xref: /illumos-gate/usr/src/uts/sun4v/cpu/niagara2.c (revision 7ec363dc481bba196d724969022171de4687989f)
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_size;
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 #ifdef SUN4V_CONTIG_MEM_PREALLOC_SIZE_MB
151 	/*
152 	 * Use CPU Makefile specific compile time define (if exists)
153 	 * to add to the contig preallocation size.
154 	 */
155 	contig_mem_prealloc_base_size = MB(SUN4V_CONTIG_MEM_PREALLOC_SIZE_MB);
156 #endif
157 }
158 
159 /*
160  * Set the magic constants of the implementation.
161  */
162 void
163 cpu_fiximp(struct cpu_node *cpunode)
164 {
165 	/*
166 	 * The Cache node is optional in MD. Therefore in case "Cache"
167 	 * node does not exists in MD, set the default L2 cache associativity,
168 	 * size, linesize.
169 	 */
170 	if (cpunode->ecache_size == 0)
171 		cpunode->ecache_size = L2CACHE_SIZE;
172 	if (cpunode->ecache_linesize == 0)
173 		cpunode->ecache_linesize = L2CACHE_LINESIZE;
174 	if (cpunode->ecache_associativity == 0)
175 		cpunode->ecache_associativity = L2CACHE_ASSOCIATIVITY;
176 }
177 
178 void
179 cpu_map_exec_units(struct cpu *cp)
180 {
181 	ASSERT(MUTEX_HELD(&cpu_lock));
182 
183 	/*
184 	 * The cpu_ipipe and cpu_fpu fields are initialized based on
185 	 * the execution unit sharing information from the MD. They
186 	 * default to the CPU id in the absence of such information.
187 	 */
188 	cp->cpu_m.cpu_ipipe = cpunodes[cp->cpu_id].exec_unit_mapping;
189 	if (cp->cpu_m.cpu_ipipe == NO_EU_MAPPING_FOUND)
190 		cp->cpu_m.cpu_ipipe = (id_t)(cp->cpu_id);
191 
192 	cp->cpu_m.cpu_fpu = cpunodes[cp->cpu_id].fpu_mapping;
193 	if (cp->cpu_m.cpu_fpu == NO_EU_MAPPING_FOUND)
194 		cp->cpu_m.cpu_fpu = (id_t)(cp->cpu_id);
195 
196 	/*
197 	 * Niagara 2 defines the core to be at the FPU level
198 	 */
199 	cp->cpu_m.cpu_core = cp->cpu_m.cpu_fpu;
200 
201 	/*
202 	 * The cpu_chip field is initialized based on the information
203 	 * in the MD and assume that all cpus within a chip
204 	 * share the same L2 cache. If no such info is available, we
205 	 * set the cpu to belong to the defacto chip 0.
206 	 */
207 	cp->cpu_m.cpu_mpipe = cpunodes[cp->cpu_id].l2_cache_mapping;
208 	if (cp->cpu_m.cpu_mpipe == NO_L2_CACHE_MAPPING_FOUND)
209 		cp->cpu_m.cpu_mpipe = CPU_L2_CACHEID_INVALID;
210 
211 	cp->cpu_m.cpu_chip = cpunodes[cp->cpu_id].l2_cache_mapping;
212 	if (cp->cpu_m.cpu_chip == NO_L2_CACHE_MAPPING_FOUND)
213 		cp->cpu_m.cpu_chip = CPU_CHIPID_INVALID;
214 }
215 
216 static int cpucnt;
217 
218 void
219 cpu_init_private(struct cpu *cp)
220 {
221 	extern void niagara_kstat_init(void);
222 
223 	ASSERT(MUTEX_HELD(&cpu_lock));
224 
225 	cpu_map_exec_units(cp);
226 
227 	if ((cpucnt++ == 0) && (cpu_hsvc_available == B_TRUE))
228 		(void) niagara_kstat_init();
229 }
230 
231 /*ARGSUSED*/
232 void
233 cpu_uninit_private(struct cpu *cp)
234 {
235 	extern void niagara_kstat_fini(void);
236 
237 	ASSERT(MUTEX_HELD(&cpu_lock));
238 	if ((--cpucnt == 0) && (cpu_hsvc_available == B_TRUE))
239 		(void) niagara_kstat_fini();
240 }
241 
242 /*
243  * On Niagara2, any flush will cause all preceding stores to be
244  * synchronized wrt the i$, regardless of address or ASI.  In fact,
245  * the address is ignored, so we always flush address 0.
