_CE_ECACHE,
426 		"UCC",
427 		FM_EREPORT_PAYLOAD_L2_DATA,
428 		FM_EREPORT_CPU_USIII_UCC,
429 
430 	/* UE, EDU:ST, EDU:BLD, WDU, CPU */
431 	C_AFSR_UE,	"UE ",		ECC_ASYNC_TRAPS, CPU_UE,
432 		"Uncorrectable system bus (UE)",
433 		FM_EREPORT_PAYLOAD_MEMORY,
434 		FM_EREPORT_CPU_USIII_UE,
435 	C_AFSR_EDU,	"EDU ",		ECC_C_TRAP,	CPU_UE_ECACHE_RETIRE,
436 		"EDU:ST",
437 		FM_EREPORT_PAYLOAD_L2_DATA,
438 		FM_EREPORT_CPU_USIII_EDUST,
439 	C_AFSR_EDU,	"EDU ",		ECC_D_TRAP,	CPU_UE_ECACHE_RETIRE,
440 		"EDU:BLD",
441 		FM_EREPORT_PAYLOAD_L2_DATA,
442 		FM_EREPORT_CPU_USIII_EDUBL,
443 	C_AFSR_WDU,	"WDU ",		ECC_C_TRAP,	CPU_UE_ECACHE_RETIRE,
444 		"WDU",
445 		FM_EREPORT_PAYLOAD_L2_DATA,
446 		FM_EREPORT_CPU_USIII_WDU,
447 	C_AFSR_CPU,	"CPU ",		ECC_C_TRAP,	CPU_UE_ECACHE,
448 		"CPU",
449 		FM_EREPORT_PAYLOAD_L2_DATA,
450 		FM_EREPORT_CPU_USIII_CPU,
451 
452 	/* CE, EDC, EMC, WDC, CPC */
453 	C_AFSR_CE,	"CE ",		ECC_C_TRAP,	CPU_CE,
454 		"Corrected system bus (CE)",
455 		FM_EREPORT_PAYLOAD_MEMORY,
456 		FM_EREPORT_CPU_USIII_CE,
457 	C_AFSR_EDC,	"EDC ",		ECC_C_TRAP,	CPU_CE_ECACHE,
458 		"EDC",
459 		FM_EREPORT_PAYLOAD_L2_DATA,
460 		FM_EREPORT_CPU_USIII_EDC,
461 	C_AFSR_EMC,	"EMC ",		ECC_C_TRAP,	CPU_EMC,
462 		"EMC",
463 		FM_EREPORT_PAYLOAD_MEMORY,
464 		FM_EREPORT_CPU_USIII_EMC,
465 	C_AFSR_WDC,	"WDC ",		ECC_C_TRAP,	CPU_CE_ECACHE,
466 		"WDC",
467 		FM_EREPORT_PAYLOAD_L2_DATA,
468 		FM_EREPORT_CPU_USIII_WDC,
469 	C_AFSR_CPC,	"CPC ",		ECC_C_TRAP,	CPU_CE_ECACHE,
470 		"CPC",
471 		FM_EREPORT_PAYLOAD_L2_DATA,
472 		FM_EREPORT_CPU_USIII_CPC,
473 
474 	/* TO, BERR */
475 	C_AFSR_TO,	"TO ",		ECC_ASYNC_TRAPS, CPU_TO,
476 		"Timeout (TO)",
477 		FM_EREPORT_PAYLOAD_IO,
478 		FM_EREPORT_CPU_USIII_TO,
479 	C_AFSR_BERR,	"BERR ",	ECC_ASYNC_TRAPS, CPU_BERR,
480 		"Bus Error (BERR)",
481 		FM_EREPORT_PAYLOAD_IO,
482 		FM_EREPORT_CPU_USIII_BERR,
483 
484 	/* IVU, IVC */
485 	C_AFSR_IVU,	"IVU ",		ECC_C_TRAP,	CPU_IV,
486 		"IVU",
487 		FM_EREPORT_PAYLOAD_SYSTEM1,
488 		FM_EREPORT_CPU_USIII_IVU,
489 	C_AFSR_IVC,	"IVC ",		ECC_C_TRAP,	CPU_IV,
490 		"IVC",
491 		FM_EREPORT_PAYLOAD_SYSTEM1,
492 		FM_EREPORT_CPU_USIII_IVC,
493 
494 	0,		NULL,		0,		0,
495 		NULL,
496 		FM_EREPORT_PAYLOAD_UNKNOWN,
497 		FM_EREPORT_CPU_USIII_UNKNOWN,
498 };
499 
500 /*
501  * Prioritized list of Error bits for AFAR overwrite.
