xref: /openbmc/linux/arch/sparc/mm/init_64.c (revision a8a28aff)
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/module.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/percpu.h>
26 #include <linux/memblock.h>
27 #include <linux/mmzone.h>
28 #include <linux/gfp.h>
29 
30 #include <asm/head.h>
31 #include <asm/page.h>
32 #include <asm/pgalloc.h>
33 #include <asm/pgtable.h>
34 #include <asm/oplib.h>
35 #include <asm/iommu.h>
36 #include <asm/io.h>
37 #include <asm/uaccess.h>
38 #include <asm/mmu_context.h>
39 #include <asm/tlbflush.h>
40 #include <asm/dma.h>
41 #include <asm/starfire.h>
42 #include <asm/tlb.h>
43 #include <asm/spitfire.h>
44 #include <asm/sections.h>
45 #include <asm/tsb.h>
46 #include <asm/hypervisor.h>
47 #include <asm/prom.h>
48 #include <asm/mdesc.h>
49 #include <asm/cpudata.h>
50 #include <asm/setup.h>
51 #include <asm/irq.h>
52 
53 #include "init_64.h"
54 
55 unsigned long kern_linear_pte_xor[4] __read_mostly;
56 
57 /* A bitmap, two bits for every 256MB of physical memory.  These two
58  * bits determine what page size we use for kernel linear
59  * translations.  They form an index into kern_linear_pte_xor[].  The
60  * value in the indexed slot is XOR'd with the TLB miss virtual
61  * address to form the resulting TTE.  The mapping is:
62  *
63  *	0	==>	4MB
64  *	1	==>	256MB
65  *	2	==>	2GB
66  *	3	==>	16GB
67  *
68  * All sun4v chips support 256MB pages.  Only SPARC-T4 and later
69  * support 2GB pages, and hopefully future cpus will support the 16GB
70  * pages as well.  For slots 2 and 3, we encode a 256MB TTE xor there
71  * if these larger page sizes are not supported by the cpu.
72  *
73  * It would be nice to determine this from the machine description
74  * 'cpu' properties, but we need to have this table setup before the
75  * MDESC is initialized.
76  */
77 unsigned long kpte_linear_bitmap[KPTE_BITMAP_BYTES / sizeof(unsigned long)];
78 
79 #ifndef CONFIG_DEBUG_PAGEALLOC
80 /* A special kernel TSB for 4MB, 256MB, 2GB and 16GB linear mappings.
81  * Space is allocated for this right after the trap table in
82  * arch/sparc64/kernel/head.S
83  */
84 extern struct tsb swapper_4m_tsb[KERNEL_TSB4M_NENTRIES];
85 #endif
86 
87 static unsigned long cpu_pgsz_mask;
88 
89 #define MAX_BANKS	32
90 
91 static struct linux_prom64_registers pavail[MAX_BANKS];
92 static int pavail_ents;
93 
94 static int cmp_p64(const void *a, const void *b)
95 {
96 	const struct linux_prom64_registers *x = a, *y = b;
97 
98 	if (x->phys_addr > y->phys_addr)
99 		return 1;
100 	if (x->phys_addr < y->phys_addr)
101 		return -1;
102 	return 0;
103 }
104 
105 static void __init read_obp_memory(const char *property,
106 				   struct linux_prom64_registers *regs,
107 				   int *num_ents)
108 {
109 	phandle node = prom_finddevice("/memory");
110 	int prop_size = prom_getproplen(node, property);
111 	int ents, ret, i;
112 
113 	ents = prop_size / sizeof(struct linux_prom64_registers);
114 	if (ents > MAX_BANKS) {
115 		prom_printf("The machine has more %s property entries than "
116 			    "this kernel can support (%d).\n",
117 			    property, MAX_BANKS);
118 		prom_halt();
119 	}
120 
121 	ret = prom_getproperty(node, property, (char *) regs, prop_size);
122 	if (ret == -1) {
123 		prom_printf("Couldn't get %s property from /memory.\n",
124 				property);
125 		prom_halt();
126 	}
127 
128 	/* Sanitize what we got from the firmware, by page aligning
129 	 * everything.
130 	 */
131 	for (i = 0; i < ents; i++) {
132 		unsigned long base, size;
133 
134 		base = regs[i].phys_addr;
135 		size = regs[i].reg_size;
136 
137 		size &= PAGE_MASK;
138 		if (base & ~PAGE_MASK) {
139 			unsigned long new_base = PAGE_ALIGN(base);
140 
141 			size -= new_base - base;
142 			if ((long) size < 0L)
143 				size = 0UL;
144 			base = new_base;
145 		}
146 		if (size == 0UL) {
147 			/* If it is empty, simply get rid of it.
148 			 * This simplifies the logic of the other
149 			 * functions that process these arrays.
150 			 */
151 			memmove(&regs[i], &regs[i + 1],
152 				(ents - i - 1) * sizeof(regs[0]));
153 			i--;
154 			ents--;
155 			continue;
156 		}
157 		regs[i].phys_addr = base;
158 		regs[i].reg_size = size;
159 	}
160 
161 	*num_ents = ents;
162 
163 	sort(regs, ents, sizeof(struct linux_prom64_registers),
164 	     cmp_p64, NULL);
165 }
166 
167 unsigned long sparc64_valid_addr_bitmap[VALID_ADDR_BITMAP_BYTES /
168 					sizeof(unsigned long)];
169 EXPORT_SYMBOL(sparc64_valid_addr_bitmap);
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 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
328 static inline bool is_hugetlb_pte(pte_t pte)
329 {
330 	if ((tlb_type == hypervisor &&
331 	     (pte_val(pte) & _PAGE_SZALL_4V) == _PAGE_SZHUGE_4V) ||
332 	    (tlb_type != hypervisor &&
333 	     (pte_val(pte) & _PAGE_SZALL_4U) == _PAGE_SZHUGE_4U))
334 		return true;
335 	return false;
336 }
337 #endif
338 
339 void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
340 {
341 	struct mm_struct *mm;
342 	unsigned long flags;
343 	pte_t pte = *ptep;
344 
345 	if (tlb_type != hypervisor) {
346 		unsigned long pfn = pte_pfn(pte);
347 
348 		if (pfn_valid(pfn))
349 			flush_dcache(pfn);
350 	}
351 
352 	mm = vma->vm_mm;
353 
354 	spin_lock_irqsave(&mm->context.lock, flags);
355 
356 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
357 	if (mm->context.huge_pte_count && is_hugetlb_pte(pte))
358 		__update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
359 					address, pte_val(pte));
360 	else
361 #endif
362 		__update_mmu_tsb_insert(mm, MM_TSB_BASE, PAGE_SHIFT,
363 					address, pte_val(pte));
364 
365 	spin_unlock_irqrestore(&mm->context.lock, flags);
366 }
367 
368 void flush_dcache_page(struct page *page)
369 {
370 	struct address_space *mapping;
371 	int this_cpu;
372 
373 	if (tlb_type == hypervisor)
374 		return;
375 
376 	/* Do not bother with the expensive D-cache flush if it
377 	 * is merely the zero page.  The 'bigcore' testcase in GDB
378 	 * causes this case to run millions of times.
379 	 */
380 	if (page == ZERO_PAGE(0))
381 		return;
382 
383 	this_cpu = get_cpu();
384 
385 	mapping = page_mapping(page);
386 	if (mapping && !mapping_mapped(mapping)) {
387 		int dirty = test_bit(PG_dcache_dirty, &page->flags);
388 		if (dirty) {
389 			int dirty_cpu = dcache_dirty_cpu(page);
390 
391 			if (dirty_cpu == this_cpu)
392 				goto out;
393 			smp_flush_dcache_page_impl(page, dirty_cpu);
394 		}
395 		set_dcache_dirty(page, this_cpu);
396 	} else {
397 		/* We could delay the flush for the !page_mapping
398 		 * case too.  But that case is for exec env/arg
399 		 * pages and those are %99 certainly going to get
400 		 * faulted into the tlb (and thus flushed) anyways.
401 		 */
402 		flush_dcache_page_impl(page);
403 	}
404 
405 out:
406 	put_cpu();
407 }
408 EXPORT_SYMBOL(flush_dcache_page);
409 
410 void __kprobes flush_icache_range(unsigned long start, unsigned long end)
411 {
412 	/* Cheetah and Hypervisor platform cpus have coherent I-cache. */
413 	if (tlb_type == spitfire) {
414 		unsigned long kaddr;
415 
416 		/* This code only runs on Spitfire cpus so this is
417 		 * why we can assume _PAGE_PADDR_4U.
418 		 */
419 		for (kaddr = start; kaddr < end; kaddr += PAGE_SIZE) {
420 			unsigned long paddr, mask = _PAGE_PADDR_4U;
421 
422 			if (kaddr >= PAGE_OFFSET)
423 				paddr = kaddr & mask;
424 			else {
425 				pgd_t *pgdp = pgd_offset_k(kaddr);
426 				pud_t *pudp = pud_offset(pgdp, kaddr);
427 				pmd_t *pmdp = pmd_offset(pudp, kaddr);
428 				pte_t *ptep = pte_offset_kernel(pmdp, kaddr);
429 
430 				paddr = pte_val(*ptep) & mask;
431 			}
432 			__flush_icache_page(paddr);
433 		}
434 	}
435 }
436 EXPORT_SYMBOL(flush_icache_range);
437 
438 void mmu_info(struct seq_file *m)
439 {
440 	static const char *pgsz_strings[] = {
441 		"8K", "64K", "512K", "4MB", "32MB",
442 		"256MB", "2GB", "16GB",
443 	};
444 	int i, printed;
445 
446 	if (tlb_type == cheetah)
447 		seq_printf(m, "MMU Type\t: Cheetah\n");
448 	else if (tlb_type == cheetah_plus)
449 		seq_printf(m, "MMU Type\t: Cheetah+\n");
450 	else if (tlb_type == spitfire)
451 		seq_printf(m, "MMU Type\t: Spitfire\n");
452 	else if (tlb_type == hypervisor)
453 		seq_printf(m, "MMU Type\t: Hypervisor (sun4v)\n");
454 	else
455 		seq_printf(m, "MMU Type\t: ???\n");
456 
457 	seq_printf(m, "MMU PGSZs\t: ");
458 	printed = 0;
459 	for (i = 0; i < ARRAY_SIZE(pgsz_strings); i++) {
460 		if (cpu_pgsz_mask & (1UL << i)) {
461 			seq_printf(m, "%s%s",
462 				   printed ? "," : "", pgsz_strings[i]);
463 			printed++;
464 		}
465 	}
466 	seq_putc(m, '\n');
467 
468 #ifdef CONFIG_DEBUG_DCFLUSH
469 	seq_printf(m, "DCPageFlushes\t: %d\n",
470 		   atomic_read(&dcpage_flushes));
471 #ifdef CONFIG_SMP
472 	seq_printf(m, "DCPageFlushesXC\t: %d\n",
473 		   atomic_read(&dcpage_flushes_xcall));
474 #endif /* CONFIG_SMP */
475 #endif /* CONFIG_DEBUG_DCFLUSH */
476 }
477 
478 struct linux_prom_translation prom_trans[512] __read_mostly;
479 unsigned int prom_trans_ents __read_mostly;
480 
481 unsigned long kern_locked_tte_data;
482 
483 /* The obp translations are saved based on 8k pagesize, since obp can
484  * use a mixture of pagesizes. Misses to the LOW_OBP_ADDRESS ->
485  * HI_OBP_ADDRESS range are handled in ktlb.S.