246  */
247 /*ARGSUSED*/
248 void
249 dtrace_flush_sec(uintptr_t addr)
250 {
251 	doflush(0);
252 }
253 
254 /*
255  * Trapstat support for Niagara2 processor
256  * The Niagara2 provides HWTW support for TSB lookup and with HWTW
257  * enabled no TSB hit information will be available. Therefore setting
258  * the time spent in TLB miss handler for TSB hits to 0.
259  */
260 int
261 cpu_trapstat_conf(int cmd)
262 {
263 	int status = 0;
264 
265 	switch (cmd) {
266 	case CPU_TSTATCONF_INIT:
267 	case CPU_TSTATCONF_FINI:
268 	case CPU_TSTATCONF_ENABLE:
269 	case CPU_TSTATCONF_DISABLE:
270 		break;
271 	default:
272 		status = EINVAL;
273 		break;
274 	}
275 	return (status);
276 }
277 
278 void
279 cpu_trapstat_data(void *buf, uint_t tstat_pgszs)
280 {
281 	tstat_pgszdata_t	*tstatp = (tstat_pgszdata_t *)buf;
282 	int	i;
283 
284 	for (i = 0; i < tstat_pgszs; i++, tstatp++) {
285 		tstatp->tpgsz_kernel.tmode_itlb.ttlb_tlb.tmiss_count = 0;
286 		tstatp->tpgsz_kernel.tmode_itlb.ttlb_tlb.tmiss_time = 0;
287 		tstatp->tpgsz_user.tmode_itlb.ttlb_tlb.tmiss_count = 0;
288 		tstatp->tpgsz_user.tmode_itlb.ttlb_tlb.tmiss_time = 0;
289 		tstatp->tpgsz_kernel.tmode_dtlb.ttlb_tlb.tmiss_count = 0;
290 		tstatp->tpgsz_kernel.tmode_dtlb.ttlb_tlb.tmiss_time = 0;
291 		tstatp->tpgsz_user.tmode_dtlb.ttlb_tlb.tmiss_count = 0;
292 		tstatp->tpgsz_user.tmode_dtlb.ttlb_tlb.tmiss_time = 0;
293 	}
294 }
295 
296 /*
297  * Page coloring support for hashed cache index mode
298  */
299 
300 /*
301  * Node id bits from machine description (MD).  Node id distinguishes
302  * local versus remote memory. Because of MPO, page allocation does
303  * not cross node boundaries. Therefore, remove the node id bits from
304  * the color, since they are fixed. Either bit 30, or 31:30 in
305  * Victoria Falls processors.
306  * The number of node id bits is always 0 in Niagara2.
307  */
308 typedef struct n2color {
309 	uchar_t nnbits;	/* number of node id bits */
310 	uchar_t nnmask; /* mask for node id bits */
311 	uchar_t	lomask;	/* mask for bits below node id */
312 	uchar_t lobits;	/* number of bits below node id */
313 } n2color_t;
314 
315 n2color_t n2color[MMU_PAGE_SIZES];
316 static uchar_t nhbits[] = {7, 7, 6, 5, 5, 5};
317 
318 /*
319  * Remove node id bits from color bits 32:28.
320  * This will reduce the number of colors.
321  * No change if number of node bits is zero.
322  */
323 static inline uint_t
324 n2_hash2color(uint_t color, uchar_t szc)
325 {
326 	n2color_t m = n2color[szc];
327 
328 	if (m.nnbits > 0) {
329 		color = ((color >> m.nnbits) & ~m.lomask) | (color & m.lomask);
330 		ASSERT((color & ~(hw_page_array[szc].hp_colors - 1)) == 0);
331 	}
332 
333 	return (color);
334 }
335 
336 /*
337  * Restore node id bits into page color.
