xref: /openbmc/linux/arch/sparc/mm/init_64.c (revision 79f08d9e)
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/irq.h>
51 
52 #include "init_64.h"
53 
54 unsigned long kern_linear_pte_xor[4] __read_mostly;
55 
56 /* A bitmap, two bits for every 256MB of physical memory.  These two
57  * bits determine what page size we use for kernel linear
58  * translations.  They form an index into kern_linear_pte_xor[].  The
59  * value in the indexed slot is XOR'd with the TLB miss virtual
60  * address to form the resulting TTE.  The mapping is:
61  *
62  *	0	==>	4MB
63  *	1	==>	256MB
64  *	2	==>	2GB
65  *	3	==>	16GB
66  *
67  * All sun4v chips support 256MB pages.  Only SPARC-T4 and later
68  * support 2GB pages, and hopefully future cpus will support the 16GB
69  * pages as well.  For slots 2 and 3, we encode a 256MB TTE xor there
70  * if these larger page sizes are not supported by the cpu.
71  *
72  * It would be nice to determine this from the machine description
73  * 'cpu' properties, but we need to have this table setup before the
74  * MDESC is initialized.
75  */
76 unsigned long kpte_linear_bitmap[KPTE_BITMAP_BYTES / sizeof(unsigned long)];
77 
78 #ifndef CONFIG_DEBUG_PAGEALLOC
79 /* A special kernel TSB for 4MB, 256MB, 2GB and 16GB linear mappings.
80  * Space is allocated for this right after the trap table in
81  * arch/sparc64/kernel/head.S
82  */
83 extern struct tsb swapper_4m_tsb[KERNEL_TSB4M_NENTRIES];
84 #endif
85 
86 static unsigned long cpu_pgsz_mask;
87 
88 #define MAX_BANKS	32
89 
90 static struct linux_prom64_registers pavail[MAX_BANKS];
91 static int pavail_ents;
92 
93 static int cmp_p64(const void *a, const void *b)
94 {
95 	const struct linux_prom64_registers *x = a, *y = b;
96 
97 	if (x->phys_addr > y->phys_addr)
98 		return 1;
99 	if (x->phys_addr < y->phys_addr)
100 		return -1;
101 	return 0;
102 }
103 
104 static void __init read_obp_memory(const char *property,
105 				   struct linux_prom64_registers *regs,
106 				   int *num_ents)
107 {
108 	phandle node = prom_finddevice("/memory");
109 	int prop_size = prom_getproplen(node, property);
110 	int ents, ret, i;
111 
112 	ents = prop_size / sizeof(struct linux_prom64_registers);
113 	if (ents > MAX_BANKS) {
114 		prom_printf("The machine has more %s property entries than "
115 			    "this kernel can support (%d).\n",
116 			    property, MAX_BANKS);
117 		prom_halt();
118 	}
119 
120 	ret = prom_getproperty(node, property, (char *) regs, prop_size);
121 	if (ret == -1) {
122 		prom_printf("Couldn't get %s property from /memory.\n",
123 				property);
124 		prom_halt();
125 	}
126 
127 	/* Sanitize what we got from the firmware, by page aligning
128 	 * everything.
129 	 */
130 	for (i = 0; i < ents; i++) {
131 		unsigned long base, size;
132 
133 		base = regs[i].phys_addr;
134 		size = regs[i].reg_size;
135 
136 		size &= PAGE_MASK;
137 		if (base & ~PAGE_MASK) {
138 			unsigned long new_base = PAGE_ALIGN(base);
139 
140 			size -= new_base - base;
141 			if ((long) size < 0L)
142 				size = 0UL;
143 			base = new_base;
144 		}
145 		if (size == 0UL) {
146 			/* If it is empty, simply get rid of it.
147 			 * This simplifies the logic of the other
148 			 * functions that process these arrays.
149 			 */
150 			memmove(&regs[i], &regs[i + 1],
151 				(ents - i - 1) * sizeof(regs[0]));
152 			i--;
153 			ents--;
154 			continue;
155 		}
156 		regs[i].phys_addr = base;
157 		regs[i].reg_size = size;
158 	}
159 
160 	*num_ents = ents;
161 
162 	sort(regs, ents, sizeof(struct linux_prom64_registers),
163 	     cmp_p64, NULL);
164 }
165 
166 unsigned long sparc64_valid_addr_bitmap[VALID_ADDR_BITMAP_BYTES /
167 					sizeof(unsigned long)];
168 EXPORT_SYMBOL(sparc64_valid_addr_bitmap);
169 
170 /* Kernel physical address base and size in bytes.  */
171 unsigned long kern_base __read_mostly;
172 unsigned long kern_size __read_mostly;
173 
174 /* Initial ramdisk setup */
175 extern unsigned long sparc_ramdisk_image64;
176 extern unsigned int sparc_ramdisk_image;
177 extern unsigned int sparc_ramdisk_size;
178 
179 struct page *mem_map_zero __read_mostly;
180 EXPORT_SYMBOL(mem_map_zero);
181 
182 unsigned int sparc64_highest_unlocked_tlb_ent __read_mostly;
183 
184 unsigned long sparc64_kern_pri_context __read_mostly;
185 unsigned long sparc64_kern_pri_nuc_bits __read_mostly;
186 unsigned long sparc64_kern_sec_context __read_mostly;
187 
188 int num_kernel_image_mappings;
189 
190 #ifdef CONFIG_DEBUG_DCFLUSH
191 atomic_t dcpage_flushes = ATOMIC_INIT(0);
192 #ifdef CONFIG_SMP
193 atomic_t dcpage_flushes_xcall = ATOMIC_INIT(0);
194 #endif
195 #endif
196 
197 inline void flush_dcache_page_impl(struct page *page)
198 {
199 	BUG_ON(tlb_type == hypervisor);
200 #ifdef CONFIG_DEBUG_DCFLUSH
201 	atomic_inc(&dcpage_flushes);
202 #endif
203 
204 #ifdef DCACHE_ALIASING_POSSIBLE
205 	__flush_dcache_page(page_address(page),
206 			    ((tlb_type == spitfire) &&
207 			     page_mapping(page) != NULL));
208 #else
209 	if (page_mapping(page) != NULL &&
210 	    tlb_type == spitfire)
211 		__flush_icache_page(__pa(page_address(page)));
212 #endif
213 }
214 
215 #define PG_dcache_dirty		PG_arch_1
216 #define PG_dcache_cpu_shift	32UL
217 #define PG_dcache_cpu_mask	\
218 	((1UL<<ilog2(roundup_pow_of_two(NR_CPUS)))-1UL)
219 
220 #define dcache_dirty_cpu(page) \
221 	(((page)->flags >> PG_dcache_cpu_shift) & PG_dcache_cpu_mask)
222 
223 static inline void set_dcache_dirty(struct page *page, int this_cpu)
224 {
225 	unsigned long mask = this_cpu;
226 	unsigned long non_cpu_bits;
227 
228 	non_cpu_bits = ~(PG_dcache_cpu_mask << PG_dcache_cpu_shift);
229 	mask = (mask << PG_dcache_cpu_shift) | (1UL << PG_dcache_dirty);
230 
231 	__asm__ __volatile__("1:\n\t"
232 			     "ldx	[%2], %%g7\n\t"
233 			     "and	%%g7, %1, %%g1\n\t"
234 			     "or	%%g1, %0, %%g1\n\t"
235 			     "casx	[%2], %%g7, %%g1\n\t"
236 			     "cmp	%%g7, %%g1\n\t"
237 			     "bne,pn	%%xcc, 1b\n\t"
238 			     " nop"
239 			     : /* no outputs */
240 			     : "r" (mask), "r" (non_cpu_bits), "r" (&page->flags)
241 			     : "g1", "g7");
242 }
243 
244 static inline void clear_dcache_dirty_cpu(struct page *page, unsigned long cpu)
245 {
246 	unsigned long mask = (1UL << PG_dcache_dirty);
247 
248 	__asm__ __volatile__("! test_and_clear_dcache_dirty\n"
249 			     "1:\n\t"
250 			     "ldx	[%2], %%g7\n\t"
251 			     "srlx	%%g7, %4, %%g1\n\t"
252 			     "and	%%g1, %3, %%g1\n\t"
253 			     "cmp	%%g1, %0\n\t"
254 			     "bne,pn	%%icc, 2f\n\t"
255 			     " andn	%%g7, %1, %%g1\n\t"
256 			     "casx	[%2], %%g7, %%g1\n\t"
257 			     "cmp	%%g7, %%g1\n\t"
258 			     "bne,pn	%%xcc, 1b\n\t"
259 			     " nop\n"
260 			     "2:"
261 			     : /* no outputs */
262 			     : "r" (cpu), "r" (mask), "r" (&page->flags),
263 			       "i" (PG_dcache_cpu_mask),
264 			       "i" (PG_dcache_cpu_shift)
265 			     : "g1", "g7");
266 }
267 
268 static inline void tsb_insert(struct tsb *ent, unsigned long tag, unsigned long pte)
269 {
270 	unsigned long tsb_addr = (unsigned long) ent;
271 
272 	if (tlb_type == cheetah_plus || tlb_type == hypervisor)
273 		tsb_addr = __pa(tsb_addr);
274 
275 	__tsb_insert(tsb_addr, tag, pte);
276 }
277 
278 unsigned long _PAGE_ALL_SZ_BITS __read_mostly;
279 
280 static void flush_dcache(unsigned long pfn)
281 {
282 	struct page *page;
283 
284 	page = pfn_to_page(pfn);
285 	if (page) {
286 		unsigned long pg_flags;
287 
288 		pg_flags = page->flags;
289 		if (pg_flags & (1UL << PG_dcache_dirty)) {
290 			int cpu = ((pg_flags >> PG_dcache_cpu_shift) &
291 				   PG_dcache_cpu_mask);
292 			int this_cpu = get_cpu();
293 
294 			/* This is just to optimize away some function calls
295 			 * in the SMP case.
296 			 */
297 			if (cpu == this_cpu)
298 				flush_dcache_page_impl(page);
299 			else
300 				smp_flush_dcache_page_impl(page, cpu);
301 
302 			clear_dcache_dirty_cpu(page, cpu);
303 
304 			put_cpu();
305 		}
306 	}
307 }
308 
309 /* mm->context.lock must be held */
310 static void __update_mmu_tsb_insert(struct mm_struct *mm, unsigned long tsb_index,
311 				    unsigned long tsb_hash_shift, unsigned long address,
312 				    unsigned long tte)
313 {
314 	struct tsb *tsb = mm->context.tsb_block[tsb_index].tsb;
315 	unsigned long tag;
316 
317 	if (unlikely(!tsb))
318 		return;
319 
320 	tsb += ((address >> tsb_hash_shift) &
321 		(mm->context.tsb_block[tsb_index].tsb_nentries - 1UL));
322 	tag = (address >> 22UL);
323 	tsb_insert(tsb, tag, tte);
324 }
325 
326 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
327 static inline bool is_hugetlb_pte(pte_t pte)
328 {
329 	if ((tlb_type == hypervisor &&
330 	     (pte_val(pte) & _PAGE_SZALL_4V) == _PAGE_SZHUGE_4V) ||
331 	    (tlb_type != hypervisor &&
332 	     (pte_val(pte) & _PAGE_SZALL_4U) == _PAGE_SZHUGE_4U))
333 		return true;
334 	return false;
335 }
336 #endif
337 
338 void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
339 {
340 	struct mm_struct *mm;
341 	unsigned long flags;
342 	pte_t pte = *ptep;
343 
344 	if (tlb_type != hypervisor) {
345 		unsigned long pfn = pte_pfn(pte);
346 
347 		if (pfn_valid(pfn))
348 			flush_dcache(pfn);
349 	}
350 
351 	mm = vma->vm_mm;
352 
353 	spin_lock_irqsave(&mm->context.lock, flags);
354 
355 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
356 	if (mm->context.huge_pte_count && is_hugetlb_pte(pte))
357 		__update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
358 					address, pte_val(pte));
359 	else
360 #endif
361 		__update_mmu_tsb_insert(mm, MM_TSB_BASE, PAGE_SHIFT,
362 					address, pte_val(pte));
363 
364 	spin_unlock_irqrestore(&mm->context.lock, flags);
365 }
366 
367 void flush_dcache_page(struct page *page)
368 {
369 	struct address_space *mapping;
370 	int this_cpu;
371 
372 	if (tlb_type == hypervisor)
373 		return;
374 
375 	/* Do not bother with the expensive D-cache flush if it
376 	 * is merely the zero page.  The 'bigcore' testcase in GDB
377 	 * causes this case to run millions of times.