502  * See Cheetah PRM P.6.1
503  *   Class 4:  UCC, UCU
504  *   Class 3:  UE, EDU, EMU, WDU, CPU
505  *   Class 2:  CE, EDC, EMC, WDC, CPC
506  *   Class 1:  TO, BERR
507  */
508 uint64_t afar_overwrite[] = {
509 	C_AFSR_UCC | C_AFSR_UCU,
510 	C_AFSR_UE | C_AFSR_EDU | C_AFSR_EMU | C_AFSR_WDU | C_AFSR_CPU,
511 	C_AFSR_CE | C_AFSR_EDC | C_AFSR_EMC | C_AFSR_WDC | C_AFSR_CPC,
512 	C_AFSR_TO | C_AFSR_BERR,
513 	0
514 };
515 
516 /*
517  * Prioritized list of Error bits for ESYND overwrite.
518  * See Cheetah PRM P.6.2
519  *   Class 2:  UE, IVU, EDU, WDU, UCU, CPU
520  *   Class 1:  CE, IVC, EDC, WDC, UCC, CPC
521  */
522 uint64_t esynd_overwrite[] = {
523 	C_AFSR_UE | C_AFSR_IVU | C_AFSR_EDU | C_AFSR_WDU | C_AFSR_UCU |
524 	    C_AFSR_CPU,
525 	C_AFSR_CE | C_AFSR_IVC | C_AFSR_EDC | C_AFSR_WDC | C_AFSR_UCC |
526 	    C_AFSR_CPC,
527 	0
528 };
529 
530 /*
531  * Prioritized list of Error bits for MSYND overwrite.
532  * See Cheetah PRM P.6.3
533  *   Class 2:  EMU
534  *   Class 1:  EMC
535  */
536 uint64_t msynd_overwrite[] = {
537 	C_AFSR_EMU,
538 	C_AFSR_EMC,
539 	0
540 };
541 
542 /*
543  * change cpu speed bits -- new speed will be normal-speed/divisor.
544  *
545  * The Jalapeno memory controllers are required to drain outstanding
546  * memory transactions within 32 JBus clocks in order to be ready
547  * to enter Estar mode.  In some corner cases however, that time
548  * fell short.
549  *
550  * A safe software solution is to force MCU to act like in Estar mode,
551  * then delay 1us (in ppm code) prior to assert J_CHNG_L signal.
552  * To reverse the effect, upon exiting Estar, software restores the
553  * MCU to its original state.
554  */
555 /* ARGSUSED1 */
556 void
557 cpu_change_speed(uint64_t divisor, uint64_t arg2)
558 {
559 	bus_config_eclk_t *bceclk;
560 	uint64_t		reg;
561 
562 	for (bceclk = bus_config_eclk; bceclk->divisor; bceclk++) {
563 		if (bceclk->divisor != divisor)
564 			continue;
565 		reg = get_safari_config();
566 		reg &= ~SAFARI_CONFIG_ECLK_MASK;
567 		reg |= bceclk->mask;
568 		set_safari_config(reg);
569 		CPU->cpu_m.divisor = (uchar_t)divisor;
570 		return;
571 	}
572 	/*
573 	 * We will reach here only if OBP and kernel don't agree on
574 	 * the speeds supported by the CPU.
575 	 */
576 	cmn_err(CE_WARN, "cpu_change_speed: bad divisor %" PRIu64, divisor);
577 }
578 
579 /*
580  * Cpu private initialization.  This includes allocating the cpu_private
581  * data structure, initializing it, and initializing the scrubber for this
582  * cpu.  This function calls cpu_init_ecache_scrub_dr to init the scrubber.
583  * We use kmem_cache_create for the cheetah private data structure because
584  * it needs to be allocated on a PAGESIZE (8192) byte boundary.
585  */
586 void
587 cpu_init_private(struct cpu *cp)
588 {
589 	cheetah_private_t *chprp;
590 	int i;
591 
592 	ASSERT(CPU_PRIVATE(cp) == NULL);
593 
594 	/* LINTED: E_TRUE_LOGICAL_EXPR */
595 	ASSERT((offsetof(cheetah_private_t, chpr_tl1_err_data) +
596 	    sizeof (ch_err_tl1_data_t) * CH_ERR_TL1_TLMAX) <= PAGESIZE);
597 
598 	/*
599 	 * Running with a Cheetah+, Jaguar, or Panther on a Cheetah CPU
600 	 * machine is not a supported configuration. Attempting to do so
601 	 * may result in unpredictable failures (e.g. running Cheetah+
602 	 * CPUs with Cheetah E$ disp flush) so don't allow it.
603 	 *
604 	 * This is just defensive code since this configuration mismatch
605 	 * should have been caught prior to OS execution.
606 	 */
607 	if (!IS_CHEETAH(cpunodes[cp->cpu_id].implementation)) {
608 		cmn_err(CE_PANIC, "CPU%d: UltraSPARC-III+/IV/IV+ not"
609 		    " supported on UltraSPARC-III code\n", cp->cpu_id);
610 	}
611 
612 	/*
613 	 * If the ch_private_cache has not been created, create it.
