xref: /illumos-gate/usr/src/uts/sun4v/cpu/niagara2.c (revision b779d3e04b6c453fc372ef1dbf8df01a8c2e10dc)
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 2008 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 #include <sys/mutex_impl.h>
63 
64 uint_t root_phys_addr_lo_mask = 0xffffffffU;
65 #if defined(NIAGARA2_IMPL)
66 char cpu_module_name[] = "SUNW,UltraSPARC-T2";
67 #elif defined(VFALLS_IMPL)
68 char cpu_module_name[] = "SUNW,UltraSPARC-T2+";
69 #endif
70 
71 /*
72  * Hypervisor services information for the NIAGARA2 and Victoria Falls
73  * CPU module
74  */
75 static boolean_t cpu_hsvc_available = B_TRUE;
76 static uint64_t cpu_sup_minor;		/* Supported minor number */
77 #if defined(NIAGARA2_IMPL)
78 static hsvc_info_t cpu_hsvc = {
79 	HSVC_REV_1, NULL, HSVC_GROUP_NIAGARA2_CPU, NIAGARA2_HSVC_MAJOR,
80 	NIAGARA2_HSVC_MINOR, cpu_module_name
81 };
82 #elif defined(VFALLS_IMPL)
83 static hsvc_info_t cpu_hsvc = {
84 	HSVC_REV_1, NULL, HSVC_GROUP_VFALLS_CPU, VFALLS_HSVC_MAJOR,
85 	VFALLS_HSVC_MINOR, cpu_module_name
86 };
87 #endif
88 
89 void
90 cpu_setup(void)
91 {
92 	extern int mmu_exported_pagesize_mask;
93 	extern int cpc_has_overflow_intr;
94 	extern size_t contig_mem_prealloc_base_size;
95 	int status;
96 
97 	/*
98 	 * Negotiate the API version for Niagara2 specific hypervisor
99 	 * services.
100 	 */
101 	status = hsvc_register(&cpu_hsvc, &cpu_sup_minor);
102 	if (status != 0) {
103 		cmn_err(CE_WARN, "%s: cannot negotiate hypervisor services "
104 		    "group: 0x%lx major: 0x%lx minor: 0x%lx errno: %d",
105 		    cpu_hsvc.hsvc_modname, cpu_hsvc.hsvc_group,
106 		    cpu_hsvc.hsvc_major, cpu_hsvc.hsvc_minor, status);
107 		cpu_hsvc_available = B_FALSE;
108 	}
109 
110 	/*
111 	 * The setup common to all CPU modules is done in cpu_setup_common
112 	 * routine.
113 	 */
114 	cpu_setup_common(NULL);
115 
116 	cache |= (CACHE_PTAG | CACHE_IOCOHERENT);
117 
118 	if ((mmu_exported_pagesize_mask &
119 	    DEFAULT_SUN4V_MMU_PAGESIZE_MASK) !=
120 	    DEFAULT_SUN4V_MMU_PAGESIZE_MASK)
121 		cmn_err(CE_PANIC, "machine description"
122 		    " does not have required sun4v page sizes"
123 		    " 8K, 64K and 4M: MD mask is 0x%x",
124 		    mmu_exported_pagesize_mask);
125 
126 	cpu_hwcap_flags = AV_SPARC_VIS | AV_SPARC_VIS2 | AV_SPARC_ASI_BLK_INIT;
127 
128 	/*
129 	 * Niagara2 supports a 48-bit subset of the full 64-bit virtual
130 	 * address space. Virtual addresses between 0x0000800000000000
131 	 * and 0xffff.7fff.ffff.ffff inclusive lie within a "VA Hole"
132 	 * and must never be mapped. In addition, software must not use
133 	 * pages within 4GB of the VA hole as instruction pages to
134 	 * avoid problems with prefetching into the VA hole.
135 	 */
136 	hole_start = (caddr_t)((1ull << (va_bits - 1)) - (1ull << 32));
137 	hole_end = (caddr_t)((0ull - (1ull << (va_bits - 1))) + (1ull << 32));
138 
139 	/*
140 	 * Niagara2 has a performance counter overflow interrupt
141 	 */
142 	cpc_has_overflow_intr = 1;
143 
144 	/*
145 	 * Enable 4M pages for OOB.
