xref: /linux/arch/sparc/mm/init_64.c (revision f3a8b6645dc2e60d11f20c1c23afd964ff4e55ae)
1 /*
2  *  arch/sparc64/mm/init.c
3  *
4  *  Copyright (C) 1996-1999 David S. Miller (davem@caip.rutgers.edu)
5  *  Copyright (C) 1997-1999 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
6  */
7 
8 #include <linux/extable.h>
9 #include <linux/kernel.h>
10 #include <linux/sched.h>
11 #include <linux/string.h>
12 #include <linux/init.h>
13 #include <linux/bootmem.h>
14 #include <linux/mm.h>
15 #include <linux/hugetlb.h>
16 #include <linux/initrd.h>
17 #include <linux/swap.h>
18 #include <linux/pagemap.h>
19 #include <linux/poison.h>
20 #include <linux/fs.h>
21 #include <linux/seq_file.h>
22 #include <linux/kprobes.h>
23 #include <linux/cache.h>
24 #include <linux/sort.h>
25 #include <linux/ioport.h>
26 #include <linux/percpu.h>
27 #include <linux/memblock.h>
28 #include <linux/mmzone.h>
29 #include <linux/gfp.h>
30 
31 #include <asm/head.h>
32 #include <asm/page.h>
33 #include <asm/pgalloc.h>
34 #include <asm/pgtable.h>
35 #include <asm/oplib.h>
36 #include <asm/iommu.h>
37 #include <asm/io.h>
38 #include <asm/uaccess.h>
39 #include <asm/mmu_context.h>
40 #include <asm/tlbflush.h>
41 #include <asm/dma.h>
42 #include <asm/starfire.h>
43 #include <asm/tlb.h>
44 #include <asm/spitfire.h>
45 #include <asm/sections.h>
46 #include <asm/tsb.h>
47 #include <asm/hypervisor.h>
48 #include <asm/prom.h>
49 #include <asm/mdesc.h>
50 #include <asm/cpudata.h>
51 #include <asm/setup.h>
52 #include <asm/irq.h>
53 
54 #include "init_64.h"
55 
56 unsigned long kern_linear_pte_xor[4] __read_mostly;
57 static unsigned long page_cache4v_flag;
58 
59 /* A bitmap, two bits for every 256MB of physical memory.  These two
60  * bits determine what page size we use for kernel linear
61  * translations.  They form an index into kern_linear_pte_xor[].  The
62  * value in the indexed slot is XOR'd with the TLB miss virtual
63  * address to form the resulting TTE.  The mapping is:
64  *
65  *	0	==>	4MB
66  *	1	==>	256MB
67  *	2	==>	2GB
68  *	3	==>	16GB
69  *
70  * All sun4v chips support 256MB pages.  Only SPARC-T4 and later
71  * support 2GB pages, and hopefully future cpus will support the 16GB
72  * pages as well.  For slots 2 and 3, we encode a 256MB TTE xor there
73  * if these larger page sizes are not supported by the cpu.
74  *
75  * It would be nice to determine this from the machine description
76  * 'cpu' properties, but we need to have this table setup before the
77  * MDESC is initialized.
78  */
79 
80 #ifndef CONFIG_DEBUG_PAGEALLOC
81 /* A special kernel TSB for 4MB, 256MB, 2GB and 16GB linear mappings.
82  * Space is allocated for this right after the trap table in
83  * arch/sparc64/kernel/head.S
84  */
85 extern struct tsb swapper_4m_tsb[KERNEL_TSB4M_NENTRIES];
86 #endif
87 extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
88 
89 static unsigned long cpu_pgsz_mask;
90 
91 #define MAX_BANKS	1024
92 
93 static struct linux_prom64_registers pavail[MAX_BANKS];
94 static int pavail_ents;
95 
96 u64 numa_latency[MAX_NUMNODES][MAX_NUMNODES];
97 
98 static int cmp_p64(const void *a, const void *b)
99 {
100 	const struct linux_prom64_registers *x = a, *y = b;
101 
102 	if (x->phys_addr > y->phys_addr)
103 		return 1;
104 	if (x->phys_addr < y->phys_addr)
105 		return -1;
106 	return 0;
107 }
108 
109 static void __init read_obp_memory(const char *property,
110 				   struct linux_prom64_registers *regs,
111 				   int *num_ents)
112 {
113 	phandle node = prom_finddevice("/memory");
114 	int prop_size = prom_getproplen(node, property);
115 	int ents, ret, i;
116 
117 	ents = prop_size / sizeof(struct linux_prom64_registers);
118 	if (ents > MAX_BANKS) {
119 		prom_printf("The machine has more %s property entries than "
120 			    "this kernel can support (%d).\n",
121 			    property, MAX_BANKS);
122 		prom_halt();
123 	}
124 
125 	ret = prom_getproperty(node, property, (char *) regs, prop_size);
126 	if (ret == -1) {
127 		prom_printf("Couldn't get %s property from /memory.\n",
128 				property);
129 		prom_halt();
130 	}
131 
132 	/* Sanitize what we got from the firmware, by page aligning
133 	 * everything.
134 	 */
135 	for (i = 0; i < ents; i++) {
136 		unsigned long base, size;
137 
138 		base = regs[i].phys_addr;
139 		size = regs[i].reg_size;
140 
141 		size &= PAGE_MASK;
142 		if (base & ~PAGE_MASK) {
143 			unsigned long new_base = PAGE_ALIGN(base);
144 
145 			size -= new_base - base;
146 			if ((long) size < 0L)
147 				size = 0UL;
148 			base = new_base;
149 		}
150 		if (size == 0UL) {
151 			/* If it is empty, simply get rid of it.
152 			 * This simplifies the logic of the other
153 			 * functions that process these arrays.
154 			 */
155 			memmove(&regs[i], &regs[i + 1],
156 				(ents - i - 1) * sizeof(regs[0]));
157 			i--;
158 			ents--;
159 			continue;
160 		}
161 		regs[i].phys_addr = base;
162 		regs[i].reg_size = size;
163 	}
164 
165 	*num_ents = ents;
166 
167 	sort(regs, ents, sizeof(struct linux_prom64_registers),
168 	     cmp_p64, NULL);
169 }
170 
171 /* Kernel physical address base and size in bytes.  */
172 unsigned long kern_base __read_mostly;
173 unsigned long kern_size __read_mostly;
174 
175 /* Initial ramdisk setup */
176 extern unsigned long sparc_ramdisk_image64;
177 extern unsigned int sparc_ramdisk_image;
178 extern unsigned int sparc_ramdisk_size;
179 
180 struct page *mem_map_zero __read_mostly;
181 EXPORT_SYMBOL(mem_map_zero);
182 
183 unsigned int sparc64_highest_unlocked_tlb_ent __read_mostly;
184 
185 unsigned long sparc64_kern_pri_context __read_mostly;
186 unsigned long sparc64_kern_pri_nuc_bits __read_mostly;
187 unsigned long sparc64_kern_sec_context __read_mostly;
188 
189 int num_kernel_image_mappings;
190 
191 #ifdef CONFIG_DEBUG_DCFLUSH
192 atomic_t dcpage_flushes = ATOMIC_INIT(0);
193 #ifdef CONFIG_SMP
194 atomic_t dcpage_flushes_xcall = ATOMIC_INIT(0);
195 #endif
196 #endif
197 
198 inline void flush_dcache_page_impl(struct page *page)
199 {
200 	BUG_ON(tlb_type == hypervisor);
201 #ifdef CONFIG_DEBUG_DCFLUSH
202 	atomic_inc(&dcpage_flushes);
203 #endif
204 
205 #ifdef DCACHE_ALIASING_POSSIBLE
206 	__flush_dcache_page(page_address(page),
207 			    ((tlb_type == spitfire) &&
208 			     page_mapping(page) != NULL));
209 #else
210 	if (page_mapping(page) != NULL &&
211 	    tlb_type == spitfire)
212 		__flush_icache_page(__pa(page_address(page)));
213 #endif
214 }
215 
216 #define PG_dcache_dirty		PG_arch_1
217 #define PG_dcache_cpu_shift	32UL
218 #define PG_dcache_cpu_mask	\
219 	((1UL<<ilog2(roundup_pow_of_two(NR_CPUS)))-1UL)
220 
221 #define dcache_dirty_cpu(page) \
222 	(((page)->flags >> PG_dcache_cpu_shift) & PG_dcache_cpu_mask)
223 
224 static inline void set_dcache_dirty(struct page *page, int this_cpu)
225 {
226 	unsigned long mask = this_cpu;
227 	unsigned long non_cpu_bits;
228 
229 	non_cpu_bits = ~(PG_dcache_cpu_mask << PG_dcache_cpu_shift);
230 	mask = (mask << PG_dcache_cpu_shift) | (1UL << PG_dcache_dirty);
231 
232 	__asm__ __volatile__("1:\n\t"
233 			     "ldx	[%2], %%g7\n\t"
234 			     "and	%%g7, %1, %%g1\n\t"
235 			     "or	%%g1, %0, %%g1\n\t"
236 			     "casx	[%2], %%g7, %%g1\n\t"
237 			     "cmp	%%g7, %%g1\n\t"
238 			     "bne,pn	%%xcc, 1b\n\t"
239 			     " nop"
240 			     : /* no outputs */
241 			     : "r" (mask), "r" (non_cpu_bits), "r" (&page->flags)
242 			     : "g1", "g7");
243 }
244 
245 static inline void clear_dcache_dirty_cpu(struct page *page, unsigned long cpu)
246 {
247 	unsigned long mask = (1UL << PG_dcache_dirty);
248 
249 	__asm__ __volatile__("! test_and_clear_dcache_dirty\n"
250 			     "1:\n\t"
251 			     "ldx	[%2], %%g7\n\t"
252 			     "srlx	%%g7, %4, %%g1\n\t"
253 			     "and	%%g1, %3, %%g1\n\t"
254 			     "cmp	%%g1, %0\n\t"
255 			     "bne,pn	%%icc, 2f\n\t"
256 			     " andn	%%g7, %1, %%g1\n\t"
257 			     "casx	[%2], %%g7, %%g1\n\t"
258 			     "cmp	%%g7, %%g1\n\t"
259 			     "bne,pn	%%xcc, 1b\n\t"
260 			     " nop\n"
261 			     "2:"
262 			     : /* no outputs */
263 			     : "r" (cpu), "r" (mask), "r" (&page->flags),
264 			       "i" (PG_dcache_cpu_mask),
265 			       "i" (PG_dcache_cpu_shift)
266 			     : "g1", "g7");
267 }
268 
269 static inline void tsb_insert(struct tsb *ent, unsigned long tag, unsigned long pte)
270 {
271 	unsigned long tsb_addr = (unsigned long) ent;
272 
273 	if (tlb_type == cheetah_plus || tlb_type == hypervisor)
274 		tsb_addr = __pa(tsb_addr);
275 
276 	__tsb_insert(tsb_addr, tag, pte);
277 }
278 
279 unsigned long _PAGE_ALL_SZ_BITS __read_mostly;
280 
281 static void flush_dcache(unsigned long pfn)
282 {
283 	struct page *page;
284 
285 	page = pfn_to_page(pfn);
286 	if (page) {
287 		unsigned long pg_flags;
288 
289 		pg_flags = page->flags;
290 		if (pg_flags & (1UL << PG_dcache_dirty)) {
291 			int cpu = ((pg_flags >> PG_dcache_cpu_shift) &
292 				   PG_dcache_cpu_mask);
293 			int this_cpu = get_cpu();
294 
295 			/* This is just to optimize away some function calls
296 			 * in the SMP case.
297 			 */
298 			if (cpu == this_cpu)
299 				flush_dcache_page_impl(page);
300 			else
301 				smp_flush_dcache_page_impl(page, cpu);
302 
303 			clear_dcache_dirty_cpu(page, cpu);
304 
305 			put_cpu();
306 		}
307 	}
308 }
309 
310 /* mm->context.lock must be held */
311 static void __update_mmu_tsb_insert(struct mm_struct *mm, unsigned long tsb_index,
312 				    unsigned long tsb_hash_shift, unsigned long address,
313 				    unsigned long tte)
314 {
315 	struct tsb *tsb = mm->context.tsb_block[tsb_index].tsb;
316 	unsigned long tag;
317 
318 	if (unlikely(!tsb))
319 		return;
320 
321 	tsb += ((address >> tsb_hash_shift) &
322 		(mm->context.tsb_block[tsb_index].tsb_nentries - 1UL));
323 	tag = (address >> 22UL);
324 	tsb_insert(tsb, tag, tte);
325 }
326 
327 void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
328 {
329 	struct mm_struct *mm;
330 	unsigned long flags;
331 	pte_t pte = *ptep;
332 
333 	if (tlb_type != hypervisor) {
334 		unsigned long pfn = pte_pfn(pte);
335 
336 		if (pfn_valid(pfn))
337 			flush_dcache(pfn);
338 	}
339 
340 	mm = vma->vm_mm;
341 
342 	/* Don't insert a non-valid PTE into the TSB, we'll deadlock.  */
343 	if (!pte_accessible(mm, pte))
344 		return;
345 
346 	spin_lock_irqsave(&mm->context.lock, flags);
347 
348 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
349 	if ((mm->context.hugetlb_pte_count || mm->context.thp_pte_count) &&
350 	    is_hugetlb_pte(pte)) {
351 		/* We are fabricating 8MB pages using 4MB real hw pages.  */
352 		pte_val(pte) |= (address & (1UL << REAL_HPAGE_SHIFT));
353 		__update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
354 					address, pte_val(pte));
355 	} else
356 #endif
357 		__update_mmu_tsb_insert(mm, MM_TSB_BASE, PAGE_SHIFT,
358 					address, pte_val(pte));
359 
360 	spin_unlock_irqrestore(&mm->context.lock, flags);
361 }
362 
363 void flush_dcache_page(struct page *page)
364 {
365 	struct address_space *mapping;
366 	int this_cpu;
367 
368 	if (tlb_type == hypervisor)
369 		return;
370 
371 	/* Do not bother with the expensive D-cache flush if it
372 	 * is merely the zero page.  The 'bigcore' testcase in GDB
373 	 * causes this case to run millions of times.
