xref: /titanic_50/usr/src/uts/sun4u/cpu/us3_cheetah.c (revision 5bc515a198c7f6b3c0036a1b6c10ac9f2119439d)
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 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #include <sys/types.h>
27 #include <sys/systm.h>
28 #include <sys/ddi.h>
29 #include <sys/sysmacros.h>
30 #include <sys/archsystm.h>
31 #include <sys/vmsystm.h>
32 #include <sys/machparam.h>
33 #include <sys/machsystm.h>
34 #include <sys/machthread.h>
35 #include <sys/cpu.h>
36 #include <sys/cmp.h>
37 #include <sys/elf_SPARC.h>
38 #include <vm/hat_sfmmu.h>
39 #include <vm/seg_kmem.h>
40 #include <sys/cpuvar.h>
41 #include <sys/cheetahregs.h>
42 #include <sys/us3_module.h>
43 #include <sys/async.h>
44 #include <sys/cmn_err.h>
45 #include <sys/debug.h>
46 #include <sys/dditypes.h>
47 #include <sys/prom_debug.h>
48 #include <sys/prom_plat.h>
49 #include <sys/cpu_module.h>
50 #include <sys/sysmacros.h>
51 #include <sys/intreg.h>
52 #include <sys/clock.h>
53 #include <sys/platform_module.h>
54 #include <sys/machtrap.h>
55 #include <sys/ontrap.h>
56 #include <sys/panic.h>
57 #include <sys/memlist.h>
58 #include <sys/bootconf.h>
59 #include <sys/ivintr.h>
60 #include <sys/atomic.h>
61 #include <sys/fm/protocol.h>
62 #include <sys/fm/cpu/UltraSPARC-III.h>
63 #include <vm/vm_dep.h>
64 
65 #ifdef	CHEETAHPLUS_ERRATUM_25
66 #include <sys/cyclic.h>
67 #endif	/* CHEETAHPLUS_ERRATUM_25 */
68 
69 /*
70  * Note that 'Cheetah PRM' refers to:
71  *   SPARC V9 JPS1 Implementation Supplement: Sun UltraSPARC-III
72  */
73 
74 /*
75  * Setup trap handlers.
76  */
77 void
78 cpu_init_trap(void)
79 {
80 	CH_SET_TRAP(pil15_epilogue, ch_pil15_interrupt_instr);
81 
82 	CH_SET_TRAP(tt0_fecc, fecc_err_instr);
83 	CH_SET_TRAP(tt1_fecc, fecc_err_tl1_instr);
84 	CH_SET_TRAP(tt1_swtrap0, fecc_err_tl1_cont_instr);
85 }
86 
87 static int
88 getintprop(pnode_t node, char *name, int deflt)
89 {
90 	int	value;
91 
92 	switch (prom_getproplen(node, name)) {
93 	case sizeof (int):
94 		(void) prom_getprop(node, name, (caddr_t)&value);
95 		break;
96 
97 	default:
98 		value = deflt;
99 		break;
100 	}
101 
102 	return (value);
103 }
104 
105 /*
106  * Set the magic constants of the implementation.
107  */
108 /*ARGSUSED*/
109 void
110 cpu_fiximp(pnode_t dnode)
111 {
112 	int i, a;
113 
114 	static struct {
115 		char	*name;
116 		int	*var;
117 		int	defval;
118 	} prop[] = {
119 		"dcache-size", &dcache_size, CH_DCACHE_SIZE,
120 		"dcache-line-size", &dcache_linesize, CH_DCACHE_LSIZE,
121 		"icache-size", &icache_size, CH_ICACHE_SIZE,
122 		"icache-line-size", &icache_linesize, CH_ICACHE_LSIZE,
123 		"ecache-size", &ecache_size, CH_ECACHE_MAX_SIZE,
124 		"ecache-line-size", &ecache_alignsize, CH_ECACHE_MAX_LSIZE,
125 		"ecache-associativity", &ecache_associativity, CH_ECACHE_NWAY
126 	};
127 
128 	for (i = 0; i < sizeof (prop) / sizeof (prop[0]); i++)
129 		*prop[i].var = getintprop(dnode, prop[i].name, prop[i].defval);
130 
131 	ecache_setsize = ecache_size / ecache_associativity;
132 
133 	vac_size = CH_VAC_SIZE;
134 	vac_mask = MMU_PAGEMASK & (vac_size - 1);
135 	i = 0; a = vac_size;
136 	while (a >>= 1)
137 		++i;
138 	vac_shift = i;
139 	shm_alignment = vac_size;
140 	vac = 1;
141 
142 	/*
143 	 * Cheetah's large page support has problems with large numbers of
144 	 * large pages, so just disable large pages out-of-the-box.
