xref: /illumos-gate/usr/src/uts/sun4v/cpu/niagara2.c (revision 437220cd296f6d8b6654d6d52508b40b1e2d1ac7)
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 /*
23  * Copyright 2007 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/archsystm.h>
32 #include <sys/machparam.h>
33 #include <sys/machsystm.h>
34 #include <sys/cpu.h>
35 #include <sys/elf_SPARC.h>
36 #include <vm/hat_sfmmu.h>
37 #include <vm/page.h>
38 #include <vm/vm_dep.h>
39 #include <sys/cpuvar.h>
40 #include <sys/async.h>
41 #include <sys/cmn_err.h>
42 #include <sys/debug.h>
43 #include <sys/dditypes.h>
44 #include <sys/sunddi.h>
45 #include <sys/cpu_module.h>
46 #include <sys/prom_debug.h>
47 #include <sys/vmsystm.h>
48 #include <sys/prom_plat.h>
49 #include <sys/sysmacros.h>
50 #include <sys/intreg.h>
51 #include <sys/machtrap.h>
52 #include <sys/ontrap.h>
53 #include <sys/ivintr.h>
54 #include <sys/atomic.h>
55 #include <sys/panic.h>
56 #include <sys/dtrace.h>
57 #include <sys/simulate.h>
58 #include <sys/fault.h>
59 #include <sys/niagara2regs.h>
60 #include <sys/hsvc.h>
61 #include <sys/trapstat.h>
62 
63 uint_t root_phys_addr_lo_mask = 0xffffffffU;
64 #if defined(NIAGARA2_IMPL)
65 char cpu_module_name[] = "SUNW,UltraSPARC-T2";
66 #elif defined(VFALLS_IMPL)
67 char cpu_module_name[] = "SUNW,UltraSPARC-T2+";
68 #endif
69 
70 /*
71  * Hypervisor services information for the NIAGARA2 and Victoria Falls
72  * CPU module
73  */
74 static boolean_t cpu_hsvc_available = B_TRUE;
75 static uint64_t cpu_sup_minor;		/* Supported minor number */
76 #if defined(NIAGARA2_IMPL)
77 static hsvc_info_t cpu_hsvc = {
78 	HSVC_REV_1, NULL, HSVC_GROUP_NIAGARA2_CPU, NIAGARA2_HSVC_MAJOR,
79 	NIAGARA2_HSVC_MINOR, cpu_module_name
80 };
81 #elif defined(VFALLS_IMPL)
82 static hsvc_info_t cpu_hsvc = {
83 	HSVC_REV_1, NULL, HSVC_GROUP_VFALLS_CPU, VFALLS_HSVC_MAJOR,
84 	VFALLS_HSVC_MINOR, cpu_module_name
85 };
86 #endif
87 
88 void
89 cpu_setup(void)
90 {
91 	extern int mmu_exported_pagesize_mask;
92 	extern int cpc_has_overflow_intr;
93 	extern size_t contig_mem_prealloc_base;
94 	int status;
95 
96 	/*
97 	 * Negotiate the API version for Niagara2 specific hypervisor
98 	 * services.
99 	 */
100 	status = hsvc_register(&cpu_hsvc, &cpu_sup_minor);
101 	if (status != 0) {
102 		cmn_err(CE_WARN, "%s: cannot negotiate hypervisor services "
103 		    "group: 0x%lx major: 0x%lx minor: 0x%lx errno: %d",
104 		    cpu_hsvc.hsvc_modname, cpu_hsvc.hsvc_group,
105 		    cpu_hsvc.hsvc_major, cpu_hsvc.hsvc_minor, status);
106 		cpu_hsvc_available = B_FALSE;
107 	}
108 
109 	/*
110 	 * The setup common to all CPU modules is done in cpu_setup_common
111 	 * routine.