374 	 */
375 	if (page == ZERO_PAGE(0))
376 		return;
377 
378 	this_cpu = get_cpu();
379 
380 	mapping = page_mapping(page);
381 	if (mapping && !mapping_mapped(mapping)) {
382 		int dirty = test_bit(PG_dcache_dirty, &page->flags);
383 		if (dirty) {
384 			int dirty_cpu = dcache_dirty_cpu(page);
385 
386 			if (dirty_cpu == this_cpu)
387 				goto out;
388 			smp_flush_dcache_page_impl(page, dirty_cpu);
389 		}
390 		set_dcache_dirty(page, this_cpu);
391 	} else {
392 		/* We could delay the flush for the !page_mapping
393 		 * case too.  But that case is for exec env/arg
394 		 * pages and those are %99 certainly going to get
395 		 * faulted into the tlb (and thus flushed) anyways.
396 		 */
397 		flush_dcache_page_impl(page);
398 	}
399 
400 out:
401 	put_cpu();
402 }
403 EXPORT_SYMBOL(flush_dcache_page);
404 
405 void __kprobes flush_icache_range(unsigned long start, unsigned long end)
406 {
407 	/* Cheetah and Hypervisor platform cpus have coherent I-cache. */
408 	if (tlb_type == spitfire) {
409 		unsigned long kaddr;
410 
411 		/* This code only runs on Spitfire cpus so this is
412 		 * why we can assume _PAGE_PADDR_4U.
413 		 */
414 		for (kaddr = start; kaddr < end; kaddr += PAGE_SIZE) {
415 			unsigned long paddr, mask = _PAGE_PADDR_4U;
416 
417 			if (kaddr >= PAGE_OFFSET)
418 				paddr = kaddr & mask;
419 			else {
420 				pgd_t *pgdp = pgd_offset_k(kaddr);
421 				pud_t *pudp = pud_offset(pgdp, kaddr);
422 				pmd_t *pmdp = pmd_offset(pudp, kaddr);
423 				pte_t *ptep = pte_offset_kernel(pmdp, kaddr);
424 
425 				paddr = pte_val(*ptep) & mask;
426 			}
427 			__flush_icache_page(paddr);
428 		}
429 	}
430 }
431 EXPORT_SYMBOL(flush_icache_range);
432 
433 void mmu_info(struct seq_file *m)
434 {
435 	static const char *pgsz_strings[] = {
436 		"8K", "64K", "512K", "4MB", "32MB",
437 		"256MB", "2GB", "16GB",
438 	};
439 	int i, printed;
440 
441 	if (tlb_type == cheetah)
442 		seq_printf(m, "MMU Type\t: Cheetah\n");
443 	else if (tlb_type == cheetah_plus)
444 		seq_printf(m, "MMU Type\t: Cheetah+\n");
445 	else if (tlb_type == spitfire)
446 		seq_printf(m, "MMU Type\t: Spitfire\n");
447 	else if (tlb_type == hypervisor)
448 		seq_printf(m, "MMU Type\t: Hypervisor (sun4v)\n");
449 	else
450 		seq_printf(m, "MMU Type\t: ???\n");
451 
452 	seq_printf(m, "MMU PGSZs\t: ");
453 	printed = 0;
454 	for (i = 0; i < ARRAY_SIZE(pgsz_strings); i++) {
455 		if (cpu_pgsz_mask & (1UL << i)) {
456 			seq_printf(m, "%s%s",
457 				   printed ? "," : "", pgsz_strings[i]);
458 			printed++;
459 		}
460 	}
461 	seq_putc(m, '\n');
462 
463 #ifdef CONFIG_DEBUG_DCFLUSH
464 	seq_printf(m, "DCPageFlushes\t: %d\n",
465 		   atomic_read(&dcpage_flushes));
466 #ifdef CONFIG_SMP
467 	seq_printf(m, "DCPageFlushesXC\t: %d\n",
468 		   atomic_read(&dcpage_flushes_xcall));
469 #endif /* CONFIG_SMP */
470 #endif /* CONFIG_DEBUG_DCFLUSH */
471 }
472 
473 struct linux_prom_translation prom_trans[512] __read_mostly;
474 unsigned int prom_trans_ents __read_mostly;
475 
476 unsigned long kern_locked_tte_data;
477 
478 /* The obp translations are saved based on 8k pagesize, since obp can
479  * use a mixture of pagesizes. Misses to the LOW_OBP_ADDRESS ->
480  * HI_OBP_ADDRESS range are handled in ktlb.S.
481  */
482 static inline int in_obp_range(unsigned long vaddr)
483 {
484 	return (vaddr >= LOW_OBP_ADDRESS &&
485 		vaddr < HI_OBP_ADDRESS);
486 }
487 
488 static int cmp_ptrans(const void *a, const void *b)
489 {
490 	const struct linux_prom_translation *x = a, *y = b;
491 
492 	if (x->virt > y->virt)
493 		return 1;
494 	if (x->virt < y->virt)
495 		return -1;
496 	return 0;
497 }
498 
499 /* Read OBP translations property into 'prom_trans[]'.  */
500 static void __init read_obp_translations(void)
501 {
502 	int n, node, ents, first, last, i;
503 
504 	node = prom_finddevice("/virtual-memory");
505 	n = prom_getproplen(node, "translations");
506 	if (unlikely(n == 0 || n == -1)) {
507 		prom_printf("prom_mappings: Couldn't get size.\n");
508 		prom_halt();
509 	}
510 	if (unlikely(n > sizeof(prom_trans))) {
511 		prom_printf("prom_mappings: Size %d is too big.\n", n);
512 		prom_halt();
513 	}
514 
515 	if ((n = prom_getproperty(node, "translations",
516 				  (char *)&prom_trans[0],
517 				  sizeof(prom_trans))) == -1) {
518 		prom_printf("prom_mappings: Couldn't get property.\n");
519 		prom_halt();
520 	}
521 
522 	n = n / sizeof(struct linux_prom_translation);
523 
524 	ents = n;
525 
526 	sort(prom_trans, ents, sizeof(struct linux_prom_translation),
527 	     cmp_ptrans, NULL);
528 
529 	/* Now kick out all the non-OBP entries.  */
530 	for (i = 0; i < ents; i++) {
531 		if (in_obp_range(prom_trans[i].virt))
532 			break;
533 	}
534 	first = i;
535 	for (; i < ents; i++) {
536 		if (!in_obp_range(prom_trans[i].virt))
537 			break;
538 	}
539 	last = i;
540 
541 	for (i = 0; i < (last - first); i++) {
542 		struct linux_prom_translation *src = &prom_trans[i + first];
543 		struct linux_prom_translation *dest = &prom_trans[i];
544 
545 		*dest = *src;
546 	}
547 	for (; i < ents; i++) {
548 		struct linux_prom_translation *dest = &prom_trans[i];
549 		dest->virt = dest->size = dest->data = 0x0UL;
550 	}
551 
552 	prom_trans_ents = last - first;
553 
554 	if (tlb_type == spitfire) {
555 		/* Clear diag TTE bits. */
556 		for (i = 0; i < prom_trans_ents; i++)
557 			prom_trans[i].data &= ~0x0003fe0000000000UL;
558 	}
559 
560 	/* Force execute bit on.  */
561 	for (i = 0; i < prom_trans_ents; i++)
562 		prom_trans[i].data |= (tlb_type == hypervisor ?
563 				       _PAGE_EXEC_4V : _PAGE_EXEC_4U);
564 }
565 
566 static void __init hypervisor_tlb_lock(unsigned long vaddr,
567 				       unsigned long pte,
568 				       unsigned long mmu)
569 {
570 	unsigned long ret = sun4v_mmu_map_perm_addr(vaddr, 0, pte, mmu);
571 
572 	if (ret != 0) {
573 		prom_printf("hypervisor_tlb_lock[%lx:%x:%lx:%lx]: "
574 			    "errors with %lx\n", vaddr, 0, pte, mmu, ret);
575 		prom_halt();
576 	}
577 }
578 
579 static unsigned long kern_large_tte(unsigned long paddr);
580 
581 static void __init remap_kernel(void)
582 {
583 	unsigned long phys_page, tte_vaddr, tte_data;
584 	int i, tlb_ent = sparc64_highest_locked_tlbent();
585 
586 	tte_vaddr = (unsigned long) KERNBASE;
587 	phys_page = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
588 	tte_data = kern_large_tte(phys_page);
589 
590 	kern_locked_tte_data = tte_data;
591 
592 	/* Now lock us into the TLBs via Hypervisor or OBP. */
593 	if (tlb_type == hypervisor) {
594 		for (i = 0; i < num_kernel_image_mappings; i++) {
595 			hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_DMMU);
596 			hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_IMMU);
597 			tte_vaddr += 0x400000;
598 			tte_data += 0x400000;
599 		}
600 	} else {
601 		for (i = 0; i < num_kernel_image_mappings; i++) {
602 			prom_dtlb_load(tlb_ent - i, tte_data, tte_vaddr);
603 			prom_itlb_load(tlb_ent - i, tte_data, tte_vaddr);
604 			tte_vaddr += 0x400000;
605 			tte_data += 0x400000;
606 		}
607 		sparc64_highest_unlocked_tlb_ent = tlb_ent - i;
608 	}
609 	if (tlb_type == cheetah_plus) {
610 		sparc64_kern_pri_context = (CTX_CHEETAH_PLUS_CTX0 |
611 					    CTX_CHEETAH_PLUS_NUC);
612 		sparc64_kern_pri_nuc_bits = CTX_CHEETAH_PLUS_NUC;
613 		sparc64_kern_sec_context = CTX_CHEETAH_PLUS_CTX0;
614 	}
615 }
616 
617 
618 static void __init inherit_prom_mappings(void)
619 {
620 	/* Now fixup OBP's idea about where we really are mapped. */
621 	printk("Remapping the kernel... ");
622 	remap_kernel();
623 	printk("done.\n");
624 }
625 
626 void prom_world(int enter)
627 {
628 	if (!enter)
629 		set_fs(get_fs());
630 
631 	__asm__ __volatile__("flushw");
632 }
633 
634 void __flush_dcache_range(unsigned long start, unsigned long end)
635 {
636 	unsigned long va;
637 
638 	if (tlb_type == spitfire) {
639 		int n = 0;
640 
641 		for (va = start; va < end; va += 32) {
642 			spitfire_put_dcache_tag(va & 0x3fe0, 0x0);
643 			if (++n >= 512)
644 				break;
645 		}
646 	} else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
647 		start = __pa(start);
648 		end = __pa(end);
649 		for (va = start; va < end; va += 32)
650 			__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
651 					     "membar #Sync"
652 					     : /* no outputs */
653 					     : "r" (va),
654 					       "i" (ASI_DCACHE_INVALIDATE));
655 	}
656 }
657 EXPORT_SYMBOL(__flush_dcache_range);
658 
659 /* get_new_mmu_context() uses "cache + 1".  */
660 DEFINE_SPINLOCK(ctx_alloc_lock);
661 unsigned long tlb_context_cache = CTX_FIRST_VERSION - 1;
662 #define MAX_CTX_NR	(1UL << CTX_NR_BITS)
663 #define CTX_BMAP_SLOTS	BITS_TO_LONGS(MAX_CTX_NR)
664 DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR);
665 
666 /* Caller does TLB context flushing on local CPU if necessary.
667  * The caller also ensures that CTX_VALID(mm->context) is false.
668  *
669  * We must be careful about boundary cases so that we never
670  * let the user have CTX 0 (nucleus) or we ever use a CTX
671  * version of zero (and thus NO_CONTEXT would not be caught
672  * by version mis-match tests in mmu_context.h).
673  *
674  * Always invoked with interrupts disabled.
675  */
676 void get_new_mmu_context(struct mm_struct *mm)
677 {
678 	unsigned long ctx, new_ctx;
679 	unsigned long orig_pgsz_bits;
680 	int new_version;
681 
682 	spin_lock(&ctx_alloc_lock);
683 	orig_pgsz_bits = (mm->context.sparc64_ctx_val & CTX_PGSZ_MASK);
684 	ctx = (tlb_context_cache + 1) & CTX_NR_MASK;
685 	new_ctx = find_next_zero_bit(mmu_context_bmap, 1 << CTX_NR_BITS, ctx);
686 	new_version = 0;
687 	if (new_ctx >= (1 << CTX_NR_BITS)) {
688 		new_ctx = find_next_zero_bit(mmu_context_bmap, ctx, 1);
689 		if (new_ctx >= ctx) {
690 			int i;
691 			new_ctx = (tlb_context_cache & CTX_VERSION_MASK) +
692 				CTX_FIRST_VERSION;
693 			if (new_ctx == 1)
694 				new_ctx = CTX_FIRST_VERSION;
695 
696 			/* Don't call memset, for 16 entries that's just
697 			 * plain silly...