145 	 * Note that the other defaults are set in sun4u/vm/mach_vm_dep.c.
146 	 */
147 	max_uheap_lpsize = MMU_PAGESIZE;
148 	max_ustack_lpsize = MMU_PAGESIZE;
149 	max_privmap_lpsize = MMU_PAGESIZE;
150 	max_utext_lpsize = MMU_PAGESIZE;
151 	max_shm_lpsize = MMU_PAGESIZE;
152 	max_bootlp_tteszc = TTE8K;
153 }
154 
155 void
156 send_mondo_set(cpuset_t set)
157 {
158 	int lo, busy, nack, shipped = 0;
159 	uint16_t i, cpuids[IDSR_BN_SETS];
160 	uint64_t idsr, nackmask = 0, busymask, curnack, curbusy;
161 	uint64_t starttick, endtick, tick, lasttick;
162 #if (NCPU > IDSR_BN_SETS)
163 	int index = 0;
164 	int ncpuids = 0;
165 #endif
166 #ifdef	CHEETAHPLUS_ERRATUM_25
167 	int recovered = 0;
168 	int cpuid;
169 #endif
170 
171 	ASSERT(!CPUSET_ISNULL(set));
172 	starttick = lasttick = gettick();
173 
174 #if (NCPU <= IDSR_BN_SETS)
175 	for (i = 0; i < NCPU; i++)
176 		if (CPU_IN_SET(set, i)) {
177 			shipit(i, shipped);
178 			nackmask |= IDSR_NACK_BIT(shipped);
179 			cpuids[shipped++] = i;
180 			CPUSET_DEL(set, i);
181 			if (CPUSET_ISNULL(set))
182 				break;
183 		}
184 	CPU_STATS_ADDQ(CPU, sys, xcalls, shipped);
185 #else
186 	for (i = 0; i < NCPU; i++)
187 		if (CPU_IN_SET(set, i)) {
188 			ncpuids++;
189 
190 			/*
191 			 * Ship only to the first (IDSR_BN_SETS) CPUs.  If we
192 			 * find we have shipped to more than (IDSR_BN_SETS)
193 			 * CPUs, set "index" to the highest numbered CPU in
194 			 * the set so we can ship to other CPUs a bit later on.
195 			 */
196 			if (shipped < IDSR_BN_SETS) {
197 				shipit(i, shipped);
198 				nackmask |= IDSR_NACK_BIT(shipped);
199 				cpuids[shipped++] = i;
200 				CPUSET_DEL(set, i);
201 				if (CPUSET_ISNULL(set))
202 					break;
203 			} else
204 				index = (int)i;
205 		}
206 
207 	CPU_STATS_ADDQ(CPU, sys, xcalls, ncpuids);
208 #endif
209 
210 	busymask = IDSR_NACK_TO_BUSY(nackmask);
211 	busy = nack = 0;
212 	endtick = starttick + xc_tick_limit;
213 	for (;;) {
214 		idsr = getidsr();
215 #if (NCPU <= IDSR_BN_SETS)
216 		if (idsr == 0)
217 			break;
218 #else
219 		if (idsr == 0 && shipped == ncpuids)
220 			break;
221 #endif
222 		tick = gettick();
223 		/*
224 		 * If there is a big jump between the current tick
225 		 * count and lasttick, we have probably hit a break
226 		 * point.  Adjust endtick accordingly to avoid panic.