486  */
487 static inline int in_obp_range(unsigned long vaddr)
488 {
489 	return (vaddr >= LOW_OBP_ADDRESS &&
490 		vaddr < HI_OBP_ADDRESS);
491 }
492 
493 static int cmp_ptrans(const void *a, const void *b)
494 {
495 	const struct linux_prom_translation *x = a, *y = b;
496 
497 	if (x->virt > y->virt)
498 		return 1;
499 	if (x->virt < y->virt)
500 		return -1;
501 	return 0;
502 }
503 
504 /* Read OBP translations property into 'prom_trans[]'.  */
505 static void __init read_obp_translations(void)
506 {
507 	int n, node, ents, first, last, i;
508 
509 	node = prom_finddevice("/virtual-memory");
510 	n = prom_getproplen(node, "translations");
511 	if (unlikely(n == 0 || n == -1)) {
512 		prom_printf("prom_mappings: Couldn't get size.\n");
513 		prom_halt();
514 	}
515 	if (unlikely(n > sizeof(prom_trans))) {
516 		prom_printf("prom_mappings: Size %d is too big.\n", n);
517 		prom_halt();
518 	}
519 
520 	if ((n = prom_getproperty(node, "translations",
521 				  (char *)&prom_trans[0],
522 				  sizeof(prom_trans))) == -1) {
523 		prom_printf("prom_mappings: Couldn't get property.\n");
524 		prom_halt();
525 	}
526 
527 	n = n / sizeof(struct linux_prom_translation);
528 
529 	ents = n;
530 
531 	sort(prom_trans, ents, sizeof(struct linux_prom_translation),
532 	     cmp_ptrans, NULL);
533 
534 	/* Now kick out all the non-OBP entries.  */
535 	for (i = 0; i < ents; i++) {
536 		if (in_obp_range(prom_trans[i].virt))
537 			break;
538 	}
539 	first = i;
540 	for (; i < ents; i++) {
541 		if (!in_obp_range(prom_trans[i].virt))
542 			break;
543 	}
544 	last = i;
545 
546 	for (i = 0; i < (last - first); i++) {
547 		struct linux_prom_translation *src = &prom_trans[i + first];
548 		struct linux_prom_translation *dest = &prom_trans[i];
549 
550 		*dest = *src;
551 	}
552 	for (; i < ents; i++) {
553 		struct linux_prom_translation *dest = &prom_trans[i];
554 		dest->virt = dest->size = dest->data = 0x0UL;
555 	}
556 
557 	prom_trans_ents = last - first;
558 
559 	if (tlb_type == spitfire) {
560 		/* Clear diag TTE bits. */
561 		for (i = 0; i < prom_trans_ents; i++)
562 			prom_trans[i].data &= ~0x0003fe0000000000UL;
563 	}
564 
565 	/* Force execute bit on.  */
566 	for (i = 0; i < prom_trans_ents; i++)
567 		prom_trans[i].data |= (tlb_type == hypervisor ?
568 				       _PAGE_EXEC_4V : _PAGE_EXEC_4U);
569 }
570 
571 static void __init hypervisor_tlb_lock(unsigned long vaddr,
572 				       unsigned long pte,
573 				       unsigned long mmu)
574 {
575 	unsigned long ret = sun4v_mmu_map_perm_addr(vaddr, 0, pte, mmu);
576 
577 	if (ret != 0) {
578 		prom_printf("hypervisor_tlb_lock[%lx:%x:%lx:%lx]: "
579 			    "errors with %lx\n", vaddr, 0, pte, mmu, ret);
580 		prom_halt();
581 	}
582 }
583 
584 static unsigned long kern_large_tte(unsigned long paddr);
585 
586 static void __init remap_kernel(void)
587 {
588 	unsigned long phys_page, tte_vaddr, tte_data;
589 	int i, tlb_ent = sparc64_highest_locked_tlbent();
590 
591 	tte_vaddr = (unsigned long) KERNBASE;
592 	phys_page = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
593 	tte_data = kern_large_tte(phys_page);
594 
595 	kern_locked_tte_data = tte_data;
596 
597 	/* Now lock us into the TLBs via Hypervisor or OBP. */
598 	if (tlb_type == hypervisor) {
599 		for (i = 0; i < num_kernel_image_mappings; i++) {
600 			hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_DMMU);
601 			hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_IMMU);
602 			tte_vaddr += 0x400000;
603 			tte_data += 0x400000;
604 		}
605 	} else {
606 		for (i = 0; i < num_kernel_image_mappings; i++) {
607 			prom_dtlb_load(tlb_ent - i, tte_data, tte_vaddr);
608 			prom_itlb_load(tlb_ent - i, tte_data, tte_vaddr);
609 			tte_vaddr += 0x400000;
610 			tte_data += 0x400000;
611 		}
612 		sparc64_highest_unlocked_tlb_ent = tlb_ent - i;
613 	}
614 	if (tlb_type == cheetah_plus) {
615 		sparc64_kern_pri_context = (CTX_CHEETAH_PLUS_CTX0 |
616 					    CTX_CHEETAH_PLUS_NUC);
617 		sparc64_kern_pri_nuc_bits = CTX_CHEETAH_PLUS_NUC;
618 		sparc64_kern_sec_context = CTX_CHEETAH_PLUS_CTX0;
619 	}
620 }
621 
622 
623 static void __init inherit_prom_mappings(void)
624 {
625 	/* Now fixup OBP's idea about where we really are mapped. */
626 	printk("Remapping the kernel... ");
627 	remap_kernel();
628 	printk("done.\n");
629 }
630 
631 void prom_world(int enter)
632 {
633 	if (!enter)
634 		set_fs(get_fs());
635 
636 	__asm__ __volatile__("flushw");
637 }
638 
639 void __flush_dcache_range(unsigned long start, unsigned long end)
640 {
641 	unsigned long va;
642 
643 	if (tlb_type == spitfire) {
644 		int n = 0;
645 
646 		for (va = start; va < end; va += 32) {
647 			spitfire_put_dcache_tag(va & 0x3fe0, 0x0);
648 			if (++n >= 512)
649 				break;
650 		}
651 	} else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
652 		start = __pa(start);
653 		end = __pa(end);
654 		for (va = start; va < end; va += 32)
655 			__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
656 					     "membar #Sync"
657 					     : /* no outputs */
658 					     : "r" (va),
659 					       "i" (ASI_DCACHE_INVALIDATE));
660 	}
661 }
662 EXPORT_SYMBOL(__flush_dcache_range);
663 
664 /* get_new_mmu_context() uses "cache + 1".  */
665 DEFINE_SPINLOCK(ctx_alloc_lock);
666 unsigned long tlb_context_cache = CTX_FIRST_VERSION - 1;
667 #define MAX_CTX_NR	(1UL << CTX_NR_BITS)
668 #define CTX_BMAP_SLOTS	BITS_TO_LONGS(MAX_CTX_NR)
669 DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR);
670 
671 /* Caller does TLB context flushing on local CPU if necessary.
672  * The caller also ensures that CTX_VALID(mm->context) is false.
673  *
674  * We must be careful about boundary cases so that we never
675  * let the user have CTX 0 (nucleus) or we ever use a CTX
676  * version of zero (and thus NO_CONTEXT would not be caught
677  * by version mis-match tests in mmu_context.h).
678  *
679  * Always invoked with interrupts disabled.
680  */
681 void get_new_mmu_context(struct mm_struct *mm)
682 {
683 	unsigned long ctx, new_ctx;
684 	unsigned long orig_pgsz_bits;
685 	int new_version;
686 
687 	spin_lock(&ctx_alloc_lock);
688 	orig_pgsz_bits = (mm->context.sparc64_ctx_val & CTX_PGSZ_MASK);
689 	ctx = (tlb_context_cache + 1) & CTX_NR_MASK;
690 	new_ctx = find_next_zero_bit(mmu_context_bmap, 1 << CTX_NR_BITS, ctx);
691 	new_version = 0;
692 	if (new_ctx >= (1 << CTX_NR_BITS)) {
693 		new_ctx = find_next_zero_bit(mmu_context_bmap, ctx, 1);
694 		if (new_ctx >= ctx) {
695 			int i;
696 			new_ctx = (tlb_context_cache & CTX_VERSION_MASK) +
697 				CTX_FIRST_VERSION;
698 			if (new_ctx == 1)
699 				new_ctx = CTX_FIRST_VERSION;
700 
701 			/* Don't call memset, for 16 entries that's just
702 			 * plain silly...
703 			 */
704 			mmu_context_bmap[0] = 3;
705 			mmu_context_bmap[1] = 0;
706 			mmu_context_bmap[2] = 0;
707 			mmu_context_bmap[3] = 0;
708 			for (i = 4; i < CTX_BMAP_SLOTS; i += 4) {
709 				mmu_context_bmap[i + 0] = 0;
710 				mmu_context_bmap[i + 1] = 0;
711 				mmu_context_bmap[i + 2] = 0;
712 				mmu_context_bmap[i + 3] = 0;
713 			}
714 			new_version = 1;
715 			goto out;
716 		}
717 	}
718 	mmu_context_bmap[new_ctx>>6] |= (1UL << (new_ctx & 63));
719 	new_ctx |= (tlb_context_cache & CTX_VERSION_MASK);
720 out:
721 	tlb_context_cache = new_ctx;
722 	mm->context.sparc64_ctx_val = new_ctx | orig_pgsz_bits;
723 	spin_unlock(&ctx_alloc_lock);
724 
725 	if (unlikely(new_version))
726 		smp_new_mmu_context_version();
727 }
728 
729 static int numa_enabled = 1;
730 static int numa_debug;
731 
732 static int __init early_numa(char *p)
733 {
734 	if (!p)
735 		return 0;
736 
737 	if (strstr(p, "off"))
738 		numa_enabled = 0;
739 
740 	if (strstr(p, "debug"))
741 		numa_debug = 1;
742 
743 	return 0;
744 }
745 early_param("numa", early_numa);
746 
747 #define numadbg(f, a...) \
748 do {	if (numa_debug) \
749 		printk(KERN_INFO f, ## a); \
750 } while (0)
751 
752 static void __init find_ramdisk(unsigned long phys_base)
753 {
754 #ifdef CONFIG_BLK_DEV_INITRD
755 	if (sparc_ramdisk_image || sparc_ramdisk_image64) {
756 		unsigned long ramdisk_image;
757 
758 		/* Older versions of the bootloader only supported a
759 		 * 32-bit physical address for the ramdisk image
760 		 * location, stored at sparc_ramdisk_image.  Newer
761 		 * SILO versions set sparc_ramdisk_image to zero and
762 		 * provide a full 64-bit physical address at
763 		 * sparc_ramdisk_image64.
764 		 */
765 		ramdisk_image = sparc_ramdisk_image;
766 		if (!ramdisk_image)
767 			ramdisk_image = sparc_ramdisk_image64;
768 
769 		/* Another bootloader quirk.  The bootloader normalizes
770 		 * the physical address to KERNBASE, so we have to
771 		 * factor that back out and add in the lowest valid
772 		 * physical page address to get the true physical address.