378 	 */
379 	if (page == ZERO_PAGE(0))
380 		return;
381 
382 	this_cpu = get_cpu();
383 
384 	mapping = page_mapping(page);
385 	if (mapping && !mapping_mapped(mapping)) {
386 		int dirty = test_bit(PG_dcache_dirty, &page->flags);
387 		if (dirty) {
388 			int dirty_cpu = dcache_dirty_cpu(page);
389 
390 			if (dirty_cpu == this_cpu)
391 				goto out;
392 			smp_flush_dcache_page_impl(page, dirty_cpu);
393 		}
394 		set_dcache_dirty(page, this_cpu);
395 	} else {
396 		/* We could delay the flush for the !page_mapping
397 		 * case too.  But that case is for exec env/arg
398 		 * pages and those are %99 certainly going to get
399 		 * faulted into the tlb (and thus flushed) anyways.
400 		 */
401 		flush_dcache_page_impl(page);
402 	}
403 
404 out:
405 	put_cpu();
406 }
407 EXPORT_SYMBOL(flush_dcache_page);
408 
409 void __kprobes flush_icache_range(unsigned long start, unsigned long end)
410 {
411 	/* Cheetah and Hypervisor platform cpus have coherent I-cache. */
412 	if (tlb_type == spitfire) {
413 		unsigned long kaddr;
414 
415 		/* This code only runs on Spitfire cpus so this is
416 		 * why we can assume _PAGE_PADDR_4U.
417 		 */
418 		for (kaddr = start; kaddr < end; kaddr += PAGE_SIZE) {
419 			unsigned long paddr, mask = _PAGE_PADDR_4U;
420 
421 			if (kaddr >= PAGE_OFFSET)
422 				paddr = kaddr & mask;
423 			else {
424 				pgd_t *pgdp = pgd_offset_k(kaddr);
425 				pud_t *pudp = pud_offset(pgdp, kaddr);
426 				pmd_t *pmdp = pmd_offset(pudp, kaddr);
427 				pte_t *ptep = pte_offset_kernel(pmdp, kaddr);
428 
429 				paddr = pte_val(*ptep) & mask;
430 			}
431 			__flush_icache_page(paddr);
432 		}
433 	}
434 }
435 EXPORT_SYMBOL(flush_icache_range);
436 
437 void mmu_info(struct seq_file *m)
438 {
439 	static const char *pgsz_strings[] = {
440 		"8K", "64K", "512K", "4MB", "32MB",
441 		"256MB", "2GB", "16GB",
442 	};
443 	int i, printed;
444 
445 	if (tlb_type == cheetah)
446 		seq_printf(m, "MMU Type\t: Cheetah\n");
447 	else if (tlb_type == cheetah_plus)
448 		seq_printf(m, "MMU Type\t: Cheetah+\n");
449 	else if (tlb_type == spitfire)
450 		seq_printf(m, "MMU Type\t: Spitfire\n");
451 	else if (tlb_type == hypervisor)
452 		seq_printf(m, "MMU Type\t: Hypervisor (sun4v)\n");
453 	else
454 		seq_printf(m, "MMU Type\t: ???\n");
455 
456 	seq_printf(m, "MMU PGSZs\t: ");
457 	printed = 0;
458 	for (i = 0; i < ARRAY_SIZE(pgsz_strings); i++) {
459 		if (cpu_pgsz_mask & (1UL << i)) {
460 			seq_printf(m, "%s%s",
461 				   printed ? "," : "", pgsz_strings[i]);
462 			printed++;
463 		}
464 	}
465 	seq_putc(m, '\n');
466 
467 #ifdef CONFIG_DEBUG_DCFLUSH
468 	seq_printf(m, "DCPageFlushes\t: %d\n",
469 		   atomic_read(&dcpage_flushes));
470 #ifdef CONFIG_SMP
471 	seq_printf(m, "DCPageFlushesXC\t: %d\n",
472 		   atomic_read(&dcpage_flushes_xcall));
473 #endif /* CONFIG_SMP */
474 #endif /* CONFIG_DEBUG_DCFLUSH */
475 }
476 
477 struct linux_prom_translation prom_trans[512] __read_mostly;
478 unsigned int prom_trans_ents __read_mostly;
479 
480 unsigned long kern_locked_tte_data;
481 
482 /* The obp translations are saved based on 8k pagesize, since obp can
483  * use a mixture of pagesizes. Misses to the LOW_OBP_ADDRESS ->
484  * HI_OBP_ADDRESS range are handled in ktlb.S.
485  */
486 static inline int in_obp_range(unsigned long vaddr)
487 {
488 	return (vaddr >= LOW_OBP_ADDRESS &&
489 		vaddr < HI_OBP_ADDRESS);
490 }
491 
492 static int cmp_ptrans(const void *a, const void *b)
493 {
494 	const struct linux_prom_translation *x = a, *y = b;
495 
496 	if (x->virt > y->virt)
497 		return 1;
498 	if (x->virt < y->virt)
499 		return -1;
500 	return 0;
501 }
502 
503 /* Read OBP translations property into 'prom_trans[]'.  */
504 static void __init read_obp_translations(void)
505 {
506 	int n, node, ents, first, last, i;
507 
508 	node = prom_finddevice("/virtual-memory");
509 	n = prom_getproplen(node, "translations");
510 	if (unlikely(n == 0 || n == -1)) {
511 		prom_printf("prom_mappings: Couldn't get size.\n");
512 		prom_halt();
513 	}
514 	if (unlikely(n > sizeof(prom_trans))) {
515 		prom_printf("prom_mappings: Size %d is too big.\n", n);
516 		prom_halt();
517 	}
518 
519 	if ((n = prom_getproperty(node, "translations",
520 				  (char *)&prom_trans[0],
521 				  sizeof(prom_trans))) == -1) {
522 		prom_printf("prom_mappings: Couldn't get property.\n");
523 		prom_halt();
524 	}
525 
526 	n = n / sizeof(struct linux_prom_translation);
527 
528 	ents = n;
529 
530 	sort(prom_trans, ents, sizeof(struct linux_prom_translation),
531 	     cmp_ptrans, NULL);
532 
533 	/* Now kick out all the non-OBP entries.  */
534 	for (i = 0; i < ents; i++) {
535 		if (in_obp_range(prom_trans[i].virt))
536 			break;
537 	}
538 	first = i;
539 	for (; i < ents; i++) {
540 		if (!in_obp_range(prom_trans[i].virt))
541 			break;
542 	}
543 	last = i;
544 
545 	for (i = 0; i < (last - first); i++) {
546 		struct linux_prom_translation *src = &prom_trans[i + first];
547 		struct linux_prom_translation *dest = &prom_trans[i];
548 
549 		*dest = *src;
550 	}
551 	for (; i < ents; i++) {
552 		struct linux_prom_translation *dest = &prom_trans[i];
553 		dest->virt = dest->size = dest->data = 0x0UL;
554 	}
555 
556 	prom_trans_ents = last - first;
557 
558 	if (tlb_type == spitfire) {
559 		/* Clear diag TTE bits. */
560 		for (i = 0; i < prom_trans_ents; i++)
561 			prom_trans[i].data &= ~0x0003fe0000000000UL;
562 	}
563 
564 	/* Force execute bit on.  */
565 	for (i = 0; i < prom_trans_ents; i++)
566 		prom_trans[i].data |= (tlb_type == hypervisor ?
567 				       _PAGE_EXEC_4V : _PAGE_EXEC_4U);
568 }
569 
570 static void __init hypervisor_tlb_lock(unsigned long vaddr,
571 				       unsigned long pte,
572 				       unsigned long mmu)
573 {
574 	unsigned long ret = sun4v_mmu_map_perm_addr(vaddr, 0, pte, mmu);
575 
576 	if (ret != 0) {
577 		prom_printf("hypervisor_tlb_lock[%lx:%x:%lx:%lx]: "
578 			    "errors with %lx\n", vaddr, 0, pte, mmu, ret);
579 		prom_halt();
580 	}
581 }
582 
583 static unsigned long kern_large_tte(unsigned long paddr);
584 
585 static void __init remap_kernel(void)
586 {
587 	unsigned long phys_page, tte_vaddr, tte_data;
588 	int i, tlb_ent = sparc64_highest_locked_tlbent();
589 
590 	tte_vaddr = (unsigned long) KERNBASE;
591 	phys_page = (prom_boot_mapping_phys_low >> 22UL) << 22UL;
592 	tte_data = kern_large_tte(phys_page);
593 
594 	kern_locked_tte_data = tte_data;
595 
596 	/* Now lock us into the TLBs via Hypervisor or OBP. */
597 	if (tlb_type == hypervisor) {
598 		for (i = 0; i < num_kernel_image_mappings; i++) {
599 			hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_DMMU);
600 			hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_IMMU);
601 			tte_vaddr += 0x400000;
602 			tte_data += 0x400000;
603 		}
604 	} else {
605 		for (i = 0; i < num_kernel_image_mappings; i++) {
606 			prom_dtlb_load(tlb_ent - i, tte_data, tte_vaddr);
607 			prom_itlb_load(tlb_ent - i, tte_data, tte_vaddr);
608 			tte_vaddr += 0x400000;
609 			tte_data += 0x400000;
610 		}
611 		sparc64_highest_unlocked_tlb_ent = tlb_ent - i;
612 	}
613 	if (tlb_type == cheetah_plus) {
614 		sparc64_kern_pri_context = (CTX_CHEETAH_PLUS_CTX0 |
615 					    CTX_CHEETAH_PLUS_NUC);
616 		sparc64_kern_pri_nuc_bits = CTX_CHEETAH_PLUS_NUC;
617 		sparc64_kern_sec_context = CTX_CHEETAH_PLUS_CTX0;
618 	}
619 }
620 
621 
622 static void __init inherit_prom_mappings(void)
623 {
624 	/* Now fixup OBP's idea about where we really are mapped. */
625 	printk("Remapping the kernel... ");
626 	remap_kernel();
627 	printk("done.\n");
628 }
629 
630 void prom_world(int enter)
631 {
632 	if (!enter)
633 		set_fs(get_fs());
634 
635 	__asm__ __volatile__("flushw");
636 }
637 
638 void __flush_dcache_range(unsigned long start, unsigned long end)
639 {
640 	unsigned long va;
641 
642 	if (tlb_type == spitfire) {
643 		int n = 0;
644 
645 		for (va = start; va < end; va += 32) {
646 			spitfire_put_dcache_tag(va & 0x3fe0, 0x0);
647 			if (++n >= 512)
648 				break;
649 		}
650 	} else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
651 		start = __pa(start);
652 		end = __pa(end);
653 		for (va = start; va < end; va += 32)
654 			__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
655 					     "membar #Sync"
656 					     : /* no outputs */
657 					     : "r" (va),
658 					       "i" (ASI_DCACHE_INVALIDATE));
659 	}
660 }
661 EXPORT_SYMBOL(__flush_dcache_range);
662 
663 /* get_new_mmu_context() uses "cache + 1".  */
664 DEFINE_SPINLOCK(ctx_alloc_lock);
665 unsigned long tlb_context_cache = CTX_FIRST_VERSION - 1;
666 #define MAX_CTX_NR	(1UL << CTX_NR_BITS)
667 #define CTX_BMAP_SLOTS	BITS_TO_LONGS(MAX_CTX_NR)
668 DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR);
669 
670 /* Caller does TLB context flushing on local CPU if necessary.
671  * The caller also ensures that CTX_VALID(mm->context) is false.
672  *
673  * We must be careful about boundary cases so that we never
674  * let the user have CTX 0 (nucleus) or we ever use a CTX
675  * version of zero (and thus NO_CONTEXT would not be caught
676  * by version mis-match tests in mmu_context.h).
677  *
678  * Always invoked with interrupts disabled.
679  */
680 void get_new_mmu_context(struct mm_struct *mm)
681 {
682 	unsigned long ctx, new_ctx;
683 	unsigned long orig_pgsz_bits;
684 	int new_version;
685 
686 	spin_lock(&ctx_alloc_lock);
687 	orig_pgsz_bits = (mm->context.sparc64_ctx_val & CTX_PGSZ_MASK);
688 	ctx = (tlb_context_cache + 1) & CTX_NR_MASK;
689 	new_ctx = find_next_zero_bit(mmu_context_bmap, 1 << CTX_NR_BITS, ctx);
690 	new_version = 0;
691 	if (new_ctx >= (1 << CTX_NR_BITS)) {
692 		new_ctx = find_next_zero_bit(mmu_context_bmap, ctx, 1);
693 		if (new_ctx >= ctx) {
694 			int i;
695 			new_ctx = (tlb_context_cache & CTX_VERSION_MASK) +
696 				CTX_FIRST_VERSION;
697 			if (new_ctx == 1)
698 				new_ctx = CTX_FIRST_VERSION;
699 
700 			/* Don't call memset, for 16 entries that's just
701 			 * plain silly...