227 		 */
228 		if (tick > (lasttick + xc_tick_jump_limit))
229 			endtick += (tick - lasttick);
230 		lasttick = tick;
231 		if (tick > endtick) {
232 			if (panic_quiesce)
233 				return;
234 #ifdef	CHEETAHPLUS_ERRATUM_25
235 			cpuid = -1;
236 			for (i = 0; i < IDSR_BN_SETS; i++) {
237 				if (idsr & (IDSR_NACK_BIT(i) |
238 				    IDSR_BUSY_BIT(i))) {
239 					cpuid = cpuids[i];
240 					break;
241 				}
242 			}
243 			if (cheetah_sendmondo_recover && cpuid != -1 &&
244 			    recovered == 0) {
245 				if (mondo_recover(cpuid, i)) {
246 					/*
247 					 * We claimed the whole memory or
248 					 * full scan is disabled.
249 					 */
250 					recovered++;
251 				}
252 				tick = gettick();
253 				endtick = tick + xc_tick_limit;
254 				lasttick = tick;
255 				/*
256 				 * Recheck idsr
257 				 */
258 				continue;
259 			} else
260 #endif	/* CHEETAHPLUS_ERRATUM_25 */
261 			{
262 				cmn_err(CE_CONT, "send mondo timeout "
263 				    "[%d NACK %d BUSY]\nIDSR 0x%"
264 				    "" PRIx64 "  cpuids:", nack, busy, idsr);
265 				for (i = 0; i < IDSR_BN_SETS; i++) {
266 					if (idsr & (IDSR_NACK_BIT(i) |
267 					    IDSR_BUSY_BIT(i))) {
268 						cmn_err(CE_CONT, " 0x%x",
269 						    cpuids[i]);
270 					}
271 				}
272 				cmn_err(CE_CONT, "\n");
273 				cmn_err(CE_PANIC, "send_mondo_set: timeout");
274 			}
275 		}
276 		curnack = idsr & nackmask;
277 		curbusy = idsr & busymask;
278 #if (NCPU > IDSR_BN_SETS)
279 		if (shipped < ncpuids) {
280 			uint64_t cpus_left;
281 			uint16_t next = (uint16_t)index;
282 
283 			cpus_left = ~(IDSR_NACK_TO_BUSY(curnack) | curbusy) &
284 			    busymask;
285 
286 			if (cpus_left) {
287 				do {
288 					/*
289 					 * Sequence through and ship to the
290 					 * remainder of the CPUs in the system
291 					 * (e.g. other than the first
292 					 * (IDSR_BN_SETS)) in reverse order.
293 					 */
294 					lo = lowbit(cpus_left) - 1;
295 					i = IDSR_BUSY_IDX(lo);
296 					shipit(next, i);
297 					shipped++;
298 					cpuids[i] = next;
299 
300 					/*
301 					 * If we've processed all the CPUs,
302 					 * exit the loop now and save
303 					 * instructions.
304 					 */
305 					if (shipped == ncpuids)
306 						break;
307 
308 					for ((index = ((int)next - 1));
309 					    index >= 0; index--)
310 						if (CPU_IN_SET(set, index)) {
311 							next = (uint16_t)index;
312 							break;
313 						}
314 
315 					cpus_left &= ~(1ull << lo);
316 				} while (cpus_left);
317 #ifdef	CHEETAHPLUS_ERRATUM_25
318 				/*
319 				 * Clear recovered because we are sending to
320 				 * a new set of targets.