698 			 */
699 			mmu_context_bmap[0] = 3;
700 			mmu_context_bmap[1] = 0;
701 			mmu_context_bmap[2] = 0;
702 			mmu_context_bmap[3] = 0;
703 			for (i = 4; i < CTX_BMAP_SLOTS; i += 4) {
704 				mmu_context_bmap[i + 0] = 0;
705 				mmu_context_bmap[i + 1] = 0;
706 				mmu_context_bmap[i + 2] = 0;
707 				mmu_context_bmap[i + 3] = 0;
708 			}
709 			new_version = 1;
710 			goto out;
711 		}
712 	}
713 	mmu_context_bmap[new_ctx>>6] |= (1UL << (new_ctx & 63));
714 	new_ctx |= (tlb_context_cache & CTX_VERSION_MASK);
715 out:
716 	tlb_context_cache = new_ctx;
717 	mm->context.sparc64_ctx_val = new_ctx | orig_pgsz_bits;
718 	spin_unlock(&ctx_alloc_lock);
719 
720 	if (unlikely(new_version))
721 		smp_new_mmu_context_version();
722 }
723 
724 static int numa_enabled = 1;
725 static int numa_debug;
726 
727 static int __init early_numa(char *p)
728 {
729 	if (!p)
730 		return 0;
731 
732 	if (strstr(p, "off"))
733 		numa_enabled = 0;
734 
735 	if (strstr(p, "debug"))
736 		numa_debug = 1;
737 
738 	return 0;
739 }
740 early_param("numa", early_numa);
741 
742 #define numadbg(f, a...) \
743 do {	if (numa_debug) \
744 		printk(KERN_INFO f, ## a); \
745 } while (0)
746 
747 static void __init find_ramdisk(unsigned long phys_base)
748 {
749 #ifdef CONFIG_BLK_DEV_INITRD
750 	if (sparc_ramdisk_image || sparc_ramdisk_image64) {
751 		unsigned long ramdisk_image;
752 
753 		/* Older versions of the bootloader only supported a
754 		 * 32-bit physical address for the ramdisk image
755 		 * location, stored at sparc_ramdisk_image.  Newer
756 		 * SILO versions set sparc_ramdisk_image to zero and
757 		 * provide a full 64-bit physical address at
758 		 * sparc_ramdisk_image64.
759 		 */
760 		ramdisk_image = sparc_ramdisk_image;
761 		if (!ramdisk_image)
762 			ramdisk_image = sparc_ramdisk_image64;
763 
764 		/* Another bootloader quirk.  The bootloader normalizes
765 		 * the physical address to KERNBASE, so we have to
766 		 * factor that back out and add in the lowest valid
767 		 * physical page address to get the true physical address.
768 		 */
769 		ramdisk_image -= KERNBASE;
770 		ramdisk_image += phys_base;
771 
772 		numadbg("Found ramdisk at physical address 0x%lx, size %u\n",
773 			ramdisk_image, sparc_ramdisk_size);
774 
775 		initrd_start = ramdisk_image;
776 		initrd_end = ramdisk_image + sparc_ramdisk_size;
777 
778 		memblock_reserve(initrd_start, sparc_ramdisk_size);
779 
780 		initrd_start += PAGE_OFFSET;
781 		initrd_end += PAGE_OFFSET;
782 	}
783 #endif
784 }
785 
786 struct node_mem_mask {
787 	unsigned long mask;
788 	unsigned long val;
789 };
790 static struct node_mem_mask node_masks[MAX_NUMNODES];
791 static int num_node_masks;
792 
793 #ifdef CONFIG_NEED_MULTIPLE_NODES
794 
795 int numa_cpu_lookup_table[NR_CPUS];
796 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
797 
798 struct mdesc_mblock {
799 	u64	base;
800 	u64	size;
801 	u64	offset; /* RA-to-PA */
802 };
803 static struct mdesc_mblock *mblocks;
804 static int num_mblocks;
805 
806 static unsigned long ra_to_pa(unsigned long addr)
807 {
808 	int i;
809 
810 	for (i = 0; i < num_mblocks; i++) {
811 		struct mdesc_mblock *m = &mblocks[i];
812 
813 		if (addr >= m->base &&
814 		    addr < (m->base + m->size)) {
815 			addr += m->offset;
816 			break;
817 		}
818 	}
819 	return addr;
820 }
821 
822 static int find_node(unsigned long addr)
823 {
824 	int i;
825 
826 	addr = ra_to_pa(addr);
827 	for (i = 0; i < num_node_masks; i++) {
828 		struct node_mem_mask *p = &node_masks[i];
829 
830 		if ((addr & p->mask) == p->val)
831 			return i;
832 	}
833 	/* The following condition has been observed on LDOM guests.*/
834 	WARN_ONCE(1, "find_node: A physical address doesn't match a NUMA node"
835 		" rule. Some physical memory will be owned by node 0.");
836 	return 0;
837 }
838 
839 static u64 memblock_nid_range(u64 start, u64 end, int *nid)
840 {
841 	*nid = find_node(start);
842 	start += PAGE_SIZE;
843 	while (start < end) {
844 		int n = find_node(start);
845 
846 		if (n != *nid)
847 			break;
848 		start += PAGE_SIZE;
849 	}
850 
851 	if (start > end)
852 		start = end;
853 
854 	return start;
855 }
856 #endif
857 
858 /* This must be invoked after performing all of the necessary
859  * memblock_set_node() calls for 'nid'.  We need to be able to get
860  * correct data from get_pfn_range_for_nid().
861  */
862 static void __init allocate_node_data(int nid)
863 {
864 	struct pglist_data *p;
865 	unsigned long start_pfn, end_pfn;
866 #ifdef CONFIG_NEED_MULTIPLE_NODES
867 	unsigned long paddr;
868 
869 	paddr = memblock_alloc_try_nid(sizeof(struct pglist_data), SMP_CACHE_BYTES, nid);
870 	if (!paddr) {
871 		prom_printf("Cannot allocate pglist_data for nid[%d]\n", nid);
872 		prom_halt();
873 	}
874 	NODE_DATA(nid) = __va(paddr);
875 	memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
876 
877 	NODE_DATA(nid)->node_id = nid;
878 #endif
879 
880 	p = NODE_DATA(nid);
881 
882 	get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
883 	p->node_start_pfn = start_pfn;
884 	p->node_spanned_pages = end_pfn - start_pfn;
885 }
886 
887 static void init_node_masks_nonnuma(void)
888 {
889 #ifdef CONFIG_NEED_MULTIPLE_NODES
890 	int i;
891 #endif
892 
893 	numadbg("Initializing tables for non-numa.\n");
894 
895 	node_masks[0].mask = node_masks[0].val = 0;
896 	num_node_masks = 1;
897 
898 #ifdef CONFIG_NEED_MULTIPLE_NODES
899 	for (i = 0; i < NR_CPUS; i++)
900 		numa_cpu_lookup_table[i] = 0;
901 
902 	cpumask_setall(&numa_cpumask_lookup_table[0]);
903 #endif
904 }
905 
906 #ifdef CONFIG_NEED_MULTIPLE_NODES
907 struct pglist_data *node_data[MAX_NUMNODES];
908 
909 EXPORT_SYMBOL(numa_cpu_lookup_table);
910 EXPORT_SYMBOL(numa_cpumask_lookup_table);
911 EXPORT_SYMBOL(node_data);
912 
913 struct mdesc_mlgroup {
914 	u64	node;
915 	u64	latency;
916 	u64	match;
917 	u64	mask;
918 };
919 static struct mdesc_mlgroup *mlgroups;
920 static int num_mlgroups;
921 
922 static int scan_pio_for_cfg_handle(struct mdesc_handle *md, u64 pio,
923 				   u32 cfg_handle)
924 {
925 	u64 arc;
926 
927 	mdesc_for_each_arc(arc, md, pio, MDESC_ARC_TYPE_FWD) {
928 		u64 target = mdesc_arc_target(md, arc);
929 		const u64 *val;
930 
931 		val = mdesc_get_property(md, target,
932 					 "cfg-handle", NULL);
933 		if (val && *val == cfg_handle)
934 			return 0;
935 	}
936 	return -ENODEV;
937 }
938 
939 static int scan_arcs_for_cfg_handle(struct mdesc_handle *md, u64 grp,
940 				    u32 cfg_handle)
941 {
942 	u64 arc, candidate, best_latency = ~(u64)0;
943 
944 	candidate = MDESC_NODE_NULL;
945 	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
946 		u64 target = mdesc_arc_target(md, arc);
947 		const char *name = mdesc_node_name(md, target);
948 		const u64 *val;
949 
950 		if (strcmp(name, "pio-latency-group"))
951 			continue;
952 
953 		val = mdesc_get_property(md, target, "latency", NULL);
954 		if (!val)
955 			continue;
956 
957 		if (*val < best_latency) {
958 			candidate = target;
959 			best_latency = *val;
960 		}
961 	}
962 
963 	if (candidate == MDESC_NODE_NULL)
964 		return -ENODEV;
965 
966 	return scan_pio_for_cfg_handle(md, candidate, cfg_handle);
967 }
968 
969 int of_node_to_nid(struct device_node *dp)
970 {
971 	const struct linux_prom64_registers *regs;
972 	struct mdesc_handle *md;
973 	u32 cfg_handle;
974 	int count, nid;
975 	u64 grp;
976 
977 	/* This is the right thing to do on currently supported
978 	 * SUN4U NUMA platforms as well, as the PCI controller does
979 	 * not sit behind any particular memory controller.
980 	 */
981 	if (!mlgroups)
982 		return -1;
983 
984 	regs = of_get_property(dp, "reg", NULL);
985 	if (!regs)
986 		return -1;
987 
988 	cfg_handle = (regs->phys_addr >> 32UL) & 0x0fffffff;
989 
990 	md = mdesc_grab();
991 
992 	count = 0;
993 	nid = -1;
994 	mdesc_for_each_node_by_name(md, grp, "group") {
995 		if (!scan_arcs_for_cfg_handle(md, grp, cfg_handle)) {
996 			nid = count;
997 			break;
998 		}
999 		count++;
1000 	}
1001 
1002 	mdesc_release(md);
1003 
1004 	return nid;
1005 }
1006 
1007 static void __init add_node_ranges(void)
1008 {
1009 	struct memblock_region *reg;
1010 
1011 	for_each_memblock(memory, reg) {
1012 		unsigned long size = reg->size;
1013 		unsigned long start, end;
1014 
1015 		start = reg->base;
1016 		end = start + size;
1017 		while (start < end) {
1018 			unsigned long this_end;
1019 			int nid;
1020 
1021 			this_end = memblock_nid_range(start, end, &nid);
1022 
1023 			numadbg("Setting memblock NUMA node nid[%d] "
1024 				"start[%lx] end[%lx]\n",
1025 				nid, start, this_end);
1026 
1027 			memblock_set_node(start, this_end - start,
1028 					  &memblock.memory, nid);
1029 			start = this_end;
1030 		}
1031 	}
1032 }
1033 
1034 static int __init grab_mlgroups(struct mdesc_handle *md)
1035 {
1036 	unsigned long paddr;
1037 	int count = 0;
1038 	u64 node;
1039 
1040 	mdesc_for_each_node_by_name(md, node, "memory-latency-group")
1041 		count++;
1042 	if (!count)
1043 		return -ENOENT;
1044 
1045 	paddr = memblock_alloc(count * sizeof(struct mdesc_mlgroup),
1046 			  SMP_CACHE_BYTES);
1047 	if (!paddr)
1048 		return -ENOMEM;
1049 
1050 	mlgroups = __va(paddr);
1051 	num_mlgroups = count;
1052 
1053 	count = 0;
1054 	mdesc_for_each_node_by_name(md, node, "memory-latency-group") {
1055 		struct mdesc_mlgroup *m = &mlgroups[count++];
1056 		const u64 *val;
1057 
1058 		m->node = node;
1059 
1060 		val = mdesc_get_property(md, node, "latency", NULL);
1061 		m->latency = *val;
1062 		val = mdesc_get_property(md, node, "address-match", NULL);
1063 		m->match = *val;
1064 		val = mdesc_get_property(md, node, "address-mask", NULL);
1065 		m->mask = *val;
1066 
1067 		numadbg("MLGROUP[%d]: node[%llx] latency[%llx] "
1068 			"match[%llx] mask[%llx]\n",
1069 			count - 1, m->node, m->latency, m->match, m->mask);
1070 	}
1071 
1072 	return 0;
1073 }
1074 
1075 static int __init grab_mblocks(struct mdesc_handle *md)
1076 {
1077 	unsigned long paddr;
1078 	int count = 0;
1079 	u64 node;
1080 
1081 	mdesc_for_each_node_by_name(md, node, "mblock")
1082 		count++;
1083 	if (!count)
1084 		return -ENOENT;
1085 
1086 	paddr = memblock_alloc(count * sizeof(struct mdesc_mblock),
1087 			  SMP_CACHE_BYTES);
1088 	if (!paddr)
1089 		return -ENOMEM;
1090 
1091 	mblocks = __va(paddr);
1092 	num_mblocks = count;
1093 
1094 	count = 0;
1095 	mdesc_for_each_node_by_name(md, node, "mblock") {
1096 		struct mdesc_mblock *m = &mblocks[count++];
1097 		const u64 *val;
1098 
1099 		val = mdesc_get_property(md, node, "base", NULL);
1100 		m->base = *val;
1101 		val = mdesc_get_property(md, node, "size", NULL);
1102 		m->size = *val;
1103 		val = mdesc_get_property(md, node,
1104 					 "address-congruence-offset", NULL);
1105 
1106 		/* The address-congruence-offset property is optional.
1107 		 * Explicity zero it be identifty this.