146 	 */
147 	max_uheap_lpsize = MMU_PAGESIZE4M;
148 	max_ustack_lpsize = MMU_PAGESIZE4M;
149 	max_privmap_lpsize = MMU_PAGESIZE4M;
150 
151 #ifdef SUN4V_CONTIG_MEM_PREALLOC_SIZE_MB
152 	/*
153 	 * Use CPU Makefile specific compile time define (if exists)
154 	 * to add to the contig preallocation size.
155 	 */
156 	contig_mem_prealloc_base_size = MB(SUN4V_CONTIG_MEM_PREALLOC_SIZE_MB);
157 #endif
158 }
159 
160 /*
161  * Set the magic constants of the implementation.
162  */
163 void
164 cpu_fiximp(struct cpu_node *cpunode)
165 {
166 	/*
167 	 * The Cache node is optional in MD. Therefore in case "Cache"
168 	 * node does not exists in MD, set the default L2 cache associativity,
169 	 * size, linesize.
170 	 */
171 	if (cpunode->ecache_size == 0)
172 		cpunode->ecache_size = L2CACHE_SIZE;
173 	if (cpunode->ecache_linesize == 0)
174 		cpunode->ecache_linesize = L2CACHE_LINESIZE;
175 	if (cpunode->ecache_associativity == 0)
176 		cpunode->ecache_associativity = L2CACHE_ASSOCIATIVITY;
177 }
178 
179 void
180 cpu_map_exec_units(struct cpu *cp)
181 {
182 	ASSERT(MUTEX_HELD(&cpu_lock));
183 
184 	/*
185 	 * The cpu_ipipe and cpu_fpu fields are initialized based on
186 	 * the execution unit sharing information from the MD. They
187 	 * default to the CPU id in the absence of such information.
188 	 */
189 	cp->cpu_m.cpu_ipipe = cpunodes[cp->cpu_id].exec_unit_mapping;
190 	if (cp->cpu_m.cpu_ipipe == NO_EU_MAPPING_FOUND)
191 		cp->cpu_m.cpu_ipipe = (id_t)(cp->cpu_id);
192 
193 	cp->cpu_m.cpu_fpu = cpunodes[cp->cpu_id].fpu_mapping;
194 	if (cp->cpu_m.cpu_fpu == NO_EU_MAPPING_FOUND)
195 		cp->cpu_m.cpu_fpu = (id_t)(cp->cpu_id);
196 
197 	/*
198 	 * Niagara 2 defines the core to be at the FPU level
199 	 */
200 	cp->cpu_m.cpu_core = cp->cpu_m.cpu_fpu;
201 
202 	/*
203 	 * The cpu_chip field is initialized based on the information
204 	 * in the MD and assume that all cpus within a chip
205 	 * share the same L2 cache. If no such info is available, we
206 	 * set the cpu to belong to the defacto chip 0.
207 	 */
208 	cp->cpu_m.cpu_mpipe = cpunodes[cp->cpu_id].l2_cache_mapping;
209 	if (cp->cpu_m.cpu_mpipe == NO_L2_CACHE_MAPPING_FOUND)
210 		cp->cpu_m.cpu_mpipe = CPU_L2_CACHEID_INVALID;
211 
212 	cp->cpu_m.cpu_chip = cpunodes[cp->cpu_id].l2_cache_mapping;
213 	if (cp->cpu_m.cpu_chip == NO_L2_CACHE_MAPPING_FOUND)
214 		cp->cpu_m.cpu_chip = CPU_CHIPID_INVALID;
215 }
216 
217 static int cpucnt;
218 
219 void
220 cpu_init_private(struct cpu *cp)
221 {
222 	extern void niagara_kstat_init(void);
223 
224 	ASSERT(MUTEX_HELD(&cpu_lock));
225 
226 	cpu_map_exec_units(cp);
227 
228 	if ((cpucnt++ == 0) && (cpu_hsvc_available == B_TRUE))
229 		(void) niagara_kstat_init();
230 
231 	mutex_delay = rdccr_delay;
232 }
233 
234 /*ARGSUSED*/
235 void
236 cpu_uninit_private(struct cpu *cp)
237 {
238 	extern void niagara_kstat_fini(void);
239 
240 	ASSERT(MUTEX_HELD(&cpu_lock));
241 	if ((--cpucnt == 0) && (cpu_hsvc_available == B_TRUE))
242 		(void) niagara_kstat_fini();
243 }
244 
245 /*
246  * On Niagara2, any flush will cause all preceding stores to be
247  * synchronized wrt the i$, regardless of address or ASI.  In fact,
248  * the address is ignored, so we always flush address 0.