702 			 */
703 			mmu_context_bmap[0] = 3;
704 			mmu_context_bmap[1] = 0;
705 			mmu_context_bmap[2] = 0;
706 			mmu_context_bmap[3] = 0;
707 			for (i = 4; i < CTX_BMAP_SLOTS; i += 4) {
708 				mmu_context_bmap[i + 0] = 0;
709 				mmu_context_bmap[i + 1] = 0;
710 				mmu_context_bmap[i + 2] = 0;
711 				mmu_context_bmap[i + 3] = 0;
712 			}
713 			new_version = 1;
714 			goto out;
715 		}
716 	}
717 	mmu_context_bmap[new_ctx>>6] |= (1UL << (new_ctx & 63));
718 	new_ctx |= (tlb_context_cache & CTX_VERSION_MASK);
719 out:
720 	tlb_context_cache = new_ctx;
721 	mm->context.sparc64_ctx_val = new_ctx | orig_pgsz_bits;
722 	spin_unlock(&ctx_alloc_lock);
723 
724 	if (unlikely(new_version))
725 		smp_new_mmu_context_version();
726 }
727 
728 static int numa_enabled = 1;
729 static int numa_debug;
730 
731 static int __init early_numa(char *p)
732 {
733 	if (!p)
734 		return 0;
735 
736 	if (strstr(p, "off"))
737 		numa_enabled = 0;
738 
739 	if (strstr(p, "debug"))
740 		numa_debug = 1;
741 
742 	return 0;
743 }
744 early_param("numa", early_numa);
745 
746 #define numadbg(f, a...) \
747 do {	if (numa_debug) \
748 		printk(KERN_INFO f, ## a); \
749 } while (0)
750 
751 static void __init find_ramdisk(unsigned long phys_base)
752 {
753 #ifdef CONFIG_BLK_DEV_INITRD
754 	if (sparc_ramdisk_image || sparc_ramdisk_image64) {
755 		unsigned long ramdisk_image;
756 
757 		/* Older versions of the bootloader only supported a
758 		 * 32-bit physical address for the ramdisk image
759 		 * location, stored at sparc_ramdisk_image.  Newer
760 		 * SILO versions set sparc_ramdisk_image to zero and
761 		 * provide a full 64-bit physical address at
762 		 * sparc_ramdisk_image64.
763 		 */
764 		ramdisk_image = sparc_ramdisk_image;
765 		if (!ramdisk_image)
766 			ramdisk_image = sparc_ramdisk_image64;
767 
768 		/* Another bootloader quirk.  The bootloader normalizes
769 		 * the physical address to KERNBASE, so we have to
770 		 * factor that back out and add in the lowest valid
771 		 * physical page address to get the true physical address.
772 		 */
773 		ramdisk_image -= KERNBASE;
774 		ramdisk_image += phys_base;
775 
776 		numadbg("Found ramdisk at physical address 0x%lx, size %u\n",
777 			ramdisk_image, sparc_ramdisk_size);
778 
779 		initrd_start = ramdisk_image;
780 		initrd_end = ramdisk_image + sparc_ramdisk_size;
781 
782 		memblock_reserve(initrd_start, sparc_ramdisk_size);
783 
784 		initrd_start += PAGE_OFFSET;
785 		initrd_end += PAGE_OFFSET;
786 	}
787 #endif
788 }
789 
790 struct node_mem_mask {
791 	unsigned long mask;
792 	unsigned long val;
793 };
794 static struct node_mem_mask node_masks[MAX_NUMNODES];
795 static int num_node_masks;
796 
797 int numa_cpu_lookup_table[NR_CPUS];
798 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
799 
800 #ifdef CONFIG_NEED_MULTIPLE_NODES
801 
802 struct mdesc_mblock {
803 	u64	base;
804 	u64	size;
805 	u64	offset; /* RA-to-PA */
806 };
807 static struct mdesc_mblock *mblocks;
808 static int num_mblocks;
809 
810 static unsigned long ra_to_pa(unsigned long addr)
811 {
812 	int i;
813 
814 	for (i = 0; i < num_mblocks; i++) {
815 		struct mdesc_mblock *m = &mblocks[i];
816 
817 		if (addr >= m->base &&
818 		    addr < (m->base + m->size)) {
819 			addr += m->offset;
820 			break;
821 		}
822 	}
823 	return addr;
824 }
825 
826 static int find_node(unsigned long addr)
827 {
828 	int i;
829 
830 	addr = ra_to_pa(addr);
831 	for (i = 0; i < num_node_masks; i++) {
832 		struct node_mem_mask *p = &node_masks[i];
833 
834 		if ((addr & p->mask) == p->val)
835 			return i;
836 	}
837 	return -1;
838 }
839 
840 static u64 memblock_nid_range(u64 start, u64 end, int *nid)
841 {
842 	*nid = find_node(start);
843 	start += PAGE_SIZE;
844 	while (start < end) {
845 		int n = find_node(start);
846 
847 		if (n != *nid)
848 			break;
849 		start += PAGE_SIZE;
850 	}
851 
852 	if (start > end)
853 		start = end;
854 
855 	return start;
856 }
857 #endif
858 
859 /* This must be invoked after performing all of the necessary
860  * memblock_set_node() calls for 'nid'.  We need to be able to get
861  * correct data from get_pfn_range_for_nid().
862  */
863 static void __init allocate_node_data(int nid)
864 {
865 	struct pglist_data *p;
866 	unsigned long start_pfn, end_pfn;
867 #ifdef CONFIG_NEED_MULTIPLE_NODES
868 	unsigned long paddr;
869 
870 	paddr = memblock_alloc_try_nid(sizeof(struct pglist_data), SMP_CACHE_BYTES, nid);
871 	if (!paddr) {
872 		prom_printf("Cannot allocate pglist_data for nid[%d]\n", nid);
873 		prom_halt();
874 	}
875 	NODE_DATA(nid) = __va(paddr);
876 	memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
877 
878 	NODE_DATA(nid)->node_id = nid;
879 #endif
880 
881 	p = NODE_DATA(nid);
882 
883 	get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
884 	p->node_start_pfn = start_pfn;
885 	p->node_spanned_pages = end_pfn - start_pfn;
886 }
887 
888 static void init_node_masks_nonnuma(void)
889 {
890 	int i;
891 
892 	numadbg("Initializing tables for non-numa.\n");
893 
894 	node_masks[0].mask = node_masks[0].val = 0;
895 	num_node_masks = 1;
896 
897 	for (i = 0; i < NR_CPUS; i++)
898 		numa_cpu_lookup_table[i] = 0;
899 
900 	cpumask_setall(&numa_cpumask_lookup_table[0]);
901 }
902 
903 #ifdef CONFIG_NEED_MULTIPLE_NODES
904 struct pglist_data *node_data[MAX_NUMNODES];
905 
906 EXPORT_SYMBOL(numa_cpu_lookup_table);
907 EXPORT_SYMBOL(numa_cpumask_lookup_table);
908 EXPORT_SYMBOL(node_data);
909 
910 struct mdesc_mlgroup {
911 	u64	node;
912 	u64	latency;
913 	u64	match;
914 	u64	mask;
915 };
916 static struct mdesc_mlgroup *mlgroups;
917 static int num_mlgroups;
918 
919 static int scan_pio_for_cfg_handle(struct mdesc_handle *md, u64 pio,
920 				   u32 cfg_handle)
921 {
922 	u64 arc;
923 
924 	mdesc_for_each_arc(arc, md, pio, MDESC_ARC_TYPE_FWD) {
925 		u64 target = mdesc_arc_target(md, arc);
926 		const u64 *val;
927 
928 		val = mdesc_get_property(md, target,
929 					 "cfg-handle", NULL);
930 		if (val && *val == cfg_handle)
931 			return 0;
932 	}
933 	return -ENODEV;
934 }
935 
936 static int scan_arcs_for_cfg_handle(struct mdesc_handle *md, u64 grp,
937 				    u32 cfg_handle)
938 {
939 	u64 arc, candidate, best_latency = ~(u64)0;
940 
941 	candidate = MDESC_NODE_NULL;
942 	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
943 		u64 target = mdesc_arc_target(md, arc);
944 		const char *name = mdesc_node_name(md, target);
945 		const u64 *val;
946 
947 		if (strcmp(name, "pio-latency-group"))
948 			continue;
949 
950 		val = mdesc_get_property(md, target, "latency", NULL);
951 		if (!val)
952 			continue;
953 
954 		if (*val < best_latency) {
955 			candidate = target;
956 			best_latency = *val;
957 		}
958 	}
959 
960 	if (candidate == MDESC_NODE_NULL)
961 		return -ENODEV;
962 
963 	return scan_pio_for_cfg_handle(md, candidate, cfg_handle);
964 }
965 
966 int of_node_to_nid(struct device_node *dp)
967 {
968 	const struct linux_prom64_registers *regs;
969 	struct mdesc_handle *md;
970 	u32 cfg_handle;
971 	int count, nid;
972 	u64 grp;
973 
974 	/* This is the right thing to do on currently supported
975 	 * SUN4U NUMA platforms as well, as the PCI controller does
976 	 * not sit behind any particular memory controller.
977 	 */
978 	if (!mlgroups)
979 		return -1;
980 
981 	regs = of_get_property(dp, "reg", NULL);
982 	if (!regs)
983 		return -1;
984 
985 	cfg_handle = (regs->phys_addr >> 32UL) & 0x0fffffff;
986 
987 	md = mdesc_grab();
988 
989 	count = 0;
990 	nid = -1;
991 	mdesc_for_each_node_by_name(md, grp, "group") {
992 		if (!scan_arcs_for_cfg_handle(md, grp, cfg_handle)) {
993 			nid = count;
994 			break;
995 		}
996 		count++;
997 	}
998 
999 	mdesc_release(md);
1000 
1001 	return nid;
1002 }
1003 
1004 static void __init add_node_ranges(void)
1005 {
1006 	struct memblock_region *reg;
1007 
1008 	for_each_memblock(memory, reg) {
1009 		unsigned long size = reg->size;
1010 		unsigned long start, end;
1011 
1012 		start = reg->base;
1013 		end = start + size;
1014 		while (start < end) {
1015 			unsigned long this_end;
1016 			int nid;
1017 
1018 			this_end = memblock_nid_range(start, end, &nid);
1019 
1020 			numadbg("Setting memblock NUMA node nid[%d] "
1021 				"start[%lx] end[%lx]\n",
1022 				nid, start, this_end);
1023 
1024 			memblock_set_node(start, this_end - start, nid);
1025 			start = this_end;
1026 		}
1027 	}
1028 }
1029 
1030 static int __init grab_mlgroups(struct mdesc_handle *md)
1031 {
1032 	unsigned long paddr;
1033 	int count = 0;
1034 	u64 node;
1035 
1036 	mdesc_for_each_node_by_name(md, node, "memory-latency-group")
1037 		count++;
1038 	if (!count)
1039 		return -ENOENT;
1040 
1041 	paddr = memblock_alloc(count * sizeof(struct mdesc_mlgroup),
1042 			  SMP_CACHE_BYTES);
1043 	if (!paddr)
1044 		return -ENOMEM;
1045 
1046 	mlgroups = __va(paddr);
1047 	num_mlgroups = count;
1048 
1049 	count = 0;
1050 	mdesc_for_each_node_by_name(md, node, "memory-latency-group") {
1051 		struct mdesc_mlgroup *m = &mlgroups[count++];
1052 		const u64 *val;
1053 
1054 		m->node = node;
1055 
1056 		val = mdesc_get_property(md, node, "latency", NULL);
1057 		m->latency = *val;
1058 		val = mdesc_get_property(md, node, "address-match", NULL);
1059 		m->match = *val;
1060 		val = mdesc_get_property(md, node, "address-mask", NULL);
1061 		m->mask = *val;
1062 
1063 		numadbg("MLGROUP[%d]: node[%llx] latency[%llx] "
1064 			"match[%llx] mask[%llx]\n",
1065 			count - 1, m->node, m->latency, m->match, m->mask);
1066 	}
1067 
1068 	return 0;
1069 }
1070 
1071 static int __init grab_mblocks(struct mdesc_handle *md)
1072 {
1073 	unsigned long paddr;
1074 	int count = 0;
1075 	u64 node;
1076 
1077 	mdesc_for_each_node_by_name(md, node, "mblock")
1078 		count++;
1079 	if (!count)
1080 		return -ENOENT;
1081 
1082 	paddr = memblock_alloc(count * sizeof(struct mdesc_mblock),
1083 			  SMP_CACHE_BYTES);
1084 	if (!paddr)
1085 		return -ENOMEM;
1086 
1087 	mblocks = __va(paddr);
1088 	num_mblocks = count;
1089 
1090 	count = 0;
1091 	mdesc_for_each_node_by_name(md, node, "mblock") {
1092 		struct mdesc_mblock *m = &mblocks[count++];
1093 		const u64 *val;
1094 
1095 		val = mdesc_get_property(md, node, "base", NULL);
1096 		m->base = *val;
1097 		val = mdesc_get_property(md, node, "size", NULL);
1098 		m->size = *val;
1099 		val = mdesc_get_property(md, node,
1100 					 "address-congruence-offset", NULL);
1101 
1102 		/* The address-congruence-offset property is optional.