1108 		 */
1109 		if (val)
1110 			m->offset = *val;
1111 		else
1112 			m->offset = 0UL;
1113 
1114 		numadbg("MBLOCK[%d]: base[%llx] size[%llx] offset[%llx]\n",
1115 			count - 1, m->base, m->size, m->offset);
1116 	}
1117 
1118 	return 0;
1119 }
1120 
1121 static void __init numa_parse_mdesc_group_cpus(struct mdesc_handle *md,
1122 					       u64 grp, cpumask_t *mask)
1123 {
1124 	u64 arc;
1125 
1126 	cpumask_clear(mask);
1127 
1128 	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_BACK) {
1129 		u64 target = mdesc_arc_target(md, arc);
1130 		const char *name = mdesc_node_name(md, target);
1131 		const u64 *id;
1132 
1133 		if (strcmp(name, "cpu"))
1134 			continue;
1135 		id = mdesc_get_property(md, target, "id", NULL);
1136 		if (*id < nr_cpu_ids)
1137 			cpumask_set_cpu(*id, mask);
1138 	}
1139 }
1140 
1141 static struct mdesc_mlgroup * __init find_mlgroup(u64 node)
1142 {
1143 	int i;
1144 
1145 	for (i = 0; i < num_mlgroups; i++) {
1146 		struct mdesc_mlgroup *m = &mlgroups[i];
1147 		if (m->node == node)
1148 			return m;
1149 	}
1150 	return NULL;
1151 }
1152 
1153 int __node_distance(int from, int to)
1154 {
1155 	if ((from >= MAX_NUMNODES) || (to >= MAX_NUMNODES)) {
1156 		pr_warn("Returning default NUMA distance value for %d->%d\n",
1157 			from, to);
1158 		return (from == to) ? LOCAL_DISTANCE : REMOTE_DISTANCE;
1159 	}
1160 	return numa_latency[from][to];
1161 }
1162 
1163 static int __init find_best_numa_node_for_mlgroup(struct mdesc_mlgroup *grp)
1164 {
1165 	int i;
1166 
1167 	for (i = 0; i < MAX_NUMNODES; i++) {
1168 		struct node_mem_mask *n = &node_masks[i];
1169 
1170 		if ((grp->mask == n->mask) && (grp->match == n->val))
1171 			break;
1172 	}
1173 	return i;
1174 }
1175 
1176 static void __init find_numa_latencies_for_group(struct mdesc_handle *md,
1177 						 u64 grp, int index)
1178 {
1179 	u64 arc;
1180 
1181 	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1182 		int tnode;
1183 		u64 target = mdesc_arc_target(md, arc);
1184 		struct mdesc_mlgroup *m = find_mlgroup(target);
1185 
1186 		if (!m)
1187 			continue;
1188 		tnode = find_best_numa_node_for_mlgroup(m);
1189 		if (tnode == MAX_NUMNODES)
1190 			continue;
1191 		numa_latency[index][tnode] = m->latency;
1192 	}
1193 }
1194 
1195 static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp,
1196 				      int index)
1197 {
1198 	struct mdesc_mlgroup *candidate = NULL;
1199 	u64 arc, best_latency = ~(u64)0;
1200 	struct node_mem_mask *n;
1201 
1202 	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1203 		u64 target = mdesc_arc_target(md, arc);
1204 		struct mdesc_mlgroup *m = find_mlgroup(target);
1205 		if (!m)
1206 			continue;
1207 		if (m->latency < best_latency) {
1208 			candidate = m;
1209 			best_latency = m->latency;
1210 		}
1211 	}
1212 	if (!candidate)
1213 		return -ENOENT;
1214 
1215 	if (num_node_masks != index) {
1216 		printk(KERN_ERR "Inconsistent NUMA state, "
1217 		       "index[%d] != num_node_masks[%d]\n",
1218 		       index, num_node_masks);
1219 		return -EINVAL;
1220 	}
1221 
1222 	n = &node_masks[num_node_masks++];
1223 
1224 	n->mask = candidate->mask;
1225 	n->val = candidate->match;
1226 
1227 	numadbg("NUMA NODE[%d]: mask[%lx] val[%lx] (latency[%llx])\n",
1228 		index, n->mask, n->val, candidate->latency);
1229 
1230 	return 0;
1231 }
1232 
1233 static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp,
1234 					 int index)
1235 {
1236 	cpumask_t mask;
1237 	int cpu;
1238 
1239 	numa_parse_mdesc_group_cpus(md, grp, &mask);
1240 
1241 	for_each_cpu(cpu, &mask)
1242 		numa_cpu_lookup_table[cpu] = index;
1243 	cpumask_copy(&numa_cpumask_lookup_table[index], &mask);
1244 
1245 	if (numa_debug) {
1246 		printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index);
1247 		for_each_cpu(cpu, &mask)
1248 			printk("%d ", cpu);
1249 		printk("]\n");
1250 	}
1251 
1252 	return numa_attach_mlgroup(md, grp, index);
1253 }
1254 
1255 static int __init numa_parse_mdesc(void)
1256 {
1257 	struct mdesc_handle *md = mdesc_grab();
1258 	int i, j, err, count;
1259 	u64 node;
1260 
1261 	node = mdesc_node_by_name(md, MDESC_NODE_NULL, "latency-groups");
1262 	if (node == MDESC_NODE_NULL) {
1263 		mdesc_release(md);
1264 		return -ENOENT;
1265 	}
1266 
1267 	err = grab_mblocks(md);
1268 	if (err < 0)
1269 		goto out;
1270 
1271 	err = grab_mlgroups(md);
1272 	if (err < 0)
1273 		goto out;
1274 
1275 	count = 0;
1276 	mdesc_for_each_node_by_name(md, node, "group") {
1277 		err = numa_parse_mdesc_group(md, node, count);
1278 		if (err < 0)
1279 			break;
1280 		count++;
1281 	}
1282 
1283 	count = 0;
1284 	mdesc_for_each_node_by_name(md, node, "group") {
1285 		find_numa_latencies_for_group(md, node, count);
1286 		count++;
1287 	}
1288 
1289 	/* Normalize numa latency matrix according to ACPI SLIT spec. */
1290 	for (i = 0; i < MAX_NUMNODES; i++) {
1291 		u64 self_latency = numa_latency[i][i];
1292 
1293 		for (j = 0; j < MAX_NUMNODES; j++) {
1294 			numa_latency[i][j] =
1295 				(numa_latency[i][j] * LOCAL_DISTANCE) /
1296 				self_latency;
1297 		}
1298 	}
1299 
1300 	add_node_ranges();
1301 
1302 	for (i = 0; i < num_node_masks; i++) {
1303 		allocate_node_data(i);
1304 		node_set_online(i);
1305 	}
1306 
1307 	err = 0;
1308 out:
1309 	mdesc_release(md);
1310 	return err;
1311 }
1312 
1313 static int __init numa_parse_jbus(void)
1314 {
1315 	unsigned long cpu, index;
1316 
1317 	/* NUMA node id is encoded in bits 36 and higher, and there is
1318 	 * a 1-to-1 mapping from CPU ID to NUMA node ID.
1319 	 */
1320 	index = 0;
1321 	for_each_present_cpu(cpu) {
1322 		numa_cpu_lookup_table[cpu] = index;
1323 		cpumask_copy(&numa_cpumask_lookup_table[index], cpumask_of(cpu));
1324 		node_masks[index].mask = ~((1UL << 36UL) - 1UL);
1325 		node_masks[index].val = cpu << 36UL;
1326 
1327 		index++;
1328 	}
1329 	num_node_masks = index;
1330 
1331 	add_node_ranges();
1332 
1333 	for (index = 0; index < num_node_masks; index++) {
1334 		allocate_node_data(index);
1335 		node_set_online(index);
1336 	}
1337 
1338 	return 0;
1339 }
1340 
1341 static int __init numa_parse_sun4u(void)
1342 {
1343 	if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1344 		unsigned long ver;
1345 
1346 		__asm__ ("rdpr %%ver, %0" : "=r" (ver));
1347 		if ((ver >> 32UL) == __JALAPENO_ID ||
1348 		    (ver >> 32UL) == __SERRANO_ID)
1349 			return numa_parse_jbus();
1350 	}
1351 	return -1;
1352 }
1353 
1354 static int __init bootmem_init_numa(void)
1355 {
1356 	int i, j;
1357 	int err = -1;
1358 
1359 	numadbg("bootmem_init_numa()\n");
1360 
1361 	/* Some sane defaults for numa latency values */
1362 	for (i = 0; i < MAX_NUMNODES; i++) {
1363 		for (j = 0; j < MAX_NUMNODES; j++)
1364 			numa_latency[i][j] = (i == j) ?
1365 				LOCAL_DISTANCE : REMOTE_DISTANCE;
1366 	}
1367 
1368 	if (numa_enabled) {
1369 		if (tlb_type == hypervisor)
1370 			err = numa_parse_mdesc();
1371 		else
1372 			err = numa_parse_sun4u();
1373 	}
1374 	return err;
1375 }
1376 
1377 #else
1378 
1379 static int bootmem_init_numa(void)
1380 {
1381 	return -1;
1382 }
1383 
1384 #endif
1385 
1386 static void __init bootmem_init_nonnuma(void)
1387 {
1388 	unsigned long top_of_ram = memblock_end_of_DRAM();
1389 	unsigned long total_ram = memblock_phys_mem_size();
1390 
1391 	numadbg("bootmem_init_nonnuma()\n");
1392 
1393 	printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
1394 	       top_of_ram, total_ram);
1395 	printk(KERN_INFO "Memory hole size: %ldMB\n",
1396 	       (top_of_ram - total_ram) >> 20);
1397 
1398 	init_node_masks_nonnuma();
1399 	memblock_set_node(0, (phys_addr_t)ULLONG_MAX, &memblock.memory, 0);
1400 	allocate_node_data(0);
1401 	node_set_online(0);
1402 }
1403 
1404 static unsigned long __init bootmem_init(unsigned long phys_base)
1405 {
1406 	unsigned long end_pfn;
1407 
1408 	end_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1409 	max_pfn = max_low_pfn = end_pfn;
1410 	min_low_pfn = (phys_base >> PAGE_SHIFT);
1411 
1412 	if (bootmem_init_numa() < 0)
1413 		bootmem_init_nonnuma();
1414 
1415 	/* Dump memblock with node info. */
1416 	memblock_dump_all();
1417 
1418 	/* XXX cpu notifier XXX */
1419 
1420 	sparse_memory_present_with_active_regions(MAX_NUMNODES);
1421 	sparse_init();
1422 
1423 	return end_pfn;
1424 }
1425 
1426 static struct linux_prom64_registers pall[MAX_BANKS] __initdata;
1427 static int pall_ents __initdata;
1428 
1429 static unsigned long max_phys_bits = 40;
1430 
1431 bool kern_addr_valid(unsigned long addr)
1432 {
1433 	pgd_t *pgd;
1434 	pud_t *pud;
1435 	pmd_t *pmd;
1436 	pte_t *pte;
1437 
1438 	if ((long)addr < 0L) {
1439 		unsigned long pa = __pa(addr);
1440 
1441 		if ((addr >> max_phys_bits) != 0UL)
1442 			return false;
1443 
1444 		return pfn_valid(pa >> PAGE_SHIFT);
1445 	}
1446 
1447 	if (addr >= (unsigned long) KERNBASE &&
1448 	    addr < (unsigned long)&_end)
1449 		return true;
1450 
1451 	pgd = pgd_offset_k(addr);
1452 	if (pgd_none(*pgd))
1453 		return 0;
1454 
1455 	pud = pud_offset(pgd, addr);
1456 	if (pud_none(*pud))
1457 		return 0;
1458 
1459 	if (pud_large(*pud))
1460 		return pfn_valid(pud_pfn(*pud));
1461 
1462 	pmd = pmd_offset(pud, addr);
1463 	if (pmd_none(*pmd))
1464 		return 0;
1465 
1466 	if (pmd_large(*pmd))
1467 		return pfn_valid(pmd_pfn(*pmd));
1468 
1469 	pte = pte_offset_kernel(pmd, addr);
1470 	if (pte_none(*pte))
1471 		return 0;
1472 
1473 	return pfn_valid(pte_pfn(*pte));
1474 }
1475 EXPORT_SYMBOL(kern_addr_valid);
1476 
1477 static unsigned long __ref kernel_map_hugepud(unsigned long vstart,
1478 					      unsigned long vend,
1479 					      pud_t *pud)
1480 {
1481 	const unsigned long mask16gb = (1UL << 34) - 1UL;
1482 	u64 pte_val = vstart;
1483 
1484 	/* Each PUD is 8GB */
1485 	if ((vstart & mask16gb) ||
1486 	    (vend - vstart <= mask16gb)) {
1487 		pte_val ^= kern_linear_pte_xor[2];
1488 		pud_val(*pud) = pte_val | _PAGE_PUD_HUGE;
1489 
1490 		return vstart + PUD_SIZE;
1491 	}
1492 
1493 	pte_val ^= kern_linear_pte_xor[3];
1494 	pte_val |= _PAGE_PUD_HUGE;
1495 
1496 	vend = vstart + mask16gb + 1UL;
1497 	while (vstart < vend) {
1498 		pud_val(*pud) = pte_val;