338  * This will increase the number of colors to match N2.
339  * No change if number of node bits is zero.
340  */
341 static inline uint_t
342 n2_color2hash(uint_t color, uchar_t szc, uint_t node)
343 {
344 	n2color_t m = n2color[szc];
345 
346 	if (m.nnbits > 0) {
347 		color = ((color & ~m.lomask) << m.nnbits) | (color & m.lomask);
348 		color |= (node & m.nnmask) << m.lobits;
349 	}
350 
351 	return (color);
352 }
353 
354 /* NI2 L2$ index is pa[32:28]^pa[17:13].pa[19:18]^pa[12:11].pa[10:6] */
355 
356 /*
357  * iterator NULL means pfn is VA, do not adjust ra_to_pa
358  * iterator (-1) means pfn is RA, need to convert to PA
359  * iterator non-null means pfn is RA, use ra_to_pa
360  */
361 uint_t
362 page_pfn_2_color_cpu(pfn_t pfn, uchar_t szc, void *cookie)
363 {
364 	mem_node_iterator_t *it = cookie;
365 	uint_t color;
366 
367 	ASSERT(szc <= TTE256M);
368 
369 	if (it == ((mem_node_iterator_t *)(-1))) {
370 		pfn = plat_rapfn_to_papfn(pfn);
371 	} else if (it != NULL) {
372 		ASSERT(pfn >= it->mi_mblock_base && pfn <= it->mi_mblock_end);
373 		pfn = pfn + it->mi_ra_to_pa;
374 	}
375 	pfn = PFN_BASE(pfn, szc);
376 	color = ((pfn >> 15) ^ pfn) & 0x1f;
377 	if (szc < TTE4M) {
378 		/* 19:18 */
379 		color = (color << 2) | ((pfn >> 5) & 0x3);
380 		if (szc > TTE64K)
381 			color >>= 1;    /* 19 */
382 	}
383 	return (n2_hash2color(color, szc));
384 }
385 
386 static uint_t
387 page_papfn_2_color_cpu(pfn_t papfn, uchar_t szc)
388 {
389 	uint_t color;
390 
391 	ASSERT(szc <= TTE256M);
392 
393 	papfn = PFN_BASE(papfn, szc);
394 	color = ((papfn >> 15) ^ papfn) & 0x1f;
395 	if (szc < TTE4M) {
396 		/* 19:18 */
397 		color = (color << 2) | ((papfn >> 5) & 0x3);
398 		if (szc > TTE64K)
399 			color >>= 1;    /* 19 */
400 	}
401 	return (color);
402 }
403 
404 #if TTE256M != 5
405 #error TTE256M is not 5
406 #endif
407 
408 uint_t
409 page_get_nsz_color_mask_cpu(uchar_t szc, uint_t mask)
410 {
411 	static uint_t ni2_color_masks[5] = {0x63, 0x1e, 0x3e, 0x1f, 0x1f};
412 	ASSERT(szc < TTE256M);
413 	mask = n2_color2hash(mask, szc, 0);
414 	mask &= ni2_color_masks[szc];
415 	if (szc == TTE64K || szc == TTE512K)
416 		mask >>= 1;
417 	return (n2_hash2color(mask, szc + 1));
418 }
419 
420 uint_t
421 page_get_nsz_color_cpu(uchar_t szc, uint_t color)
422 {
423 	ASSERT(szc < TTE256M);
424 	color = n2_color2hash(color, szc, 0);
425 	if (szc == TTE64K || szc == TTE512K)
426 		color >>= 1;
427 	return (n2_hash2color(color, szc + 1));
428 }
429 
430 uint_t
431 page_get_color_shift_cpu(uchar_t szc, uchar_t nszc)
432 {
433 	uint_t s;
434 	ASSERT(nszc >= szc);
435 	ASSERT(nszc <= TTE256M);
436 
437 	s = nhbits[szc] - n2color[szc].nnbits;
438 	s -= nhbits[nszc] - n2color[nszc].nnbits;
439 
440 	return (s);
441 }
442 
443 uint_t