112 	 */
113 	cpu_setup_common(NULL);
114 
115 	cache |= (CACHE_PTAG | CACHE_IOCOHERENT);
116 
117 	if ((mmu_exported_pagesize_mask &
118 	    DEFAULT_SUN4V_MMU_PAGESIZE_MASK) !=
119 	    DEFAULT_SUN4V_MMU_PAGESIZE_MASK)
120 		cmn_err(CE_PANIC, "machine description"
121 		    " does not have required sun4v page sizes"
122 		    " 8K, 64K and 4M: MD mask is 0x%x",
123 		    mmu_exported_pagesize_mask);
124 
125 	cpu_hwcap_flags = AV_SPARC_VIS | AV_SPARC_VIS2 | AV_SPARC_ASI_BLK_INIT;
126 
127 	/*
128 	 * Niagara2 supports a 48-bit subset of the full 64-bit virtual
129 	 * address space. Virtual addresses between 0x0000800000000000
130 	 * and 0xffff.7fff.ffff.ffff inclusive lie within a "VA Hole"
131 	 * and must never be mapped. In addition, software must not use
132 	 * pages within 4GB of the VA hole as instruction pages to
133 	 * avoid problems with prefetching into the VA hole.
134 	 */
135 	hole_start = (caddr_t)((1ull << (va_bits - 1)) - (1ull << 32));
136 	hole_end = (caddr_t)((0ull - (1ull << (va_bits - 1))) + (1ull << 32));
137 
138 	/*
139 	 * Niagara2 has a performance counter overflow interrupt
140 	 */
141 	cpc_has_overflow_intr = 1;
142 
143 	/*
144 	 * Enable 4M pages for OOB.
145 	 */
146 	max_uheap_lpsize = MMU_PAGESIZE4M;
147 	max_ustack_lpsize = MMU_PAGESIZE4M;
148 	max_privmap_lpsize = MMU_PAGESIZE4M;
149 
150 	contig_mem_prealloc_base = NIAGARA2_PREALLOC_BASE;
151 }
152 
153 /*
154  * Set the magic constants of the implementation.
155  */
156 void
157 cpu_fiximp(struct cpu_node *cpunode)
158 {
159 	/*
160 	 * The Cache node is optional in MD. Therefore in case "Cache"
161 	 * node does not exists in MD, set the default L2 cache associativity,
162 	 * size, linesize.
163 	 */
164 	if (cpunode->ecache_size == 0)
165 		cpunode->ecache_size = L2CACHE_SIZE;
166 	if (cpunode->ecache_linesize == 0)
167 		cpunode->ecache_linesize = L2CACHE_LINESIZE;
168 	if (cpunode->ecache_associativity == 0)
169 		cpunode->ecache_associativity = L2CACHE_ASSOCIATIVITY;
170 }
171 
172 void
173 cpu_map_exec_units(struct cpu *cp)
174 {
175 	ASSERT(MUTEX_HELD(&cpu_lock));
176 
177 	/*
178 	 * The cpu_ipipe and cpu_fpu fields are initialized based on
179 	 * the execution unit sharing information from the MD. They
180 	 * default to the CPU id in the absence of such information.
181 	 */
182 	cp->cpu_m.cpu_ipipe = cpunodes[cp->cpu_id].exec_unit_mapping;
183 	if (cp->cpu_m.cpu_ipipe == NO_EU_MAPPING_FOUND)
184 		cp->cpu_m.cpu_ipipe = (id_t)(cp->cpu_id);
185 
186 	cp->cpu_m.cpu_fpu = cpunodes[cp->cpu_id].fpu_mapping;
187 	if (cp->cpu_m.cpu_fpu == NO_EU_MAPPING_FOUND)
188 		cp->cpu_m.cpu_fpu = (id_t)(cp->cpu_id);
189 
190 	/*
191 	 * Niagara 2 defines the core to be at the FPU level
192 	 */
193 	cp->cpu_m.cpu_core = cp->cpu_m.cpu_fpu;
194 
195 	/*
196 	 * The cpu_chip field is initialized based on the information
197 	 * in the MD and assume that all cpus within a chip
198 	 * share the same L2 cache. If no such info is available, we
199 	 * set the cpu to belong to the defacto chip 0.