1499 
1500 		pte_val += PUD_SIZE;
1501 		vstart += PUD_SIZE;
1502 		pud++;
1503 	}
1504 	return vstart;
1505 }
1506 
1507 static bool kernel_can_map_hugepud(unsigned long vstart, unsigned long vend,
1508 				   bool guard)
1509 {
1510 	if (guard && !(vstart & ~PUD_MASK) && (vend - vstart) >= PUD_SIZE)
1511 		return true;
1512 
1513 	return false;
1514 }
1515 
1516 static unsigned long __ref kernel_map_hugepmd(unsigned long vstart,
1517 					      unsigned long vend,
1518 					      pmd_t *pmd)
1519 {
1520 	const unsigned long mask256mb = (1UL << 28) - 1UL;
1521 	const unsigned long mask2gb = (1UL << 31) - 1UL;
1522 	u64 pte_val = vstart;
1523 
1524 	/* Each PMD is 8MB */
1525 	if ((vstart & mask256mb) ||
1526 	    (vend - vstart <= mask256mb)) {
1527 		pte_val ^= kern_linear_pte_xor[0];
1528 		pmd_val(*pmd) = pte_val | _PAGE_PMD_HUGE;
1529 
1530 		return vstart + PMD_SIZE;
1531 	}
1532 
1533 	if ((vstart & mask2gb) ||
1534 	    (vend - vstart <= mask2gb)) {
1535 		pte_val ^= kern_linear_pte_xor[1];
1536 		pte_val |= _PAGE_PMD_HUGE;
1537 		vend = vstart + mask256mb + 1UL;
1538 	} else {
1539 		pte_val ^= kern_linear_pte_xor[2];
1540 		pte_val |= _PAGE_PMD_HUGE;
1541 		vend = vstart + mask2gb + 1UL;
1542 	}
1543 
1544 	while (vstart < vend) {
1545 		pmd_val(*pmd) = pte_val;
1546 
1547 		pte_val += PMD_SIZE;
1548 		vstart += PMD_SIZE;
1549 		pmd++;
1550 	}
1551 
1552 	return vstart;
1553 }
1554 
1555 static bool kernel_can_map_hugepmd(unsigned long vstart, unsigned long vend,
1556 				   bool guard)
1557 {
1558 	if (guard && !(vstart & ~PMD_MASK) && (vend - vstart) >= PMD_SIZE)
1559 		return true;
1560 
1561 	return false;
1562 }
1563 
1564 static unsigned long __ref kernel_map_range(unsigned long pstart,
1565 					    unsigned long pend, pgprot_t prot,
1566 					    bool use_huge)
1567 {
1568 	unsigned long vstart = PAGE_OFFSET + pstart;
1569 	unsigned long vend = PAGE_OFFSET + pend;
1570 	unsigned long alloc_bytes = 0UL;
1571 
1572 	if ((vstart & ~PAGE_MASK) || (vend & ~PAGE_MASK)) {
1573 		prom_printf("kernel_map: Unaligned physmem[%lx:%lx]\n",
1574 			    vstart, vend);
1575 		prom_halt();
1576 	}
1577 
1578 	while (vstart < vend) {
1579 		unsigned long this_end, paddr = __pa(vstart);
1580 		pgd_t *pgd = pgd_offset_k(vstart);
1581 		pud_t *pud;
1582 		pmd_t *pmd;
1583 		pte_t *pte;
1584 
1585 		if (pgd_none(*pgd)) {
1586 			pud_t *new;
1587 
1588 			new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1589 			alloc_bytes += PAGE_SIZE;
1590 			pgd_populate(&init_mm, pgd, new);
1591 		}
1592 		pud = pud_offset(pgd, vstart);
1593 		if (pud_none(*pud)) {
1594 			pmd_t *new;
1595 
1596 			if (kernel_can_map_hugepud(vstart, vend, use_huge)) {
1597 				vstart = kernel_map_hugepud(vstart, vend, pud);
1598 				continue;
1599 			}
1600 			new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1601 			alloc_bytes += PAGE_SIZE;
1602 			pud_populate(&init_mm, pud, new);
1603 		}
1604 
1605 		pmd = pmd_offset(pud, vstart);
1606 		if (pmd_none(*pmd)) {
1607 			pte_t *new;
1608 
1609 			if (kernel_can_map_hugepmd(vstart, vend, use_huge)) {
1610 				vstart = kernel_map_hugepmd(vstart, vend, pmd);
1611 				continue;
1612 			}
1613 			new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1614 			alloc_bytes += PAGE_SIZE;
1615 			pmd_populate_kernel(&init_mm, pmd, new);
1616 		}
1617 
1618 		pte = pte_offset_kernel(pmd, vstart);
1619 		this_end = (vstart + PMD_SIZE) & PMD_MASK;
1620 		if (this_end > vend)
1621 			this_end = vend;
1622 
1623 		while (vstart < this_end) {
1624 			pte_val(*pte) = (paddr | pgprot_val(prot));
1625 
1626 			vstart += PAGE_SIZE;
1627 			paddr += PAGE_SIZE;
1628 			pte++;
1629 		}
1630 	}
1631 
1632 	return alloc_bytes;
1633 }
1634 
1635 static void __init flush_all_kernel_tsbs(void)
1636 {
1637 	int i;
1638 
1639 	for (i = 0; i < KERNEL_TSB_NENTRIES; i++) {
1640 		struct tsb *ent = &swapper_tsb[i];
1641 
1642 		ent->tag = (1UL << TSB_TAG_INVALID_BIT);
1643 	}
1644 #ifndef CONFIG_DEBUG_PAGEALLOC
1645 	for (i = 0; i < KERNEL_TSB4M_NENTRIES; i++) {
1646 		struct tsb *ent = &swapper_4m_tsb[i];
1647 
1648 		ent->tag = (1UL << TSB_TAG_INVALID_BIT);
1649 	}
1650 #endif
1651 }
1652 
1653 extern unsigned int kvmap_linear_patch[1];
1654 
1655 static void __init kernel_physical_mapping_init(void)
1656 {
1657 	unsigned long i, mem_alloced = 0UL;
1658 	bool use_huge = true;
1659 
1660 #ifdef CONFIG_DEBUG_PAGEALLOC
1661 	use_huge = false;
1662 #endif
1663 	for (i = 0; i < pall_ents; i++) {
1664 		unsigned long phys_start, phys_end;
1665 
1666 		phys_start = pall[i].phys_addr;
1667 		phys_end = phys_start + pall[i].reg_size;
1668 
1669 		mem_alloced += kernel_map_range(phys_start, phys_end,
1670 						PAGE_KERNEL, use_huge);
1671 	}
1672 
1673 	printk("Allocated %ld bytes for kernel page tables.\n",
1674 	       mem_alloced);
1675 
1676 	kvmap_linear_patch[0] = 0x01000000; /* nop */
1677 	flushi(&kvmap_linear_patch[0]);
1678 
1679 	flush_all_kernel_tsbs();
1680 
1681 	__flush_tlb_all();
1682 }
1683 
1684 #ifdef CONFIG_DEBUG_PAGEALLOC
1685 void __kernel_map_pages(struct page *page, int numpages, int enable)
1686 {
1687 	unsigned long phys_start = page_to_pfn(page) << PAGE_SHIFT;
1688 	unsigned long phys_end = phys_start + (numpages * PAGE_SIZE);
1689 
1690 	kernel_map_range(phys_start, phys_end,
1691 			 (enable ? PAGE_KERNEL : __pgprot(0)), false);
1692 
1693 	flush_tsb_kernel_range(PAGE_OFFSET + phys_start,
1694 			       PAGE_OFFSET + phys_end);
1695 
1696 	/* we should perform an IPI and flush all tlbs,
1697 	 * but that can deadlock->flush only current cpu.
1698 	 */
1699 	__flush_tlb_kernel_range(PAGE_OFFSET + phys_start,
1700 				 PAGE_OFFSET + phys_end);
1701 }
1702 #endif
1703 
1704 unsigned long __init find_ecache_flush_span(unsigned long size)
1705 {
1706 	int i;
1707 
1708 	for (i = 0; i < pavail_ents; i++) {
1709 		if (pavail[i].reg_size >= size)
1710 			return pavail[i].phys_addr;
1711 	}
1712 
1713 	return ~0UL;
1714 }
1715 
1716 unsigned long PAGE_OFFSET;
1717 EXPORT_SYMBOL(PAGE_OFFSET);
1718 
1719 unsigned long VMALLOC_END   = 0x0000010000000000UL;
1720 EXPORT_SYMBOL(VMALLOC_END);
1721 
1722 unsigned long sparc64_va_hole_top =    0xfffff80000000000UL;
1723 unsigned long sparc64_va_hole_bottom = 0x0000080000000000UL;
1724 
1725 static void __init setup_page_offset(void)
1726 {
1727 	if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1728 		/* Cheetah/Panther support a full 64-bit virtual
1729 		 * address, so we can use all that our page tables
1730 		 * support.
1731 		 */
1732 		sparc64_va_hole_top =    0xfff0000000000000UL;
1733 		sparc64_va_hole_bottom = 0x0010000000000000UL;
1734 
1735 		max_phys_bits = 42;
1736 	} else if (tlb_type == hypervisor) {
1737 		switch (sun4v_chip_type) {
1738 		case SUN4V_CHIP_NIAGARA1:
1739 		case SUN4V_CHIP_NIAGARA2:
1740 			/* T1 and T2 support 48-bit virtual addresses.  */
1741 			sparc64_va_hole_top =    0xffff800000000000UL;
1742 			sparc64_va_hole_bottom = 0x0000800000000000UL;
1743 
1744 			max_phys_bits = 39;
1745 			break;
1746 		case SUN4V_CHIP_NIAGARA3:
1747 			/* T3 supports 48-bit virtual addresses.  */
1748 			sparc64_va_hole_top =    0xffff800000000000UL;
1749 			sparc64_va_hole_bottom = 0x0000800000000000UL;
1750 
1751 			max_phys_bits = 43;
1752 			break;
1753 		case SUN4V_CHIP_NIAGARA4:
1754 		case SUN4V_CHIP_NIAGARA5:
1755 		case SUN4V_CHIP_SPARC64X:
1756 		case SUN4V_CHIP_SPARC_M6:
1757 			/* T4 and later support 52-bit virtual addresses.  */
1758 			sparc64_va_hole_top =    0xfff8000000000000UL;
1759 			sparc64_va_hole_bottom = 0x0008000000000000UL;
1760 			max_phys_bits = 47;
1761 			break;
1762 		case SUN4V_CHIP_SPARC_M7:
1763 		case SUN4V_CHIP_SPARC_SN:
1764 		default:
1765 			/* M7 and later support 52-bit virtual addresses.  */
1766 			sparc64_va_hole_top =    0xfff8000000000000UL;
1767 			sparc64_va_hole_bottom = 0x0008000000000000UL;
1768 			max_phys_bits = 49;
1769 			break;
1770 		}
1771 	}
1772 
1773 	if (max_phys_bits > MAX_PHYS_ADDRESS_BITS) {
1774 		prom_printf("MAX_PHYS_ADDRESS_BITS is too small, need %lu\n",
1775 			    max_phys_bits);
1776 		prom_halt();
1777 	}
1778 
1779 	PAGE_OFFSET = sparc64_va_hole_top;
1780 	VMALLOC_END = ((sparc64_va_hole_bottom >> 1) +
1781 		       (sparc64_va_hole_bottom >> 2));
1782 
1783 	pr_info("MM: PAGE_OFFSET is 0x%016lx (max_phys_bits == %lu)\n",
1784 		PAGE_OFFSET, max_phys_bits);
1785 	pr_info("MM: VMALLOC [0x%016lx --> 0x%016lx]\n",
1786 		VMALLOC_START, VMALLOC_END);
1787 	pr_info("MM: VMEMMAP [0x%016lx --> 0x%016lx]\n",
1788 		VMEMMAP_BASE, VMEMMAP_BASE << 1);
1789 }
1790 
1791 static void __init tsb_phys_patch(void)
1792 {
1793 	struct tsb_ldquad_phys_patch_entry *pquad;
1794 	struct tsb_phys_patch_entry *p;
1795 
1796 	pquad = &__tsb_ldquad_phys_patch;
1797 	while (pquad < &__tsb_ldquad_phys_patch_end) {
1798 		unsigned long addr = pquad->addr;
1799 
1800 		if (tlb_type == hypervisor)
1801 			*(unsigned int *) addr = pquad->sun4v_insn;
1802 		else
1803 			*(unsigned int *) addr = pquad->sun4u_insn;
1804 		wmb();
1805 		__asm__ __volatile__("flush	%0"
1806 				     : /* no outputs */
1807 				     : "r" (addr));
1808 
1809 		pquad++;
1810 	}
1811 
1812 	p = &__tsb_phys_patch;
1813 	while (p < &__tsb_phys_patch_end) {
1814 		unsigned long addr = p->addr;
1815 
1816 		*(unsigned int *) addr = p->insn;
1817 		wmb();
1818 		__asm__ __volatile__("flush	%0"
1819 				     : /* no outputs */
1820 				     : "r" (addr));
1821 
1822 		p++;
1823 	}
1824 }
1825 
1826 /* Don't mark as init, we give this to the Hypervisor.  */
1827 #ifndef CONFIG_DEBUG_PAGEALLOC
1828 #define NUM_KTSB_DESCR	2
1829 #else
1830 #define NUM_KTSB_DESCR	1
1831 #endif
1832 static struct hv_tsb_descr ktsb_descr[NUM_KTSB_DESCR];
1833 
1834 /* The swapper TSBs are loaded with a base sequence of:
1835  *
1836  *	sethi	%uhi(SYMBOL), REG1
1837  *	sethi	%hi(SYMBOL), REG2
1838  *	or	REG1, %ulo(SYMBOL), REG1
1839  *	or	REG2, %lo(SYMBOL), REG2
1840  *	sllx	REG1, 32, REG1
1841  *	or	REG1, REG2, REG1
1842  *
1843  * When we use physical addressing for the TSB accesses, we patch the
1844  * first four instructions in the above sequence.
