xref: /linux/arch/sparc/mm/init_64.c (revision 9938b04472d5c59f8bd8152a548533a8599596a2)
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/ioport.h>
26 #include <linux/percpu.h>
27 #include <linux/memblock.h>
28 #include <linux/mmzone.h>
29 #include <linux/gfp.h>
30 
31 #include <asm/head.h>
32 #include <asm/page.h>
33 #include <asm/pgalloc.h>
34 #include <asm/pgtable.h>
35 #include <asm/oplib.h>
36 #include <asm/iommu.h>
37 #include <asm/io.h>
38 #include <asm/uaccess.h>
39 #include <asm/mmu_context.h>
40 #include <asm/tlbflush.h>
41 #include <asm/dma.h>
42 #include <asm/starfire.h>
43 #include <asm/tlb.h>
44 #include <asm/spitfire.h>
45 #include <asm/sections.h>
46 #include <asm/tsb.h>
47 #include <asm/hypervisor.h>
48 #include <asm/prom.h>
49 #include <asm/mdesc.h>
50 #include <asm/cpudata.h>
51 #include <asm/setup.h>
52 #include <asm/irq.h>
53 
54 #include "init_64.h"
55 
56 unsigned long kern_linear_pte_xor[4] __read_mostly;
57 static unsigned long page_cache4v_flag;
58 
59 /* A bitmap, two bits for every 256MB of physical memory.  These two
60  * bits determine what page size we use for kernel linear
61  * translations.  They form an index into kern_linear_pte_xor[].  The
62  * value in the indexed slot is XOR'd with the TLB miss virtual
63  * address to form the resulting TTE.  The mapping is:
64  *
65  *	0	==>	4MB
66  *	1	==>	256MB
67  *	2	==>	2GB
68  *	3	==>	16GB
69  *
70  * All sun4v chips support 256MB pages.  Only SPARC-T4 and later
71  * support 2GB pages, and hopefully future cpus will support the 16GB
72  * pages as well.  For slots 2 and 3, we encode a 256MB TTE xor there
73  * if these larger page sizes are not supported by the cpu.
74  *
75  * It would be nice to determine this from the machine description
76  * 'cpu' properties, but we need to have this table setup before the
77  * MDESC is initialized.
78  */
79 
80 #ifndef CONFIG_DEBUG_PAGEALLOC
81 /* A special kernel TSB for 4MB, 256MB, 2GB and 16GB linear mappings.
82  * Space is allocated for this right after the trap table in
83  * arch/sparc64/kernel/head.S
84  */
85 extern struct tsb swapper_4m_tsb[KERNEL_TSB4M_NENTRIES];
86 #endif
87 extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
88 
89 static unsigned long cpu_pgsz_mask;
90 
91 #define MAX_BANKS	1024
92 
93 static struct linux_prom64_registers pavail[MAX_BANKS];
94 static int pavail_ents;
95 
96 u64 numa_latency[MAX_NUMNODES][MAX_NUMNODES];
97 
98 static int cmp_p64(const void *a, const void *b)
99 {
100 	const struct linux_prom64_registers *x = a, *y = b;
101 
102 	if (x->phys_addr > y->phys_addr)
103 		return 1;
104 	if (x->phys_addr < y->phys_addr)
105 		return -1;
106 	return 0;
107 }
108 
109 static void __init read_obp_memory(const char *property,
110 				   struct linux_prom64_registers *regs,
111 				   int *num_ents)
112 {
113 	phandle node = prom_finddevice("/memory");
114 	int prop_size = prom_getproplen(node, property);
115 	int ents, ret, i;
116 
117 	ents = prop_size / sizeof(struct linux_prom64_registers);
118 	if (ents > MAX_BANKS) {
119 		prom_printf("The machine has more %s property entries than "
120 			    "this kernel can support (%d).\n",
121 			    property, MAX_BANKS);
122 		prom_halt();
123 	}
124 
125 	ret = prom_getproperty(node, property, (char *) regs, prop_size);
126 	if (ret == -1) {
127 		prom_printf("Couldn't get %s property from /memory.\n",
128 				property);
129 		prom_halt();
130 	}
131 
132 	/* Sanitize what we got from the firmware, by page aligning
133 	 * everything.