773 		 */
774 		ramdisk_image -= KERNBASE;
775 		ramdisk_image += phys_base;
776 
777 		numadbg("Found ramdisk at physical address 0x%lx, size %u\n",
778 			ramdisk_image, sparc_ramdisk_size);
779 
780 		initrd_start = ramdisk_image;
781 		initrd_end = ramdisk_image + sparc_ramdisk_size;
782 
783 		memblock_reserve(initrd_start, sparc_ramdisk_size);
784 
785 		initrd_start += PAGE_OFFSET;
786 		initrd_end += PAGE_OFFSET;
787 	}
788 #endif
789 }
790 
791 struct node_mem_mask {
792 	unsigned long mask;
793 	unsigned long val;
794 };
795 static struct node_mem_mask node_masks[MAX_NUMNODES];
796 static int num_node_masks;
797 
798 #ifdef CONFIG_NEED_MULTIPLE_NODES
799 
800 int numa_cpu_lookup_table[NR_CPUS];
801 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
802 
803 struct mdesc_mblock {
804 	u64	base;
805 	u64	size;
806 	u64	offset; /* RA-to-PA */
807 };
808 static struct mdesc_mblock *mblocks;
809 static int num_mblocks;
810 
811 static unsigned long ra_to_pa(unsigned long addr)
812 {
813 	int i;
814 
815 	for (i = 0; i < num_mblocks; i++) {
816 		struct mdesc_mblock *m = &mblocks[i];
817 
818 		if (addr >= m->base &&
819 		    addr < (m->base + m->size)) {
820 			addr += m->offset;
821 			break;
822 		}
823 	}
824 	return addr;
825 }
826 
827 static int find_node(unsigned long addr)
828 {
829 	int i;
830 
831 	addr = ra_to_pa(addr);
832 	for (i = 0; i < num_node_masks; i++) {
833 		struct node_mem_mask *p = &node_masks[i];
834 
835 		if ((addr & p->mask) == p->val)
836 			return i;
837 	}
838 	return -1;
839 }
840 
841 static u64 memblock_nid_range(u64 start, u64 end, int *nid)
842 {
843 	*nid = find_node(start);
844 	start += PAGE_SIZE;
845 	while (start < end) {
846 		int n = find_node(start);
847 
848 		if (n != *nid)
849 			break;
850 		start += PAGE_SIZE;
851 	}
852 
853 	if (start > end)
854 		start = end;
855 
856 	return start;
857 }
858 #endif
859 
860 /* This must be invoked after performing all of the necessary
861  * memblock_set_node() calls for 'nid'.  We need to be able to get
862  * correct data from get_pfn_range_for_nid().
863  */
864 static void __init allocate_node_data(int nid)
865 {
866 	struct pglist_data *p;
867 	unsigned long start_pfn, end_pfn;
868 #ifdef CONFIG_NEED_MULTIPLE_NODES
869 	unsigned long paddr;
870 
871 	paddr = memblock_alloc_try_nid(sizeof(struct pglist_data), SMP_CACHE_BYTES, nid);
872 	if (!paddr) {
873 		prom_printf("Cannot allocate pglist_data for nid[%d]\n", nid);
874 		prom_halt();
875 	}
876 	NODE_DATA(nid) = __va(paddr);
877 	memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
878 
879 	NODE_DATA(nid)->node_id = nid;
880 #endif
881 
882 	p = NODE_DATA(nid);
883 
884 	get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
885 	p->node_start_pfn = start_pfn;
886 	p->node_spanned_pages = end_pfn - start_pfn;
887 }
888 
889 static void init_node_masks_nonnuma(void)
890 {
891 #ifdef CONFIG_NEED_MULTIPLE_NODES
892 	int i;
893 #endif
894 
895 	numadbg("Initializing tables for non-numa.\n");
896 
897 	node_masks[0].mask = node_masks[0].val = 0;
898 	num_node_masks = 1;
899 
900 #ifdef CONFIG_NEED_MULTIPLE_NODES
901 	for (i = 0; i < NR_CPUS; i++)
902 		numa_cpu_lookup_table[i] = 0;
903 
904 	cpumask_setall(&numa_cpumask_lookup_table[0]);
905 #endif
906 }
907 
908 #ifdef CONFIG_NEED_MULTIPLE_NODES
909 struct pglist_data *node_data[MAX_NUMNODES];
910 
911 EXPORT_SYMBOL(numa_cpu_lookup_table);
912 EXPORT_SYMBOL(numa_cpumask_lookup_table);
913 EXPORT_SYMBOL(node_data);
914 
915 struct mdesc_mlgroup {
916 	u64	node;
917 	u64	latency;
918 	u64	match;
919 	u64	mask;
920 };
921 static struct mdesc_mlgroup *mlgroups;
922 static int num_mlgroups;
923 
924 static int scan_pio_for_cfg_handle(struct mdesc_handle *md, u64 pio,
925 				   u32 cfg_handle)
926 {
927 	u64 arc;
928 
929 	mdesc_for_each_arc(arc, md, pio, MDESC_ARC_TYPE_FWD) {
930 		u64 target = mdesc_arc_target(md, arc);
931 		const u64 *val;
932 
933 		val = mdesc_get_property(md, target,
934 					 "cfg-handle", NULL);
935 		if (val && *val == cfg_handle)
936 			return 0;
937 	}
938 	return -ENODEV;
939 }
940 
941 static int scan_arcs_for_cfg_handle(struct mdesc_handle *md, u64 grp,
942 				    u32 cfg_handle)
943 {
944 	u64 arc, candidate, best_latency = ~(u64)0;
945 
946 	candidate = MDESC_NODE_NULL;
947 	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
948 		u64 target = mdesc_arc_target(md, arc);
949 		const char *name = mdesc_node_name(md, target);
950 		const u64 *val;
951 
952 		if (strcmp(name, "pio-latency-group"))
953 			continue;
954 
955 		val = mdesc_get_property(md, target, "latency", NULL);
956 		if (!val)
957 			continue;
958 
959 		if (*val < best_latency) {
960 			candidate = target;
961 			best_latency = *val;
962 		}
963 	}
964 
965 	if (candidate == MDESC_NODE_NULL)
966 		return -ENODEV;
967 
968 	return scan_pio_for_cfg_handle(md, candidate, cfg_handle);
969 }
970 
971 int of_node_to_nid(struct device_node *dp)
972 {
973 	const struct linux_prom64_registers *regs;
974 	struct mdesc_handle *md;
975 	u32 cfg_handle;
976 	int count, nid;
977 	u64 grp;
978 
979 	/* This is the right thing to do on currently supported
980 	 * SUN4U NUMA platforms as well, as the PCI controller does
981 	 * not sit behind any particular memory controller.
982 	 */
983 	if (!mlgroups)
984 		return -1;
985 
986 	regs = of_get_property(dp, "reg", NULL);
987 	if (!regs)
988 		return -1;
989 
990 	cfg_handle = (regs->phys_addr >> 32UL) & 0x0fffffff;
991 
992 	md = mdesc_grab();
993 
994 	count = 0;
995 	nid = -1;
996 	mdesc_for_each_node_by_name(md, grp, "group") {
997 		if (!scan_arcs_for_cfg_handle(md, grp, cfg_handle)) {
998 			nid = count;
999 			break;
1000 		}
1001 		count++;
1002 	}
1003 
1004 	mdesc_release(md);
1005 
1006 	return nid;
1007 }
1008 
1009 static void __init add_node_ranges(void)
1010 {
1011 	struct memblock_region *reg;
1012 
1013 	for_each_memblock(memory, reg) {
1014 		unsigned long size = reg->size;
1015 		unsigned long start, end;
1016 
1017 		start = reg->base;
1018 		end = start + size;
1019 		while (start < end) {
1020 			unsigned long this_end;
1021 			int nid;
1022 
1023 			this_end = memblock_nid_range(start, end, &nid);
1024 
1025 			numadbg("Setting memblock NUMA node nid[%d] "
1026 				"start[%lx] end[%lx]\n",
1027 				nid, start, this_end);
1028 
1029 			memblock_set_node(start, this_end - start,
1030 					  &memblock.memory, nid);
1031 			start = this_end;
1032 		}
1033 	}
1034 }
1035 
1036 static int __init grab_mlgroups(struct mdesc_handle *md)
1037 {
1038 	unsigned long paddr;
1039 	int count = 0;
1040 	u64 node;
1041 
1042 	mdesc_for_each_node_by_name(md, node, "memory-latency-group")
1043 		count++;
1044 	if (!count)
1045 		return -ENOENT;
1046 
1047 	paddr = memblock_alloc(count * sizeof(struct mdesc_mlgroup),
1048 			  SMP_CACHE_BYTES);
1049 	if (!paddr)
1050 		return -ENOMEM;
1051 
1052 	mlgroups = __va(paddr);
1053 	num_mlgroups = count;
1054 
1055 	count = 0;
1056 	mdesc_for_each_node_by_name(md, node, "memory-latency-group") {
1057 		struct mdesc_mlgroup *m = &mlgroups[count++];
1058 		const u64 *val;
1059 
1060 		m->node = node;
1061 
1062 		val = mdesc_get_property(md, node, "latency", NULL);
1063 		m->latency = *val;
1064 		val = mdesc_get_property(md, node, "address-match", NULL);
1065 		m->match = *val;
1066 		val = mdesc_get_property(md, node, "address-mask", NULL);
1067 		m->mask = *val;
1068 
1069 		numadbg("MLGROUP[%d]: node[%llx] latency[%llx] "
1070 			"match[%llx] mask[%llx]\n",
1071 			count - 1, m->node, m->latency, m->match, m->mask);
1072 	}
1073 
1074 	return 0;
1075 }
1076 
1077 static int __init grab_mblocks(struct mdesc_handle *md)
1078 {
1079 	unsigned long paddr;
1080 	int count = 0;
1081 	u64 node;
1082 
1083 	mdesc_for_each_node_by_name(md, node, "mblock")
1084 		count++;
1085 	if (!count)
1086 		return -ENOENT;
1087 
1088 	paddr = memblock_alloc(count * sizeof(struct mdesc_mblock),
1089 			  SMP_CACHE_BYTES);
1090 	if (!paddr)
1091 		return -ENOMEM;
1092 
1093 	mblocks = __va(paddr);
1094 	num_mblocks = count;
1095 
1096 	count = 0;
1097 	mdesc_for_each_node_by_name(md, node, "mblock") {
1098 		struct mdesc_mblock *m = &mblocks[count++];
1099 		const u64 *val;
1100 
1101 		val = mdesc_get_property(md, node, "base", NULL);
1102 		m->base = *val;
1103 		val = mdesc_get_property(md, node, "size", NULL);
1104 		m->size = *val;
1105 		val = mdesc_get_property(md, node,
1106 					 "address-congruence-offset", NULL);
1107 
1108 		/* The address-congruence-offset property is optional.
1109 		 * Explicity zero it be identifty this.