1103 		 * Explicity zero it be identifty this.
1104 		 */
1105 		if (val)
1106 			m->offset = *val;
1107 		else
1108 			m->offset = 0UL;
1109 
1110 		numadbg("MBLOCK[%d]: base[%llx] size[%llx] offset[%llx]\n",
1111 			count - 1, m->base, m->size, m->offset);
1112 	}
1113 
1114 	return 0;
1115 }
1116 
1117 static void __init numa_parse_mdesc_group_cpus(struct mdesc_handle *md,
1118 					       u64 grp, cpumask_t *mask)
1119 {
1120 	u64 arc;
1121 
1122 	cpumask_clear(mask);
1123 
1124 	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_BACK) {
1125 		u64 target = mdesc_arc_target(md, arc);
1126 		const char *name = mdesc_node_name(md, target);
1127 		const u64 *id;
1128 
1129 		if (strcmp(name, "cpu"))
1130 			continue;
1131 		id = mdesc_get_property(md, target, "id", NULL);
1132 		if (*id < nr_cpu_ids)
1133 			cpumask_set_cpu(*id, mask);
1134 	}
1135 }
1136 
1137 static struct mdesc_mlgroup * __init find_mlgroup(u64 node)
1138 {
1139 	int i;
1140 
1141 	for (i = 0; i < num_mlgroups; i++) {
1142 		struct mdesc_mlgroup *m = &mlgroups[i];
1143 		if (m->node == node)
1144 			return m;
1145 	}
1146 	return NULL;
1147 }
1148 
1149 static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp,
1150 				      int index)
1151 {
1152 	struct mdesc_mlgroup *candidate = NULL;
1153 	u64 arc, best_latency = ~(u64)0;
1154 	struct node_mem_mask *n;
1155 
1156 	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1157 		u64 target = mdesc_arc_target(md, arc);
1158 		struct mdesc_mlgroup *m = find_mlgroup(target);
1159 		if (!m)
1160 			continue;
1161 		if (m->latency < best_latency) {
1162 			candidate = m;
1163 			best_latency = m->latency;
1164 		}
1165 	}
1166 	if (!candidate)
1167 		return -ENOENT;
1168 
1169 	if (num_node_masks != index) {
1170 		printk(KERN_ERR "Inconsistent NUMA state, "
1171 		       "index[%d] != num_node_masks[%d]\n",
1172 		       index, num_node_masks);
1173 		return -EINVAL;
1174 	}
1175 
1176 	n = &node_masks[num_node_masks++];
1177 
1178 	n->mask = candidate->mask;
1179 	n->val = candidate->match;
1180 
1181 	numadbg("NUMA NODE[%d]: mask[%lx] val[%lx] (latency[%llx])\n",
1182 		index, n->mask, n->val, candidate->latency);
1183 
1184 	return 0;
1185 }
1186 
1187 static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp,
1188 					 int index)
1189 {
1190 	cpumask_t mask;
1191 	int cpu;
1192 
1193 	numa_parse_mdesc_group_cpus(md, grp, &mask);
1194 
1195 	for_each_cpu(cpu, &mask)
1196 		numa_cpu_lookup_table[cpu] = index;
1197 	cpumask_copy(&numa_cpumask_lookup_table[index], &mask);
1198 
1199 	if (numa_debug) {
1200 		printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index);
1201 		for_each_cpu(cpu, &mask)
1202 			printk("%d ", cpu);
1203 		printk("]\n");
1204 	}
1205 
1206 	return numa_attach_mlgroup(md, grp, index);
1207 }
1208 
1209 static int __init numa_parse_mdesc(void)
1210 {
1211 	struct mdesc_handle *md = mdesc_grab();
1212 	int i, err, count;
1213 	u64 node;
1214 
1215 	node = mdesc_node_by_name(md, MDESC_NODE_NULL, "latency-groups");
1216 	if (node == MDESC_NODE_NULL) {
1217 		mdesc_release(md);
1218 		return -ENOENT;
1219 	}
1220 
1221 	err = grab_mblocks(md);
1222 	if (err < 0)
1223 		goto out;
1224 
1225 	err = grab_mlgroups(md);
1226 	if (err < 0)
1227 		goto out;
1228 
1229 	count = 0;
1230 	mdesc_for_each_node_by_name(md, node, "group") {
1231 		err = numa_parse_mdesc_group(md, node, count);
1232 		if (err < 0)
1233 			break;
1234 		count++;
1235 	}
1236 
1237 	add_node_ranges();
1238 
1239 	for (i = 0; i < num_node_masks; i++) {
1240 		allocate_node_data(i);
1241 		node_set_online(i);
1242 	}
1243 
1244 	err = 0;
1245 out:
1246 	mdesc_release(md);
1247 	return err;
1248 }
1249 
1250 static int __init numa_parse_jbus(void)
1251 {
1252 	unsigned long cpu, index;
1253 
1254 	/* NUMA node id is encoded in bits 36 and higher, and there is
1255 	 * a 1-to-1 mapping from CPU ID to NUMA node ID.
1256 	 */
1257 	index = 0;
1258 	for_each_present_cpu(cpu) {
1259 		numa_cpu_lookup_table[cpu] = index;
1260 		cpumask_copy(&numa_cpumask_lookup_table[index], cpumask_of(cpu));
1261 		node_masks[index].mask = ~((1UL << 36UL) - 1UL);
1262 		node_masks[index].val = cpu << 36UL;
1263 
1264 		index++;
1265 	}
1266 	num_node_masks = index;
1267 
1268 	add_node_ranges();
1269 
1270 	for (index = 0; index < num_node_masks; index++) {
1271 		allocate_node_data(index);
1272 		node_set_online(index);
1273 	}
1274 
1275 	return 0;
1276 }
1277 
1278 static int __init numa_parse_sun4u(void)
1279 {
1280 	if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1281 		unsigned long ver;
1282 
1283 		__asm__ ("rdpr %%ver, %0" : "=r" (ver));
1284 		if ((ver >> 32UL) == __JALAPENO_ID ||
1285 		    (ver >> 32UL) == __SERRANO_ID)
1286 			return numa_parse_jbus();
1287 	}
1288 	return -1;
1289 }
1290 
1291 static int __init bootmem_init_numa(void)
1292 {
1293 	int err = -1;
1294 
1295 	numadbg("bootmem_init_numa()\n");
1296 
1297 	if (numa_enabled) {
1298 		if (tlb_type == hypervisor)
1299 			err = numa_parse_mdesc();
1300 		else
1301 			err = numa_parse_sun4u();
1302 	}
1303 	return err;
1304 }
1305 
1306 #else
1307 
1308 static int bootmem_init_numa(void)
1309 {
1310 	return -1;
1311 }
1312 
1313 #endif
1314 
1315 static void __init bootmem_init_nonnuma(void)
1316 {
1317 	unsigned long top_of_ram = memblock_end_of_DRAM();
1318 	unsigned long total_ram = memblock_phys_mem_size();
1319 
1320 	numadbg("bootmem_init_nonnuma()\n");
1321 
1322 	printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
1323 	       top_of_ram, total_ram);
1324 	printk(KERN_INFO "Memory hole size: %ldMB\n",
1325 	       (top_of_ram - total_ram) >> 20);
1326 
1327 	init_node_masks_nonnuma();
1328 	memblock_set_node(0, (phys_addr_t)ULLONG_MAX, 0);
1329 	allocate_node_data(0);
1330 	node_set_online(0);
1331 }
1332 
1333 static unsigned long __init bootmem_init(unsigned long phys_base)
1334 {
1335 	unsigned long end_pfn;
1336 
1337 	end_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1338 	max_pfn = max_low_pfn = end_pfn;
1339 	min_low_pfn = (phys_base >> PAGE_SHIFT);
1340 
1341 	if (bootmem_init_numa() < 0)
1342 		bootmem_init_nonnuma();
1343 
1344 	/* Dump memblock with node info. */
1345 	memblock_dump_all();
1346 
1347 	/* XXX cpu notifier XXX */
1348 
1349 	sparse_memory_present_with_active_regions(MAX_NUMNODES);
1350 	sparse_init();
1351 
1352 	return end_pfn;
1353 }
1354 
1355 static struct linux_prom64_registers pall[MAX_BANKS] __initdata;
1356 static int pall_ents __initdata;
1357 
1358 #ifdef CONFIG_DEBUG_PAGEALLOC
1359 static unsigned long __ref kernel_map_range(unsigned long pstart,
1360 					    unsigned long pend, pgprot_t prot)
1361 {
1362 	unsigned long vstart = PAGE_OFFSET + pstart;
1363 	unsigned long vend = PAGE_OFFSET + pend;
1364 	unsigned long alloc_bytes = 0UL;
1365 
1366 	if ((vstart & ~PAGE_MASK) || (vend & ~PAGE_MASK)) {
1367 		prom_printf("kernel_map: Unaligned physmem[%lx:%lx]\n",
1368 			    vstart, vend);
1369 		prom_halt();
1370 	}
1371 
1372 	while (vstart < vend) {
1373 		unsigned long this_end, paddr = __pa(vstart);
1374 		pgd_t *pgd = pgd_offset_k(vstart);
1375 		pud_t *pud;
1376 		pmd_t *pmd;
1377 		pte_t *pte;
1378 
1379 		pud = pud_offset(pgd, vstart);
1380 		if (pud_none(*pud)) {
1381 			pmd_t *new;
1382 
1383 			new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1384 			alloc_bytes += PAGE_SIZE;
1385 			pud_populate(&init_mm, pud, new);
1386 		}
1387 
1388 		pmd = pmd_offset(pud, vstart);
1389 		if (!pmd_present(*pmd)) {
1390 			pte_t *new;
1391 
1392 			new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1393 			alloc_bytes += PAGE_SIZE;
1394 			pmd_populate_kernel(&init_mm, pmd, new);
1395 		}
1396 
1397 		pte = pte_offset_kernel(pmd, vstart);
1398 		this_end = (vstart + PMD_SIZE) & PMD_MASK;
1399 		if (this_end > vend)
1400 			this_end = vend;
1401 
1402 		while (vstart < this_end) {
1403 			pte_val(*pte) = (paddr | pgprot_val(prot));
1404 
1405 			vstart += PAGE_SIZE;
1406 			paddr += PAGE_SIZE;
1407 			pte++;
1408 		}
1409 	}
1410 
1411 	return alloc_bytes;
1412 }
1413 
1414 extern unsigned int kvmap_linear_patch[1];
1415 #endif /* CONFIG_DEBUG_PAGEALLOC */
1416 
1417 static void __init kpte_set_val(unsigned long index, unsigned long val)
1418 {
1419 	unsigned long *ptr = kpte_linear_bitmap;
1420 
1421 	val <<= ((index % (BITS_PER_LONG / 2)) * 2);
1422 	ptr += (index / (BITS_PER_LONG / 2));
1423 
1424 	*ptr |= val;
1425 }
1426 
1427 static const unsigned long kpte_shift_min = 28; /* 256MB */
1428 static const unsigned long kpte_shift_max = 34; /* 16GB */
1429 static const unsigned long kpte_shift_incr = 3;
1430 
1431 static unsigned long kpte_mark_using_shift(unsigned long start, unsigned long end,
1432 					   unsigned long shift)
1433 {
1434 	unsigned long size = (1UL << shift);
1435 	unsigned long mask = (size - 1UL);
1436 	unsigned long remains = end - start;
1437 	unsigned long val;
1438 
1439 	if (remains < size || (start & mask))
1440 		return start;
1441 
1442 	/* VAL maps:
1443 	 *
1444 	 *	shift 28 --> kern_linear_pte_xor index 1
1445 	 *	shift 31 --> kern_linear_pte_xor index 2
1446 	 *	shift 34 --> kern_linear_pte_xor index 3
1447 	 */
1448 	val = ((shift - kpte_shift_min) / kpte_shift_incr) + 1;
1449 
1450 	remains &= ~mask;
1451 	if (shift != kpte_shift_max)
1452 		remains = size;
1453 
1454 	while (remains) {
1455 		unsigned long index = start >> kpte_shift_min;
1456 
1457 		kpte_set_val(index, val);
1458 
1459 		start += 1UL << kpte_shift_min;
1460 		remains -= 1UL << kpte_shift_min;
1461 	}
1462 
1463 	return start;
1464 }
1465 
1466 static void __init mark_kpte_bitmap(unsigned long start, unsigned long end)
1467 {
1468 	unsigned long smallest_size, smallest_mask;
1469 	unsigned long s;
1470 
1471 	smallest_size = (1UL << kpte_shift_min);
1472 	smallest_mask = (smallest_size - 1UL);
1473 
1474 	while (start < end) {
1475 		unsigned long orig_start = start;
1476 
1477 		for (s = kpte_shift_max; s >= kpte_shift_min; s -= kpte_shift_incr) {
1478 			start = kpte_mark_using_shift(start, end, s);
1479 
1480 			if (start != orig_start)
1481 				break;
1482 		}
1483 
1484 		if (start == orig_start)
1485 			start = (start + smallest_size) & ~smallest_mask;
1486 	}
1487 }
1488 
1489 static void __init init_kpte_bitmap(void)
1490 {
1491 	unsigned long i;
1492 
1493 	for (i = 0; i < pall_ents; i++) {
1494 		unsigned long phys_start, phys_end;
1495 
1496 		phys_start = pall[i].phys_addr;
1497 		phys_end = phys_start + pall[i].reg_size;
1498 
1499 		mark_kpte_bitmap(phys_start, phys_end);
1500 	}
1501 }
1502 
1503 static void __init kernel_physical_mapping_init(void)
1504 {
1505 #ifdef CONFIG_DEBUG_PAGEALLOC
1506 	unsigned long i, mem_alloced = 0UL;
1507 
1508 	for (i = 0; i < pall_ents; i++) {
1509 		unsigned long phys_start, phys_end;
1510 
1511 		phys_start = pall[i].phys_addr;
1512 		phys_end = phys_start + pall[i].reg_size;
1513 
1514 		mem_alloced += kernel_map_range(phys_start, phys_end,
1515 						PAGE_KERNEL);
1516 	}
1517 
1518 	printk("Allocated %ld bytes for kernel page tables.\n",
1519 	       mem_alloced);
1520 
1521 	kvmap_linear_patch[0] = 0x01000000; /* nop */
1522 	flushi(&kvmap_linear_patch[0]);
1523 
1524 	__flush_tlb_all();
1525 #endif
1526 }
1527 
1528 #ifdef CONFIG_DEBUG_PAGEALLOC
1529 void kernel_map_pages(struct page *page, int numpages, int enable)
1530 {
1531 	unsigned long phys_start = page_to_pfn(page) << PAGE_SHIFT;
1532 	unsigned long phys_end = phys_start + (numpages * PAGE_SIZE);
1533 
1534 	kernel_map_range(phys_start, phys_end,
1535 			 (enable ? PAGE_KERNEL : __pgprot(0)));
1536 
1537 	flush_tsb_kernel_range(PAGE_OFFSET + phys_start,
1538 			       PAGE_OFFSET + phys_end);
1539 
1540 	/* we should perform an IPI and flush all tlbs,
1541 	 * but that can deadlock->flush only current cpu.