200 	 */
201 	cp->cpu_m.cpu_mpipe = cpunodes[cp->cpu_id].l2_cache_mapping;
202 	if (cp->cpu_m.cpu_mpipe == NO_L2_CACHE_MAPPING_FOUND)
203 		cp->cpu_m.cpu_mpipe = CPU_L2_CACHEID_INVALID;
204 }
205 
206 static int cpucnt;
207 
208 void
209 cpu_init_private(struct cpu *cp)
210 {
211 	extern void niagara_kstat_init(void);
212 
213 	ASSERT(MUTEX_HELD(&cpu_lock));
214 
215 	cpu_map_exec_units(cp);
216 
217 	if ((cpucnt++ == 0) && (cpu_hsvc_available == B_TRUE))
218 		(void) niagara_kstat_init();
219 }
220 
221 /*ARGSUSED*/
222 void
223 cpu_uninit_private(struct cpu *cp)
224 {
225 	extern void niagara_kstat_fini(void);
226 
227 	ASSERT(MUTEX_HELD(&cpu_lock));
228 	if ((--cpucnt == 0) && (cpu_hsvc_available == B_TRUE))
229 		(void) niagara_kstat_fini();
230 }
231 
232 /*
233  * On Niagara2, any flush will cause all preceding stores to be
234  * synchronized wrt the i$, regardless of address or ASI.  In fact,
235  * the address is ignored, so we always flush address 0.
236  */
237 /*ARGSUSED*/
238 void
239 dtrace_flush_sec(uintptr_t addr)
240 {
241 	doflush(0);
242 }
243 
244 /*
245  * Trapstat support for Niagara2 processor
246  * The Niagara2 provides HWTW support for TSB lookup and with HWTW
247  * enabled no TSB hit information will be available. Therefore setting
248  * the time spent in TLB miss handler for TSB hits to 0.
249  */
250 int
251 cpu_trapstat_conf(int cmd)
252 {
253 	int status = 0;
254 
255 	switch (cmd) {
256 	case CPU_TSTATCONF_INIT:
257 	case CPU_TSTATCONF_FINI:
258 	case CPU_TSTATCONF_ENABLE:
259 	case CPU_TSTATCONF_DISABLE:
260 		break;
261 	default:
262 		status = EINVAL;
263 		break;
264 	}
265 	return (status);
266 }
267 
268 void
269 cpu_trapstat_data(void *buf, uint_t tstat_pgszs)
270 {
271 	tstat_pgszdata_t	*tstatp = (tstat_pgszdata_t *)buf;
272 	int	i;
273 
274 	for (i = 0; i < tstat_pgszs; i++, tstatp++) {
275 		tstatp->tpgsz_kernel.tmode_itlb.ttlb_tlb.tmiss_count = 0;
276 		tstatp->tpgsz_kernel.tmode_itlb.ttlb_tlb.tmiss_time = 0;
277 		tstatp->tpgsz_user.tmode_itlb.ttlb_tlb.tmiss_count = 0;
278 		tstatp->tpgsz_user.tmode_itlb.ttlb_tlb.tmiss_time = 0;
279 		tstatp->tpgsz_kernel.tmode_dtlb.ttlb_tlb.tmiss_count = 0;
280 		tstatp->tpgsz_kernel.tmode_dtlb.ttlb_tlb.tmiss_time = 0;
281 		tstatp->tpgsz_user.tmode_dtlb.ttlb_tlb.tmiss_count = 0;
282 		tstatp->tpgsz_user.tmode_dtlb.ttlb_tlb.tmiss_time = 0;
283 	}
284 }
285 
286 /*
287  * Page coloring support for hashed cache index mode
288  */
289 
290 /*
291  * Node id bits from machine description (MD).  Node id distinguishes
292  * local versus remote memory. Because of MPO, page allocation does
293  * not cross node boundaries. Therefore, remove the node id bits from
294  * the color, since they are fixed. Either bit 30, or 31:30 in
295  * Victoria Falls processors.
296  * The number of node id bits is always 0 in Niagara2.
297  */
298 typedef struct n2color {
299 	uchar_t nnbits;	/* number of node id bits */
300 	uchar_t nnmask; /* mask for node id bits */
301 	uchar_t	lomask;	/* mask for bits below node id */
302 	uchar_t lobits;	/* number of bits below node id */
303 } n2color_t;
304 
305 n2color_t n2color[MMU_PAGE_SIZES];
306 static uchar_t nhbits[] = {7, 7, 6, 5, 5, 5};
307 
308 /*
309  * Remove node id bits from color bits 32:28.