1110 		 */
1111 		if (val)
1112 			m->offset = *val;
1113 		else
1114 			m->offset = 0UL;
1115 
1116 		numadbg("MBLOCK[%d]: base[%llx] size[%llx] offset[%llx]\n",
1117 			count - 1, m->base, m->size, m->offset);
1118 	}
1119 
1120 	return 0;
1121 }
1122 
1123 static void __init numa_parse_mdesc_group_cpus(struct mdesc_handle *md,
1124 					       u64 grp, cpumask_t *mask)
1125 {
1126 	u64 arc;
1127 
1128 	cpumask_clear(mask);
1129 
1130 	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_BACK) {
1131 		u64 target = mdesc_arc_target(md, arc);
1132 		const char *name = mdesc_node_name(md, target);
1133 		const u64 *id;
1134 
1135 		if (strcmp(name, "cpu"))
1136 			continue;
1137 		id = mdesc_get_property(md, target, "id", NULL);
1138 		if (*id < nr_cpu_ids)
1139 			cpumask_set_cpu(*id, mask);
1140 	}
1141 }
1142 
1143 static struct mdesc_mlgroup * __init find_mlgroup(u64 node)
1144 {
1145 	int i;
1146 
1147 	for (i = 0; i < num_mlgroups; i++) {
1148 		struct mdesc_mlgroup *m = &mlgroups[i];
1149 		if (m->node == node)
1150 			return m;
1151 	}
1152 	return NULL;
1153 }
1154 
1155 static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp,
1156 				      int index)
1157 {
1158 	struct mdesc_mlgroup *candidate = NULL;
1159 	u64 arc, best_latency = ~(u64)0;
1160 	struct node_mem_mask *n;
1161 
1162 	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1163 		u64 target = mdesc_arc_target(md, arc);
1164 		struct mdesc_mlgroup *m = find_mlgroup(target);
1165 		if (!m)
1166 			continue;
1167 		if (m->latency < best_latency) {
1168 			candidate = m;
1169 			best_latency = m->latency;
1170 		}
1171 	}
1172 	if (!candidate)
1173 		return -ENOENT;
1174 
1175 	if (num_node_masks != index) {
1176 		printk(KERN_ERR "Inconsistent NUMA state, "
1177 		       "index[%d] != num_node_masks[%d]\n",
1178 		       index, num_node_masks);
1179 		return -EINVAL;
1180 	}
1181 
1182 	n = &node_masks[num_node_masks++];
1183 
1184 	n->mask = candidate->mask;
1185 	n->val = candidate->match;
1186 
1187 	numadbg("NUMA NODE[%d]: mask[%lx] val[%lx] (latency[%llx])\n",
1188 		index, n->mask, n->val, candidate->latency);
1189 
1190 	return 0;
1191 }
1192 
1193 static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp,
1194 					 int index)
1195 {
1196 	cpumask_t mask;
1197 	int cpu;
1198 
1199 	numa_parse_mdesc_group_cpus(md, grp, &mask);
1200 
1201 	for_each_cpu(cpu, &mask)
1202 		numa_cpu_lookup_table[cpu] = index;
1203 	cpumask_copy(&numa_cpumask_lookup_table[index], &mask);
1204 
1205 	if (numa_debug) {
1206 		printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index);
1207 		for_each_cpu(cpu, &mask)
1208 			printk("%d ", cpu);
1209 		printk("]\n");
1210 	}
1211 
1212 	return numa_attach_mlgroup(md, grp, index);
1213 }
1214 
1215 static int __init numa_parse_mdesc(void)
1216 {
1217 	struct mdesc_handle *md = mdesc_grab();
1218 	int i, err, count;
1219 	u64 node;
1220 
1221 	node = mdesc_node_by_name(md, MDESC_NODE_NULL, "latency-groups");
1222 	if (node == MDESC_NODE_NULL) {
1223 		mdesc_release(md);
1224 		return -ENOENT;
1225 	}
1226 
1227 	err = grab_mblocks(md);
1228 	if (err < 0)
1229 		goto out;
1230 
1231 	err = grab_mlgroups(md);
1232 	if (err < 0)
1233 		goto out;
1234 
1235 	count = 0;
1236 	mdesc_for_each_node_by_name(md, node, "group") {
1237 		err = numa_parse_mdesc_group(md, node, count);
1238 		if (err < 0)
1239 			break;
1240 		count++;
1241 	}
1242 
1243 	add_node_ranges();
1244 
1245 	for (i = 0; i < num_node_masks; i++) {
1246 		allocate_node_data(i);
1247 		node_set_online(i);
1248 	}
1249 
1250 	err = 0;
1251 out:
1252 	mdesc_release(md);
1253 	return err;
1254 }
1255 
1256 static int __init numa_parse_jbus(void)
1257 {
1258 	unsigned long cpu, index;
1259 
1260 	/* NUMA node id is encoded in bits 36 and higher, and there is
1261 	 * a 1-to-1 mapping from CPU ID to NUMA node ID.
1262 	 */
1263 	index = 0;
1264 	for_each_present_cpu(cpu) {
1265 		numa_cpu_lookup_table[cpu] = index;
1266 		cpumask_copy(&numa_cpumask_lookup_table[index], cpumask_of(cpu));
1267 		node_masks[index].mask = ~((1UL << 36UL) - 1UL);
1268 		node_masks[index].val = cpu << 36UL;
1269 
1270 		index++;
1271 	}
1272 	num_node_masks = index;
1273 
1274 	add_node_ranges();
1275 
1276 	for (index = 0; index < num_node_masks; index++) {
1277 		allocate_node_data(index);
1278 		node_set_online(index);
1279 	}
1280 
1281 	return 0;
1282 }
1283 
1284 static int __init numa_parse_sun4u(void)
1285 {
1286 	if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1287 		unsigned long ver;
1288 
1289 		__asm__ ("rdpr %%ver, %0" : "=r" (ver));
1290 		if ((ver >> 32UL) == __JALAPENO_ID ||
1291 		    (ver >> 32UL) == __SERRANO_ID)
1292 			return numa_parse_jbus();
1293 	}
1294 	return -1;
1295 }
1296 
1297 static int __init bootmem_init_numa(void)
1298 {
1299 	int err = -1;
1300 
1301 	numadbg("bootmem_init_numa()\n");
1302 
1303 	if (numa_enabled) {
1304 		if (tlb_type == hypervisor)
1305 			err = numa_parse_mdesc();
1306 		else
1307 			err = numa_parse_sun4u();
1308 	}
1309 	return err;
1310 }
1311 
1312 #else
1313 
1314 static int bootmem_init_numa(void)
1315 {
1316 	return -1;
1317 }
1318 
1319 #endif
1320 
1321 static void __init bootmem_init_nonnuma(void)
1322 {
1323 	unsigned long top_of_ram = memblock_end_of_DRAM();
1324 	unsigned long total_ram = memblock_phys_mem_size();
1325 
1326 	numadbg("bootmem_init_nonnuma()\n");
1327 
1328 	printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
1329 	       top_of_ram, total_ram);
1330 	printk(KERN_INFO "Memory hole size: %ldMB\n",
1331 	       (top_of_ram - total_ram) >> 20);
1332 
1333 	init_node_masks_nonnuma();
1334 	memblock_set_node(0, (phys_addr_t)ULLONG_MAX, &memblock.memory, 0);
1335 	allocate_node_data(0);
1336 	node_set_online(0);
1337 }
1338 
1339 static unsigned long __init bootmem_init(unsigned long phys_base)
1340 {
1341 	unsigned long end_pfn;
1342 
1343 	end_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1344 	max_pfn = max_low_pfn = end_pfn;
1345 	min_low_pfn = (phys_base >> PAGE_SHIFT);
1346 
1347 	if (bootmem_init_numa() < 0)
1348 		bootmem_init_nonnuma();
1349 
1350 	/* Dump memblock with node info. */
1351 	memblock_dump_all();
1352 
1353 	/* XXX cpu notifier XXX */
1354 
1355 	sparse_memory_present_with_active_regions(MAX_NUMNODES);
1356 	sparse_init();
1357 
1358 	return end_pfn;
1359 }
1360 
1361 static struct linux_prom64_registers pall[MAX_BANKS] __initdata;
1362 static int pall_ents __initdata;
1363 
1364 #ifdef CONFIG_DEBUG_PAGEALLOC
1365 static unsigned long __ref kernel_map_range(unsigned long pstart,
1366 					    unsigned long pend, pgprot_t prot)
1367 {
1368 	unsigned long vstart = PAGE_OFFSET + pstart;
1369 	unsigned long vend = PAGE_OFFSET + pend;
1370 	unsigned long alloc_bytes = 0UL;
1371 
1372 	if ((vstart & ~PAGE_MASK) || (vend & ~PAGE_MASK)) {
1373 		prom_printf("kernel_map: Unaligned physmem[%lx:%lx]\n",
1374 			    vstart, vend);
1375 		prom_halt();
1376 	}
1377 
1378 	while (vstart < vend) {
1379 		unsigned long this_end, paddr = __pa(vstart);
1380 		pgd_t *pgd = pgd_offset_k(vstart);
1381 		pud_t *pud;
1382 		pmd_t *pmd;
1383 		pte_t *pte;
1384 
1385 		pud = pud_offset(pgd, vstart);
1386 		if (pud_none(*pud)) {
1387 			pmd_t *new;
1388 
1389 			new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1390 			alloc_bytes += PAGE_SIZE;
1391 			pud_populate(&init_mm, pud, new);
1392 		}
1393 
1394 		pmd = pmd_offset(pud, vstart);
1395 		if (!pmd_present(*pmd)) {
1396 			pte_t *new;
1397 
1398 			new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1399 			alloc_bytes += PAGE_SIZE;
1400 			pmd_populate_kernel(&init_mm, pmd, new);
1401 		}
1402 
1403 		pte = pte_offset_kernel(pmd, vstart);
1404 		this_end = (vstart + PMD_SIZE) & PMD_MASK;
1405 		if (this_end > vend)
1406 			this_end = vend;
1407 
1408 		while (vstart < this_end) {
1409 			pte_val(*pte) = (paddr | pgprot_val(prot));
1410 
1411 			vstart += PAGE_SIZE;
1412 			paddr += PAGE_SIZE;
1413 			pte++;
1414 		}
1415 	}
1416 
1417 	return alloc_bytes;
1418 }
1419 
1420 extern unsigned int kvmap_linear_patch[1];
1421 #endif /* CONFIG_DEBUG_PAGEALLOC */
1422 
1423 static void __init kpte_set_val(unsigned long index, unsigned long val)
1424 {
1425 	unsigned long *ptr = kpte_linear_bitmap;
1426 
1427 	val <<= ((index % (BITS_PER_LONG / 2)) * 2);
1428 	ptr += (index / (BITS_PER_LONG / 2));
1429 
1430 	*ptr |= val;
1431 }
1432 
1433 static const unsigned long kpte_shift_min = 28; /* 256MB */
1434 static const unsigned long kpte_shift_max = 34; /* 16GB */
1435 static const unsigned long kpte_shift_incr = 3;
1436 
1437 static unsigned long kpte_mark_using_shift(unsigned long start, unsigned long end,
1438 					   unsigned long shift)
1439 {
1440 	unsigned long size = (1UL << shift);
1441 	unsigned long mask = (size - 1UL);
1442 	unsigned long remains = end - start;
1443 	unsigned long val;
1444 
1445 	if (remains < size || (start & mask))
1446 		return start;
1447 
1448 	/* VAL maps:
1449 	 *
1450 	 *	shift 28 --> kern_linear_pte_xor index 1
1451 	 *	shift 31 --> kern_linear_pte_xor index 2
1452 	 *	shift 34 --> kern_linear_pte_xor index 3
1453 	 */
1454 	val = ((shift - kpte_shift_min) / kpte_shift_incr) + 1;
1455 
1456 	remains &= ~mask;
1457 	if (shift != kpte_shift_max)
1458 		remains = size;
1459 
1460 	while (remains) {
1461 		unsigned long index = start >> kpte_shift_min;
1462 
1463 		kpte_set_val(index, val);
1464 
1465 		start += 1UL << kpte_shift_min;
1466 		remains -= 1UL << kpte_shift_min;
1467 	}
1468 
1469 	return start;
1470 }
1471 
1472 static void __init mark_kpte_bitmap(unsigned long start, unsigned long end)
1473 {
1474 	unsigned long smallest_size, smallest_mask;
1475 	unsigned long s;
1476 
1477 	smallest_size = (1UL << kpte_shift_min);
1478 	smallest_mask = (smallest_size - 1UL);
1479 
1480 	while (start < end) {
1481 		unsigned long orig_start = start;
1482 
1483 		for (s = kpte_shift_max; s >= kpte_shift_min; s -= kpte_shift_incr) {
1484 			start = kpte_mark_using_shift(start, end, s);
1485 
1486 			if (start != orig_start)
1487 				break;
1488 		}
1489 
1490 		if (start == orig_start)
1491 			start = (start + smallest_size) & ~smallest_mask;
1492 	}
1493 }
1494 
1495 static void __init init_kpte_bitmap(void)
1496 {
1497 	unsigned long i;
1498 
1499 	for (i = 0; i < pall_ents; i++) {
1500 		unsigned long phys_start, phys_end;
1501 
1502 		phys_start = pall[i].phys_addr;
1503 		phys_end = phys_start + pall[i].reg_size;
1504 
1505 		mark_kpte_bitmap(phys_start, phys_end);
1506 	}
1507 }
1508 
1509 static void __init kernel_physical_mapping_init(void)
1510 {
1511 #ifdef CONFIG_DEBUG_PAGEALLOC
1512 	unsigned long i, mem_alloced = 0UL;
1513 
1514 	for (i = 0; i < pall_ents; i++) {
1515 		unsigned long phys_start, phys_end;
1516 
1517 		phys_start = pall[i].phys_addr;
1518 		phys_end = phys_start + pall[i].reg_size;
1519 
1520 		mem_alloced += kernel_map_range(phys_start, phys_end,
1521 						PAGE_KERNEL);
1522 	}
1523 
1524 	printk("Allocated %ld bytes for kernel page tables.\n",
1525 	       mem_alloced);
1526 
1527 	kvmap_linear_patch[0] = 0x01000000; /* nop */
1528 	flushi(&kvmap_linear_patch[0]);
1529 
1530 	__flush_tlb_all();
1531 #endif
1532 }
1533 
1534 #ifdef CONFIG_DEBUG_PAGEALLOC
1535 void kernel_map_pages(struct page *page, int numpages, int enable)
1536 {
1537 	unsigned long phys_start = page_to_pfn(page) << PAGE_SHIFT;
1538 	unsigned long phys_end = phys_start + (numpages * PAGE_SIZE);
1539 
1540 	kernel_map_range(phys_start, phys_end,
1541 			 (enable ? PAGE_KERNEL : __pgprot(0)));
1542 
1543 	flush_tsb_kernel_range(PAGE_OFFSET + phys_start,
1544 			       PAGE_OFFSET + phys_end);
1545 
1546 	/* we should perform an IPI and flush all tlbs,
1547 	 * but that can deadlock->flush only current cpu.