1845  */
1846 
1847 static void patch_one_ktsb_phys(unsigned int *start, unsigned int *end, unsigned long pa)
1848 {
1849 	unsigned long high_bits, low_bits;
1850 
1851 	high_bits = (pa >> 32) & 0xffffffff;
1852 	low_bits = (pa >> 0) & 0xffffffff;
1853 
1854 	while (start < end) {
1855 		unsigned int *ia = (unsigned int *)(unsigned long)*start;
1856 
1857 		ia[0] = (ia[0] & ~0x3fffff) | (high_bits >> 10);
1858 		__asm__ __volatile__("flush	%0" : : "r" (ia));
1859 
1860 		ia[1] = (ia[1] & ~0x3fffff) | (low_bits >> 10);
1861 		__asm__ __volatile__("flush	%0" : : "r" (ia + 1));
1862 
1863 		ia[2] = (ia[2] & ~0x1fff) | (high_bits & 0x3ff);
1864 		__asm__ __volatile__("flush	%0" : : "r" (ia + 2));
1865 
1866 		ia[3] = (ia[3] & ~0x1fff) | (low_bits & 0x3ff);
1867 		__asm__ __volatile__("flush	%0" : : "r" (ia + 3));
1868 
1869 		start++;
1870 	}
1871 }
1872 
1873 static void ktsb_phys_patch(void)
1874 {
1875 	extern unsigned int __swapper_tsb_phys_patch;
1876 	extern unsigned int __swapper_tsb_phys_patch_end;
1877 	unsigned long ktsb_pa;
1878 
1879 	ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
1880 	patch_one_ktsb_phys(&__swapper_tsb_phys_patch,
1881 			    &__swapper_tsb_phys_patch_end, ktsb_pa);
1882 #ifndef CONFIG_DEBUG_PAGEALLOC
1883 	{
1884 	extern unsigned int __swapper_4m_tsb_phys_patch;
1885 	extern unsigned int __swapper_4m_tsb_phys_patch_end;
1886 	ktsb_pa = (kern_base +
1887 		   ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
1888 	patch_one_ktsb_phys(&__swapper_4m_tsb_phys_patch,
1889 			    &__swapper_4m_tsb_phys_patch_end, ktsb_pa);
1890 	}
1891 #endif
1892 }
1893 
1894 static void __init sun4v_ktsb_init(void)
1895 {
1896 	unsigned long ktsb_pa;
1897 
1898 	/* First KTSB for PAGE_SIZE mappings.  */
1899 	ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
1900 
1901 	switch (PAGE_SIZE) {
1902 	case 8 * 1024:
1903 	default:
1904 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_8K;
1905 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_8K;
1906 		break;
1907 
1908 	case 64 * 1024:
1909 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_64K;
1910 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_64K;
1911 		break;
1912 
1913 	case 512 * 1024:
1914 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_512K;
1915 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_512K;
1916 		break;
1917 
1918 	case 4 * 1024 * 1024:
1919 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_4MB;
1920 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_4MB;
1921 		break;
1922 	}
1923 
1924 	ktsb_descr[0].assoc = 1;
1925 	ktsb_descr[0].num_ttes = KERNEL_TSB_NENTRIES;
1926 	ktsb_descr[0].ctx_idx = 0;
1927 	ktsb_descr[0].tsb_base = ktsb_pa;
1928 	ktsb_descr[0].resv = 0;
1929 
1930 #ifndef CONFIG_DEBUG_PAGEALLOC
1931 	/* Second KTSB for 4MB/256MB/2GB/16GB mappings.  */
1932 	ktsb_pa = (kern_base +
1933 		   ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
1934 
1935 	ktsb_descr[1].pgsz_idx = HV_PGSZ_IDX_4MB;
1936 	ktsb_descr[1].pgsz_mask = ((HV_PGSZ_MASK_4MB |
1937 				    HV_PGSZ_MASK_256MB |
1938 				    HV_PGSZ_MASK_2GB |
1939 				    HV_PGSZ_MASK_16GB) &
1940 				   cpu_pgsz_mask);
1941 	ktsb_descr[1].assoc = 1;
1942 	ktsb_descr[1].num_ttes = KERNEL_TSB4M_NENTRIES;
1943 	ktsb_descr[1].ctx_idx = 0;
1944 	ktsb_descr[1].tsb_base = ktsb_pa;
1945 	ktsb_descr[1].resv = 0;
1946 #endif
1947 }
1948 
1949 void sun4v_ktsb_register(void)
1950 {
1951 	unsigned long pa, ret;
1952 
1953 	pa = kern_base + ((unsigned long)&ktsb_descr[0] - KERNBASE);
1954 
1955 	ret = sun4v_mmu_tsb_ctx0(NUM_KTSB_DESCR, pa);
1956 	if (ret != 0) {
1957 		prom_printf("hypervisor_mmu_tsb_ctx0[%lx]: "
1958 			    "errors with %lx\n", pa, ret);
1959 		prom_halt();
1960 	}
1961 }
1962 
1963 static void __init sun4u_linear_pte_xor_finalize(void)
1964 {
1965 #ifndef CONFIG_DEBUG_PAGEALLOC
1966 	/* This is where we would add Panther support for
1967 	 * 32MB and 256MB pages.
1968 	 */
1969 #endif
1970 }
1971 
1972 static void __init sun4v_linear_pte_xor_finalize(void)
1973 {
1974 	unsigned long pagecv_flag;
1975 
1976 	/* Bit 9 of TTE is no longer CV bit on M7 processor and it instead
1977 	 * enables MCD error. Do not set bit 9 on M7 processor.
1978 	 */
1979 	switch (sun4v_chip_type) {
1980 	case SUN4V_CHIP_SPARC_M7:
1981 	case SUN4V_CHIP_SPARC_SN:
1982 		pagecv_flag = 0x00;
1983 		break;
1984 	default:
1985 		pagecv_flag = _PAGE_CV_4V;
1986 		break;
1987 	}
1988 #ifndef CONFIG_DEBUG_PAGEALLOC
1989 	if (cpu_pgsz_mask & HV_PGSZ_MASK_256MB) {
1990 		kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZ256MB_4V) ^
1991 			PAGE_OFFSET;
1992 		kern_linear_pte_xor[1] |= (_PAGE_CP_4V | pagecv_flag |
1993 					   _PAGE_P_4V | _PAGE_W_4V);
1994 	} else {
1995 		kern_linear_pte_xor[1] = kern_linear_pte_xor[0];
1996 	}
1997 
1998 	if (cpu_pgsz_mask & HV_PGSZ_MASK_2GB) {
1999 		kern_linear_pte_xor[2] = (_PAGE_VALID | _PAGE_SZ2GB_4V) ^
2000 			PAGE_OFFSET;
2001 		kern_linear_pte_xor[2] |= (_PAGE_CP_4V | pagecv_flag |
2002 					   _PAGE_P_4V | _PAGE_W_4V);
2003 	} else {
2004 		kern_linear_pte_xor[2] = kern_linear_pte_xor[1];
2005 	}
2006 
2007 	if (cpu_pgsz_mask & HV_PGSZ_MASK_16GB) {
2008 		kern_linear_pte_xor[3] = (_PAGE_VALID | _PAGE_SZ16GB_4V) ^
2009 			PAGE_OFFSET;
2010 		kern_linear_pte_xor[3] |= (_PAGE_CP_4V | pagecv_flag |
2011 					   _PAGE_P_4V | _PAGE_W_4V);
2012 	} else {
2013 		kern_linear_pte_xor[3] = kern_linear_pte_xor[2];
2014 	}
2015 #endif
2016 }
2017 
2018 /* paging_init() sets up the page tables */
2019 
2020 static unsigned long last_valid_pfn;
2021 
2022 static void sun4u_pgprot_init(void);
2023 static void sun4v_pgprot_init(void);
2024 
2025 static phys_addr_t __init available_memory(void)
2026 {
2027 	phys_addr_t available = 0ULL;
2028 	phys_addr_t pa_start, pa_end;
2029 	u64 i;
2030 
2031 	for_each_free_mem_range(i, NUMA_NO_NODE, MEMBLOCK_NONE, &pa_start,
2032 				&pa_end, NULL)
2033 		available = available + (pa_end  - pa_start);
2034 
2035 	return available;
2036 }
2037 
2038 #define _PAGE_CACHE_4U	(_PAGE_CP_4U | _PAGE_CV_4U)
2039 #define _PAGE_CACHE_4V	(_PAGE_CP_4V | _PAGE_CV_4V)
2040 #define __DIRTY_BITS_4U	 (_PAGE_MODIFIED_4U | _PAGE_WRITE_4U | _PAGE_W_4U)
2041 #define __DIRTY_BITS_4V	 (_PAGE_MODIFIED_4V | _PAGE_WRITE_4V | _PAGE_W_4V)
2042 #define __ACCESS_BITS_4U (_PAGE_ACCESSED_4U | _PAGE_READ_4U | _PAGE_R)
2043 #define __ACCESS_BITS_4V (_PAGE_ACCESSED_4V | _PAGE_READ_4V | _PAGE_R)
2044 
2045 /* We need to exclude reserved regions. This exclusion will include
2046  * vmlinux and initrd. To be more precise the initrd size could be used to
2047  * compute a new lower limit because it is freed later during initialization.
2048  */
2049 static void __init reduce_memory(phys_addr_t limit_ram)
2050 {
2051 	phys_addr_t avail_ram = available_memory();
2052 	phys_addr_t pa_start, pa_end;
2053 	u64 i;
2054 
2055 	if (limit_ram >= avail_ram)
2056 		return;
2057 
2058 	for_each_free_mem_range(i, NUMA_NO_NODE, MEMBLOCK_NONE, &pa_start,
2059 				&pa_end, NULL) {
2060 		phys_addr_t region_size = pa_end - pa_start;
2061 		phys_addr_t clip_start = pa_start;
2062 
2063 		avail_ram = avail_ram - region_size;
2064 		/* Are we consuming too much? */
2065 		if (avail_ram < limit_ram) {
2066 			phys_addr_t give_back = limit_ram - avail_ram;
2067 
2068 			region_size = region_size - give_back;
2069 			clip_start = clip_start + give_back;
2070 		}
2071 
2072 		memblock_remove(clip_start, region_size);
2073 
2074 		if (avail_ram <= limit_ram)
2075 			break;
2076 		i = 0UL;
2077 	}
2078 }
2079 
2080 void __init paging_init(void)
2081 {
2082 	unsigned long end_pfn, shift, phys_base;
2083 	unsigned long real_end, i;
2084 
2085 	setup_page_offset();
2086 
2087 	/* These build time checkes make sure that the dcache_dirty_cpu()
2088 	 * page->flags usage will work.
2089 	 *
2090 	 * When a page gets marked as dcache-dirty, we store the
2091 	 * cpu number starting at bit 32 in the page->flags.  Also,
2092 	 * functions like clear_dcache_dirty_cpu use the cpu mask
2093 	 * in 13-bit signed-immediate instruction fields.
2094 	 */
2095 
2096 	/*
2097 	 * Page flags must not reach into upper 32 bits that are used
2098 	 * for the cpu number
2099 	 */
2100 	BUILD_BUG_ON(NR_PAGEFLAGS > 32);
2101 
2102 	/*
2103 	 * The bit fields placed in the high range must not reach below
2104 	 * the 32 bit boundary. Otherwise we cannot place the cpu field
2105 	 * at the 32 bit boundary.
2106 	 */
2107 	BUILD_BUG_ON(SECTIONS_WIDTH + NODES_WIDTH + ZONES_WIDTH +
2108 		ilog2(roundup_pow_of_two(NR_CPUS)) > 32);
2109 
2110 	BUILD_BUG_ON(NR_CPUS > 4096);
2111 
2112 	kern_base = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
2113 	kern_size = (unsigned long)&_end - (unsigned long)KERNBASE;
2114 
2115 	/* Invalidate both kernel TSBs.  */
2116 	memset(swapper_tsb, 0x40, sizeof(swapper_tsb));
2117 #ifndef CONFIG_DEBUG_PAGEALLOC
2118 	memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2119 #endif
2120 
2121 	/* TTE.cv bit on sparc v9 occupies the same position as TTE.mcde
2122 	 * bit on M7 processor. This is a conflicting usage of the same
2123 	 * bit. Enabling TTE.cv on M7 would turn on Memory Corruption
2124 	 * Detection error on all pages and this will lead to problems
2125 	 * later. Kernel does not run with MCD enabled and hence rest
2126 	 * of the required steps to fully configure memory corruption
2127 	 * detection are not taken. We need to ensure TTE.mcde is not
2128 	 * set on M7 processor. Compute the value of cacheability
2129 	 * flag for use later taking this into consideration.
2130 	 */
2131 	switch (sun4v_chip_type) {
2132 	case SUN4V_CHIP_SPARC_M7:
2133 	case SUN4V_CHIP_SPARC_SN:
2134 		page_cache4v_flag = _PAGE_CP_4V;
2135 		break;
2136 	default:
2137 		page_cache4v_flag = _PAGE_CACHE_4V;
2138 		break;
2139 	}
2140 
2141 	if (tlb_type == hypervisor)
2142 		sun4v_pgprot_init();
2143 	else
2144 		sun4u_pgprot_init();
2145 
2146 	if (tlb_type == cheetah_plus ||
2147 	    tlb_type == hypervisor) {
2148 		tsb_phys_patch();
2149 		ktsb_phys_patch();
2150 	}
2151 
2152 	if (tlb_type == hypervisor)
2153 		sun4v_patch_tlb_handlers();
2154 
2155 	/* Find available physical memory...
2156 	 *
2157 	 * Read it twice in order to work around a bug in openfirmware.
2158 	 * The call to grab this table itself can cause openfirmware to
2159 	 * allocate memory, which in turn can take away some space from
2160 	 * the list of available memory.  Reading it twice makes sure
2161 	 * we really do get the final value.
2162 	 */
2163 	read_obp_translations();
2164 	read_obp_memory("reg", &pall[0], &pall_ents);
2165 	read_obp_memory("available", &pavail[0], &pavail_ents);
2166 	read_obp_memory("available", &pavail[0], &pavail_ents);
2167 
2168 	phys_base = 0xffffffffffffffffUL;
2169 	for (i = 0; i < pavail_ents; i++) {
2170 		phys_base = min(phys_base, pavail[i].phys_addr);
2171 		memblock_add(pavail[i].phys_addr, pavail[i].reg_size);
2172 	}
2173 
2174 	memblock_reserve(kern_base, kern_size);
2175 
2176 	find_ramdisk(phys_base);
2177 
2178 	if (cmdline_memory_size)
2179 		reduce_memory(cmdline_memory_size);
2180 
2181 	memblock_allow_resize();
2182 	memblock_dump_all();
2183 
2184 	set_bit(0, mmu_context_bmap);
2185 
2186 	shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE);
2187 
2188 	real_end = (unsigned long)_end;
2189 	num_kernel_image_mappings = DIV_ROUND_UP(real_end - KERNBASE, 1 << ILOG2_4MB);
2190 	printk("Kernel: Using %d locked TLB entries for main kernel image.\n",
2191 	       num_kernel_image_mappings);
2192 
2193 	/* Set kernel pgd to upper alias so physical page computations
2194 	 * work.