249  */
250 /*ARGSUSED*/
251 void
252 dtrace_flush_sec(uintptr_t addr)
253 {
254 	doflush(0);
255 }
256 
257 /*
258  * Trapstat support for Niagara2 processor
259  * The Niagara2 provides HWTW support for TSB lookup and with HWTW
260  * enabled no TSB hit information will be available. Therefore setting
261  * the time spent in TLB miss handler for TSB hits to 0.
262  */
263 int
264 cpu_trapstat_conf(int cmd)
265 {
266 	int status = 0;
267 
268 	switch (cmd) {
269 	case CPU_TSTATCONF_INIT:
270 	case CPU_TSTATCONF_FINI:
271 	case CPU_TSTATCONF_ENABLE:
272 	case CPU_TSTATCONF_DISABLE:
273 		break;
274 	default:
275 		status = EINVAL;
276 		break;
277 	}
278 	return (status);
279 }
280 
281 void
282 cpu_trapstat_data(void *buf, uint_t tstat_pgszs)
283 {
284 	tstat_pgszdata_t	*tstatp = (tstat_pgszdata_t *)buf;
285 	int	i;
286 
287 	for (i = 0; i < tstat_pgszs; i++, tstatp++) {
288 		tstatp->tpgsz_kernel.tmode_itlb.ttlb_tlb.tmiss_count = 0;
289 		tstatp->tpgsz_kernel.tmode_itlb.ttlb_tlb.tmiss_time = 0;
290 		tstatp->tpgsz_user.tmode_itlb.ttlb_tlb.tmiss_count = 0;
291 		tstatp->tpgsz_user.tmode_itlb.ttlb_tlb.tmiss_time = 0;
292 		tstatp->tpgsz_kernel.tmode_dtlb.ttlb_tlb.tmiss_count = 0;
293 		tstatp->tpgsz_kernel.tmode_dtlb.ttlb_tlb.tmiss_time = 0;
294 		tstatp->tpgsz_user.tmode_dtlb.ttlb_tlb.tmiss_count = 0;
295 		tstatp->tpgsz_user.tmode_dtlb.ttlb_tlb.tmiss_time = 0;
296 	}
297 }
298 
299 /*
300  * Page coloring support for hashed cache index mode
301  */
302 
303 /*
304  * Node id bits from machine description (MD).  Node id distinguishes
305  * local versus remote memory. Because of MPO, page allocation does
306  * not cross node boundaries. Therefore, remove the node id bits from
307  * the color, since they are fixed. Either bit 30, or 31:30 in
308  * Victoria Falls processors.
309  * The number of node id bits is always 0 in Niagara2.
310  */
311 typedef struct n2color {
312 	uchar_t nnbits;	/* number of node id bits */
313 	uchar_t nnmask; /* mask for node id bits */
314 	uchar_t	lomask;	/* mask for bits below node id */
315 	uchar_t lobits;	/* number of bits below node id */
316 } n2color_t;
317 
318 n2color_t n2color[MMU_PAGE_SIZES];
319 static uchar_t nhbits[] = {7, 7, 6, 5, 5, 5};
320 
321 /*
322  * Remove node id bits from color bits 32:28.
323  * This will reduce the number of colors.
324  * No change if number of node bits is zero.