1542 	 */
1543 	__flush_tlb_kernel_range(PAGE_OFFSET + phys_start,
1544 				 PAGE_OFFSET + phys_end);
1545 }
1546 #endif
1547 
1548 unsigned long __init find_ecache_flush_span(unsigned long size)
1549 {
1550 	int i;
1551 
1552 	for (i = 0; i < pavail_ents; i++) {
1553 		if (pavail[i].reg_size >= size)
1554 			return pavail[i].phys_addr;
1555 	}
1556 
1557 	return ~0UL;
1558 }
1559 
1560 unsigned long PAGE_OFFSET;
1561 EXPORT_SYMBOL(PAGE_OFFSET);
1562 
1563 static void __init page_offset_shift_patch_one(unsigned int *insn, unsigned long phys_bits)
1564 {
1565 	unsigned long final_shift;
1566 	unsigned int val = *insn;
1567 	unsigned int cnt;
1568 
1569 	/* We are patching in ilog2(max_supported_phys_address), and
1570 	 * we are doing so in a manner similar to a relocation addend.
1571 	 * That is, we are adding the shift value to whatever value
1572 	 * is in the shift instruction count field already.
1573 	 */
1574 	cnt = (val & 0x3f);
1575 	val &= ~0x3f;
1576 
1577 	/* If we are trying to shift >= 64 bits, clear the destination
1578 	 * register.  This can happen when phys_bits ends up being equal
1579 	 * to MAX_PHYS_ADDRESS_BITS.
1580 	 */
1581 	final_shift = (cnt + (64 - phys_bits));
1582 	if (final_shift >= 64) {
1583 		unsigned int rd = (val >> 25) & 0x1f;
1584 
1585 		val = 0x80100000 | (rd << 25);
1586 	} else {
1587 		val |= final_shift;
1588 	}
1589 	*insn = val;
1590 
1591 	__asm__ __volatile__("flush	%0"
1592 			     : /* no outputs */
1593 			     : "r" (insn));
1594 }
1595 
1596 static void __init page_offset_shift_patch(unsigned long phys_bits)
1597 {
1598 	extern unsigned int __page_offset_shift_patch;
1599 	extern unsigned int __page_offset_shift_patch_end;
1600 	unsigned int *p;
1601 
1602 	p = &__page_offset_shift_patch;
1603 	while (p < &__page_offset_shift_patch_end) {
1604 		unsigned int *insn = (unsigned int *)(unsigned long)*p;
1605 
1606 		page_offset_shift_patch_one(insn, phys_bits);
1607 
1608 		p++;
1609 	}
1610 }
1611 
1612 static void __init setup_page_offset(void)
1613 {
1614 	unsigned long max_phys_bits = 40;
1615 
1616 	if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1617 		max_phys_bits = 42;
1618 	} else if (tlb_type == hypervisor) {
1619 		switch (sun4v_chip_type) {
1620 		case SUN4V_CHIP_NIAGARA1:
1621 		case SUN4V_CHIP_NIAGARA2:
1622 			max_phys_bits = 39;
1623 			break;
1624 		case SUN4V_CHIP_NIAGARA3:
1625 			max_phys_bits = 43;
1626 			break;
1627 		case SUN4V_CHIP_NIAGARA4:
1628 		case SUN4V_CHIP_NIAGARA5:
1629 		case SUN4V_CHIP_SPARC64X:
1630 		default:
1631 			max_phys_bits = 47;
1632 			break;
1633 		}
1634 	}
1635 
1636 	if (max_phys_bits > MAX_PHYS_ADDRESS_BITS) {
1637 		prom_printf("MAX_PHYS_ADDRESS_BITS is too small, need %lu\n",
1638 			    max_phys_bits);
1639 		prom_halt();
1640 	}
1641 
1642 	PAGE_OFFSET = PAGE_OFFSET_BY_BITS(max_phys_bits);
1643 
1644 	pr_info("PAGE_OFFSET is 0x%016lx (max_phys_bits == %lu)\n",
1645 		PAGE_OFFSET, max_phys_bits);
1646 
1647 	page_offset_shift_patch(max_phys_bits);
1648 }
1649 
1650 static void __init tsb_phys_patch(void)
1651 {
1652 	struct tsb_ldquad_phys_patch_entry *pquad;
1653 	struct tsb_phys_patch_entry *p;
1654 
1655 	pquad = &__tsb_ldquad_phys_patch;
1656 	while (pquad < &__tsb_ldquad_phys_patch_end) {
1657 		unsigned long addr = pquad->addr;
1658 
1659 		if (tlb_type == hypervisor)
1660 			*(unsigned int *) addr = pquad->sun4v_insn;
1661 		else
1662 			*(unsigned int *) addr = pquad->sun4u_insn;
1663 		wmb();
1664 		__asm__ __volatile__("flush	%0"
1665 				     : /* no outputs */
1666 				     : "r" (addr));
1667 
1668 		pquad++;
1669 	}
1670 
1671 	p = &__tsb_phys_patch;
1672 	while (p < &__tsb_phys_patch_end) {
1673 		unsigned long addr = p->addr;
1674 
1675 		*(unsigned int *) addr = p->insn;
1676 		wmb();
1677 		__asm__ __volatile__("flush	%0"
1678 				     : /* no outputs */
1679 				     : "r" (addr));
1680 
1681 		p++;
1682 	}
1683 }
1684 
1685 /* Don't mark as init, we give this to the Hypervisor.  */
1686 #ifndef CONFIG_DEBUG_PAGEALLOC
1687 #define NUM_KTSB_DESCR	2
1688 #else
1689 #define NUM_KTSB_DESCR	1
1690 #endif
1691 static struct hv_tsb_descr ktsb_descr[NUM_KTSB_DESCR];
1692 extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
1693 
1694 static void patch_one_ktsb_phys(unsigned int *start, unsigned int *end, unsigned long pa)
1695 {
1696 	pa >>= KTSB_PHYS_SHIFT;
1697 
1698 	while (start < end) {
1699 		unsigned int *ia = (unsigned int *)(unsigned long)*start;
1700 
1701 		ia[0] = (ia[0] & ~0x3fffff) | (pa >> 10);
1702 		__asm__ __volatile__("flush	%0" : : "r" (ia));
1703 
1704 		ia[1] = (ia[1] & ~0x3ff) | (pa & 0x3ff);
1705 		__asm__ __volatile__("flush	%0" : : "r" (ia + 1));
1706 
1707 		start++;
1708 	}
1709 }
1710 
1711 static void ktsb_phys_patch(void)
1712 {
1713 	extern unsigned int __swapper_tsb_phys_patch;
1714 	extern unsigned int __swapper_tsb_phys_patch_end;
1715 	unsigned long ktsb_pa;
1716 
1717 	ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
1718 	patch_one_ktsb_phys(&__swapper_tsb_phys_patch,
1719 			    &__swapper_tsb_phys_patch_end, ktsb_pa);
1720 #ifndef CONFIG_DEBUG_PAGEALLOC
1721 	{
1722 	extern unsigned int __swapper_4m_tsb_phys_patch;
1723 	extern unsigned int __swapper_4m_tsb_phys_patch_end;
1724 	ktsb_pa = (kern_base +
1725 		   ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
1726 	patch_one_ktsb_phys(&__swapper_4m_tsb_phys_patch,
1727 			    &__swapper_4m_tsb_phys_patch_end, ktsb_pa);
1728 	}
1729 #endif
1730 }
1731 
1732 static void __init sun4v_ktsb_init(void)
1733 {
1734 	unsigned long ktsb_pa;
1735 
1736 	/* First KTSB for PAGE_SIZE mappings.  */
1737 	ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
1738 
1739 	switch (PAGE_SIZE) {
1740 	case 8 * 1024:
1741 	default:
1742 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_8K;
1743 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_8K;
1744 		break;
1745 
1746 	case 64 * 1024:
1747 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_64K;
1748 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_64K;
1749 		break;
1750 
1751 	case 512 * 1024:
1752 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_512K;
1753 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_512K;
1754 		break;
1755 
1756 	case 4 * 1024 * 1024:
1757 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_4MB;
1758 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_4MB;
1759 		break;
1760 	}
1761 
1762 	ktsb_descr[0].assoc = 1;
1763 	ktsb_descr[0].num_ttes = KERNEL_TSB_NENTRIES;
1764 	ktsb_descr[0].ctx_idx = 0;
1765 	ktsb_descr[0].tsb_base = ktsb_pa;
1766 	ktsb_descr[0].resv = 0;
1767 
1768 #ifndef CONFIG_DEBUG_PAGEALLOC
1769 	/* Second KTSB for 4MB/256MB/2GB/16GB mappings.  */
1770 	ktsb_pa = (kern_base +
1771 		   ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
1772 
1773 	ktsb_descr[1].pgsz_idx = HV_PGSZ_IDX_4MB;
1774 	ktsb_descr[1].pgsz_mask = ((HV_PGSZ_MASK_4MB |
1775 				    HV_PGSZ_MASK_256MB |
1776 				    HV_PGSZ_MASK_2GB |
1777 				    HV_PGSZ_MASK_16GB) &
1778 				   cpu_pgsz_mask);
1779 	ktsb_descr[1].assoc = 1;
1780 	ktsb_descr[1].num_ttes = KERNEL_TSB4M_NENTRIES;
1781 	ktsb_descr[1].ctx_idx = 0;
1782 	ktsb_descr[1].tsb_base = ktsb_pa;
1783 	ktsb_descr[1].resv = 0;
1784 #endif
1785 }
1786 
1787 void sun4v_ktsb_register(void)
1788 {
1789 	unsigned long pa, ret;
1790 
1791 	pa = kern_base + ((unsigned long)&ktsb_descr[0] - KERNBASE);
1792 
1793 	ret = sun4v_mmu_tsb_ctx0(NUM_KTSB_DESCR, pa);
1794 	if (ret != 0) {
1795 		prom_printf("hypervisor_mmu_tsb_ctx0[%lx]: "
1796 			    "errors with %lx\n", pa, ret);
1797 		prom_halt();
1798 	}
1799 }
1800 
1801 static void __init sun4u_linear_pte_xor_finalize(void)
1802 {
1803 #ifndef CONFIG_DEBUG_PAGEALLOC
1804 	/* This is where we would add Panther support for
1805 	 * 32MB and 256MB pages.