444 page_convert_color_cpu(uint_t ncolor, uchar_t szc, uchar_t nszc)
445 {
446 	uint_t color;
447 
448 	ASSERT(nszc > szc);
449 	ASSERT(nszc <= TTE256M);
450 	ncolor = n2_color2hash(ncolor, nszc, 0);
451 	color = ncolor << (nhbits[szc] - nhbits[nszc]);
452 	color = n2_hash2color(color, szc);
453 	return (color);
454 }
455 
456 #define	PAPFN_2_MNODE(pfn) \
457 	(((pfn) & it->mi_mnode_pfn_mask) >> it->mi_mnode_pfn_shift)
458 
459 /*ARGSUSED*/
460 pfn_t
461 page_next_pfn_for_color_cpu(pfn_t pfn, uchar_t szc, uint_t color,
462     uint_t ceq_mask, uint_t color_mask, void *cookie)
463 {
464 	mem_node_iterator_t *it = cookie;
465 	pfn_t pstep = PNUM_SIZE(szc);
466 	pfn_t npfn, pfn_ceq_mask, pfn_color;
467 	pfn_t tmpmask, mask = (pfn_t)-1;
468 	uint_t pfnmn;
469 
470 	ASSERT((color & ~ceq_mask) == 0);
471 	ASSERT(pfn >= it->mi_mblock_base && pfn <= it->mi_mblock_end);
472 
473 	/* convert RA to PA for accurate color calculation */
474 	if (it->mi_init) {
475 		/* first call after it, so cache these values */
476 		it->mi_hash_ceq_mask =
477 		    n2_color2hash(ceq_mask, szc, it->mi_mnode_mask);
478 		it->mi_hash_color =
479 		    n2_color2hash(color, szc, it->mi_mnode);
480 		it->mi_init = 0;
481 	} else {
482 		ASSERT(it->mi_hash_ceq_mask ==
483 		    n2_color2hash(ceq_mask, szc, it->mi_mnode_mask));
484 		ASSERT(it->mi_hash_color ==
485 		    n2_color2hash(color, szc, it->mi_mnode));
486 	}
487 	ceq_mask = it->mi_hash_ceq_mask;
488 	color = it->mi_hash_color;
489 	pfn += it->mi_ra_to_pa;
490 
491 	/* restart here when we switch memblocks */
492 next_mem_block:
493 	if (szc <= TTE64K) {
494 		pfnmn = PAPFN_2_MNODE(pfn);
495 	}
496 	if (((page_papfn_2_color_cpu(pfn, szc) ^ color) & ceq_mask) == 0 &&
497 	    (szc > TTE64K || pfnmn == it->mi_mnode)) {
498 
499 		/* we start from the page with correct color */
500 		if (szc >= TTE512K) {
501 			if (szc >= TTE4M) {
502 				/* page color is PA[32:28] */
503 				pfn_ceq_mask = ceq_mask << 15;
504 			} else {
505 				/* page color is PA[32:28].PA[19:19] */
506 				pfn_ceq_mask = ((ceq_mask & 1) << 6) |
507 				    ((ceq_mask >> 1) << 15);
508 			}
509 			npfn = ADD_MASKED(pfn, pstep, pfn_ceq_mask, mask);
510 			goto done;
511 		} else {
512 			/*
513 			 * We deal 64K or 8K page. Check if we could the
514 			 * satisfy the request without changing PA[32:28]
515 			 */
516 			pfn_ceq_mask = ((ceq_mask & 3) << 5) | (ceq_mask >> 2);
517 			pfn_ceq_mask |= it->mi_mnode_pfn_mask;
518 			npfn = ADD_MASKED(pfn, pstep, pfn_ceq_mask, mask);
519 
520 			if ((((npfn ^ pfn) >> 15) & 0x1f) == 0)
521 				goto done;
522 
523 			/*
524 			 * for next pfn we have to change bits PA[32:28]
525 			 * set PA[63:28] and PA[19:18] of the next pfn
526 			 */
527 			npfn = (pfn >> 15) << 15;