2195 	 */
2196 	init_mm.pgd += ((shift) / (sizeof(pgd_t)));
2197 
2198 	memset(swapper_pg_dir, 0, sizeof(swapper_pg_dir));
2199 
2200 	inherit_prom_mappings();
2201 
2202 	/* Ok, we can use our TLB miss and window trap handlers safely.  */
2203 	setup_tba();
2204 
2205 	__flush_tlb_all();
2206 
2207 	prom_build_devicetree();
2208 	of_populate_present_mask();
2209 #ifndef CONFIG_SMP
2210 	of_fill_in_cpu_data();
2211 #endif
2212 
2213 	if (tlb_type == hypervisor) {
2214 		sun4v_mdesc_init();
2215 		mdesc_populate_present_mask(cpu_all_mask);
2216 #ifndef CONFIG_SMP
2217 		mdesc_fill_in_cpu_data(cpu_all_mask);
2218 #endif
2219 		mdesc_get_page_sizes(cpu_all_mask, &cpu_pgsz_mask);
2220 
2221 		sun4v_linear_pte_xor_finalize();
2222 
2223 		sun4v_ktsb_init();
2224 		sun4v_ktsb_register();
2225 	} else {
2226 		unsigned long impl, ver;
2227 
2228 		cpu_pgsz_mask = (HV_PGSZ_MASK_8K | HV_PGSZ_MASK_64K |
2229 				 HV_PGSZ_MASK_512K | HV_PGSZ_MASK_4MB);
2230 
2231 		__asm__ __volatile__("rdpr %%ver, %0" : "=r" (ver));
2232 		impl = ((ver >> 32) & 0xffff);
2233 		if (impl == PANTHER_IMPL)
2234 			cpu_pgsz_mask |= (HV_PGSZ_MASK_32MB |
2235 					  HV_PGSZ_MASK_256MB);
2236 
2237 		sun4u_linear_pte_xor_finalize();
2238 	}
2239 
2240 	/* Flush the TLBs and the 4M TSB so that the updated linear
2241 	 * pte XOR settings are realized for all mappings.
2242 	 */
2243 	__flush_tlb_all();
2244 #ifndef CONFIG_DEBUG_PAGEALLOC
2245 	memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2246 #endif
2247 	__flush_tlb_all();
2248 
2249 	/* Setup bootmem... */
2250 	last_valid_pfn = end_pfn = bootmem_init(phys_base);
2251 
2252 	kernel_physical_mapping_init();
2253 
2254 	{
2255 		unsigned long max_zone_pfns[MAX_NR_ZONES];
2256 
2257 		memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
2258 
2259 		max_zone_pfns[ZONE_NORMAL] = end_pfn;
2260 
2261 		free_area_init_nodes(max_zone_pfns);
2262 	}
2263 
2264 	printk("Booting Linux...\n");
2265 }
2266 
2267 int page_in_phys_avail(unsigned long paddr)
2268 {
2269 	int i;
2270 
2271 	paddr &= PAGE_MASK;
2272 
2273 	for (i = 0; i < pavail_ents; i++) {
2274 		unsigned long start, end;
2275 
2276 		start = pavail[i].phys_addr;
2277 		end = start + pavail[i].reg_size;
2278 
2279 		if (paddr >= start && paddr < end)
2280 			return 1;
2281 	}
2282 	if (paddr >= kern_base && paddr < (kern_base + kern_size))
2283 		return 1;
2284 #ifdef CONFIG_BLK_DEV_INITRD
2285 	if (paddr >= __pa(initrd_start) &&
2286 	    paddr < __pa(PAGE_ALIGN(initrd_end)))
2287 		return 1;
2288 #endif
2289 
2290 	return 0;
2291 }
2292 
2293 static void __init register_page_bootmem_info(void)
2294 {
2295 #ifdef CONFIG_NEED_MULTIPLE_NODES
2296 	int i;
2297 
2298 	for_each_online_node(i)
2299 		if (NODE_DATA(i)->node_spanned_pages)
2300 			register_page_bootmem_info_node(NODE_DATA(i));
2301 #endif
2302 }
2303 void __init mem_init(void)
2304 {
2305 	high_memory = __va(last_valid_pfn << PAGE_SHIFT);
2306 
2307 	register_page_bootmem_info();
2308 	free_all_bootmem();
2309 
2310 	/*
2311 	 * Set up the zero page, mark it reserved, so that page count
2312 	 * is not manipulated when freeing the page from user ptes.
2313 	 */
2314 	mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0);
2315 	if (mem_map_zero == NULL) {
2316 		prom_printf("paging_init: Cannot alloc zero page.\n");
2317 		prom_halt();
2318 	}
2319 	mark_page_reserved(mem_map_zero);
2320 
2321 	mem_init_print_info(NULL);
2322 
2323 	if (tlb_type == cheetah || tlb_type == cheetah_plus)
2324 		cheetah_ecache_flush_init();
2325 }
2326 
2327 void free_initmem(void)
2328 {
2329 	unsigned long addr, initend;
2330 	int do_free = 1;
2331 
2332 	/* If the physical memory maps were trimmed by kernel command
2333 	 * line options, don't even try freeing this initmem stuff up.
2334 	 * The kernel image could have been in the trimmed out region
2335 	 * and if so the freeing below will free invalid page structs.
2336 	 */
2337 	if (cmdline_memory_size)
2338 		do_free = 0;
2339 
2340 	/*
2341 	 * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes.
2342 	 */
2343 	addr = PAGE_ALIGN((unsigned long)(__init_begin));
2344 	initend = (unsigned long)(__init_end) & PAGE_MASK;
2345 	for (; addr < initend; addr += PAGE_SIZE) {
2346 		unsigned long page;
2347 
2348 		page = (addr +
2349 			((unsigned long) __va(kern_base)) -
2350 			((unsigned long) KERNBASE));
2351 		memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
2352 
2353 		if (do_free)
2354 			free_reserved_page(virt_to_page(page));
2355 	}
2356 }
2357 
2358 #ifdef CONFIG_BLK_DEV_INITRD
2359 void free_initrd_mem(unsigned long start, unsigned long end)
2360 {
2361 	free_reserved_area((void *)start, (void *)end, POISON_FREE_INITMEM,
2362 			   "initrd");
2363 }
2364 #endif
2365 
2366 pgprot_t PAGE_KERNEL __read_mostly;
2367 EXPORT_SYMBOL(PAGE_KERNEL);
2368 
2369 pgprot_t PAGE_KERNEL_LOCKED __read_mostly;
2370 pgprot_t PAGE_COPY __read_mostly;
2371 
2372 pgprot_t PAGE_SHARED __read_mostly;
2373 EXPORT_SYMBOL(PAGE_SHARED);
2374 
2375 unsigned long pg_iobits __read_mostly;
2376 
2377 unsigned long _PAGE_IE __read_mostly;
2378 EXPORT_SYMBOL(_PAGE_IE);
2379 
2380 unsigned long _PAGE_E __read_mostly;
2381 EXPORT_SYMBOL(_PAGE_E);
2382 
2383 unsigned long _PAGE_CACHE __read_mostly;
2384 EXPORT_SYMBOL(_PAGE_CACHE);
2385 
2386 #ifdef CONFIG_SPARSEMEM_VMEMMAP
2387 int __meminit vmemmap_populate(unsigned long vstart, unsigned long vend,
2388 			       int node)
2389 {
2390 	unsigned long pte_base;
2391 
2392 	pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2393 		    _PAGE_CP_4U | _PAGE_CV_4U |
2394 		    _PAGE_P_4U | _PAGE_W_4U);
2395 	if (tlb_type == hypervisor)
2396 		pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2397 			    page_cache4v_flag | _PAGE_P_4V | _PAGE_W_4V);
2398 
2399 	pte_base |= _PAGE_PMD_HUGE;
2400 
2401 	vstart = vstart & PMD_MASK;
2402 	vend = ALIGN(vend, PMD_SIZE);
2403 	for (; vstart < vend; vstart += PMD_SIZE) {
2404 		pgd_t *pgd = pgd_offset_k(vstart);
2405 		unsigned long pte;
2406 		pud_t *pud;
2407 		pmd_t *pmd;
2408 
2409 		if (pgd_none(*pgd)) {
2410 			pud_t *new = vmemmap_alloc_block(PAGE_SIZE, node);
2411 
2412 			if (!new)
2413 				return -ENOMEM;
2414 			pgd_populate(&init_mm, pgd, new);
2415 		}
2416 
2417 		pud = pud_offset(pgd, vstart);
2418 		if (pud_none(*pud)) {
2419 			pmd_t *new = vmemmap_alloc_block(PAGE_SIZE, node);
2420 
2421 			if (!new)
2422 				return -ENOMEM;
2423 			pud_populate(&init_mm, pud, new);
2424 		}
2425 
2426 		pmd = pmd_offset(pud, vstart);
2427 
2428 		pte = pmd_val(*pmd);
2429 		if (!(pte & _PAGE_VALID)) {
2430 			void *block = vmemmap_alloc_block(PMD_SIZE, node);
2431 
2432 			if (!block)
2433 				return -ENOMEM;
2434 
2435 			pmd_val(*pmd) = pte_base | __pa(block);
2436 		}
2437 	}
2438 
2439 	return 0;
2440 }
2441 
2442 void vmemmap_free(unsigned long start, unsigned long end)
2443 {
2444 }
2445 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
2446 
2447 static void prot_init_common(unsigned long page_none,
2448 			     unsigned long page_shared,
2449 			     unsigned long page_copy,
2450 			     unsigned long page_readonly,
2451 			     unsigned long page_exec_bit)
2452 {
2453 	PAGE_COPY = __pgprot(page_copy);
2454 	PAGE_SHARED = __pgprot(page_shared);
2455 
2456 	protection_map[0x0] = __pgprot(page_none);
2457 	protection_map[0x1] = __pgprot(page_readonly & ~page_exec_bit);
2458 	protection_map[0x2] = __pgprot(page_copy & ~page_exec_bit);
2459 	protection_map[0x3] = __pgprot(page_copy & ~page_exec_bit);
2460 	protection_map[0x4] = __pgprot(page_readonly);
2461 	protection_map[0x5] = __pgprot(page_readonly);
2462 	protection_map[0x6] = __pgprot(page_copy);
2463 	protection_map[0x7] = __pgprot(page_copy);
2464 	protection_map[0x8] = __pgprot(page_none);
2465 	protection_map[0x9] = __pgprot(page_readonly & ~page_exec_bit);
2466 	protection_map[0xa] = __pgprot(page_shared & ~page_exec_bit);
2467 	protection_map[0xb] = __pgprot(page_shared & ~page_exec_bit);
2468 	protection_map[0xc] = __pgprot(page_readonly);
2469 	protection_map[0xd] = __pgprot(page_readonly);
2470 	protection_map[0xe] = __pgprot(page_shared);
2471 	protection_map[0xf] = __pgprot(page_shared);
2472 }
2473 
2474 static void __init sun4u_pgprot_init(void)
2475 {
2476 	unsigned long page_none, page_shared, page_copy, page_readonly;
2477 	unsigned long page_exec_bit;
2478 	int i;
2479 
2480 	PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2481 				_PAGE_CACHE_4U | _PAGE_P_4U |
2482 				__ACCESS_BITS_4U | __DIRTY_BITS_4U |
2483 				_PAGE_EXEC_4U);
2484 	PAGE_KERNEL_LOCKED = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2485 				       _PAGE_CACHE_4U | _PAGE_P_4U |
2486 				       __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2487 				       _PAGE_EXEC_4U | _PAGE_L_4U);
2488 
2489 	_PAGE_IE = _PAGE_IE_4U;
2490 	_PAGE_E = _PAGE_E_4U;
2491 	_PAGE_CACHE = _PAGE_CACHE_4U;
2492 
2493 	pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4U | __DIRTY_BITS_4U |
2494 		     __ACCESS_BITS_4U | _PAGE_E_4U);
2495 
2496 #ifdef CONFIG_DEBUG_PAGEALLOC
2497 	kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2498 #else
2499 	kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4U) ^
2500 		PAGE_OFFSET;
2501 #endif
2502 	kern_linear_pte_xor[0] |= (_PAGE_CP_4U | _PAGE_CV_4U |
2503 				   _PAGE_P_4U | _PAGE_W_4U);
2504 
2505 	for (i = 1; i < 4; i++)
2506 		kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2507 
2508 	_PAGE_ALL_SZ_BITS =  (_PAGE_SZ4MB_4U | _PAGE_SZ512K_4U |
2509 			      _PAGE_SZ64K_4U | _PAGE_SZ8K_4U |
2510 			      _PAGE_SZ32MB_4U | _PAGE_SZ256MB_4U);
2511 
2512 
2513 	page_none = _PAGE_PRESENT_4U | _PAGE_ACCESSED_4U | _PAGE_CACHE_4U;
2514 	page_shared = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2515 		       __ACCESS_BITS_4U | _PAGE_WRITE_4U | _PAGE_EXEC_4U);
2516 	page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2517 		       __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2518 	page_readonly   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2519 			   __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2520 
2521 	page_exec_bit = _PAGE_EXEC_4U;
2522 
2523 	prot_init_common(page_none, page_shared, page_copy, page_readonly,
2524 			 page_exec_bit);
2525 }
2526 
2527 static void __init sun4v_pgprot_init(void)
2528 {
2529 	unsigned long page_none, page_shared, page_copy, page_readonly;
2530 	unsigned long page_exec_bit;
2531 	int i;
2532 
2533 	PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4V | _PAGE_VALID |
2534 				page_cache4v_flag | _PAGE_P_4V |
2535 				__ACCESS_BITS_4V | __DIRTY_BITS_4V |
2536 				_PAGE_EXEC_4V);
2537 	PAGE_KERNEL_LOCKED = PAGE_KERNEL;
2538 
2539 	_PAGE_IE = _PAGE_IE_4V;
2540 	_PAGE_E = _PAGE_E_4V;
2541 	_PAGE_CACHE = page_cache4v_flag;
2542 
2543 #ifdef CONFIG_DEBUG_PAGEALLOC
2544 	kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2545 #else
2546 	kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4V) ^
2547 		PAGE_OFFSET;
2548 #endif
2549 	kern_linear_pte_xor[0] |= (page_cache4v_flag | _PAGE_P_4V |
2550 				   _PAGE_W_4V);
2551 
2552 	for (i = 1; i < 4; i++)
2553 		kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2554 
2555 	pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4V | __DIRTY_BITS_4V |
2556 		     __ACCESS_BITS_4V | _PAGE_E_4V);
2557 
2558 	_PAGE_ALL_SZ_BITS = (_PAGE_SZ16GB_4V | _PAGE_SZ2GB_4V |
2559 			     _PAGE_SZ256MB_4V | _PAGE_SZ32MB_4V |
2560 			     _PAGE_SZ4MB_4V | _PAGE_SZ512K_4V |
2561 			     _PAGE_SZ64K_4V | _PAGE_SZ8K_4V);
2562 
2563 	page_none = _PAGE_PRESENT_4V | _PAGE_ACCESSED_4V | page_cache4v_flag;
2564 	page_shared = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2565 		       __ACCESS_BITS_4V | _PAGE_WRITE_4V | _PAGE_EXEC_4V);
2566 	page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2567 		       __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2568 	page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2569 			 __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2570 
2571 	page_exec_bit = _PAGE_EXEC_4V;
2572 
2573 	prot_init_common(page_none, page_shared, page_copy, page_readonly,
2574 			 page_exec_bit);
2575 }
2576 
2577 unsigned long pte_sz_bits(unsigned long sz)
2578 {
2579 	if (tlb_type == hypervisor) {
2580 		switch (sz) {
2581 		case 8 * 1024:
2582 		default:
2583 			return _PAGE_SZ8K_4V;
2584 		case 64 * 1024:
2585 			return _PAGE_SZ64K_4V;
2586 		case 512 * 1024:
2587 			return _PAGE_SZ512K_4V;
2588 		case 4 * 1024 * 1024:
2589 			return _PAGE_SZ4MB_4V;
2590 		}
2591 	} else {
2592 		switch (sz) {
2593 		case 8 * 1024:
2594 		default:
2595 			return _PAGE_SZ8K_4U;
2596 		case 64 * 1024:
2597 			return _PAGE_SZ64K_4U;
2598 		case 512 * 1024:
2599 			return _PAGE_SZ512K_4U;
2600 		case 4 * 1024 * 1024:
2601 			return _PAGE_SZ4MB_4U;
2602 		}
2603 	}
2604 }
2605 
2606 pte_t mk_pte_io(unsigned long page, pgprot_t prot, int space, unsigned long page_size)
2607 {
2608 	pte_t pte;
2609 
2610 	pte_val(pte)  = page | pgprot_val(pgprot_noncached(prot));
2611 	pte_val(pte) |= (((unsigned long)space) << 32);
2612 	pte_val(pte) |= pte_sz_bits(page_size);
2613 
2614 	return pte;
2615 }
2616 
2617 static unsigned long kern_large_tte(unsigned long paddr)
2618 {
2619 	unsigned long val;
2620 
2621 	val = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2622 	       _PAGE_CP_4U | _PAGE_CV_4U | _PAGE_P_4U |
2623 	       _PAGE_EXEC_4U | _PAGE_L_4U | _PAGE_W_4U);
2624 	if (tlb_type == hypervisor)
2625 		val = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2626 		       page_cache4v_flag | _PAGE_P_4V |
2627 		       _PAGE_EXEC_4V | _PAGE_W_4V);
2628 
2629 	return val | paddr;
2630 }
2631 
2632 /* If not locked, zap it. */
2633 void __flush_tlb_all(void)
2634 {
2635 	unsigned long pstate;
2636 	int i;
2637 
2638 	__asm__ __volatile__("flushw\n\t"
2639 			     "rdpr	%%pstate, %0\n\t"
2640 			     "wrpr	%0, %1, %%pstate"
2641 			     : "=r" (pstate)
2642 			     : "i" (PSTATE_IE));
2643 	if (tlb_type == hypervisor) {
2644 		sun4v_mmu_demap_all();
2645 	} else if (tlb_type == spitfire) {
2646 		for (i = 0; i < 64; i++) {
2647 			/* Spitfire Errata #32 workaround */
2648 			/* NOTE: Always runs on spitfire, so no
2649 			 *       cheetah+ page size encodings.