321 				 */
322 				recovered = 0;
323 #endif
324 				continue;
325 			}
326 		}
327 #endif
328 		if (curbusy) {
329 			busy++;
330 			continue;
331 		}
332 
333 #ifdef SEND_MONDO_STATS
334 		{
335 			int n = gettick() - starttick;
336 			if (n < 8192)
337 				x_nack_stimes[n >> 7]++;
338 		}
339 #endif
340 		while (gettick() < (tick + sys_clock_mhz))
341 			;
342 		do {
343 			lo = lowbit(curnack) - 1;
344 			i = IDSR_NACK_IDX(lo);
345 			shipit(cpuids[i], i);
346 			curnack &= ~(1ull << lo);
347 		} while (curnack);
348 		nack++;
349 		busy = 0;
350 	}
351 #ifdef SEND_MONDO_STATS
352 	{
353 		int n = gettick() - starttick;
354 		if (n < 8192)
355 			x_set_stimes[n >> 7]++;
356 		else
357 			x_set_ltimes[(n >> 13) & 0xf]++;
358 	}
359 	x_set_cpus[shipped]++;
360 #endif
361 }
362 
363 /*
364  * Handles error logging for implementation specific error types.
365  */
366 /*ARGSUSED*/
367 int
368 cpu_impl_async_log_err(void *flt, errorq_elem_t *eqep)
369 {
370 	/* There aren't any error types which are specific to cheetah only */
371 	return (CH_ASYNC_LOG_UNKNOWN);
372 }
373 
374 /*
375  * Figure out if Ecache is direct-mapped (Cheetah or Cheetah+ with Ecache
376  * control ECCR_ASSOC bit off or 2-way (Cheetah+ with ECCR_ASSOC on).
377  * We need to do this on the fly because we may have mixed Cheetah+'s with
378  * both direct and 2-way Ecaches.
379  */
380 int
381 cpu_ecache_nway(void)
382 {
383 	return (CH_ECACHE_NWAY);
384 }
385 
386 /*
387  * Note that these are entered into the table: Fatal Errors (PERR, IERR,
388  * ISAP, EMU) first, orphaned UCU/UCC, AFAR Overwrite policy, finally IVU, IVC.
389  * Afar overwrite policy is:
390  *   UCU,UCC > UE,EDU,WDU,CPU > CE,EDC,EMC,WDC,CPC > TO,BERR
391  */
392 ecc_type_to_info_t ecc_type_to_info[] = {
393 
394 	/* Fatal Errors */
395 	C_AFSR_PERR,	"PERR ",	ECC_ALL_TRAPS,	CPU_FATAL,
396 		"PERR Fatal",
397 		FM_EREPORT_PAYLOAD_SYSTEM2,
398 		FM_EREPORT_CPU_USIII_PERR,
399 	C_AFSR_IERR,	"IERR ", 	ECC_ALL_TRAPS,	CPU_FATAL,
400 		"IERR Fatal",
401 		FM_EREPORT_PAYLOAD_SYSTEM2,
402 		FM_EREPORT_CPU_USIII_IERR,
403 	C_AFSR_ISAP,	"ISAP ",	ECC_ALL_TRAPS,	CPU_FATAL,
404 		"ISAP Fatal",
405 		FM_EREPORT_PAYLOAD_SYSTEM1,
406 		FM_EREPORT_CPU_USIII_ISAP,
407 	C_AFSR_EMU,	"EMU ",		ECC_ASYNC_TRAPS, CPU_FATAL,