528 			npfn |= (ceq_mask & color & 3) << 5;
529 			pfn_ceq_mask = (szc == TTE8K) ? 0 :
530 			    (ceq_mask & 0x1c) << 13;
531 			pfn_ceq_mask |= it->mi_mnode_pfn_mask;
532 			npfn = ADD_MASKED(npfn, (1 << 15), pfn_ceq_mask, mask);
533 
534 			/*
535 			 * set bits PA[17:13] to match the color
536 			 */
537 			npfn |= ((npfn >> 15) ^ (color >> 2)) & (ceq_mask >> 2);
538 			goto done;
539 		}
540 	}
541 
542 	/*
543 	 * we start from the page with incorrect color - rare case
544 	 */
545 	if (szc >= TTE512K) {
546 		if (szc >= TTE4M) {
547 			/* page color is in bits PA[32:28] */
548 			npfn = ((pfn >> 20) << 20) | (color << 15);
549 			pfn_ceq_mask = (ceq_mask << 15) | 0x7fff;
550 		} else {
551 			/* try get the right color by changing bit PA[19:19] */
552 			npfn = pfn + pstep;
553 			if (((page_papfn_2_color_cpu(npfn, szc) ^ color) &
554 			    ceq_mask) == 0)
555 				goto done;
556 
557 			/* page color is PA[32:28].PA[19:19] */
558 			pfn_ceq_mask = ((ceq_mask & 1) << 6) |
559 			    ((ceq_mask >> 1) << 15) | (0xff << 7);
560 			pfn_color = ((color & 1) << 6) | ((color >> 1) << 15);
561 			npfn = ((pfn >> 20) << 20) | pfn_color;
562 		}
563 
564 		while (npfn <= pfn) {
565 			npfn = ADD_MASKED(npfn, pstep, pfn_ceq_mask, mask);
566 		}
567 		goto done;
568 	}
569 
570 	/*
571 	 *  We deal 64K or 8K page of incorrect color.
572 	 * Try correcting color without changing PA[32:28]
573 	 */
574 	pfn_ceq_mask = ((ceq_mask & 3) << 5) | (ceq_mask >> 2);
575 	pfn_color = ((color & 3) << 5) | (color >> 2);
576 	if (pfnmn == it->mi_mnode) {
577 		npfn = (pfn & ~(pfn_t)0x7f);
578 		npfn |= (((pfn >> 15) & 0x1f) ^ pfn_color) & pfn_ceq_mask;
579 		npfn = (szc == TTE64K) ? (npfn & ~(pfn_t)0x7) : npfn;
580 
581 		if (((page_papfn_2_color_cpu(npfn, szc) ^ color) &
582 		    ceq_mask) == 0) {
583 			/* the color is fixed - find the next page */
584 			pfn_ceq_mask |= it->mi_mnode_pfn_mask;
585 			while (npfn <= pfn) {
586 				npfn = ADD_MASKED(npfn, pstep, pfn_ceq_mask,
587 				    mask);
588 			}
589 			if ((((npfn ^ pfn) >> 15) & 0x1f) == 0)
590 				goto done;
591 		}
592 	}
593 
594 	/* to fix the color need to touch PA[32:28] */
595 	npfn = (szc == TTE8K) ? ((pfn >> 15) << 15) :
596 	    (((pfn >> 18) << 18) | ((color & 0x1c) << 13));
597 
598 	/* fix mnode if input pfn is in the wrong mnode. */
599 	if ((pfnmn = PAPFN_2_MNODE(npfn)) != it->mi_mnode) {
600 		npfn += ((it->mi_mnode - pfnmn) & it->mi_mnode_mask) <<
601 		    it->mi_mnode_pfn_shift;
602 	}
603 
604 	tmpmask = (szc == TTE8K) ? 0 : (ceq_mask & 0x1c) << 13;
605 	tmpmask |= it->mi_mnode_pfn_mask;
606 
607 	while (npfn <= pfn) {
608 		npfn = ADD_MASKED(npfn, (1 << 15), tmpmask, mask);
609 	}
610 
611 	/* set bits PA[19:13] to match the color */