1806 	 */
1807 #endif
1808 }
1809 
1810 static void __init sun4v_linear_pte_xor_finalize(void)
1811 {
1812 #ifndef CONFIG_DEBUG_PAGEALLOC
1813 	if (cpu_pgsz_mask & HV_PGSZ_MASK_256MB) {
1814 		kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZ256MB_4V) ^
1815 			PAGE_OFFSET;
1816 		kern_linear_pte_xor[1] |= (_PAGE_CP_4V | _PAGE_CV_4V |
1817 					   _PAGE_P_4V | _PAGE_W_4V);
1818 	} else {
1819 		kern_linear_pte_xor[1] = kern_linear_pte_xor[0];
1820 	}
1821 
1822 	if (cpu_pgsz_mask & HV_PGSZ_MASK_2GB) {
1823 		kern_linear_pte_xor[2] = (_PAGE_VALID | _PAGE_SZ2GB_4V) ^
1824 			PAGE_OFFSET;
1825 		kern_linear_pte_xor[2] |= (_PAGE_CP_4V | _PAGE_CV_4V |
1826 					   _PAGE_P_4V | _PAGE_W_4V);
1827 	} else {
1828 		kern_linear_pte_xor[2] = kern_linear_pte_xor[1];
1829 	}
1830 
1831 	if (cpu_pgsz_mask & HV_PGSZ_MASK_16GB) {
1832 		kern_linear_pte_xor[3] = (_PAGE_VALID | _PAGE_SZ16GB_4V) ^
1833 			PAGE_OFFSET;
1834 		kern_linear_pte_xor[3] |= (_PAGE_CP_4V | _PAGE_CV_4V |
1835 					   _PAGE_P_4V | _PAGE_W_4V);
1836 	} else {
1837 		kern_linear_pte_xor[3] = kern_linear_pte_xor[2];
1838 	}
1839 #endif
1840 }
1841 
1842 /* paging_init() sets up the page tables */
1843 
1844 static unsigned long last_valid_pfn;
1845 pgd_t swapper_pg_dir[PTRS_PER_PGD];
1846 
1847 static void sun4u_pgprot_init(void);
1848 static void sun4v_pgprot_init(void);
1849 
1850 void __init paging_init(void)
1851 {
1852 	unsigned long end_pfn, shift, phys_base;
1853 	unsigned long real_end, i;
1854 	int node;
1855 
1856 	setup_page_offset();
1857 
1858 	/* These build time checkes make sure that the dcache_dirty_cpu()
1859 	 * page->flags usage will work.
1860 	 *
1861 	 * When a page gets marked as dcache-dirty, we store the
1862 	 * cpu number starting at bit 32 in the page->flags.  Also,
1863 	 * functions like clear_dcache_dirty_cpu use the cpu mask
1864 	 * in 13-bit signed-immediate instruction fields.
1865 	 */
1866 
1867 	/*
1868 	 * Page flags must not reach into upper 32 bits that are used
1869 	 * for the cpu number
1870 	 */
1871 	BUILD_BUG_ON(NR_PAGEFLAGS > 32);
1872 
1873 	/*
1874 	 * The bit fields placed in the high range must not reach below
1875 	 * the 32 bit boundary. Otherwise we cannot place the cpu field
1876 	 * at the 32 bit boundary.
1877 	 */
1878 	BUILD_BUG_ON(SECTIONS_WIDTH + NODES_WIDTH + ZONES_WIDTH +
1879 		ilog2(roundup_pow_of_two(NR_CPUS)) > 32);
1880 
1881 	BUILD_BUG_ON(NR_CPUS > 4096);
1882 
1883 	kern_base = (prom_boot_mapping_phys_low >> 22UL) << 22UL;
1884 	kern_size = (unsigned long)&_end - (unsigned long)KERNBASE;
1885 
1886 	/* Invalidate both kernel TSBs.  */
1887 	memset(swapper_tsb, 0x40, sizeof(swapper_tsb));
1888 #ifndef CONFIG_DEBUG_PAGEALLOC
1889 	memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
1890 #endif
1891 
1892 	if (tlb_type == hypervisor)
1893 		sun4v_pgprot_init();
1894 	else
1895 		sun4u_pgprot_init();
1896 
1897 	if (tlb_type == cheetah_plus ||
1898 	    tlb_type == hypervisor) {
1899 		tsb_phys_patch();
1900 		ktsb_phys_patch();
1901 	}
1902 
1903 	if (tlb_type == hypervisor)
1904 		sun4v_patch_tlb_handlers();
1905 
1906 	/* Find available physical memory...
1907 	 *
1908 	 * Read it twice in order to work around a bug in openfirmware.
1909 	 * The call to grab this table itself can cause openfirmware to
1910 	 * allocate memory, which in turn can take away some space from
1911 	 * the list of available memory.  Reading it twice makes sure
1912 	 * we really do get the final value.
1913 	 */
1914 	read_obp_translations();
1915 	read_obp_memory("reg", &pall[0], &pall_ents);
1916 	read_obp_memory("available", &pavail[0], &pavail_ents);
1917 	read_obp_memory("available", &pavail[0], &pavail_ents);
1918 
1919 	phys_base = 0xffffffffffffffffUL;
1920 	for (i = 0; i < pavail_ents; i++) {
1921 		phys_base = min(phys_base, pavail[i].phys_addr);
1922 		memblock_add(pavail[i].phys_addr, pavail[i].reg_size);
1923 	}
1924 
1925 	memblock_reserve(kern_base, kern_size);
1926 
1927 	find_ramdisk(phys_base);
1928 
1929 	memblock_enforce_memory_limit(cmdline_memory_size);
1930 
1931 	memblock_allow_resize();
1932 	memblock_dump_all();
1933 
1934 	set_bit(0, mmu_context_bmap);
1935 
1936 	shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE);
1937 
1938 	real_end = (unsigned long)_end;
1939 	num_kernel_image_mappings = DIV_ROUND_UP(real_end - KERNBASE, 1 << 22);
1940 	printk("Kernel: Using %d locked TLB entries for main kernel image.\n",
1941 	       num_kernel_image_mappings);
1942 
1943 	/* Set kernel pgd to upper alias so physical page computations
1944 	 * work.
1945 	 */
1946 	init_mm.pgd += ((shift) / (sizeof(pgd_t)));
1947 
1948 	memset(swapper_low_pmd_dir, 0, sizeof(swapper_low_pmd_dir));
1949 
1950 	/* Now can init the kernel/bad page tables. */
1951 	pud_set(pud_offset(&swapper_pg_dir[0], 0),
1952 		swapper_low_pmd_dir + (shift / sizeof(pgd_t)));
1953 
1954 	inherit_prom_mappings();
1955 
1956 	init_kpte_bitmap();
1957 
1958 	/* Ok, we can use our TLB miss and window trap handlers safely.  */
1959 	setup_tba();
1960 
1961 	__flush_tlb_all();
1962 
1963 	prom_build_devicetree();
1964 	of_populate_present_mask();
1965 #ifndef CONFIG_SMP
1966 	of_fill_in_cpu_data();
1967 #endif
1968 
1969 	if (tlb_type == hypervisor) {
1970 		sun4v_mdesc_init();
1971 		mdesc_populate_present_mask(cpu_all_mask);
1972 #ifndef CONFIG_SMP
1973 		mdesc_fill_in_cpu_data(cpu_all_mask);
1974 #endif
1975 		mdesc_get_page_sizes(cpu_all_mask, &cpu_pgsz_mask);
1976 
1977 		sun4v_linear_pte_xor_finalize();
1978 
1979 		sun4v_ktsb_init();
1980 		sun4v_ktsb_register();
1981 	} else {
1982 		unsigned long impl, ver;
1983 
1984 		cpu_pgsz_mask = (HV_PGSZ_MASK_8K | HV_PGSZ_MASK_64K |
1985 				 HV_PGSZ_MASK_512K | HV_PGSZ_MASK_4MB);
1986 
1987 		__asm__ __volatile__("rdpr %%ver, %0" : "=r" (ver));
1988 		impl = ((ver >> 32) & 0xffff);
1989 		if (impl == PANTHER_IMPL)
1990 			cpu_pgsz_mask |= (HV_PGSZ_MASK_32MB |
1991 					  HV_PGSZ_MASK_256MB);
1992 
1993 		sun4u_linear_pte_xor_finalize();
1994 	}
1995 
1996 	/* Flush the TLBs and the 4M TSB so that the updated linear
1997 	 * pte XOR settings are realized for all mappings.
1998 	 */
1999 	__flush_tlb_all();
2000 #ifndef CONFIG_DEBUG_PAGEALLOC
2001 	memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2002 #endif
2003 	__flush_tlb_all();
2004 
2005 	/* Setup bootmem... */
2006 	last_valid_pfn = end_pfn = bootmem_init(phys_base);
2007 
2008 	/* Once the OF device tree and MDESC have been setup, we know
2009 	 * the list of possible cpus.  Therefore we can allocate the
2010 	 * IRQ stacks.
2011 	 */
2012 	for_each_possible_cpu(i) {
2013 		node = cpu_to_node(i);
2014 
2015 		softirq_stack[i] = __alloc_bootmem_node(NODE_DATA(node),
2016 							THREAD_SIZE,
2017 							THREAD_SIZE, 0);
2018 		hardirq_stack[i] = __alloc_bootmem_node(NODE_DATA(node),
2019 							THREAD_SIZE,
2020 							THREAD_SIZE, 0);
2021 	}
2022 
2023 	kernel_physical_mapping_init();
2024 
2025 	{
2026 		unsigned long max_zone_pfns[MAX_NR_ZONES];
2027 
2028 		memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
2029 
2030 		max_zone_pfns[ZONE_NORMAL] = end_pfn;
2031 
2032 		free_area_init_nodes(max_zone_pfns);
2033 	}
2034 
2035 	printk("Booting Linux...\n");
2036 }
2037 
2038 int page_in_phys_avail(unsigned long paddr)
2039 {
2040 	int i;
2041 
2042 	paddr &= PAGE_MASK;
2043 
2044 	for (i = 0; i < pavail_ents; i++) {
2045 		unsigned long start, end;
2046 
2047 		start = pavail[i].phys_addr;
2048 		end = start + pavail[i].reg_size;
2049 
2050 		if (paddr >= start && paddr < end)
2051 			return 1;
2052 	}
2053 	if (paddr >= kern_base && paddr < (kern_base + kern_size))
2054 		return 1;
2055 #ifdef CONFIG_BLK_DEV_INITRD
2056 	if (paddr >= __pa(initrd_start) &&
2057 	    paddr < __pa(PAGE_ALIGN(initrd_end)))
2058 		return 1;
2059 #endif
2060 
2061 	return 0;
2062 }
2063 
2064 static struct linux_prom64_registers pavail_rescan[MAX_BANKS] __initdata;
2065 static int pavail_rescan_ents __initdata;
2066 
2067 /* Certain OBP calls, such as fetching "available" properties, can
2068  * claim physical memory.  So, along with initializing the valid
2069  * address bitmap, what we do here is refetch the physical available
2070  * memory list again, and make sure it provides at least as much
2071  * memory as 'pavail' does.