408 		"EMU Fatal",
409 		FM_EREPORT_PAYLOAD_MEMORY,
410 		FM_EREPORT_CPU_USIII_EMU,
411 
412 	/* Orphaned UCC/UCU Errors */
413 	C_AFSR_UCU,	"OUCU ",	ECC_ORPH_TRAPS, CPU_ORPH,
414 		"Orphaned UCU",
415 		FM_EREPORT_PAYLOAD_L2_DATA,
416 		FM_EREPORT_CPU_USIII_UCU,
417 	C_AFSR_UCC,	"OUCC ",	ECC_ORPH_TRAPS, CPU_ORPH,
418 		"Orphaned UCC",
419 		FM_EREPORT_PAYLOAD_L2_DATA,
420 		FM_EREPORT_CPU_USIII_UCC,
421 
422 	/* UCU, UCC */
423 	C_AFSR_UCU,	"UCU ",		ECC_F_TRAP,	CPU_UE_ECACHE,
424 		"UCU",
425 		FM_EREPORT_PAYLOAD_L2_DATA,
426 		FM_EREPORT_CPU_USIII_UCU,
427 	C_AFSR_UCC,	"UCC ",		ECC_F_TRAP,	CPU_CE_ECACHE,
428 		"UCC",
429 		FM_EREPORT_PAYLOAD_L2_DATA,
430 		FM_EREPORT_CPU_USIII_UCC,
431 
432 	/* UE, EDU:ST, EDU:BLD, WDU, CPU */
433 	C_AFSR_UE,	"UE ",		ECC_ASYNC_TRAPS, CPU_UE,
434 		"Uncorrectable system bus (UE)",
435 		FM_EREPORT_PAYLOAD_MEMORY,
436 		FM_EREPORT_CPU_USIII_UE,
437 	C_AFSR_EDU,	"EDU ",		ECC_C_TRAP,	CPU_UE_ECACHE_RETIRE,
438 		"EDU:ST",
439 		FM_EREPORT_PAYLOAD_L2_DATA,
440 		FM_EREPORT_CPU_USIII_EDUST,
441 	C_AFSR_EDU,	"EDU ",		ECC_D_TRAP,	CPU_UE_ECACHE_RETIRE,
442 		"EDU:BLD",
443 		FM_EREPORT_PAYLOAD_L2_DATA,
444 		FM_EREPORT_CPU_USIII_EDUBL,
445 	C_AFSR_WDU,	"WDU ",		ECC_C_TRAP,	CPU_UE_ECACHE_RETIRE,
446 		"WDU",
447 		FM_EREPORT_PAYLOAD_L2_DATA,
448 		FM_EREPORT_CPU_USIII_WDU,
449 	C_AFSR_CPU,	"CPU ",		ECC_C_TRAP,	CPU_UE_ECACHE,
450 		"CPU",
451 		FM_EREPORT_PAYLOAD_L2_DATA,
452 		FM_EREPORT_CPU_USIII_CPU,
453 
454 	/* CE, EDC, EMC, WDC, CPC */
455 	C_AFSR_CE,	"CE ",		ECC_C_TRAP,	CPU_CE,
456 		"Corrected system bus (CE)",
457 		FM_EREPORT_PAYLOAD_MEMORY,
458 		FM_EREPORT_CPU_USIII_CE,
459 	C_AFSR_EDC,	"EDC ",		ECC_C_TRAP,	CPU_CE_ECACHE,
460 		"EDC",
461 		FM_EREPORT_PAYLOAD_L2_DATA,
462 		FM_EREPORT_CPU_USIII_EDC,
463 	C_AFSR_EMC,	"EMC ",		ECC_C_TRAP,	CPU_EMC,
464 		"EMC",
465 		FM_EREPORT_PAYLOAD_MEMORY,
466 		FM_EREPORT_CPU_USIII_EMC,
467 	C_AFSR_WDC,	"WDC ",		ECC_C_TRAP,	CPU_CE_ECACHE,
468 		"WDC",
469 		FM_EREPORT_PAYLOAD_L2_DATA,
470 		FM_EREPORT_CPU_USIII_WDC,
471 	C_AFSR_CPC,	"CPC ",		ECC_C_TRAP,	CPU_CE_ECACHE,
472 		"CPC",
473 		FM_EREPORT_PAYLOAD_L2_DATA,
474 		FM_EREPORT_CPU_USIII_CPC,
475 
476 	/* TO, BERR */
477 	C_AFSR_TO,	"TO ",		ECC_ASYNC_TRAPS, CPU_TO,
478 		"Timeout (TO)",
479 		FM_EREPORT_PAYLOAD_IO,
480 		FM_EREPORT_CPU_USIII_TO,