1548 	 */
1549 	__flush_tlb_kernel_range(PAGE_OFFSET + phys_start,
1550 				 PAGE_OFFSET + phys_end);
1551 }
1552 #endif
1553 
1554 unsigned long __init find_ecache_flush_span(unsigned long size)
1555 {
1556 	int i;
1557 
1558 	for (i = 0; i < pavail_ents; i++) {
1559 		if (pavail[i].reg_size >= size)
1560 			return pavail[i].phys_addr;
1561 	}
1562 
1563 	return ~0UL;
1564 }
1565 
1566 unsigned long PAGE_OFFSET;
1567 EXPORT_SYMBOL(PAGE_OFFSET);
1568 
1569 static void __init page_offset_shift_patch_one(unsigned int *insn, unsigned long phys_bits)
1570 {
1571 	unsigned long final_shift;
1572 	unsigned int val = *insn;
1573 	unsigned int cnt;
1574 
1575 	/* We are patching in ilog2(max_supported_phys_address), and
1576 	 * we are doing so in a manner similar to a relocation addend.
1577 	 * That is, we are adding the shift value to whatever value
1578 	 * is in the shift instruction count field already.
1579 	 */
1580 	cnt = (val & 0x3f);
1581 	val &= ~0x3f;
1582 
1583 	/* If we are trying to shift >= 64 bits, clear the destination
1584 	 * register.  This can happen when phys_bits ends up being equal
1585 	 * to MAX_PHYS_ADDRESS_BITS.
1586 	 */
1587 	final_shift = (cnt + (64 - phys_bits));
1588 	if (final_shift >= 64) {
1589 		unsigned int rd = (val >> 25) & 0x1f;
1590 
1591 		val = 0x80100000 | (rd << 25);
1592 	} else {
1593 		val |= final_shift;
1594 	}
1595 	*insn = val;
1596 
1597 	__asm__ __volatile__("flush	%0"
1598 			     : /* no outputs */
1599 			     : "r" (insn));
1600 }
1601 
1602 static void __init page_offset_shift_patch(unsigned long phys_bits)
1603 {
1604 	extern unsigned int __page_offset_shift_patch;
1605 	extern unsigned int __page_offset_shift_patch_end;
1606 	unsigned int *p;
1607 
1608 	p = &__page_offset_shift_patch;
1609 	while (p < &__page_offset_shift_patch_end) {
1610 		unsigned int *insn = (unsigned int *)(unsigned long)*p;
1611 
1612 		page_offset_shift_patch_one(insn, phys_bits);
1613 
1614 		p++;
1615 	}
1616 }
1617 
1618 static void __init setup_page_offset(void)
1619 {
1620 	unsigned long max_phys_bits = 40;
1621 
1622 	if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1623 		max_phys_bits = 42;
1624 	} else if (tlb_type == hypervisor) {
1625 		switch (sun4v_chip_type) {
1626 		case SUN4V_CHIP_NIAGARA1:
1627 		case SUN4V_CHIP_NIAGARA2:
1628 			max_phys_bits = 39;
1629 			break;
1630 		case SUN4V_CHIP_NIAGARA3:
1631 			max_phys_bits = 43;
1632 			break;
1633 		case SUN4V_CHIP_NIAGARA4:
1634 		case SUN4V_CHIP_NIAGARA5:
1635 		case SUN4V_CHIP_SPARC64X:
1636 		default:
1637 			max_phys_bits = 47;
1638 			break;
1639 		}
1640 	}
1641 
1642 	if (max_phys_bits > MAX_PHYS_ADDRESS_BITS) {
1643 		prom_printf("MAX_PHYS_ADDRESS_BITS is too small, need %lu\n",
1644 			    max_phys_bits);
1645 		prom_halt();
1646 	}
1647 
1648 	PAGE_OFFSET = PAGE_OFFSET_BY_BITS(max_phys_bits);
1649 
1650 	pr_info("PAGE_OFFSET is 0x%016lx (max_phys_bits == %lu)\n",
1651 		PAGE_OFFSET, max_phys_bits);
1652 
1653 	page_offset_shift_patch(max_phys_bits);
1654 }
1655 
1656 static void __init tsb_phys_patch(void)
1657 {
1658 	struct tsb_ldquad_phys_patch_entry *pquad;
1659 	struct tsb_phys_patch_entry *p;
1660 
1661 	pquad = &__tsb_ldquad_phys_patch;
1662 	while (pquad < &__tsb_ldquad_phys_patch_end) {
1663 		unsigned long addr = pquad->addr;
1664 
1665 		if (tlb_type == hypervisor)
1666 			*(unsigned int *) addr = pquad->sun4v_insn;
1667 		else
1668 			*(unsigned int *) addr = pquad->sun4u_insn;
1669 		wmb();
1670 		__asm__ __volatile__("flush	%0"
1671 				     : /* no outputs */
1672 				     : "r" (addr));
1673 
1674 		pquad++;
1675 	}
1676 
1677 	p = &__tsb_phys_patch;
1678 	while (p < &__tsb_phys_patch_end) {
1679 		unsigned long addr = p->addr;
1680 
1681 		*(unsigned int *) addr = p->insn;
1682 		wmb();
1683 		__asm__ __volatile__("flush	%0"
1684 				     : /* no outputs */
1685 				     : "r" (addr));
1686 
1687 		p++;
1688 	}
1689 }
1690 
1691 /* Don't mark as init, we give this to the Hypervisor.  */
1692 #ifndef CONFIG_DEBUG_PAGEALLOC
1693 #define NUM_KTSB_DESCR	2
1694 #else
1695 #define NUM_KTSB_DESCR	1
1696 #endif
1697 static struct hv_tsb_descr ktsb_descr[NUM_KTSB_DESCR];
1698 extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
1699 
1700 static void patch_one_ktsb_phys(unsigned int *start, unsigned int *end, unsigned long pa)
1701 {
1702 	pa >>= KTSB_PHYS_SHIFT;
1703 
1704 	while (start < end) {
1705 		unsigned int *ia = (unsigned int *)(unsigned long)*start;
1706 
1707 		ia[0] = (ia[0] & ~0x3fffff) | (pa >> 10);
1708 		__asm__ __volatile__("flush	%0" : : "r" (ia));
1709 
1710 		ia[1] = (ia[1] & ~0x3ff) | (pa & 0x3ff);
1711 		__asm__ __volatile__("flush	%0" : : "r" (ia + 1));
1712 
1713 		start++;
1714 	}
1715 }
1716 
1717 static void ktsb_phys_patch(void)
1718 {
1719 	extern unsigned int __swapper_tsb_phys_patch;
1720 	extern unsigned int __swapper_tsb_phys_patch_end;
1721 	unsigned long ktsb_pa;
1722 
1723 	ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
1724 	patch_one_ktsb_phys(&__swapper_tsb_phys_patch,
1725 			    &__swapper_tsb_phys_patch_end, ktsb_pa);
1726 #ifndef CONFIG_DEBUG_PAGEALLOC
1727 	{
1728 	extern unsigned int __swapper_4m_tsb_phys_patch;
1729 	extern unsigned int __swapper_4m_tsb_phys_patch_end;
1730 	ktsb_pa = (kern_base +
1731 		   ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
1732 	patch_one_ktsb_phys(&__swapper_4m_tsb_phys_patch,
1733 			    &__swapper_4m_tsb_phys_patch_end, ktsb_pa);
1734 	}
1735 #endif
1736 }
1737 
1738 static void __init sun4v_ktsb_init(void)
1739 {
1740 	unsigned long ktsb_pa;
1741 
1742 	/* First KTSB for PAGE_SIZE mappings.  */
1743 	ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
1744 
1745 	switch (PAGE_SIZE) {
1746 	case 8 * 1024:
1747 	default:
1748 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_8K;
1749 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_8K;
1750 		break;
1751 
1752 	case 64 * 1024:
1753 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_64K;
1754 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_64K;
1755 		break;
1756 
1757 	case 512 * 1024:
1758 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_512K;
1759 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_512K;
1760 		break;
1761 
1762 	case 4 * 1024 * 1024:
1763 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_4MB;
1764 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_4MB;
1765 		break;
1766 	}
1767 
1768 	ktsb_descr[0].assoc = 1;
1769 	ktsb_descr[0].num_ttes = KERNEL_TSB_NENTRIES;
1770 	ktsb_descr[0].ctx_idx = 0;
1771 	ktsb_descr[0].tsb_base = ktsb_pa;
1772 	ktsb_descr[0].resv = 0;
1773 
1774 #ifndef CONFIG_DEBUG_PAGEALLOC
1775 	/* Second KTSB for 4MB/256MB/2GB/16GB mappings.  */
1776 	ktsb_pa = (kern_base +
1777 		   ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
1778 
1779 	ktsb_descr[1].pgsz_idx = HV_PGSZ_IDX_4MB;
1780 	ktsb_descr[1].pgsz_mask = ((HV_PGSZ_MASK_4MB |
1781 				    HV_PGSZ_MASK_256MB |
1782 				    HV_PGSZ_MASK_2GB |
1783 				    HV_PGSZ_MASK_16GB) &
1784 				   cpu_pgsz_mask);
1785 	ktsb_descr[1].assoc = 1;
1786 	ktsb_descr[1].num_ttes = KERNEL_TSB4M_NENTRIES;
1787 	ktsb_descr[1].ctx_idx = 0;
1788 	ktsb_descr[1].tsb_base = ktsb_pa;
1789 	ktsb_descr[1].resv = 0;
1790 #endif
1791 }
1792 
1793 void sun4v_ktsb_register(void)
1794 {
1795 	unsigned long pa, ret;
1796 
1797 	pa = kern_base + ((unsigned long)&ktsb_descr[0] - KERNBASE);
1798 
1799 	ret = sun4v_mmu_tsb_ctx0(NUM_KTSB_DESCR, pa);
1800 	if (ret != 0) {
1801 		prom_printf("hypervisor_mmu_tsb_ctx0[%lx]: "
1802 			    "errors with %lx\n", pa, ret);
1803 		prom_halt();
1804 	}
1805 }
1806 
1807 static void __init sun4u_linear_pte_xor_finalize(void)
1808 {
1809 #ifndef CONFIG_DEBUG_PAGEALLOC
1810 	/* This is where we would add Panther support for
1811 	 * 32MB and 256MB pages.