325  */
326 static inline uint_t
327 n2_hash2color(uint_t color, uchar_t szc)
328 {
329 	n2color_t m = n2color[szc];
330 
331 	if (m.nnbits > 0) {
332 		color = ((color >> m.nnbits) & ~m.lomask) | (color & m.lomask);
333 		ASSERT((color & ~(hw_page_array[szc].hp_colors - 1)) == 0);
334 	}
335 
336 	return (color);
337 }
338 
339 /*
340  * Restore node id bits into page color.
341  * This will increase the number of colors to match N2.
342  * No change if number of node bits is zero.
343  */
344 static inline uint_t
345 n2_color2hash(uint_t color, uchar_t szc, uint_t node)
346 {
347 	n2color_t m = n2color[szc];
348 
349 	if (m.nnbits > 0) {
350 		color = ((color & ~m.lomask) << m.nnbits) | (color & m.lomask);
351 		color |= (node & m.nnmask) << m.lobits;
352 	}
353 
354 	return (color);
355 }
356 
357 /* NI2 L2$ index is pa[32:28]^pa[17:13].pa[19:18]^pa[12:11].pa[10:6] */
358 
359 /*
360  * iterator NULL means pfn is VA, do not adjust ra_to_pa
361  * iterator (-1) means pfn is RA, need to convert to PA
362  * iterator non-null means pfn is RA, use ra_to_pa
363  */
364 uint_t
365 page_pfn_2_color_cpu(pfn_t pfn, uchar_t szc, void *cookie)
366 {
367 	mem_node_iterator_t *it = cookie;
368 	uint_t color;
369 
370 	ASSERT(szc <= TTE256M);
371 
372 	if (it == ((mem_node_iterator_t *)(-1))) {
373 		pfn = plat_rapfn_to_papfn(pfn);
374 	} else if (it != NULL) {
375 		ASSERT(pfn >= it->mi_mblock_base && pfn <= it->mi_mblock_end);
376 		pfn = pfn + it->mi_ra_to_pa;
377 	}
378 	pfn = PFN_BASE(pfn, szc);
379 	color = ((pfn >> 15) ^ pfn) & 0x1f;
380 	if (szc < TTE4M) {
381 		/* 19:18 */
382 		color = (color << 2) | ((pfn >> 5) & 0x3);
383 		if (szc > TTE64K)
384 			color >>= 1;    /* 19 */
385 	}
386 	return (n2_hash2color(color, szc));
387 }
388 
389 static uint_t
390 page_papfn_2_color_cpu(pfn_t papfn, uchar_t szc)
391 {
392 	uint_t color;
393 
394 	ASSERT(szc <= TTE256M);
395 
396 	papfn = PFN_BASE(papfn, szc);
397 	color = ((papfn >> 15) ^ papfn) & 0x1f;
398 	if (szc < TTE4M) {
399 		/* 19:18 */
400 		color = (color << 2) | ((papfn >> 5) & 0x3);
401 		if (szc > TTE64K)
402 			color >>= 1;    /* 19 */
403 	}
404 	return (color);
405 }
406 
407 #if TTE256M != 5
408 #error TTE256M is not 5
409 #endif
410 
411 uint_t
412 page_get_nsz_color_mask_cpu(uchar_t szc, uint_t mask)
413 {
414 	static uint_t ni2_color_masks[5] = {0x63, 0x1e, 0x3e, 0x1f, 0x1f};
415 	ASSERT(szc < TTE256M);
416 	mask = n2_color2hash(mask, szc, 0);
417 	mask &= ni2_color_masks[szc];
418 	if (szc == TTE64K || szc == TTE512K)
419 		mask >>= 1;
420 	return (n2_hash2color(mask, szc + 1));
421 }
422 
423 uint_t
424 page_get_nsz_color_cpu(uchar_t szc, uint_t color)
425 {
426 	ASSERT(szc < TTE256M);
427 	color = n2_color2hash(color, szc, 0);
428 	if (szc == TTE64K || szc == TTE512K)
429 		color >>= 1;
430 	return (n2_hash2color(color, szc + 1));
431 }
432 