481 	C_AFSR_BERR,	"BERR ",	ECC_ASYNC_TRAPS, CPU_BERR,
482 		"Bus Error (BERR)",
483 		FM_EREPORT_PAYLOAD_IO,
484 		FM_EREPORT_CPU_USIII_BERR,
485 
486 	/* IVU, IVC */
487 	C_AFSR_IVU,	"IVU ",		ECC_C_TRAP,	CPU_IV,
488 		"IVU",
489 		FM_EREPORT_PAYLOAD_SYSTEM1,
490 		FM_EREPORT_CPU_USIII_IVU,
491 	C_AFSR_IVC,	"IVC ",		ECC_C_TRAP,	CPU_IV,
492 		"IVC",
493 		FM_EREPORT_PAYLOAD_SYSTEM1,
494 		FM_EREPORT_CPU_USIII_IVC,
495 
496 	0,		NULL,		0,		0,
497 		NULL,
498 		FM_EREPORT_PAYLOAD_UNKNOWN,
499 		FM_EREPORT_CPU_USIII_UNKNOWN,
500 };
501 
502 /*
503  * Prioritized list of Error bits for AFAR overwrite.
504  * See Cheetah PRM P.6.1
505  *   Class 4:  UCC, UCU
506  *   Class 3:  UE, EDU, EMU, WDU, CPU
507  *   Class 2:  CE, EDC, EMC, WDC, CPC
508  *   Class 1:  TO, BERR
509  */
510 uint64_t afar_overwrite[] = {
511 	C_AFSR_UCC | C_AFSR_UCU,
512 	C_AFSR_UE | C_AFSR_EDU | C_AFSR_EMU | C_AFSR_WDU | C_AFSR_CPU,
513 	C_AFSR_CE | C_AFSR_EDC | C_AFSR_EMC | C_AFSR_WDC | C_AFSR_CPC,
514 	C_AFSR_TO | C_AFSR_BERR,
515 	0
516 };
517 
518 /*
519  * Prioritized list of Error bits for ESYND overwrite.
520  * See Cheetah PRM P.6.2
521  *   Class 2:  UE, IVU, EDU, WDU, UCU, CPU
522  *   Class 1:  CE, IVC, EDC, WDC, UCC, CPC
523  */
524 uint64_t esynd_overwrite[] = {
525 	C_AFSR_UE | C_AFSR_IVU | C_AFSR_EDU | C_AFSR_WDU | C_AFSR_UCU |
526 	    C_AFSR_CPU,
527 	C_AFSR_CE | C_AFSR_IVC | C_AFSR_EDC | C_AFSR_WDC | C_AFSR_UCC |
528 	    C_AFSR_CPC,
529 	0
530 };
531 
532 /*
533  * Prioritized list of Error bits for MSYND overwrite.
534  * See Cheetah PRM P.6.3
535  *   Class 2:  EMU
536  *   Class 1:  EMC
537  */
538 uint64_t msynd_overwrite[] = {
539 	C_AFSR_EMU,
540 	C_AFSR_EMC,
541 	0
542 };
543 
544 /*
545  * change cpu speed bits -- new speed will be normal-speed/divisor.
546  *
547  * The Jalapeno memory controllers are required to drain outstanding
548  * memory transactions within 32 JBus clocks in order to be ready
549  * to enter Estar mode.  In some corner cases however, that time
550  * fell short.
551  *
552  * A safe software solution is to force MCU to act like in Estar mode,
553  * then delay 1us (in ppm code) prior to assert J_CHNG_L signal.
554  * To reverse the effect, upon exiting Estar, software restores the
555  * MCU to its original state.