1812 	 */
1813 #endif
1814 }
1815 
1816 static void __init sun4v_linear_pte_xor_finalize(void)
1817 {
1818 #ifndef CONFIG_DEBUG_PAGEALLOC
1819 	if (cpu_pgsz_mask & HV_PGSZ_MASK_256MB) {
1820 		kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZ256MB_4V) ^
1821 			PAGE_OFFSET;
1822 		kern_linear_pte_xor[1] |= (_PAGE_CP_4V | _PAGE_CV_4V |
1823 					   _PAGE_P_4V | _PAGE_W_4V);
1824 	} else {
1825 		kern_linear_pte_xor[1] = kern_linear_pte_xor[0];
1826 	}
1827 
1828 	if (cpu_pgsz_mask & HV_PGSZ_MASK_2GB) {
1829 		kern_linear_pte_xor[2] = (_PAGE_VALID | _PAGE_SZ2GB_4V) ^
1830 			PAGE_OFFSET;
1831 		kern_linear_pte_xor[2] |= (_PAGE_CP_4V | _PAGE_CV_4V |
1832 					   _PAGE_P_4V | _PAGE_W_4V);
1833 	} else {
1834 		kern_linear_pte_xor[2] = kern_linear_pte_xor[1];
1835 	}
1836 
1837 	if (cpu_pgsz_mask & HV_PGSZ_MASK_16GB) {
1838 		kern_linear_pte_xor[3] = (_PAGE_VALID | _PAGE_SZ16GB_4V) ^
1839 			PAGE_OFFSET;
1840 		kern_linear_pte_xor[3] |= (_PAGE_CP_4V | _PAGE_CV_4V |
1841 					   _PAGE_P_4V | _PAGE_W_4V);
1842 	} else {
1843 		kern_linear_pte_xor[3] = kern_linear_pte_xor[2];
1844 	}
1845 #endif
1846 }
1847 
1848 /* paging_init() sets up the page tables */
1849 
1850 static unsigned long last_valid_pfn;
1851 pgd_t swapper_pg_dir[PTRS_PER_PGD];
1852 
1853 static void sun4u_pgprot_init(void);
1854 static void sun4v_pgprot_init(void);
1855 
1856 void __init paging_init(void)
1857 {
1858 	unsigned long end_pfn, shift, phys_base;
1859 	unsigned long real_end, i;
1860 	int node;
1861 
1862 	setup_page_offset();
1863 
1864 	/* These build time checkes make sure that the dcache_dirty_cpu()
1865 	 * page->flags usage will work.
1866 	 *
1867 	 * When a page gets marked as dcache-dirty, we store the
1868 	 * cpu number starting at bit 32 in the page->flags.  Also,
1869 	 * functions like clear_dcache_dirty_cpu use the cpu mask
1870 	 * in 13-bit signed-immediate instruction fields.
1871 	 */
1872 
1873 	/*
1874 	 * Page flags must not reach into upper 32 bits that are used
1875 	 * for the cpu number
1876 	 */
1877 	BUILD_BUG_ON(NR_PAGEFLAGS > 32);
1878 
1879 	/*
1880 	 * The bit fields placed in the high range must not reach below
1881 	 * the 32 bit boundary. Otherwise we cannot place the cpu field
1882 	 * at the 32 bit boundary.
1883 	 */
1884 	BUILD_BUG_ON(SECTIONS_WIDTH + NODES_WIDTH + ZONES_WIDTH +
1885 		ilog2(roundup_pow_of_two(NR_CPUS)) > 32);
1886 
1887 	BUILD_BUG_ON(NR_CPUS > 4096);
1888 
1889 	kern_base = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
1890 	kern_size = (unsigned long)&_end - (unsigned long)KERNBASE;
1891 
1892 	/* Invalidate both kernel TSBs.  */
1893 	memset(swapper_tsb, 0x40, sizeof(swapper_tsb));
1894 #ifndef CONFIG_DEBUG_PAGEALLOC
1895 	memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
1896 #endif
1897 
1898 	if (tlb_type == hypervisor)
1899 		sun4v_pgprot_init();
1900 	else
1901 		sun4u_pgprot_init();
1902 
1903 	if (tlb_type == cheetah_plus ||
1904 	    tlb_type == hypervisor) {
1905 		tsb_phys_patch();
1906 		ktsb_phys_patch();
1907 	}
1908 
1909 	if (tlb_type == hypervisor)
1910 		sun4v_patch_tlb_handlers();
1911 
1912 	/* Find available physical memory...
1913 	 *
1914 	 * Read it twice in order to work around a bug in openfirmware.
1915 	 * The call to grab this table itself can cause openfirmware to
1916 	 * allocate memory, which in turn can take away some space from
1917 	 * the list of available memory.  Reading it twice makes sure
1918 	 * we really do get the final value.
1919 	 */
1920 	read_obp_translations();
1921 	read_obp_memory("reg", &pall[0], &pall_ents);
1922 	read_obp_memory("available", &pavail[0], &pavail_ents);
1923 	read_obp_memory("available", &pavail[0], &pavail_ents);
1924 
1925 	phys_base = 0xffffffffffffffffUL;
1926 	for (i = 0; i < pavail_ents; i++) {
1927 		phys_base = min(phys_base, pavail[i].phys_addr);
1928 		memblock_add(pavail[i].phys_addr, pavail[i].reg_size);
1929 	}
1930 
1931 	memblock_reserve(kern_base, kern_size);
1932 
1933 	find_ramdisk(phys_base);
1934 
1935 	memblock_enforce_memory_limit(cmdline_memory_size);
1936 
1937 	memblock_allow_resize();
1938 	memblock_dump_all();
1939 
1940 	set_bit(0, mmu_context_bmap);
1941 
1942 	shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE);
1943 
1944 	real_end = (unsigned long)_end;
1945 	num_kernel_image_mappings = DIV_ROUND_UP(real_end - KERNBASE, 1 << ILOG2_4MB);
1946 	printk("Kernel: Using %d locked TLB entries for main kernel image.\n",
1947 	       num_kernel_image_mappings);
1948 
1949 	/* Set kernel pgd to upper alias so physical page computations
1950 	 * work.
1951 	 */
1952 	init_mm.pgd += ((shift) / (sizeof(pgd_t)));
1953 
1954 	memset(swapper_low_pmd_dir, 0, sizeof(swapper_low_pmd_dir));
1955 
1956 	/* Now can init the kernel/bad page tables. */
1957 	pud_set(pud_offset(&swapper_pg_dir[0], 0),
1958 		swapper_low_pmd_dir + (shift / sizeof(pgd_t)));
1959 
1960 	inherit_prom_mappings();
1961 
1962 	init_kpte_bitmap();
1963 
1964 	/* Ok, we can use our TLB miss and window trap handlers safely.  */
1965 	setup_tba();
1966 
1967 	__flush_tlb_all();
1968 
1969 	prom_build_devicetree();
1970 	of_populate_present_mask();
1971 #ifndef CONFIG_SMP
1972 	of_fill_in_cpu_data();
1973 #endif
1974 
1975 	if (tlb_type == hypervisor) {
1976 		sun4v_mdesc_init();
1977 		mdesc_populate_present_mask(cpu_all_mask);
1978 #ifndef CONFIG_SMP
1979 		mdesc_fill_in_cpu_data(cpu_all_mask);
1980 #endif
1981 		mdesc_get_page_sizes(cpu_all_mask, &cpu_pgsz_mask);
1982 
1983 		sun4v_linear_pte_xor_finalize();
1984 
1985 		sun4v_ktsb_init();
1986 		sun4v_ktsb_register();
1987 	} else {
1988 		unsigned long impl, ver;
1989 
1990 		cpu_pgsz_mask = (HV_PGSZ_MASK_8K | HV_PGSZ_MASK_64K |
1991 				 HV_PGSZ_MASK_512K | HV_PGSZ_MASK_4MB);
1992 
1993 		__asm__ __volatile__("rdpr %%ver, %0" : "=r" (ver));
1994 		impl = ((ver >> 32) & 0xffff);
1995 		if (impl == PANTHER_IMPL)
1996 			cpu_pgsz_mask |= (HV_PGSZ_MASK_32MB |
1997 					  HV_PGSZ_MASK_256MB);
1998 
1999 		sun4u_linear_pte_xor_finalize();
2000 	}
2001 
2002 	/* Flush the TLBs and the 4M TSB so that the updated linear
2003 	 * pte XOR settings are realized for all mappings.
2004 	 */
2005 	__flush_tlb_all();
2006 #ifndef CONFIG_DEBUG_PAGEALLOC
2007 	memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2008 #endif
2009 	__flush_tlb_all();
2010 
2011 	/* Setup bootmem... */
2012 	last_valid_pfn = end_pfn = bootmem_init(phys_base);
2013 
2014 	/* Once the OF device tree and MDESC have been setup, we know
2015 	 * the list of possible cpus.  Therefore we can allocate the
2016 	 * IRQ stacks.
2017 	 */
2018 	for_each_possible_cpu(i) {
2019 		node = cpu_to_node(i);
2020 
2021 		softirq_stack[i] = __alloc_bootmem_node(NODE_DATA(node),
2022 							THREAD_SIZE,
2023 							THREAD_SIZE, 0);
2024 		hardirq_stack[i] = __alloc_bootmem_node(NODE_DATA(node),
2025 							THREAD_SIZE,
2026 							THREAD_SIZE, 0);
2027 	}
2028 
2029 	kernel_physical_mapping_init();
2030 
2031 	{
2032 		unsigned long max_zone_pfns[MAX_NR_ZONES];
2033 
2034 		memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
2035 
2036 		max_zone_pfns[ZONE_NORMAL] = end_pfn;
2037 
2038 		free_area_init_nodes(max_zone_pfns);
2039 	}
2040 
2041 	printk("Booting Linux...\n");
2042 }
2043 
2044 int page_in_phys_avail(unsigned long paddr)
2045 {
2046 	int i;
2047 
2048 	paddr &= PAGE_MASK;
2049 
2050 	for (i = 0; i < pavail_ents; i++) {
2051 		unsigned long start, end;
2052 
2053 		start = pavail[i].phys_addr;
2054 		end = start + pavail[i].reg_size;
2055 
2056 		if (paddr >= start && paddr < end)
2057 			return 1;
2058 	}
2059 	if (paddr >= kern_base && paddr < (kern_base + kern_size))
2060 		return 1;
2061 #ifdef CONFIG_BLK_DEV_INITRD
2062 	if (paddr >= __pa(initrd_start) &&
2063 	    paddr < __pa(PAGE_ALIGN(initrd_end)))
2064 		return 1;
2065 #endif
2066 
2067 	return 0;
2068 }
2069 
2070 static struct linux_prom64_registers pavail_rescan[MAX_BANKS] __initdata;
2071 static int pavail_rescan_ents __initdata;
2072 
2073 /* Certain OBP calls, such as fetching "available" properties, can
2074  * claim physical memory.  So, along with initializing the valid
2075  * address bitmap, what we do here is refetch the physical available
2076  * memory list again, and make sure it provides at least as much
2077  * memory as 'pavail' does.