2650 			 */
2651 			__asm__ __volatile__("stxa	%0, [%1] %2\n\t"
2652 					     "flush	%%g6"
2653 					     : /* No outputs */
2654 					     : "r" (0),
2655 					     "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2656 
2657 			if (!(spitfire_get_dtlb_data(i) & _PAGE_L_4U)) {
2658 				__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2659 						     "membar #Sync"
2660 						     : /* no outputs */
2661 						     : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
2662 				spitfire_put_dtlb_data(i, 0x0UL);
2663 			}
2664 
2665 			/* Spitfire Errata #32 workaround */
2666 			/* NOTE: Always runs on spitfire, so no
2667 			 *       cheetah+ page size encodings.
2668 			 */
2669 			__asm__ __volatile__("stxa	%0, [%1] %2\n\t"
2670 					     "flush	%%g6"
2671 					     : /* No outputs */
2672 					     : "r" (0),
2673 					     "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2674 
2675 			if (!(spitfire_get_itlb_data(i) & _PAGE_L_4U)) {
2676 				__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2677 						     "membar #Sync"
2678 						     : /* no outputs */
2679 						     : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
2680 				spitfire_put_itlb_data(i, 0x0UL);
2681 			}
2682 		}
2683 	} else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
2684 		cheetah_flush_dtlb_all();
2685 		cheetah_flush_itlb_all();
2686 	}
2687 	__asm__ __volatile__("wrpr	%0, 0, %%pstate"
2688 			     : : "r" (pstate));
2689 }
2690 
2691 pte_t *pte_alloc_one_kernel(struct mm_struct *mm,
2692 			    unsigned long address)
2693 {
2694 	struct page *page = alloc_page(GFP_KERNEL | __GFP_NOTRACK | __GFP_ZERO);
2695 	pte_t *pte = NULL;
2696 
2697 	if (page)
2698 		pte = (pte_t *) page_address(page);
2699 
2700 	return pte;
2701 }
2702 
2703 pgtable_t pte_alloc_one(struct mm_struct *mm,
2704 			unsigned long address)
2705 {
2706 	struct page *page = alloc_page(GFP_KERNEL | __GFP_NOTRACK | __GFP_ZERO);
2707 	if (!page)
2708 		return NULL;
2709 	if (!pgtable_page_ctor(page)) {
2710 		free_hot_cold_page(page, 0);
2711 		return NULL;
2712 	}
2713 	return (pte_t *) page_address(page);
2714 }
2715 
2716 void pte_free_kernel(struct mm_struct *mm, pte_t *pte)
2717 {
2718 	free_page((unsigned long)pte);
2719 }
2720 
2721 static void __pte_free(pgtable_t pte)
2722 {
2723 	struct page *page = virt_to_page(pte);
2724 
2725 	pgtable_page_dtor(page);
2726 	__free_page(page);
2727 }
2728 
2729 void pte_free(struct mm_struct *mm, pgtable_t pte)
2730 {
2731 	__pte_free(pte);
2732 }
2733 
2734 void pgtable_free(void *table, bool is_page)
2735 {
2736 	if (is_page)
2737 		__pte_free(table);
2738 	else
2739 		kmem_cache_free(pgtable_cache, table);
2740 }
2741 
2742 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2743 void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
2744 			  pmd_t *pmd)
2745 {
2746 	unsigned long pte, flags;
2747 	struct mm_struct *mm;
2748 	pmd_t entry = *pmd;
2749 
2750 	if (!pmd_large(entry) || !pmd_young(entry))
2751 		return;
2752 
2753 	pte = pmd_val(entry);
2754 
2755 	/* Don't insert a non-valid PMD into the TSB, we'll deadlock.  */
2756 	if (!(pte & _PAGE_VALID))
2757 		return;
2758 
2759 	/* We are fabricating 8MB pages using 4MB real hw pages.  */
2760 	pte |= (addr & (1UL << REAL_HPAGE_SHIFT));
2761 
2762 	mm = vma->vm_mm;
2763 
2764 	spin_lock_irqsave(&mm->context.lock, flags);
2765 
2766 	if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL)
2767 		__update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
2768 					addr, pte);
2769 
2770 	spin_unlock_irqrestore(&mm->context.lock, flags);
2771 }
2772 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2773 
2774 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
2775 static void context_reload(void *__data)
2776 {
2777 	struct mm_struct *mm = __data;
2778 
2779 	if (mm == current->mm)
2780 		load_secondary_context(mm);
2781 }
2782 
2783 void hugetlb_setup(struct pt_regs *regs)
2784 {
2785 	struct mm_struct *mm = current->mm;
2786 	struct tsb_config *tp;
2787 
2788 	if (faulthandler_disabled() || !mm) {
2789 		const struct exception_table_entry *entry;
2790 
2791 		entry = search_exception_tables(regs->tpc);
2792 		if (entry) {
2793 			regs->tpc = entry->fixup;
2794 			regs->tnpc = regs->tpc + 4;
2795 			return;
2796 		}
2797 		pr_alert("Unexpected HugeTLB setup in atomic context.\n");
2798 		die_if_kernel("HugeTSB in atomic", regs);
2799 	}
2800 
2801 	tp = &mm->context.tsb_block[MM_TSB_HUGE];
2802 	if (likely(tp->tsb == NULL))
2803 		tsb_grow(mm, MM_TSB_HUGE, 0);
2804 
2805 	tsb_context_switch(mm);
2806 	smp_tsb_sync(mm);
2807 
2808 	/* On UltraSPARC-III+ and later, configure the second half of
2809 	 * the Data-TLB for huge pages.
2810 	 */
2811 	if (tlb_type == cheetah_plus) {
2812 		bool need_context_reload = false;
2813 		unsigned long ctx;
2814 
2815 		spin_lock_irq(&ctx_alloc_lock);
2816 		ctx = mm->context.sparc64_ctx_val;
2817 		ctx &= ~CTX_PGSZ_MASK;
2818 		ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT;
2819 		ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT;
2820 
2821 		if (ctx != mm->context.sparc64_ctx_val) {
2822 			/* When changing the page size fields, we
2823 			 * must perform a context flush so that no
2824 			 * stale entries match.  This flush must
2825 			 * occur with the original context register
2826 			 * settings.
2827 			 */
2828 			do_flush_tlb_mm(mm);
2829 
2830 			/* Reload the context register of all processors
2831 			 * also executing in this address space.
2832 			 */
2833 			mm->context.sparc64_ctx_val = ctx;
2834 			need_context_reload = true;
2835 		}
2836 		spin_unlock_irq(&ctx_alloc_lock);
2837 
2838 		if (need_context_reload)
2839 			on_each_cpu(context_reload, mm, 0);
2840 	}
2841 }
2842 #endif
2843 
2844 static struct resource code_resource = {
2845 	.name	= "Kernel code",
2846 	.flags	= IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
2847 };
2848 
2849 static struct resource data_resource = {
2850 	.name	= "Kernel data",
2851 	.flags	= IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
2852 };
2853 
2854 static struct resource bss_resource = {
2855 	.name	= "Kernel bss",
2856 	.flags	= IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
2857 };
2858 
2859 static inline resource_size_t compute_kern_paddr(void *addr)
2860 {
2861 	return (resource_size_t) (addr - KERNBASE + kern_base);
2862 }
2863 
2864 static void __init kernel_lds_init(void)
2865 {
2866 	code_resource.start = compute_kern_paddr(_text);
2867 	code_resource.end   = compute_kern_paddr(_etext - 1);
2868 	data_resource.start = compute_kern_paddr(_etext);
2869 	data_resource.end   = compute_kern_paddr(_edata - 1);
2870 	bss_resource.start  = compute_kern_paddr(__bss_start);
2871 	bss_resource.end    = compute_kern_paddr(_end - 1);
2872 }
2873 
2874 static int __init report_memory(void)
2875 {
2876 	int i;
2877 	struct resource *res;
2878 
2879 	kernel_lds_init();
2880 
2881 	for (i = 0; i < pavail_ents; i++) {
2882 		res = kzalloc(sizeof(struct resource), GFP_KERNEL);
2883 
2884 		if (!res) {
2885 			pr_warn("Failed to allocate source.\n");
2886 			break;
2887 		}
2888 
2889 		res->name = "System RAM";
2890 		res->start = pavail[i].phys_addr;
2891 		res->end = pavail[i].phys_addr + pavail[i].reg_size - 1;
2892 		res->flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM;
2893 
2894 		if (insert_resource(&iomem_resource, res) < 0) {
2895 			pr_warn("Resource insertion failed.\n");
2896 			break;
2897 		}
2898 
2899 		insert_resource(res, &code_resource);
2900 		insert_resource(res, &data_resource);
2901 		insert_resource(res, &bss_resource);
2902 	}
2903 
2904 	return 0;
2905 }
2906 arch_initcall(report_memory);
2907 
2908 #ifdef CONFIG_SMP
2909 #define do_flush_tlb_kernel_range	smp_flush_tlb_kernel_range
2910 #else
2911 #define do_flush_tlb_kernel_range	__flush_tlb_kernel_range
2912 #endif
2913 
2914 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
2915 {
2916 	if (start < HI_OBP_ADDRESS && end > LOW_OBP_ADDRESS) {
2917 		if (start < LOW_OBP_ADDRESS) {
2918 			flush_tsb_kernel_range(start, LOW_OBP_ADDRESS);
2919 			do_flush_tlb_kernel_range(start, LOW_OBP_ADDRESS);
2920 		}
2921 		if (end > HI_OBP_ADDRESS) {
2922 			flush_tsb_kernel_range(HI_OBP_ADDRESS, end);
2923 			do_flush_tlb_kernel_range(HI_OBP_ADDRESS, end);
2924 		}
2925 	} else {
2926 		flush_tsb_kernel_range(start, end);
2927 		do_flush_tlb_kernel_range(start, end);
2928 	}
2929 }
2930