433 uint_t
434 page_get_color_shift_cpu(uchar_t szc, uchar_t nszc)
435 {
436 	uint_t s;
437 	ASSERT(nszc >= szc);
438 	ASSERT(nszc <= TTE256M);
439 
440 	s = nhbits[szc] - n2color[szc].nnbits;
441 	s -= nhbits[nszc] - n2color[nszc].nnbits;
442 
443 	return (s);
444 }
445 
446 uint_t
447 page_convert_color_cpu(uint_t ncolor, uchar_t szc, uchar_t nszc)
448 {
449 	uint_t color;
450 
451 	ASSERT(nszc > szc);
452 	ASSERT(nszc <= TTE256M);
453 	ncolor = n2_color2hash(ncolor, nszc, 0);
454 	color = ncolor << (nhbits[szc] - nhbits[nszc]);
455 	color = n2_hash2color(color, szc);
456 	return (color);
457 }
458 
459 #define	PAPFN_2_MNODE(pfn) \
460 	(((pfn) & it->mi_mnode_pfn_mask) >> it->mi_mnode_pfn_shift)
461 
462 /*ARGSUSED*/
463 pfn_t
464 page_next_pfn_for_color_cpu(pfn_t pfn, uchar_t szc, uint_t color,
465     uint_t ceq_mask, uint_t color_mask, void *cookie)
466 {
467 	mem_node_iterator_t *it = cookie;
468 	pfn_t pstep = PNUM_SIZE(szc);
469 	pfn_t npfn, pfn_ceq_mask, pfn_color;
470 	pfn_t tmpmask, mask = (pfn_t)-1;
471 	uint_t pfnmn;
472 
473 	ASSERT((color & ~ceq_mask) == 0);
474 	ASSERT(pfn >= it->mi_mblock_base && pfn <= it->mi_mblock_end);
475 
476 	/* convert RA to PA for accurate color calculation */
477 	if (it->mi_init) {
478 		/* first call after it, so cache these values */
479 		it->mi_hash_ceq_mask =
480 		    n2_color2hash(ceq_mask, szc, it->mi_mnode_mask);
481 		it->mi_hash_color =
482 		    n2_color2hash(color, szc, it->mi_mnode);
483 		it->mi_init = 0;
484 	} else {
485 		ASSERT(it->mi_hash_ceq_mask ==
486 		    n2_color2hash(ceq_mask, szc, it->mi_mnode_mask));
487 		ASSERT(it->mi_hash_color ==
488 		    n2_color2hash(color, szc, it->mi_mnode));
489 	}
490 	ceq_mask = it->mi_hash_ceq_mask;
491 	color = it->mi_hash_color;
492 	pfn += it->mi_ra_to_pa;
493 
494 	/* restart here when we switch memblocks */
495 next_mem_block:
496 	if (szc <= TTE64K) {
497 		pfnmn = PAPFN_2_MNODE(pfn);
498 	}
499 	if (((page_papfn_2_color_cpu(pfn, szc) ^ color) & ceq_mask) == 0 &&
500 	    (szc > TTE64K || pfnmn == it->mi_mnode)) {
501 
502 		/* we start from the page with correct color */
503 		if (szc >= TTE512K) {
504 			if (szc >= TTE4M) {
505 				/* page color is PA[32:28] */
506 				pfn_ceq_mask = ceq_mask << 15;
507 			} else {
508 				/* page color is PA[32:28].PA[19:19] */
509 				pfn_ceq_mask = ((ceq_mask & 1) << 6) |
510 				    ((ceq_mask >> 1) << 15);
511 			}
512 			npfn = ADD_MASKED(pfn, pstep, pfn_ceq_mask, mask);
513 			goto done;
514 		} else {
515 			/*
516 			 * We deal 64K or 8K page. Check if we could the
517 			 * satisfy the request without changing PA[32:28]
518 			 */
519 			pfn_ceq_mask = ((ceq_mask & 3) << 5) | (ceq_mask >> 2);
520 			pfn_ceq_mask |= it->mi_mnode_pfn_mask;
521 			npfn = ADD_MASKED(pfn, pstep, pfn_ceq_mask, mask);
522 
523 			if ((((npfn ^ pfn) >> 15) & 0x1f) == 0)