2078  */
2079 static void __init setup_valid_addr_bitmap_from_pavail(unsigned long *bitmap)
2080 {
2081 	int i;
2082 
2083 	read_obp_memory("available", &pavail_rescan[0], &pavail_rescan_ents);
2084 
2085 	for (i = 0; i < pavail_ents; i++) {
2086 		unsigned long old_start, old_end;
2087 
2088 		old_start = pavail[i].phys_addr;
2089 		old_end = old_start + pavail[i].reg_size;
2090 		while (old_start < old_end) {
2091 			int n;
2092 
2093 			for (n = 0; n < pavail_rescan_ents; n++) {
2094 				unsigned long new_start, new_end;
2095 
2096 				new_start = pavail_rescan[n].phys_addr;
2097 				new_end = new_start +
2098 					pavail_rescan[n].reg_size;
2099 
2100 				if (new_start <= old_start &&
2101 				    new_end >= (old_start + PAGE_SIZE)) {
2102 					set_bit(old_start >> ILOG2_4MB, bitmap);
2103 					goto do_next_page;
2104 				}
2105 			}
2106 
2107 			prom_printf("mem_init: Lost memory in pavail\n");
2108 			prom_printf("mem_init: OLD start[%lx] size[%lx]\n",
2109 				    pavail[i].phys_addr,
2110 				    pavail[i].reg_size);
2111 			prom_printf("mem_init: NEW start[%lx] size[%lx]\n",
2112 				    pavail_rescan[i].phys_addr,
2113 				    pavail_rescan[i].reg_size);
2114 			prom_printf("mem_init: Cannot continue, aborting.\n");
2115 			prom_halt();
2116 
2117 		do_next_page:
2118 			old_start += PAGE_SIZE;
2119 		}
2120 	}
2121 }
2122 
2123 static void __init patch_tlb_miss_handler_bitmap(void)
2124 {
2125 	extern unsigned int valid_addr_bitmap_insn[];
2126 	extern unsigned int valid_addr_bitmap_patch[];
2127 
2128 	valid_addr_bitmap_insn[1] = valid_addr_bitmap_patch[1];
2129 	mb();
2130 	valid_addr_bitmap_insn[0] = valid_addr_bitmap_patch[0];
2131 	flushi(&valid_addr_bitmap_insn[0]);
2132 }
2133 
2134 static void __init register_page_bootmem_info(void)
2135 {
2136 #ifdef CONFIG_NEED_MULTIPLE_NODES
2137 	int i;
2138 
2139 	for_each_online_node(i)
2140 		if (NODE_DATA(i)->node_spanned_pages)
2141 			register_page_bootmem_info_node(NODE_DATA(i));
2142 #endif
2143 }
2144 void __init mem_init(void)
2145 {
2146 	unsigned long addr, last;
2147 
2148 	addr = PAGE_OFFSET + kern_base;
2149 	last = PAGE_ALIGN(kern_size) + addr;
2150 	while (addr < last) {
2151 		set_bit(__pa(addr) >> ILOG2_4MB, sparc64_valid_addr_bitmap);
2152 		addr += PAGE_SIZE;
2153 	}
2154 
2155 	setup_valid_addr_bitmap_from_pavail(sparc64_valid_addr_bitmap);
2156 	patch_tlb_miss_handler_bitmap();
2157 
2158 	high_memory = __va(last_valid_pfn << PAGE_SHIFT);
2159 
2160 	register_page_bootmem_info();
2161 	free_all_bootmem();
2162 
2163 	/*
2164 	 * Set up the zero page, mark it reserved, so that page count
2165 	 * is not manipulated when freeing the page from user ptes.
2166 	 */
2167 	mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0);
2168 	if (mem_map_zero == NULL) {
2169 		prom_printf("paging_init: Cannot alloc zero page.\n");
2170 		prom_halt();
2171 	}
2172 	mark_page_reserved(mem_map_zero);
2173 
2174 	mem_init_print_info(NULL);
2175 
2176 	if (tlb_type == cheetah || tlb_type == cheetah_plus)
2177 		cheetah_ecache_flush_init();
2178 }
2179 
2180 void free_initmem(void)
2181 {
2182 	unsigned long addr, initend;
2183 	int do_free = 1;
2184 
2185 	/* If the physical memory maps were trimmed by kernel command
2186 	 * line options, don't even try freeing this initmem stuff up.
2187 	 * The kernel image could have been in the trimmed out region
2188 	 * and if so the freeing below will free invalid page structs.
2189 	 */
2190 	if (cmdline_memory_size)
2191 		do_free = 0;
2192 
2193 	/*
2194 	 * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes.
2195 	 */
2196 	addr = PAGE_ALIGN((unsigned long)(__init_begin));
2197 	initend = (unsigned long)(__init_end) & PAGE_MASK;
2198 	for (; addr < initend; addr += PAGE_SIZE) {
2199 		unsigned long page;
2200 
2201 		page = (addr +
2202 			((unsigned long) __va(kern_base)) -
2203 			((unsigned long) KERNBASE));
2204 		memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
2205 
2206 		if (do_free)
2207 			free_reserved_page(virt_to_page(page));
2208 	}
2209 }
2210 
2211 #ifdef CONFIG_BLK_DEV_INITRD
2212 void free_initrd_mem(unsigned long start, unsigned long end)
2213 {
2214 	free_reserved_area((void *)start, (void *)end, POISON_FREE_INITMEM,
2215 			   "initrd");
2216 }
2217 #endif
2218 
2219 #define _PAGE_CACHE_4U	(_PAGE_CP_4U | _PAGE_CV_4U)
2220 #define _PAGE_CACHE_4V	(_PAGE_CP_4V | _PAGE_CV_4V)
2221 #define __DIRTY_BITS_4U	 (_PAGE_MODIFIED_4U | _PAGE_WRITE_4U | _PAGE_W_4U)
2222 #define __DIRTY_BITS_4V	 (_PAGE_MODIFIED_4V | _PAGE_WRITE_4V | _PAGE_W_4V)
2223 #define __ACCESS_BITS_4U (_PAGE_ACCESSED_4U | _PAGE_READ_4U | _PAGE_R)
2224 #define __ACCESS_BITS_4V (_PAGE_ACCESSED_4V | _PAGE_READ_4V | _PAGE_R)
2225 
2226 pgprot_t PAGE_KERNEL __read_mostly;
2227 EXPORT_SYMBOL(PAGE_KERNEL);
2228 
2229 pgprot_t PAGE_KERNEL_LOCKED __read_mostly;
2230 pgprot_t PAGE_COPY __read_mostly;
2231 
2232 pgprot_t PAGE_SHARED __read_mostly;
2233 EXPORT_SYMBOL(PAGE_SHARED);
2234 
2235 unsigned long pg_iobits __read_mostly;
2236 
2237 unsigned long _PAGE_IE __read_mostly;
2238 EXPORT_SYMBOL(_PAGE_IE);
2239 
2240 unsigned long _PAGE_E __read_mostly;
2241 EXPORT_SYMBOL(_PAGE_E);
2242 
2243 unsigned long _PAGE_CACHE __read_mostly;
2244 EXPORT_SYMBOL(_PAGE_CACHE);
2245 
2246 #ifdef CONFIG_SPARSEMEM_VMEMMAP
2247 unsigned long vmemmap_table[VMEMMAP_SIZE];
2248 
2249 static long __meminitdata addr_start, addr_end;
2250 static int __meminitdata node_start;
2251 
2252 int __meminit vmemmap_populate(unsigned long vstart, unsigned long vend,
2253 			       int node)
2254 {
2255 	unsigned long phys_start = (vstart - VMEMMAP_BASE);
2256 	unsigned long phys_end = (vend - VMEMMAP_BASE);
2257 	unsigned long addr = phys_start & VMEMMAP_CHUNK_MASK;
2258 	unsigned long end = VMEMMAP_ALIGN(phys_end);
2259 	unsigned long pte_base;
2260 
2261 	pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2262 		    _PAGE_CP_4U | _PAGE_CV_4U |
2263 		    _PAGE_P_4U | _PAGE_W_4U);
2264 	if (tlb_type == hypervisor)
2265 		pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2266 			    _PAGE_CP_4V | _PAGE_CV_4V |
2267 			    _PAGE_P_4V | _PAGE_W_4V);
2268 
2269 	for (; addr < end; addr += VMEMMAP_CHUNK) {
2270 		unsigned long *vmem_pp =
2271 			vmemmap_table + (addr >> VMEMMAP_CHUNK_SHIFT);
2272 		void *block;
2273 
2274 		if (!(*vmem_pp & _PAGE_VALID)) {
2275 			block = vmemmap_alloc_block(1UL << ILOG2_4MB, node);
2276 			if (!block)
2277 				return -ENOMEM;
2278 
2279 			*vmem_pp = pte_base | __pa(block);
2280 
2281 			/* check to see if we have contiguous blocks */
2282 			if (addr_end != addr || node_start != node) {
2283 				if (addr_start)
2284 					printk(KERN_DEBUG " [%lx-%lx] on node %d\n",
2285 					       addr_start, addr_end-1, node_start);
2286 				addr_start = addr;
2287 				node_start = node;
2288 			}
2289 			addr_end = addr + VMEMMAP_CHUNK;
2290 		}
2291 	}
2292 	return 0;
2293 }
2294 
2295 void __meminit vmemmap_populate_print_last(void)
2296 {
2297 	if (addr_start) {
2298 		printk(KERN_DEBUG " [%lx-%lx] on node %d\n",
2299 		       addr_start, addr_end-1, node_start);
2300 		addr_start = 0;
2301 		addr_end = 0;
2302 		node_start = 0;
2303 	}
2304 }
2305 
2306 void vmemmap_free(unsigned long start, unsigned long end)
2307 {
2308 }
2309 
2310 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
2311 
2312 static void prot_init_common(unsigned long page_none,
2313 			     unsigned long page_shared,
2314 			     unsigned long page_copy,
2315 			     unsigned long page_readonly,
2316 			     unsigned long page_exec_bit)
2317 {
2318 	PAGE_COPY = __pgprot(page_copy);
2319 	PAGE_SHARED = __pgprot(page_shared);
2320 
2321 	protection_map[0x0] = __pgprot(page_none);
2322 	protection_map[0x1] = __pgprot(page_readonly & ~page_exec_bit);
2323 	protection_map[0x2] = __pgprot(page_copy & ~page_exec_bit);
2324 	protection_map[0x3] = __pgprot(page_copy & ~page_exec_bit);
2325 	protection_map[0x4] = __pgprot(page_readonly);
2326 	protection_map[0x5] = __pgprot(page_readonly);
2327 	protection_map[0x6] = __pgprot(page_copy);
2328 	protection_map[0x7] = __pgprot(page_copy);
2329 	protection_map[0x8] = __pgprot(page_none);
2330 	protection_map[0x9] = __pgprot(page_readonly & ~page_exec_bit);
2331 	protection_map[0xa] = __pgprot(page_shared & ~page_exec_bit);
2332 	protection_map[0xb] = __pgprot(page_shared & ~page_exec_bit);
2333 	protection_map[0xc] = __pgprot(page_readonly);
2334 	protection_map[0xd] = __pgprot(page_readonly);
2335 	protection_map[0xe] = __pgprot(page_shared);
2336 	protection_map[0xf] = __pgprot(page_shared);
2337 }
2338 
2339 static void __init sun4u_pgprot_init(void)
2340 {
2341 	unsigned long page_none, page_shared, page_copy, page_readonly;
2342 	unsigned long page_exec_bit;
2343 	int i;
2344 
2345 	PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2346 				_PAGE_CACHE_4U | _PAGE_P_4U |
2347 				__ACCESS_BITS_4U | __DIRTY_BITS_4U |
2348 				_PAGE_EXEC_4U);
2349 	PAGE_KERNEL_LOCKED = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2350 				       _PAGE_CACHE_4U | _PAGE_P_4U |
2351 				       __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2352 				       _PAGE_EXEC_4U | _PAGE_L_4U);
2353 
2354 	_PAGE_IE = _PAGE_IE_4U;
2355 	_PAGE_E = _PAGE_E_4U;
2356 	_PAGE_CACHE = _PAGE_CACHE_4U;
2357 
2358 	pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4U | __DIRTY_BITS_4U |
2359 		     __ACCESS_BITS_4U | _PAGE_E_4U);
2360 
2361 #ifdef CONFIG_DEBUG_PAGEALLOC
2362 	kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2363 #else
2364 	kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4U) ^
2365 		PAGE_OFFSET;
2366 #endif
2367 	kern_linear_pte_xor[0] |= (_PAGE_CP_4U | _PAGE_CV_4U |
2368 				   _PAGE_P_4U | _PAGE_W_4U);
2369 
2370 	for (i = 1; i < 4; i++)
2371 		kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2372 
2373 	_PAGE_ALL_SZ_BITS =  (_PAGE_SZ4MB_4U | _PAGE_SZ512K_4U |
2374 			      _PAGE_SZ64K_4U | _PAGE_SZ8K_4U |
2375 			      _PAGE_SZ32MB_4U | _PAGE_SZ256MB_4U);
2376 
2377 
2378 	page_none = _PAGE_PRESENT_4U | _PAGE_ACCESSED_4U | _PAGE_CACHE_4U;