556  */
557 /* ARGSUSED1 */
558 void
559 cpu_change_speed(uint64_t divisor, uint64_t arg2)
560 {
561 	bus_config_eclk_t	*bceclk;
562 	uint64_t		reg;
563 	processor_info_t	*pi = &(CPU->cpu_type_info);
564 
565 	for (bceclk = bus_config_eclk; bceclk->divisor; bceclk++) {
566 		if (bceclk->divisor != divisor)
567 			continue;
568 		reg = get_safari_config();
569 		reg &= ~SAFARI_CONFIG_ECLK_MASK;
570 		reg |= bceclk->mask;
571 		set_safari_config(reg);
572 		CPU->cpu_m.divisor = (uchar_t)divisor;
573 		cpu_set_curr_clock(((uint64_t)pi->pi_clock * 1000000) /
574 		    divisor);
575 		return;
576 	}
577 	/*
578 	 * We will reach here only if OBP and kernel don't agree on
579 	 * the speeds supported by the CPU.
580 	 */
581 	cmn_err(CE_WARN, "cpu_change_speed: bad divisor %" PRIu64, divisor);
582 }
583 
584 /*
585  * Cpu private initialization.  This includes allocating the cpu_private
586  * data structure, initializing it, and initializing the scrubber for this
587  * cpu.  This function calls cpu_init_ecache_scrub_dr to init the scrubber.
588  * We use kmem_cache_create for the cheetah private data structure because
589  * it needs to be allocated on a PAGESIZE (8192) byte boundary.
590  */
591 void
592 cpu_init_private(struct cpu *cp)
593 {
594 	cheetah_private_t *chprp;
595 	int i;
596 
597 	ASSERT(CPU_PRIVATE(cp) == NULL);
598 
599 	/* LINTED: E_TRUE_LOGICAL_EXPR */
600 	ASSERT((offsetof(cheetah_private_t, chpr_tl1_err_data) +
601 	    sizeof (ch_err_tl1_data_t) * CH_ERR_TL1_TLMAX) <= PAGESIZE);
602 
603 	/*
604 	 * Running with a Cheetah+, Jaguar, or Panther on a Cheetah CPU
605 	 * machine is not a supported configuration. Attempting to do so
606 	 * may result in unpredictable failures (e.g. running Cheetah+
607 	 * CPUs with Cheetah E$ disp flush) so don't allow it.
608 	 *
609 	 * This is just defensive code since this configuration mismatch
610 	 * should have been caught prior to OS execution.
611 	 */
612 	if (!IS_CHEETAH(cpunodes[cp->cpu_id].implementation)) {
613 		cmn_err(CE_PANIC, "CPU%d: UltraSPARC-III+/IV/IV+ not"
614 		    " supported on UltraSPARC-III code\n", cp->cpu_id);
615 	}
616 
617 	/*
618 	 * If the ch_private_cache has not been created, create it.
619 	 */
620 	if (ch_private_cache == NULL) {
621 		ch_private_cache = kmem_cache_create("ch_private_cache",
622 		    sizeof (cheetah_private_t), PAGESIZE, NULL, NULL,
623 		    NULL, NULL, static_arena, 0);
624 	}
625 
626 	chprp = CPU_PRIVATE(cp) = kmem_cache_alloc(ch_private_cache, KM_SLEEP);
627 
628 	bzero(chprp, sizeof (cheetah_private_t));
629 	chprp->chpr_fecctl0_logout.clo_data.chd_afar = LOGOUT_INVALID;
630 	chprp->chpr_cecc_logout.clo_data.chd_afar = LOGOUT_INVALID;
631 	chprp->chpr_async_logout.clo_data.chd_afar = LOGOUT_INVALID;
632 	for (i = 0; i < CH_ERR_TL1_TLMAX; i++)
633 		chprp->chpr_tl1_err_data[i].ch_err_tl1_logout.clo_data.chd_afar
634 		    = LOGOUT_INVALID;
635 
636 	chprp->chpr_icache_size = CH_ICACHE_SIZE;
637 	chprp->chpr_icache_linesize = CH_ICACHE_LSIZE;
638 
639 	cpu_init_ecache_scrub_dr(cp);
640 
641 	chprp->chpr_ec_set_size = cpunodes[cp->cpu_id].ecache_size /
642 	    cpu_ecache_nway();
643 
644 	adjust_hw_copy_limits(cpunodes[cp->cpu_id].ecache_size);
645 	ch_err_tl1_paddrs[cp->cpu_id] = va_to_pa(chprp);
646 	ASSERT(ch_err_tl1_paddrs[cp->cpu_id] != -1);
647 }
648 
649 /*
650  * Clear the error state registers for this CPU.