310  * This will reduce the number of colors.
311  * No change if number of node bits is zero.
312  */
313 static inline uint_t
314 n2_hash2color(uint_t color, uchar_t szc)
315 {
316 	n2color_t m = n2color[szc];
317 
318 	if (m.nnbits > 0) {
319 		color = ((color >> m.nnbits) & ~m.lomask) | (color & m.lomask);
320 		ASSERT((color & ~(hw_page_array[szc].hp_colors - 1)) == 0);
321 	}
322 
323 	return (color);
324 }
325 
326 /*
327  * Restore node id bits into page color.
328  * This will increase the number of colors to match N2.
329  * No change if number of node bits is zero.
330  */
331 static inline uint_t
332 n2_color2hash(uint_t color, uchar_t szc, uint_t node)
333 {
334 	n2color_t m = n2color[szc];
335 
336 	if (m.nnbits > 0) {
337 		color = ((color & ~m.lomask) << m.nnbits) | (color & m.lomask);
338 		color |= (node & m.nnmask) << m.lobits;
339 	}
340 
341 	return (color);
342 }
343 
344 /* NI2 L2$ index is pa[32:28]^pa[17:13].pa[19:18]^pa[12:11].pa[10:6] */
345 
346 /*
347  * iterator NULL means pfn is VA, do not adjust ra_to_pa
348  * iterator (-1) means pfn is RA, need to convert to PA
349  * iterator non-null means pfn is RA, use ra_to_pa
350  */
351 uint_t
352 page_pfn_2_color_cpu(pfn_t pfn, uchar_t szc, void *cookie)
353 {
354 	mem_node_iterator_t *it = cookie;
355 	uint_t color;
356 
357 	ASSERT(szc <= TTE256M);
358 
359 	if (it == ((mem_node_iterator_t *)(-1))) {
360 		pfn = plat_rapfn_to_papfn(pfn);
361 	} else if (it != NULL) {
362 		ASSERT(pfn >= it->mi_mblock_base && pfn <= it->mi_mblock_end);
363 		pfn = pfn + it->mi_ra_to_pa;
364 	}
365 	pfn = PFN_BASE(pfn, szc);
366 	color = ((pfn >> 15) ^ pfn) & 0x1f;
367 	if (szc < TTE4M) {
368 		/* 19:18 */
369 		color = (color << 2) | ((pfn >> 5) & 0x3);
370 		if (szc > TTE64K)
371 			color >>= 1;    /* 19 */
372 	}
373 	return (n2_hash2color(color, szc));
374 }
375 
376 static uint_t
377 page_papfn_2_color_cpu(pfn_t papfn, uchar_t szc)
378 {
379 	uint_t color;
380 
381 	ASSERT(szc <= TTE256M);
382 
383 	papfn = PFN_BASE(papfn, szc);
384 	color = ((papfn >> 15) ^ papfn) & 0x1f;
385 	if (szc < TTE4M) {
386 		/* 19:18 */
387 		color = (color << 2) | ((papfn >> 5) & 0x3);
388 		if (szc > TTE64K)
389 			color >>= 1;    /* 19 */
390 	}
391 	return (color);
392 }
393 
394 #if TTE256M != 5
395 #error TTE256M is not 5
396 #endif
397 
398 uint_t
399 page_get_nsz_color_mask_cpu(uchar_t szc, uint_t mask)
400 {
401 	static uint_t ni2_color_masks[5] = {0x63, 0x1e, 0x3e, 0x1f, 0x1f};
402 	ASSERT(szc < TTE256M);
403 	mask = n2_color2hash(mask, szc, 0);
404 	mask &= ni2_color_masks[szc];
405 	if (szc == TTE64K || szc == TTE512K)
406 		mask >>= 1;
407 	return (n2_hash2color(mask, szc + 1));
408 }
409 
410 uint_t
411 page_get_nsz_color_cpu(uchar_t szc, uint_t color)
412 {
413 	ASSERT(szc < TTE256M);
414 	color = n2_color2hash(color, szc, 0);