524 				goto done;
525 
526 			/*
527 			 * for next pfn we have to change bits PA[32:28]
528 			 * set PA[63:28] and PA[19:18] of the next pfn
529 			 */
530 			npfn = (pfn >> 15) << 15;
531 			npfn |= (ceq_mask & color & 3) << 5;
532 			pfn_ceq_mask = (szc == TTE8K) ? 0 :
533 			    (ceq_mask & 0x1c) << 13;
534 			pfn_ceq_mask |= it->mi_mnode_pfn_mask;
535 			npfn = ADD_MASKED(npfn, (1 << 15), pfn_ceq_mask, mask);
536 
537 			/*
538 			 * set bits PA[17:13] to match the color
539 			 */
540 			npfn |= ((npfn >> 15) ^ (color >> 2)) & (ceq_mask >> 2);
541 			goto done;
542 		}
543 	}
544 
545 	/*
546 	 * we start from the page with incorrect color - rare case
547 	 */
548 	if (szc >= TTE512K) {
549 		if (szc >= TTE4M) {
550 			/* page color is in bits PA[32:28] */
551 			npfn = ((pfn >> 20) << 20) | (color << 15);
552 			pfn_ceq_mask = (ceq_mask << 15) | 0x7fff;
553 		} else {
554 			/* try get the right color by changing bit PA[19:19] */
555 			npfn = pfn + pstep;
556 			if (((page_papfn_2_color_cpu(npfn, szc) ^ color) &
557 			    ceq_mask) == 0)
558 				goto done;
559 
560 			/* page color is PA[32:28].PA[19:19] */
561 			pfn_ceq_mask = ((ceq_mask & 1) << 6) |
562 			    ((ceq_mask >> 1) << 15) | (0xff << 7);
563 			pfn_color = ((color & 1) << 6) | ((color >> 1) << 15);
564 			npfn = ((pfn >> 20) << 20) | pfn_color;
565 		}
566 
567 		while (npfn <= pfn) {
568 			npfn = ADD_MASKED(npfn, pstep, pfn_ceq_mask, mask);
569 		}
570 		goto done;
571 	}
572 
573 	/*
574 	 *  We deal 64K or 8K page of incorrect color.
575 	 * Try correcting color without changing PA[32:28]
576 	 */
577 	pfn_ceq_mask = ((ceq_mask & 3) << 5) | (ceq_mask >> 2);
578 	pfn_color = ((color & 3) << 5) | (color >> 2);
579 	if (pfnmn == it->mi_mnode) {
580 		npfn = (pfn & ~(pfn_t)0x7f);
581 		npfn |= (((pfn >> 15) & 0x1f) ^ pfn_color) & pfn_ceq_mask;
582 		npfn = (szc == TTE64K) ? (npfn & ~(pfn_t)0x7) : npfn;
583 
584 		if (((page_papfn_2_color_cpu(npfn, szc) ^ color) &
585 		    ceq_mask) == 0) {
586 			/* the color is fixed - find the next page */
587 			pfn_ceq_mask |= it->mi_mnode_pfn_mask;
588 			while (npfn <= pfn) {
589 				npfn = ADD_MASKED(npfn, pstep, pfn_ceq_mask,
590 				    mask);
591 			}
592 			if ((((npfn ^ pfn) >> 15) & 0x1f) == 0)
593 				goto done;
594 		}
595 	}
596 
597 	/* to fix the color need to touch PA[32:28] */
598 	npfn = (szc == TTE8K) ? ((pfn >> 15) << 15) :
599 	    (((pfn >> 18) << 18) | ((color & 0x1c) << 13));
600 
601 	/* fix mnode if input pfn is in the wrong mnode. */
602 	if ((pfnmn = PAPFN_2_MNODE(npfn)) != it->mi_mnode) {
603 		npfn += ((it->mi_mnode - pfnmn) & it->mi_mnode_mask) <<
604 		    it->mi_mnode_pfn_shift;
605 	}
606 
607 	tmpmask = (szc == TTE8K) ? 0 : (ceq_mask & 0x1c) << 13;
608 	tmpmask |= it->mi_mnode_pfn_mask;
609 
610 	while (npfn <= pfn) {
611 		npfn = ADD_MASKED(npfn, (1 << 15), tmpmask, mask);