612 	npfn |= (((npfn >> 15) & 0x1f) ^ pfn_color) & pfn_ceq_mask;
613 	npfn = (szc == TTE64K) ? (npfn & ~(pfn_t)0x7) : npfn;
614 
615 done:
616 	ASSERT(((page_papfn_2_color_cpu(npfn, szc) ^ color) & ceq_mask) == 0);
617 	ASSERT(PAPFN_2_MNODE(npfn) == it->mi_mnode);
618 
619 	/* PA to RA */
620 	npfn -= it->mi_ra_to_pa;
621 
622 	/* check for possible memblock switch */
623 	if (npfn > it->mi_mblock_end) {
624 		pfn = plat_mem_node_iterator_init(npfn, it->mi_mnode, it, 0);
625 		if (pfn == (pfn_t)-1)
626 			return (pfn);
627 		ASSERT(pfn >= it->mi_mblock_base && pfn <= it->mi_mblock_end);
628 		pfn += it->mi_ra_to_pa;
629 		goto next_mem_block;
630 	}
631 
632 	return (npfn);
633 }
634 
635 /*
636  * init page coloring
637  * VF encodes node_id for an L-group in either bit 30 or 31:30,
638  * which effectively reduces the number of colors available per mnode.
639  */
640 void
641 page_coloring_init_cpu()
642 {
643 	int i;
644 	uchar_t id;
645 	uchar_t lo;
646 	uchar_t hi;
647 	n2color_t m;
648 	mem_node_iterator_t it;
649 	static uchar_t idmask[] = {0, 0x7, 0x1f, 0x1f, 0x1f, 0x1f};
650 
651 	for (i = 0; i < max_mem_nodes; i++) {
652 		memset(&it, 0, sizeof (it));
653 		if (plat_mem_node_iterator_init(0, i, &it, 1) != (pfn_t)-1)
654 			break;
655 	}
656 	ASSERT(i < max_mem_nodes);
657 	for (i = 0; i < mmu_page_sizes; i++) {
658 		(void) memset(&m, 0, sizeof (m));
659 		id = it.mi_mnode_pfn_mask >> 15;	/* node id mask */
660 		id &= idmask[i];
661 		lo = lowbit(id);
662 		if (lo > 0) {
663 			hi = highbit(id);
664 			m.nnbits = hi - lo + 1;
665 			m.nnmask = (1 << m.nnbits) - 1;
666 			lo += nhbits[i] - 5;
667 			m.lomask = (1 << (lo - 1)) - 1;
668 			m.lobits = lo - 1;
669 		}
670 		hw_page_array[i].hp_colors = 1 << (nhbits[i] - m.nnbits);
671 		n2color[i] = m;
672 	}
673 }
674 
675 /*
676  * group colorequiv colors on N2 by low order bits of the color first
677  */
678 void
679 page_set_colorequiv_arr_cpu(void)
680 {
681 	static uint_t nequiv_shades_log2[MMU_PAGE_SIZES] = {2, 5, 0, 0, 0, 0};
682 
683 	nequiv_shades_log2[1] -= n2color[1].nnbits;
684 	if (colorequiv > 1) {
685 		int i;
686 		uint_t sv_a = lowbit(colorequiv) - 1;
687 
688 		if (sv_a > 15)
689 			sv_a = 15;
690 
691 		for (i = 0; i < MMU_PAGE_SIZES; i++) {
692 			uint_t colors;
693 			uint_t a = sv_a;
694 
695 			if ((colors = hw_page_array[i].hp_colors) <= 1)
696 				continue;
697 			while ((colors >> a) == 0)
698 				a--;
699 			if (a > (colorequivszc[i] & 0xf) +
700 			    (colorequivszc[i] >> 4)) {
701 				if (a <= nequiv_shades_log2[i]) {
702 					colorequivszc[i] = (uchar_t)a;
703 				} else {
704 					colorequivszc[i] =
705 					    ((a - nequiv_shades_log2[i]) << 4) |
706 					    nequiv_shades_log2[i];
707 				}
708 			}
709 		}
710 	}
711 }
712