415 	if (szc == TTE64K || szc == TTE512K)
416 		color >>= 1;
417 	return (n2_hash2color(color, szc + 1));
418 }
419 
420 uint_t
421 page_get_color_shift_cpu(uchar_t szc, uchar_t nszc)
422 {
423 	uint_t s;
424 	ASSERT(nszc >= szc);
425 	ASSERT(nszc <= TTE256M);
426 
427 	s = nhbits[szc] - n2color[szc].nnbits;
428 	s -= nhbits[nszc] - n2color[nszc].nnbits;
429 
430 	return (s);
431 }
432 
433 uint_t
434 page_convert_color_cpu(uint_t ncolor, uchar_t szc, uchar_t nszc)
435 {
436 	uint_t color;
437 
438 	ASSERT(nszc > szc);
439 	ASSERT(nszc <= TTE256M);
440 	ncolor = n2_color2hash(ncolor, nszc, 0);
441 	color = ncolor << (nhbits[szc] - nhbits[nszc]);
442 	color = n2_hash2color(color, szc);
443 	return (color);
444 }
445 
446 #define	PAPFN_2_MNODE(pfn) \
447 	(((pfn) & it->mi_mnode_pfn_mask) >> it->mi_mnode_pfn_shift)
448 
449 /*ARGSUSED*/
450 pfn_t
451 page_next_pfn_for_color_cpu(pfn_t pfn, uchar_t szc, uint_t color,
452     uint_t ceq_mask, uint_t color_mask, void *cookie)
453 {
454 	mem_node_iterator_t *it = cookie;
455 	pfn_t pstep = PNUM_SIZE(szc);
456 	pfn_t npfn, pfn_ceq_mask, pfn_color;
457 	pfn_t tmpmask, mask = (pfn_t)-1;
458 	uint_t pfnmn;
459 
460 	ASSERT((color & ~ceq_mask) == 0);
461 	ASSERT(pfn >= it->mi_mblock_base && pfn <= it->mi_mblock_end);
462 
463 	/* convert RA to PA for accurate color calculation */
464 	if (it->mi_init) {
465 		/* first call after it, so cache these values */
466 		it->mi_hash_ceq_mask =
467 		    n2_color2hash(ceq_mask, szc, it->mi_mnode_mask);
468 		it->mi_hash_color =
469 		    n2_color2hash(color, szc, it->mi_mnode);
470 		it->mi_init = 0;
471 	} else {
472 		ASSERT(it->mi_hash_ceq_mask ==
473 		    n2_color2hash(ceq_mask, szc, it->mi_mnode_mask));
474 		ASSERT(it->mi_hash_color ==
475 		    n2_color2hash(color, szc, it->mi_mnode));
476 	}
477 	ceq_mask = it->mi_hash_ceq_mask;
478 	color = it->mi_hash_color;
479 	pfn += it->mi_ra_to_pa;
480 
481 	/* restart here when we switch memblocks */
482 next_mem_block:
483 	if (szc <= TTE64K) {
484 		pfnmn = PAPFN_2_MNODE(pfn);
485 	}
486 	if (((page_papfn_2_color_cpu(pfn, szc) ^ color) & ceq_mask) == 0 &&
487 	    (szc > TTE64K || pfnmn == it->mi_mnode)) {
488 
489 		/* we start from the page with correct color */
490 		if (szc >= TTE512K) {
491 			if (szc >= TTE4M) {
492 				/* page color is PA[32:28] */
493 				pfn_ceq_mask = ceq_mask << 15;
494 			} else {
495 				/* page color is PA[32:28].PA[19:19] */
496 				pfn_ceq_mask = ((ceq_mask & 1) << 6) |
497 				    ((ceq_mask >> 1) << 15);
498 			}
499 			npfn = ADD_MASKED(pfn, pstep, pfn_ceq_mask, mask);
500 			goto done;
501 		} else {
502 			/*
503 			 * We deal 64K or 8K page. Check if we could the
504 			 * satisfy the request without changing PA[32:28]
505 			 */
506 			pfn_ceq_mask = ((ceq_mask & 3) << 5) | (ceq_mask >> 2);
507 			pfn_ceq_mask |= it->mi_mnode_pfn_mask;