2072  */
2073 static void __init setup_valid_addr_bitmap_from_pavail(unsigned long *bitmap)
2074 {
2075 	int i;
2076 
2077 	read_obp_memory("available", &pavail_rescan[0], &pavail_rescan_ents);
2078 
2079 	for (i = 0; i < pavail_ents; i++) {
2080 		unsigned long old_start, old_end;
2081 
2082 		old_start = pavail[i].phys_addr;
2083 		old_end = old_start + pavail[i].reg_size;
2084 		while (old_start < old_end) {
2085 			int n;
2086 
2087 			for (n = 0; n < pavail_rescan_ents; n++) {
2088 				unsigned long new_start, new_end;
2089 
2090 				new_start = pavail_rescan[n].phys_addr;
2091 				new_end = new_start +
2092 					pavail_rescan[n].reg_size;
2093 
2094 				if (new_start <= old_start &&
2095 				    new_end >= (old_start + PAGE_SIZE)) {
2096 					set_bit(old_start >> 22, bitmap);
2097 					goto do_next_page;
2098 				}
2099 			}
2100 
2101 			prom_printf("mem_init: Lost memory in pavail\n");
2102 			prom_printf("mem_init: OLD start[%lx] size[%lx]\n",
2103 				    pavail[i].phys_addr,
2104 				    pavail[i].reg_size);
2105 			prom_printf("mem_init: NEW start[%lx] size[%lx]\n",
2106 				    pavail_rescan[i].phys_addr,
2107 				    pavail_rescan[i].reg_size);
2108 			prom_printf("mem_init: Cannot continue, aborting.\n");
2109 			prom_halt();
2110 
2111 		do_next_page:
2112 			old_start += PAGE_SIZE;
2113 		}
2114 	}
2115 }
2116 
2117 static void __init patch_tlb_miss_handler_bitmap(void)
2118 {
2119 	extern unsigned int valid_addr_bitmap_insn[];
2120 	extern unsigned int valid_addr_bitmap_patch[];
2121 
2122 	valid_addr_bitmap_insn[1] = valid_addr_bitmap_patch[1];
2123 	mb();
2124 	valid_addr_bitmap_insn[0] = valid_addr_bitmap_patch[0];
2125 	flushi(&valid_addr_bitmap_insn[0]);
2126 }
2127 
2128 static void __init register_page_bootmem_info(void)
2129 {
2130 #ifdef CONFIG_NEED_MULTIPLE_NODES
2131 	int i;
2132 
2133 	for_each_online_node(i)
2134 		if (NODE_DATA(i)->node_spanned_pages)
2135 			register_page_bootmem_info_node(NODE_DATA(i));
2136 #endif
2137 }
2138 void __init mem_init(void)
2139 {
2140 	unsigned long addr, last;
2141 
2142 	addr = PAGE_OFFSET + kern_base;
2143 	last = PAGE_ALIGN(kern_size) + addr;
2144 	while (addr < last) {
2145 		set_bit(__pa(addr) >> 22, sparc64_valid_addr_bitmap);
2146 		addr += PAGE_SIZE;
2147 	}
2148 
2149 	setup_valid_addr_bitmap_from_pavail(sparc64_valid_addr_bitmap);
2150 	patch_tlb_miss_handler_bitmap();
2151 
2152 	high_memory = __va(last_valid_pfn << PAGE_SHIFT);
2153 
2154 	register_page_bootmem_info();
2155 	free_all_bootmem();
2156 
2157 	/*
2158 	 * Set up the zero page, mark it reserved, so that page count
2159 	 * is not manipulated when freeing the page from user ptes.
2160 	 */
2161 	mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0);
2162 	if (mem_map_zero == NULL) {
2163 		prom_printf("paging_init: Cannot alloc zero page.\n");
2164 		prom_halt();
2165 	}
2166 	mark_page_reserved(mem_map_zero);
2167 
2168 	mem_init_print_info(NULL);
2169 
2170 	if (tlb_type == cheetah || tlb_type == cheetah_plus)
2171 		cheetah_ecache_flush_init();
2172 }
2173 
2174 void free_initmem(void)
2175 {
2176 	unsigned long addr, initend;
2177 	int do_free = 1;
2178 
2179 	/* If the physical memory maps were trimmed by kernel command
2180 	 * line options, don't even try freeing this initmem stuff up.
2181 	 * The kernel image could have been in the trimmed out region
2182 	 * and if so the freeing below will free invalid page structs.
2183 	 */
2184 	if (cmdline_memory_size)
2185 		do_free = 0;
2186 
2187 	/*
2188 	 * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes.
2189 	 */
2190 	addr = PAGE_ALIGN((unsigned long)(__init_begin));
2191 	initend = (unsigned long)(__init_end) & PAGE_MASK;
2192 	for (; addr < initend; addr += PAGE_SIZE) {
2193 		unsigned long page;
2194 
2195 		page = (addr +
2196 			((unsigned long) __va(kern_base)) -
2197 			((unsigned long) KERNBASE));
2198 		memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
2199 
2200 		if (do_free)
2201 			free_reserved_page(virt_to_page(page));
2202 	}
2203 }
2204 
2205 #ifdef CONFIG_BLK_DEV_INITRD
2206 void free_initrd_mem(unsigned long start, unsigned long end)
2207 {
2208 	free_reserved_area((void *)start, (void *)end, POISON_FREE_INITMEM,
2209 			   "initrd");
2210 }
2211 #endif
2212 
2213 #define _PAGE_CACHE_4U	(_PAGE_CP_4U | _PAGE_CV_4U)
2214 #define _PAGE_CACHE_4V	(_PAGE_CP_4V | _PAGE_CV_4V)
2215 #define __DIRTY_BITS_4U	 (_PAGE_MODIFIED_4U | _PAGE_WRITE_4U | _PAGE_W_4U)
2216 #define __DIRTY_BITS_4V	 (_PAGE_MODIFIED_4V | _PAGE_WRITE_4V | _PAGE_W_4V)
2217 #define __ACCESS_BITS_4U (_PAGE_ACCESSED_4U | _PAGE_READ_4U | _PAGE_R)
2218 #define __ACCESS_BITS_4V (_PAGE_ACCESSED_4V | _PAGE_READ_4V | _PAGE_R)
2219 
2220 pgprot_t PAGE_KERNEL __read_mostly;
2221 EXPORT_SYMBOL(PAGE_KERNEL);
2222 
2223 pgprot_t PAGE_KERNEL_LOCKED __read_mostly;
2224 pgprot_t PAGE_COPY __read_mostly;
2225 
2226 pgprot_t PAGE_SHARED __read_mostly;
2227 EXPORT_SYMBOL(PAGE_SHARED);
2228 
2229 unsigned long pg_iobits __read_mostly;
2230 
2231 unsigned long _PAGE_IE __read_mostly;
2232 EXPORT_SYMBOL(_PAGE_IE);
2233 
2234 unsigned long _PAGE_E __read_mostly;
2235 EXPORT_SYMBOL(_PAGE_E);
2236 
2237 unsigned long _PAGE_CACHE __read_mostly;
2238 EXPORT_SYMBOL(_PAGE_CACHE);
2239 
2240 #ifdef CONFIG_SPARSEMEM_VMEMMAP
2241 unsigned long vmemmap_table[VMEMMAP_SIZE];
2242 
2243 static long __meminitdata addr_start, addr_end;
2244 static int __meminitdata node_start;
2245 
2246 int __meminit vmemmap_populate(unsigned long vstart, unsigned long vend,
2247 			       int node)
2248 {
2249 	unsigned long phys_start = (vstart - VMEMMAP_BASE);
2250 	unsigned long phys_end = (vend - VMEMMAP_BASE);
2251 	unsigned long addr = phys_start & VMEMMAP_CHUNK_MASK;
2252 	unsigned long end = VMEMMAP_ALIGN(phys_end);
2253 	unsigned long pte_base;
2254 
2255 	pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2256 		    _PAGE_CP_4U | _PAGE_CV_4U |
2257 		    _PAGE_P_4U | _PAGE_W_4U);
2258 	if (tlb_type == hypervisor)
2259 		pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2260 			    _PAGE_CP_4V | _PAGE_CV_4V |
2261 			    _PAGE_P_4V | _PAGE_W_4V);
2262 
2263 	for (; addr < end; addr += VMEMMAP_CHUNK) {
2264 		unsigned long *vmem_pp =
2265 			vmemmap_table + (addr >> VMEMMAP_CHUNK_SHIFT);
2266 		void *block;
2267 
2268 		if (!(*vmem_pp & _PAGE_VALID)) {
2269 			block = vmemmap_alloc_block(1UL << 22, node);
2270 			if (!block)
2271 				return -ENOMEM;
2272 
2273 			*vmem_pp = pte_base | __pa(block);
2274 
2275 			/* check to see if we have contiguous blocks */
2276 			if (addr_end != addr || node_start != node) {
2277 				if (addr_start)
2278 					printk(KERN_DEBUG " [%lx-%lx] on node %d\n",
2279 					       addr_start, addr_end-1, node_start);
2280 				addr_start = addr;
2281 				node_start = node;
2282 			}
2283 			addr_end = addr + VMEMMAP_CHUNK;
2284 		}
2285 	}
2286 	return 0;
2287 }
2288 
2289 void __meminit vmemmap_populate_print_last(void)
2290 {
2291 	if (addr_start) {
2292 		printk(KERN_DEBUG " [%lx-%lx] on node %d\n",
2293 		       addr_start, addr_end-1, node_start);
2294 		addr_start = 0;
2295 		addr_end = 0;
2296 		node_start = 0;
2297 	}
2298 }
2299 
2300 void vmemmap_free(unsigned long start, unsigned long end)
2301 {
2302 }
2303 
2304 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
2305 
2306 static void prot_init_common(unsigned long page_none,
2307 			     unsigned long page_shared,
2308 			     unsigned long page_copy,
2309 			     unsigned long page_readonly,
2310 			     unsigned long page_exec_bit)
2311 {
2312 	PAGE_COPY = __pgprot(page_copy);
2313 	PAGE_SHARED = __pgprot(page_shared);
2314 
2315 	protection_map[0x0] = __pgprot(page_none);
2316 	protection_map[0x1] = __pgprot(page_readonly & ~page_exec_bit);
2317 	protection_map[0x2] = __pgprot(page_copy & ~page_exec_bit);
2318 	protection_map[0x3] = __pgprot(page_copy & ~page_exec_bit);
2319 	protection_map[0x4] = __pgprot(page_readonly);
2320 	protection_map[0x5] = __pgprot(page_readonly);
2321 	protection_map[0x6] = __pgprot(page_copy);
2322 	protection_map[0x7] = __pgprot(page_copy);
2323 	protection_map[0x8] = __pgprot(page_none);
2324 	protection_map[0x9] = __pgprot(page_readonly & ~page_exec_bit);
2325 	protection_map[0xa] = __pgprot(page_shared & ~page_exec_bit);
2326 	protection_map[0xb] = __pgprot(page_shared & ~page_exec_bit);
2327 	protection_map[0xc] = __pgprot(page_readonly);
2328 	protection_map[0xd] = __pgprot(page_readonly);
2329 	protection_map[0xe] = __pgprot(page_shared);
2330 	protection_map[0xf] = __pgprot(page_shared);
2331 }
2332 
2333 static void __init sun4u_pgprot_init(void)
2334 {
2335 	unsigned long page_none, page_shared, page_copy, page_readonly;
2336 	unsigned long page_exec_bit;
2337 	int i;
2338 
2339 	PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2340 				_PAGE_CACHE_4U | _PAGE_P_4U |
2341 				__ACCESS_BITS_4U | __DIRTY_BITS_4U |
2342 				_PAGE_EXEC_4U);
2343 	PAGE_KERNEL_LOCKED = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2344 				       _PAGE_CACHE_4U | _PAGE_P_4U |
2345 				       __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2346 				       _PAGE_EXEC_4U | _PAGE_L_4U);
2347 
2348 	_PAGE_IE = _PAGE_IE_4U;
2349 	_PAGE_E = _PAGE_E_4U;
2350 	_PAGE_CACHE = _PAGE_CACHE_4U;
2351 
2352 	pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4U | __DIRTY_BITS_4U |
2353 		     __ACCESS_BITS_4U | _PAGE_E_4U);
2354 
2355 #ifdef CONFIG_DEBUG_PAGEALLOC
2356 	kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2357 #else
2358 	kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4U) ^
2359 		PAGE_OFFSET;
2360 #endif
2361 	kern_linear_pte_xor[0] |= (_PAGE_CP_4U | _PAGE_CV_4U |
2362 				   _PAGE_P_4U | _PAGE_W_4U);
2363 
2364 	for (i = 1; i < 4; i++)
2365 		kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2366 
2367 	_PAGE_ALL_SZ_BITS =  (_PAGE_SZ4MB_4U | _PAGE_SZ512K_4U |
2368 			      _PAGE_SZ64K_4U | _PAGE_SZ8K_4U |
2369 			      _PAGE_SZ32MB_4U | _PAGE_SZ256MB_4U);
2370 
2371 
2372 	page_none = _PAGE_PRESENT_4U | _PAGE_ACCESSED_4U | _PAGE_CACHE_4U;