614 	 */
615 	if (ch_private_cache == NULL) {
616 		ch_private_cache = kmem_cache_create("ch_private_cache",
617 		    sizeof (cheetah_private_t), PAGESIZE, NULL, NULL,
618 		    NULL, NULL, static_arena, 0);
619 	}
620 
621 	chprp = CPU_PRIVATE(cp) = kmem_cache_alloc(ch_private_cache, KM_SLEEP);
622 
623 	bzero(chprp, sizeof (cheetah_private_t));
624 	chprp->chpr_fecctl0_logout.clo_data.chd_afar = LOGOUT_INVALID;
625 	chprp->chpr_cecc_logout.clo_data.chd_afar = LOGOUT_INVALID;
626 	chprp->chpr_async_logout.clo_data.chd_afar = LOGOUT_INVALID;
627 	for (i = 0; i < CH_ERR_TL1_TLMAX; i++)
628 		chprp->chpr_tl1_err_data[i].ch_err_tl1_logout.clo_data.chd_afar
629 		    = LOGOUT_INVALID;
630 
631 	chprp->chpr_icache_size = CH_ICACHE_SIZE;
632 	chprp->chpr_icache_linesize = CH_ICACHE_LSIZE;
633 
634 	cpu_init_ecache_scrub_dr(cp);
635 
636 	chprp->chpr_ec_set_size = cpunodes[cp->cpu_id].ecache_size /
637 	    cpu_ecache_nway();
638 
639 	adjust_hw_copy_limits(cpunodes[cp->cpu_id].ecache_size);
640 	ch_err_tl1_paddrs[cp->cpu_id] = va_to_pa(chprp);
641 	ASSERT(ch_err_tl1_paddrs[cp->cpu_id] != -1);
642 }
643 
644 /*
645  * Clear the error state registers for this CPU.
646  * For Cheetah, just clear the AFSR
647  */
648 void
649 set_cpu_error_state(ch_cpu_errors_t *cpu_error_regs)
650 {
651 	set_asyncflt(cpu_error_regs->afsr & ~C_AFSR_FATAL_ERRS);
652 }
653 
654 /*
655  * For Cheetah, the error recovery code uses an alternate flush area in the
656  * TL>0 fast ECC handler.  ecache_tl1_flushaddr is the physical address of
657  * this exclusive displacement flush area.
658  */
659 uint64_t ecache_tl1_flushaddr = (uint64_t)-1; /* physaddr for E$ flushing */
660 
661 /*
662  * Allocate and initialize the exclusive displacement flush area.
663  * Must be called before startup_bop_gone().
664  */
665 caddr_t
666 ecache_init_scrub_flush_area(caddr_t alloc_base)
667 {
668 	unsigned size = 2 * CH_ECACHE_8M_SIZE;
669 	caddr_t tmp_alloc_base = alloc_base;
670 	caddr_t flush_alloc_base =
671 	    (caddr_t)roundup((uintptr_t)alloc_base, size);
672 	caddr_t ecache_tl1_virtaddr;
673 
674 	/*
675 	 * Allocate the physical memory for the exclusive flush area
676 	 *
677 	 * Need to allocate an exclusive flush area that is twice the
678 	 * largest supported E$ size, physically contiguous, and
679 	 * aligned on twice the largest E$ size boundary.
680 	 *
681 	 * Memory allocated via BOP_ALLOC is included in the "cage"
682 	 * from the DR perspective and due to this, its physical
683 	 * address will never change and the memory will not be
684 	 * removed.
685 	 *
686 	 * BOP_ALLOC takes 4 arguments: bootops, virtual address hint,
687 	 * size of the area to allocate, and alignment of the area to
688 	 * allocate. It returns zero if the allocation fails, or the
689 	 * virtual address for a successful allocation. Memory BOP_ALLOC'd
690 	 * is physically contiguous.
691 	 */
692 	if ((ecache_tl1_virtaddr = (caddr_t)BOP_ALLOC(bootops,
693 	    flush_alloc_base, size, size)) != NULL) {
694 
695 		tmp_alloc_base =
696 		    (caddr_t)roundup((uintptr_t)(ecache_tl1_virtaddr + size),
697 		    ecache_alignsize);
698 
699 		/*
700 		 * get the physical address of the exclusive flush area
701 		 */
702 		ecache_tl1_flushaddr = va_to_pa(ecache_tl1_virtaddr);
703 
704 	} else {
705 		ecache_tl1_virtaddr = (caddr_t)-1;
706 		cmn_err(CE_NOTE, "!ecache_init_scrub_flush_area failed\n");
707 	}
708 
709 	return (tmp_alloc_base);
710 }
711 
712 /*
713  * Update cpu_offline_set so the scrubber knows which cpus are offline
714  */
715 /*ARGSUSED*/
716 int
717 cpu_scrub_cpu_setup(cpu_setup_t what, int cpuid, void *arg)
718 {
719 	switch (what) {
720 	case CPU_ON:
721 	case CPU_INIT:
722 		CPUSET_DEL(cpu_offline_set, cpuid);
723 		break;
724 	case CPU_OFF:
725 		CPUSET_ADD(cpu_offline_set, cpuid);
726 		break;
727 	default:
728 		break;
729 	}
730 	return (0);
731 }
732