508 			npfn = ADD_MASKED(pfn, pstep, pfn_ceq_mask, mask);
509 
510 			if ((((npfn ^ pfn) >> 15) & 0x1f) == 0)
511 				goto done;
512 
513 			/*
514 			 * for next pfn we have to change bits PA[32:28]
515 			 * set PA[63:28] and PA[19:18] of the next pfn
516 			 */
517 			npfn = (pfn >> 15) << 15;
518 			npfn |= (ceq_mask & color & 3) << 5;
519 			pfn_ceq_mask = (szc == TTE8K) ? 0 :
520 			    (ceq_mask & 0x1c) << 13;
521 			pfn_ceq_mask |= it->mi_mnode_pfn_mask;
522 			npfn = ADD_MASKED(npfn, (1 << 15), pfn_ceq_mask, mask);
523 
524 			/*
525 			 * set bits PA[17:13] to match the color
526 			 */
527 			npfn |= ((npfn >> 15) ^ (color >> 2)) & (ceq_mask >> 2);
528 			goto done;
529 		}
530 	}
531 
532 	/*
533 	 * we start from the page with incorrect color - rare case
534 	 */
535 	if (szc >= TTE512K) {
536 		if (szc >= TTE4M) {
537 			/* page color is in bits PA[32:28] */
538 			npfn = ((pfn >> 20) << 20) | (color << 15);
539 			pfn_ceq_mask = (ceq_mask << 15) | 0x7fff;
540 		} else {
541 			/* try get the right color by changing bit PA[19:19] */
542 			npfn = pfn + pstep;
543 			if (((page_papfn_2_color_cpu(npfn, szc) ^ color) &
544 			    ceq_mask) == 0)
545 				goto done;
546 
547 			/* page color is PA[32:28].PA[19:19] */
548 			pfn_ceq_mask = ((ceq_mask & 1) << 6) |
549 			    ((ceq_mask >> 1) << 15) | (0xff << 7);
550 			pfn_color = ((color & 1) << 6) | ((color >> 1) << 15);
551 			npfn = ((pfn >> 20) << 20) | pfn_color;
552 		}
553 
554 		while (npfn <= pfn) {
555 			npfn = ADD_MASKED(npfn, pstep, pfn_ceq_mask, mask);
556 		}
557 		goto done;
558 	}
559 
560 	/*
561 	 *  We deal 64K or 8K page of incorrect color.
562 	 * Try correcting color without changing PA[32:28]
563 	 */
564 	pfn_ceq_mask = ((ceq_mask & 3) << 5) | (ceq_mask >> 2);
565 	pfn_color = ((color & 3) << 5) | (color >> 2);
566 	if (pfnmn == it->mi_mnode) {
567 		npfn = (pfn & ~(pfn_t)0x7f);
568 		npfn |= (((pfn >> 15) & 0x1f) ^ pfn_color) & pfn_ceq_mask;
569 		npfn = (szc == TTE64K) ? (npfn & ~(pfn_t)0x7) : npfn;
570 
571 		if (((page_papfn_2_color_cpu(npfn, szc) ^ color) &
572 		    ceq_mask) == 0) {
573 			/* the color is fixed - find the next page */
574 			pfn_ceq_mask |= it->mi_mnode_pfn_mask;
575 			while (npfn <= pfn) {
576 				npfn = ADD_MASKED(npfn, pstep, pfn_ceq_mask,
577 				    mask);
578 			}
579 			if ((((npfn ^ pfn) >> 15) & 0x1f) == 0)
580 				goto done;
581 		}
582 	}
583 
584 	/* to fix the color need to touch PA[32:28] */
585 	npfn = (szc == TTE8K) ? ((pfn >> 15) << 15) :
586 	    (((pfn >> 18) << 18) | ((color & 0x1c) << 13));
587 
588 	/* fix mnode if input pfn is in the wrong mnode. */
589 	if ((pfnmn = PAPFN_2_MNODE(npfn)) != it->mi_mnode) {
590 		npfn += ((it->mi_mnode - pfnmn) & it->mi_mnode_mask) <<
591 		    it->mi_mnode_pfn_shift;
592 	}
593 