2373 	page_shared = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2374 		       __ACCESS_BITS_4U | _PAGE_WRITE_4U | _PAGE_EXEC_4U);
2375 	page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2376 		       __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2377 	page_readonly   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2378 			   __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2379 
2380 	page_exec_bit = _PAGE_EXEC_4U;
2381 
2382 	prot_init_common(page_none, page_shared, page_copy, page_readonly,
2383 			 page_exec_bit);
2384 }
2385 
2386 static void __init sun4v_pgprot_init(void)
2387 {
2388 	unsigned long page_none, page_shared, page_copy, page_readonly;
2389 	unsigned long page_exec_bit;
2390 	int i;
2391 
2392 	PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4V | _PAGE_VALID |
2393 				_PAGE_CACHE_4V | _PAGE_P_4V |
2394 				__ACCESS_BITS_4V | __DIRTY_BITS_4V |
2395 				_PAGE_EXEC_4V);
2396 	PAGE_KERNEL_LOCKED = PAGE_KERNEL;
2397 
2398 	_PAGE_IE = _PAGE_IE_4V;
2399 	_PAGE_E = _PAGE_E_4V;
2400 	_PAGE_CACHE = _PAGE_CACHE_4V;
2401 
2402 #ifdef CONFIG_DEBUG_PAGEALLOC
2403 	kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2404 #else
2405 	kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4V) ^
2406 		PAGE_OFFSET;
2407 #endif
2408 	kern_linear_pte_xor[0] |= (_PAGE_CP_4V | _PAGE_CV_4V |
2409 				   _PAGE_P_4V | _PAGE_W_4V);
2410 
2411 	for (i = 1; i < 4; i++)
2412 		kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2413 
2414 	pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4V | __DIRTY_BITS_4V |
2415 		     __ACCESS_BITS_4V | _PAGE_E_4V);
2416 
2417 	_PAGE_ALL_SZ_BITS = (_PAGE_SZ16GB_4V | _PAGE_SZ2GB_4V |
2418 			     _PAGE_SZ256MB_4V | _PAGE_SZ32MB_4V |
2419 			     _PAGE_SZ4MB_4V | _PAGE_SZ512K_4V |
2420 			     _PAGE_SZ64K_4V | _PAGE_SZ8K_4V);
2421 
2422 	page_none = _PAGE_PRESENT_4V | _PAGE_ACCESSED_4V | _PAGE_CACHE_4V;
2423 	page_shared = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2424 		       __ACCESS_BITS_4V | _PAGE_WRITE_4V | _PAGE_EXEC_4V);
2425 	page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2426 		       __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2427 	page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2428 			 __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2429 
2430 	page_exec_bit = _PAGE_EXEC_4V;
2431 
2432 	prot_init_common(page_none, page_shared, page_copy, page_readonly,
2433 			 page_exec_bit);
2434 }
2435 
2436 unsigned long pte_sz_bits(unsigned long sz)
2437 {
2438 	if (tlb_type == hypervisor) {
2439 		switch (sz) {
2440 		case 8 * 1024:
2441 		default:
2442 			return _PAGE_SZ8K_4V;
2443 		case 64 * 1024:
2444 			return _PAGE_SZ64K_4V;
2445 		case 512 * 1024:
2446 			return _PAGE_SZ512K_4V;
2447 		case 4 * 1024 * 1024:
2448 			return _PAGE_SZ4MB_4V;
2449 		}
2450 	} else {
2451 		switch (sz) {
2452 		case 8 * 1024:
2453 		default:
2454 			return _PAGE_SZ8K_4U;
2455 		case 64 * 1024:
2456 			return _PAGE_SZ64K_4U;
2457 		case 512 * 1024:
2458 			return _PAGE_SZ512K_4U;
2459 		case 4 * 1024 * 1024:
2460 			return _PAGE_SZ4MB_4U;
2461 		}
2462 	}
2463 }
2464 
2465 pte_t mk_pte_io(unsigned long page, pgprot_t prot, int space, unsigned long page_size)
2466 {
2467 	pte_t pte;
2468 
2469 	pte_val(pte)  = page | pgprot_val(pgprot_noncached(prot));
2470 	pte_val(pte) |= (((unsigned long)space) << 32);
2471 	pte_val(pte) |= pte_sz_bits(page_size);
2472 
2473 	return pte;
2474 }
2475 
2476 static unsigned long kern_large_tte(unsigned long paddr)
2477 {
2478 	unsigned long val;
2479 
2480 	val = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2481 	       _PAGE_CP_4U | _PAGE_CV_4U | _PAGE_P_4U |
2482 	       _PAGE_EXEC_4U | _PAGE_L_4U | _PAGE_W_4U);
2483 	if (tlb_type == hypervisor)
2484 		val = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2485 		       _PAGE_CP_4V | _PAGE_CV_4V | _PAGE_P_4V |
2486 		       _PAGE_EXEC_4V | _PAGE_W_4V);
2487 
2488 	return val | paddr;
2489 }
2490 
2491 /* If not locked, zap it. */
2492 void __flush_tlb_all(void)
2493 {
2494 	unsigned long pstate;
2495 	int i;
2496 
2497 	__asm__ __volatile__("flushw\n\t"
2498 			     "rdpr	%%pstate, %0\n\t"
2499 			     "wrpr	%0, %1, %%pstate"
2500 			     : "=r" (pstate)
2501 			     : "i" (PSTATE_IE));
2502 	if (tlb_type == hypervisor) {
2503 		sun4v_mmu_demap_all();
2504 	} else if (tlb_type == spitfire) {
2505 		for (i = 0; i < 64; i++) {
2506 			/* Spitfire Errata #32 workaround */
2507 			/* NOTE: Always runs on spitfire, so no
2508 			 *       cheetah+ page size encodings.
2509 			 */
2510 			__asm__ __volatile__("stxa	%0, [%1] %2\n\t"
2511 					     "flush	%%g6"
2512 					     : /* No outputs */
2513 					     : "r" (0),
2514 					     "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2515 
2516 			if (!(spitfire_get_dtlb_data(i) & _PAGE_L_4U)) {
2517 				__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2518 						     "membar #Sync"
2519 						     : /* no outputs */
2520 						     : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
2521 				spitfire_put_dtlb_data(i, 0x0UL);
2522 			}
2523 
2524 			/* Spitfire Errata #32 workaround */
2525 			/* NOTE: Always runs on spitfire, so no
2526 			 *       cheetah+ page size encodings.
2527 			 */
2528 			__asm__ __volatile__("stxa	%0, [%1] %2\n\t"
2529 					     "flush	%%g6"
2530 					     : /* No outputs */
2531 					     : "r" (0),
2532 					     "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2533 
2534 			if (!(spitfire_get_itlb_data(i) & _PAGE_L_4U)) {
2535 				__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2536 						     "membar #Sync"
2537 						     : /* no outputs */
2538 						     : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
2539 				spitfire_put_itlb_data(i, 0x0UL);
2540 			}
2541 		}
2542 	} else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
2543 		cheetah_flush_dtlb_all();
2544 		cheetah_flush_itlb_all();
2545 	}
2546 	__asm__ __volatile__("wrpr	%0, 0, %%pstate"
2547 			     : : "r" (pstate));
2548 }
2549 
2550 pte_t *pte_alloc_one_kernel(struct mm_struct *mm,
2551 			    unsigned long address)
2552 {
2553 	struct page *page = alloc_page(GFP_KERNEL | __GFP_NOTRACK |
2554 				       __GFP_REPEAT | __GFP_ZERO);
2555 	pte_t *pte = NULL;
2556 
2557 	if (page)
2558 		pte = (pte_t *) page_address(page);
2559 
2560 	return pte;
2561 }
2562 
2563 pgtable_t pte_alloc_one(struct mm_struct *mm,
2564 			unsigned long address)
2565 {
2566 	struct page *page = alloc_page(GFP_KERNEL | __GFP_NOTRACK |
2567 				       __GFP_REPEAT | __GFP_ZERO);
2568 	if (!page)
2569 		return NULL;
2570 	if (!pgtable_page_ctor(page)) {
2571 		free_hot_cold_page(page, 0);
2572 		return NULL;
2573 	}
2574 	return (pte_t *) page_address(page);
2575 }
2576 
2577 void pte_free_kernel(struct mm_struct *mm, pte_t *pte)
2578 {
2579 	free_page((unsigned long)pte);
2580 }
2581 
2582 static void __pte_free(pgtable_t pte)
2583 {
2584 	struct page *page = virt_to_page(pte);
2585 
2586 	pgtable_page_dtor(page);
2587 	__free_page(page);
2588 }
2589 
2590 void pte_free(struct mm_struct *mm, pgtable_t pte)
2591 {
2592 	__pte_free(pte);
2593 }
2594 
2595 void pgtable_free(void *table, bool is_page)
2596 {
2597 	if (is_page)
2598 		__pte_free(table);
2599 	else
2600 		kmem_cache_free(pgtable_cache, table);
2601 }
2602 
2603 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2604 void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
2605 			  pmd_t *pmd)
2606 {
2607 	unsigned long pte, flags;
2608 	struct mm_struct *mm;
2609 	pmd_t entry = *pmd;
2610 
2611 	if (!pmd_large(entry) || !pmd_young(entry))
2612 		return;
2613 
2614 	pte = pmd_val(entry);
2615 
2616 	/* We are fabricating 8MB pages using 4MB real hw pages.  */
2617 	pte |= (addr & (1UL << REAL_HPAGE_SHIFT));
2618 
2619 	mm = vma->vm_mm;
2620 
2621 	spin_lock_irqsave(&mm->context.lock, flags);
2622 
2623 	if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL)
2624 		__update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
2625 					addr, pte);
2626 
2627 	spin_unlock_irqrestore(&mm->context.lock, flags);
2628 }
2629 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2630 
2631 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
2632 static void context_reload(void *__data)
2633 {
2634 	struct mm_struct *mm = __data;
2635 
2636 	if (mm == current->mm)
2637 		load_secondary_context(mm);
2638 }
2639 
2640 void hugetlb_setup(struct pt_regs *regs)
2641 {
2642 	struct mm_struct *mm = current->mm;
2643 	struct tsb_config *tp;
2644 
2645 	if (in_atomic() || !mm) {
2646 		const struct exception_table_entry *entry;
2647 
2648 		entry = search_exception_tables(regs->tpc);
2649 		if (entry) {
2650 			regs->tpc = entry->fixup;
2651 			regs->tnpc = regs->tpc + 4;
2652 			return;
2653 		}
2654 		pr_alert("Unexpected HugeTLB setup in atomic context.\n");
2655 		die_if_kernel("HugeTSB in atomic", regs);
2656 	}
2657 
2658 	tp = &mm->context.tsb_block[MM_TSB_HUGE];
2659 	if (likely(tp->tsb == NULL))
2660 		tsb_grow(mm, MM_TSB_HUGE, 0);
2661 
2662 	tsb_context_switch(mm);
2663 	smp_tsb_sync(mm);
2664 
2665 	/* On UltraSPARC-III+ and later, configure the second half of
2666 	 * the Data-TLB for huge pages.
2667 	 */
2668 	if (tlb_type == cheetah_plus) {
2669 		unsigned long ctx;
2670 
2671 		spin_lock(&ctx_alloc_lock);
2672 		ctx = mm->context.sparc64_ctx_val;
2673 		ctx &= ~CTX_PGSZ_MASK;
2674 		ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT;
2675 		ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT;
2676 
2677 		if (ctx != mm->context.sparc64_ctx_val) {
2678 			/* When changing the page size fields, we
2679 			 * must perform a context flush so that no
2680 			 * stale entries match.  This flush must
2681 			 * occur with the original context register
2682 			 * settings.
2683 			 */
2684 			do_flush_tlb_mm(mm);
2685 
2686 			/* Reload the context register of all processors
2687 			 * also executing in this address space.
2688 			 */
2689 			mm->context.sparc64_ctx_val = ctx;
2690 			on_each_cpu(context_reload, mm, 0);
2691 		}
2692 		spin_unlock(&ctx_alloc_lock);
2693 	}
2694 }
2695 #endif
2696