2379 	page_shared = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2380 		       __ACCESS_BITS_4U | _PAGE_WRITE_4U | _PAGE_EXEC_4U);
2381 	page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2382 		       __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2383 	page_readonly   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2384 			   __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2385 
2386 	page_exec_bit = _PAGE_EXEC_4U;
2387 
2388 	prot_init_common(page_none, page_shared, page_copy, page_readonly,
2389 			 page_exec_bit);
2390 }
2391 
2392 static void __init sun4v_pgprot_init(void)
2393 {
2394 	unsigned long page_none, page_shared, page_copy, page_readonly;
2395 	unsigned long page_exec_bit;
2396 	int i;
2397 
2398 	PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4V | _PAGE_VALID |
2399 				_PAGE_CACHE_4V | _PAGE_P_4V |
2400 				__ACCESS_BITS_4V | __DIRTY_BITS_4V |
2401 				_PAGE_EXEC_4V);
2402 	PAGE_KERNEL_LOCKED = PAGE_KERNEL;
2403 
2404 	_PAGE_IE = _PAGE_IE_4V;
2405 	_PAGE_E = _PAGE_E_4V;
2406 	_PAGE_CACHE = _PAGE_CACHE_4V;
2407 
2408 #ifdef CONFIG_DEBUG_PAGEALLOC
2409 	kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2410 #else
2411 	kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4V) ^
2412 		PAGE_OFFSET;
2413 #endif
2414 	kern_linear_pte_xor[0] |= (_PAGE_CP_4V | _PAGE_CV_4V |
2415 				   _PAGE_P_4V | _PAGE_W_4V);
2416 
2417 	for (i = 1; i < 4; i++)
2418 		kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2419 
2420 	pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4V | __DIRTY_BITS_4V |
2421 		     __ACCESS_BITS_4V | _PAGE_E_4V);
2422 
2423 	_PAGE_ALL_SZ_BITS = (_PAGE_SZ16GB_4V | _PAGE_SZ2GB_4V |
2424 			     _PAGE_SZ256MB_4V | _PAGE_SZ32MB_4V |
2425 			     _PAGE_SZ4MB_4V | _PAGE_SZ512K_4V |
2426 			     _PAGE_SZ64K_4V | _PAGE_SZ8K_4V);
2427 
2428 	page_none = _PAGE_PRESENT_4V | _PAGE_ACCESSED_4V | _PAGE_CACHE_4V;
2429 	page_shared = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2430 		       __ACCESS_BITS_4V | _PAGE_WRITE_4V | _PAGE_EXEC_4V);
2431 	page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2432 		       __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2433 	page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2434 			 __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2435 
2436 	page_exec_bit = _PAGE_EXEC_4V;
2437 
2438 	prot_init_common(page_none, page_shared, page_copy, page_readonly,
2439 			 page_exec_bit);
2440 }
2441 
2442 unsigned long pte_sz_bits(unsigned long sz)
2443 {
2444 	if (tlb_type == hypervisor) {
2445 		switch (sz) {
2446 		case 8 * 1024:
2447 		default:
2448 			return _PAGE_SZ8K_4V;
2449 		case 64 * 1024:
2450 			return _PAGE_SZ64K_4V;
2451 		case 512 * 1024:
2452 			return _PAGE_SZ512K_4V;
2453 		case 4 * 1024 * 1024:
2454 			return _PAGE_SZ4MB_4V;
2455 		}
2456 	} else {
2457 		switch (sz) {
2458 		case 8 * 1024:
2459 		default:
2460 			return _PAGE_SZ8K_4U;
2461 		case 64 * 1024:
2462 			return _PAGE_SZ64K_4U;
2463 		case 512 * 1024:
2464 			return _PAGE_SZ512K_4U;
2465 		case 4 * 1024 * 1024:
2466 			return _PAGE_SZ4MB_4U;
2467 		}
2468 	}
2469 }
2470 
2471 pte_t mk_pte_io(unsigned long page, pgprot_t prot, int space, unsigned long page_size)
2472 {
2473 	pte_t pte;
2474 
2475 	pte_val(pte)  = page | pgprot_val(pgprot_noncached(prot));
2476 	pte_val(pte) |= (((unsigned long)space) << 32);
2477 	pte_val(pte) |= pte_sz_bits(page_size);
2478 
2479 	return pte;
2480 }
2481 
2482 static unsigned long kern_large_tte(unsigned long paddr)
2483 {
2484 	unsigned long val;
2485 
2486 	val = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2487 	       _PAGE_CP_4U | _PAGE_CV_4U | _PAGE_P_4U |
2488 	       _PAGE_EXEC_4U | _PAGE_L_4U | _PAGE_W_4U);
2489 	if (tlb_type == hypervisor)
2490 		val = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2491 		       _PAGE_CP_4V | _PAGE_CV_4V | _PAGE_P_4V |
2492 		       _PAGE_EXEC_4V | _PAGE_W_4V);
2493 
2494 	return val | paddr;
2495 }
2496 
2497 /* If not locked, zap it. */
2498 void __flush_tlb_all(void)
2499 {
2500 	unsigned long pstate;
2501 	int i;
2502 
2503 	__asm__ __volatile__("flushw\n\t"
2504 			     "rdpr	%%pstate, %0\n\t"
2505 			     "wrpr	%0, %1, %%pstate"
2506 			     : "=r" (pstate)
2507 			     : "i" (PSTATE_IE));
2508 	if (tlb_type == hypervisor) {
2509 		sun4v_mmu_demap_all();
2510 	} else if (tlb_type == spitfire) {
2511 		for (i = 0; i < 64; i++) {
2512 			/* Spitfire Errata #32 workaround */
2513 			/* NOTE: Always runs on spitfire, so no
2514 			 *       cheetah+ page size encodings.
2515 			 */
2516 			__asm__ __volatile__("stxa	%0, [%1] %2\n\t"
2517 					     "flush	%%g6"
2518 					     : /* No outputs */
2519 					     : "r" (0),
2520 					     "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2521 
2522 			if (!(spitfire_get_dtlb_data(i) & _PAGE_L_4U)) {
2523 				__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2524 						     "membar #Sync"
2525 						     : /* no outputs */
2526 						     : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
2527 				spitfire_put_dtlb_data(i, 0x0UL);
2528 			}
2529 
2530 			/* Spitfire Errata #32 workaround */
2531 			/* NOTE: Always runs on spitfire, so no
2532 			 *       cheetah+ page size encodings.
2533 			 */
2534 			__asm__ __volatile__("stxa	%0, [%1] %2\n\t"
2535 					     "flush	%%g6"
2536 					     : /* No outputs */
2537 					     : "r" (0),
2538 					     "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2539 
2540 			if (!(spitfire_get_itlb_data(i) & _PAGE_L_4U)) {
2541 				__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2542 						     "membar #Sync"
2543 						     : /* no outputs */
2544 						     : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
2545 				spitfire_put_itlb_data(i, 0x0UL);
2546 			}
2547 		}
2548 	} else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
2549 		cheetah_flush_dtlb_all();
2550 		cheetah_flush_itlb_all();
2551 	}
2552 	__asm__ __volatile__("wrpr	%0, 0, %%pstate"
2553 			     : : "r" (pstate));
2554 }
2555 
2556 pte_t *pte_alloc_one_kernel(struct mm_struct *mm,
2557 			    unsigned long address)
2558 {
2559 	struct page *page = alloc_page(GFP_KERNEL | __GFP_NOTRACK |
2560 				       __GFP_REPEAT | __GFP_ZERO);
2561 	pte_t *pte = NULL;
2562 
2563 	if (page)
2564 		pte = (pte_t *) page_address(page);
2565 
2566 	return pte;
2567 }
2568 
2569 pgtable_t pte_alloc_one(struct mm_struct *mm,
2570 			unsigned long address)
2571 {
2572 	struct page *page = alloc_page(GFP_KERNEL | __GFP_NOTRACK |
2573 				       __GFP_REPEAT | __GFP_ZERO);
2574 	if (!page)
2575 		return NULL;
2576 	if (!pgtable_page_ctor(page)) {
2577 		free_hot_cold_page(page, 0);
2578 		return NULL;
2579 	}
2580 	return (pte_t *) page_address(page);
2581 }
2582 
2583 void pte_free_kernel(struct mm_struct *mm, pte_t *pte)
2584 {
2585 	free_page((unsigned long)pte);
2586 }
2587 
2588 static void __pte_free(pgtable_t pte)
2589 {
2590 	struct page *page = virt_to_page(pte);
2591 
2592 	pgtable_page_dtor(page);
2593 	__free_page(page);
2594 }
2595 
2596 void pte_free(struct mm_struct *mm, pgtable_t pte)
2597 {
2598 	__pte_free(pte);
2599 }
2600 
2601 void pgtable_free(void *table, bool is_page)
2602 {
2603 	if (is_page)
2604 		__pte_free(table);
2605 	else
2606 		kmem_cache_free(pgtable_cache, table);
2607 }
2608 
2609 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2610 void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
2611 			  pmd_t *pmd)
2612 {
2613 	unsigned long pte, flags;
2614 	struct mm_struct *mm;
2615 	pmd_t entry = *pmd;
2616 
2617 	if (!pmd_large(entry) || !pmd_young(entry))
2618 		return;
2619 
2620 	pte = pmd_val(entry);
2621 
2622 	/* We are fabricating 8MB pages using 4MB real hw pages.  */
2623 	pte |= (addr & (1UL << REAL_HPAGE_SHIFT));
2624 
2625 	mm = vma->vm_mm;
2626 
2627 	spin_lock_irqsave(&mm->context.lock, flags);
2628 
2629 	if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL)
2630 		__update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
2631 					addr, pte);
2632 
2633 	spin_unlock_irqrestore(&mm->context.lock, flags);
2634 }
2635 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2636 
2637 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
2638 static void context_reload(void *__data)
2639 {
2640 	struct mm_struct *mm = __data;
2641 
2642 	if (mm == current->mm)
2643 		load_secondary_context(mm);
2644 }
2645 
2646 void hugetlb_setup(struct pt_regs *regs)
2647 {
2648 	struct mm_struct *mm = current->mm;
2649 	struct tsb_config *tp;
2650 
2651 	if (in_atomic() || !mm) {
2652 		const struct exception_table_entry *entry;
2653 
2654 		entry = search_exception_tables(regs->tpc);
2655 		if (entry) {
2656 			regs->tpc = entry->fixup;
2657 			regs->tnpc = regs->tpc + 4;
2658 			return;
2659 		}
2660 		pr_alert("Unexpected HugeTLB setup in atomic context.\n");
2661 		die_if_kernel("HugeTSB in atomic", regs);
2662 	}
2663 
2664 	tp = &mm->context.tsb_block[MM_TSB_HUGE];
2665 	if (likely(tp->tsb == NULL))
2666 		tsb_grow(mm, MM_TSB_HUGE, 0);
2667 
2668 	tsb_context_switch(mm);
2669 	smp_tsb_sync(mm);
2670 
2671 	/* On UltraSPARC-III+ and later, configure the second half of
2672 	 * the Data-TLB for huge pages.
2673 	 */
2674 	if (tlb_type == cheetah_plus) {
2675 		unsigned long ctx;
2676 
2677 		spin_lock(&ctx_alloc_lock);
2678 		ctx = mm->context.sparc64_ctx_val;
2679 		ctx &= ~CTX_PGSZ_MASK;
2680 		ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT;
2681 		ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT;
2682 
2683 		if (ctx != mm->context.sparc64_ctx_val) {
2684 			/* When changing the page size fields, we
2685 			 * must perform a context flush so that no
2686 			 * stale entries match.  This flush must
2687 			 * occur with the original context register
2688 			 * settings.
2689 			 */
2690 			do_flush_tlb_mm(mm);
2691 
2692 			/* Reload the context register of all processors
2693 			 * also executing in this address space.
2694 			 */
2695 			mm->context.sparc64_ctx_val = ctx;
2696 			on_each_cpu(context_reload, mm, 0);
2697 		}
2698 		spin_unlock(&ctx_alloc_lock);
2699 	}
2700 }
2701 #endif
2702