594 	tmpmask = (szc == TTE8K) ? 0 : (ceq_mask & 0x1c) << 13;
595 	tmpmask |= it->mi_mnode_pfn_mask;
596 
597 	while (npfn <= pfn) {
598 		npfn = ADD_MASKED(npfn, (1 << 15), tmpmask, mask);
599 	}
600 
601 	/* set bits PA[19:13] to match the color */
602 	npfn |= (((npfn >> 15) & 0x1f) ^ pfn_color) & pfn_ceq_mask;
603 	npfn = (szc == TTE64K) ? (npfn & ~(pfn_t)0x7) : npfn;
604 
605 done:
606 	ASSERT(((page_papfn_2_color_cpu(npfn, szc) ^ color) & ceq_mask) == 0);
607 	ASSERT(PAPFN_2_MNODE(npfn) == it->mi_mnode);
608 
609 	/* PA to RA */
610 	npfn -= it->mi_ra_to_pa;
611 
612 	/* check for possible memblock switch */
613 	if (npfn > it->mi_mblock_end) {
614 		pfn = plat_mem_node_iterator_init(npfn, it->mi_mnode, it, 0);
615 		if (pfn == (pfn_t)-1)
616 			return (pfn);
617 		ASSERT(pfn >= it->mi_mblock_base && pfn <= it->mi_mblock_end);
618 		pfn += it->mi_ra_to_pa;
619 		goto next_mem_block;
620 	}
621 
622 	return (npfn);
623 }
624 
625 /*
626  * init page coloring
627  * VF encodes node_id for an L-group in either bit 30 or 31:30,
628  * which effectively reduces the number of colors available per mnode.
629  */
630 void
631 page_coloring_init_cpu()
632 {
633 	int i;
634 	uchar_t id;
635 	uchar_t lo;
636 	uchar_t hi;
637 	n2color_t m;
638 	mem_node_iterator_t it;
639 	static uchar_t idmask[] = {0, 0x7, 0x1f, 0x1f, 0x1f, 0x1f};
640 
641 	(void) plat_mem_node_iterator_init(0, 0, &it, 1);
642 	for (i = 0; i < mmu_page_sizes; i++) {
643 		memset(&m, 0, sizeof (m));
644 		id = it.mi_mnode_pfn_mask >> 15;	/* node id mask */
645 		id &= idmask[i];
646 		lo = lowbit(id);
647 		if (lo > 0) {
648 			hi = highbit(id);
649 			m.nnbits = hi - lo + 1;
650 			m.nnmask = (1 << m.nnbits) - 1;
651 			lo += nhbits[i] - 5;
652 			m.lomask = (1 << (lo - 1)) - 1;
653 			m.lobits = lo - 1;
654 		}
655 		hw_page_array[i].hp_colors = 1 << (nhbits[i] - m.nnbits);
656 		n2color[i] = m;
657 	}
658 }
659 
660 /*
661  * group colorequiv colors on N2 by low order bits of the color first
662  */
663 void
664 page_set_colorequiv_arr_cpu(void)
665 {
666 	static uint_t nequiv_shades_log2[MMU_PAGE_SIZES] = {2, 5, 0, 0, 0, 0};
667 
668 	nequiv_shades_log2[1] -= n2color[1].nnbits;
669 	if (colorequiv > 1) {
670 		int i;
671 		uint_t sv_a = lowbit(colorequiv) - 1;
672 
673 		if (sv_a > 15)
674 			sv_a = 15;
675 
676 		for (i = 0; i < MMU_PAGE_SIZES; i++) {
677 			uint_t colors;
678 			uint_t a = sv_a;
679 
680 			if ((colors = hw_page_array[i].hp_colors) <= 1)
681 				continue;
682 			while ((colors >> a) == 0)
683 				a--;
684 			if (a > (colorequivszc[i] & 0xf) +
685 			    (colorequivszc[i] >> 4)) {
686 				if (a <= nequiv_shades_log2[i]) {
687 					colorequivszc[i] = (uchar_t)a;
688 				} else {
689 					colorequivszc[i] =
690 					    ((a - nequiv_shades_log2[i]) << 4) |
691 					    nequiv_shades_log2[i];
692 				}
693 			}
694 		}
695 	}
696 }
697