612 	}
613 
614 	/* set bits PA[19:13] to match the color */
615 	npfn |= (((npfn >> 15) & 0x1f) ^ pfn_color) & pfn_ceq_mask;
616 	npfn = (szc == TTE64K) ? (npfn & ~(pfn_t)0x7) : npfn;
617 
618 done:
619 	ASSERT(((page_papfn_2_color_cpu(npfn, szc) ^ color) & ceq_mask) == 0);
620 	ASSERT(PAPFN_2_MNODE(npfn) == it->mi_mnode);
621 
622 	/* PA to RA */
623 	npfn -= it->mi_ra_to_pa;
624 
625 	/* check for possible memblock switch */
626 	if (npfn > it->mi_mblock_end) {
627 		pfn = plat_mem_node_iterator_init(npfn, it->mi_mnode, szc, it,
628 		    0);
629 		if (pfn == (pfn_t)-1)
630 			return (pfn);
631 		ASSERT(pfn >= it->mi_mblock_base && pfn <= it->mi_mblock_end);
632 		pfn += it->mi_ra_to_pa;
633 		goto next_mem_block;
634 	}
635 
636 	return (npfn);
637 }
638 
639 /*
640  * init page coloring
641  * VF encodes node_id for an L-group in either bit 30 or 31:30,
642  * which effectively reduces the number of colors available per mnode.
643  */
644 void
645 page_coloring_init_cpu()
646 {
647 	int i;
648 	uchar_t id;
649 	uchar_t lo;
650 	uchar_t hi;
651 	n2color_t m;
652 	mem_node_iterator_t it;
653 	static uchar_t idmask[] = {0, 0x7, 0x1f, 0x1f, 0x1f, 0x1f};
654 
655 	for (i = 0; i < max_mem_nodes; i++) {
656 		memset(&it, 0, sizeof (it));
657 		if (plat_mem_node_iterator_init(0, i, 0, &it, 1) != (pfn_t)-1)
658 			break;
659 	}
660 	ASSERT(i < max_mem_nodes);
661 	for (i = 0; i < mmu_page_sizes; i++) {
662 		(void) memset(&m, 0, sizeof (m));
663 		id = it.mi_mnode_pfn_mask >> 15;	/* node id mask */
664 		id &= idmask[i];
665 		lo = lowbit(id);
666 		if (lo > 0) {
667 			hi = highbit(id);
668 			m.nnbits = hi - lo + 1;
669 			m.nnmask = (1 << m.nnbits) - 1;
670 			lo += nhbits[i] - 5;
671 			m.lomask = (1 << (lo - 1)) - 1;
672 			m.lobits = lo - 1;
673 		}
674 		hw_page_array[i].hp_colors = 1 << (nhbits[i] - m.nnbits);
675 		n2color[i] = m;
676 	}
677 }
678 
679 /*
680  * group colorequiv colors on N2 by low order bits of the color first
681  */
682 void
683 page_set_colorequiv_arr_cpu(void)
684 {
685 	static uint_t nequiv_shades_log2[MMU_PAGE_SIZES] = {2, 5, 0, 0, 0, 0};
686 
687 	nequiv_shades_log2[1] -= n2color[1].nnbits;
688 	if (colorequiv > 1) {
689 		int i;
690 		uint_t sv_a = lowbit(colorequiv) - 1;
691 
692 		if (sv_a > 15)
693 			sv_a = 15;
694 
695 		for (i = 0; i < MMU_PAGE_SIZES; i++) {
696 			uint_t colors;
697 			uint_t a = sv_a;
698 
699 			if ((colors = hw_page_array[i].hp_colors) <= 1)
700 				continue;
701 			while ((colors >> a) == 0)
702 				a--;
703 			if (a > (colorequivszc[i] & 0xf) +
704 			    (colorequivszc[i] >> 4)) {
705 				if (a <= nequiv_shades_log2[i]) {
706 					colorequivszc[i] = (uchar_t)a;
707 				} else {
708 					colorequivszc[i] =
709 					    ((a - nequiv_shades_log2[i]) << 4) |
710 					    nequiv_shades_log2[i];
711 				}
712 			}
713 		}
714 	}
715 }
716