651  * For Cheetah, just clear the AFSR
652  */
653 void
654 set_cpu_error_state(ch_cpu_errors_t *cpu_error_regs)
655 {
656 	set_asyncflt(cpu_error_regs->afsr & ~C_AFSR_FATAL_ERRS);
657 }
658 
659 /*
660  * For Cheetah, the error recovery code uses an alternate flush area in the
661  * TL>0 fast ECC handler.  ecache_tl1_flushaddr is the physical address of
662  * this exclusive displacement flush area.
663  */
664 uint64_t ecache_tl1_flushaddr = (uint64_t)-1; /* physaddr for E$ flushing */
665 
666 /*
667  * Allocate and initialize the exclusive displacement flush area.
668  */
669 caddr_t
670 ecache_init_scrub_flush_area(caddr_t alloc_base)
671 {
672 	unsigned size = 2 * CH_ECACHE_8M_SIZE;
673 	caddr_t tmp_alloc_base = alloc_base;
674 	caddr_t flush_alloc_base =
675 	    (caddr_t)roundup((uintptr_t)alloc_base, size);
676 	caddr_t ecache_tl1_virtaddr;
677 
678 	/*
679 	 * Allocate the physical memory for the exclusive flush area
680 	 *
681 	 * Need to allocate an exclusive flush area that is twice the
682 	 * largest supported E$ size, physically contiguous, and
683 	 * aligned on twice the largest E$ size boundary.
684 	 *
685 	 * Memory allocated via prom_alloc is included in the "cage"
686 	 * from the DR perspective and due to this, its physical
687 	 * address will never change and the memory will not be
688 	 * removed.
689 	 *
690 	 * prom_alloc takes 3 arguments: bootops, virtual address hint,
691 	 * size of the area to allocate, and alignment of the area to
692 	 * allocate. It returns zero if the allocation fails, or the
693 	 * virtual address for a successful allocation. Memory prom_alloc'd
694 	 * is physically contiguous.
695 	 */
696 	if ((ecache_tl1_virtaddr =
697 	    prom_alloc(flush_alloc_base, size, size)) != NULL) {
698 
699 		tmp_alloc_base =
700 		    (caddr_t)roundup((uintptr_t)(ecache_tl1_virtaddr + size),
701 		    ecache_alignsize);
702 
703 		/*
704 		 * get the physical address of the exclusive flush area
705 		 */
706 		ecache_tl1_flushaddr = va_to_pa(ecache_tl1_virtaddr);
707 
708 	} else {
709 		ecache_tl1_virtaddr = (caddr_t)-1;
710 		cmn_err(CE_NOTE, "!ecache_init_scrub_flush_area failed\n");
711 	}
712 
713 	return (tmp_alloc_base);
714 }
715 
716 /*
717  * Update cpu_offline_set so the scrubber knows which cpus are offline
718  */
719 /*ARGSUSED*/
720 int
721 cpu_scrub_cpu_setup(cpu_setup_t what, int cpuid, void *arg)
722 {
723 	switch (what) {
724 	case CPU_ON:
725 	case CPU_INIT:
726 		CPUSET_DEL(cpu_offline_set, cpuid);
727 		break;
728 	case CPU_OFF:
729 		CPUSET_ADD(cpu_offline_set, cpuid);
730 		break;
731 	default:
732 		break;
733 	}
734 	return (0);
735 }
736