134 	 */
135 	for (i = 0; i < ents; i++) {
136 		unsigned long base, size;
137 
138 		base = regs[i].phys_addr;
139 		size = regs[i].reg_size;
140 
141 		size &= PAGE_MASK;
142 		if (base & ~PAGE_MASK) {
143 			unsigned long new_base = PAGE_ALIGN(base);
144 
145 			size -= new_base - base;
146 			if ((long) size < 0L)
147 				size = 0UL;
148 			base = new_base;
149 		}
150 		if (size == 0UL) {
151 			/* If it is empty, simply get rid of it.
152 			 * This simplifies the logic of the other
153 			 * functions that process these arrays.
154 			 */
155 			memmove(&regs[i], &regs[i + 1],
156 				(ents - i - 1) * sizeof(regs[0]));
157 			i--;
158 			ents--;
159 			continue;
160 		}
161 		regs[i].phys_addr = base;
162 		regs[i].reg_size = size;
163 	}
164 
165 	*num_ents = ents;
166 
167 	sort(regs, ents, sizeof(struct linux_prom64_registers),
168 	     cmp_p64, NULL);
169 }
170 
171 /* Kernel physical address base and size in bytes.  */
172 unsigned long kern_base __read_mostly;
173 unsigned long kern_size __read_mostly;
174 
175 /* Initial ramdisk setup */
176 extern unsigned long sparc_ramdisk_image64;
177 extern unsigned int sparc_ramdisk_image;
178 extern unsigned int sparc_ramdisk_size;
179 
180 struct page *mem_map_zero __read_mostly;
181 EXPORT_SYMBOL(mem_map_zero);
182 
183 unsigned int sparc64_highest_unlocked_tlb_ent __read_mostly;
184 
185 unsigned long sparc64_kern_pri_context __read_mostly;
186 unsigned long sparc64_kern_pri_nuc_bits __read_mostly;
187 unsigned long sparc64_kern_sec_context __read_mostly;
188 
189 int num_kernel_image_mappings;
190 
191 #ifdef CONFIG_DEBUG_DCFLUSH
192 atomic_t dcpage_flushes = ATOMIC_INIT(0);
193 #ifdef CONFIG_SMP
194 atomic_t dcpage_flushes_xcall = ATOMIC_INIT(0);
195 #endif
196 #endif
197 
198 inline void flush_dcache_page_impl(struct page *page)
199 {
200 	BUG_ON(tlb_type == hypervisor);
201 #ifdef CONFIG_DEBUG_DCFLUSH
202 	atomic_inc(&dcpage_flushes);
203 #endif
204 
205 #ifdef DCACHE_ALIASING_POSSIBLE
206 	__flush_dcache_page(page_address(page),
207 			    ((tlb_type == spitfire) &&
208 			     page_mapping(page) != NULL));
209 #else
210 	if (page_mapping(page) != NULL &&
211 	    tlb_type == spitfire)
212 		__flush_icache_page(__pa(page_address(page)));
213 #endif
214 }
215 
216 #define PG_dcache_dirty		PG_arch_1
217 #define PG_dcache_cpu_shift	32UL
218 #define PG_dcache_cpu_mask	\
219 	((1UL<<ilog2(roundup_pow_of_two(NR_CPUS)))-1UL)
220 
221 #define dcache_dirty_cpu(page) \
222 	(((page)->flags >> PG_dcache_cpu_shift) & PG_dcache_cpu_mask)
223 
224 static inline void set_dcache_dirty(struct page *page, int this_cpu)
225 {
226 	unsigned long mask = this_cpu;
227 	unsigned long non_cpu_bits;
228 
229 	non_cpu_bits = ~(PG_dcache_cpu_mask << PG_dcache_cpu_shift);
230 	mask = (mask << PG_dcache_cpu_shift) | (1UL << PG_dcache_dirty);
231 
232 	__asm__ __volatile__("1:\n\t"
233 			     "ldx	[%2], %%g7\n\t"
234 			     "and	%%g7, %1, %%g1\n\t"
235 			     "or	%%g1, %0, %%g1\n\t"
236 			     "casx	[%2], %%g7, %%g1\n\t"
237 			     "cmp	%%g7, %%g1\n\t"
238 			     "bne,pn	%%xcc, 1b\n\t"
239 			     " nop"
240 			     : /* no outputs */
241 			     : "r" (mask), "r" (non_cpu_bits), "r" (&page->flags)
242 			     : "g1", "g7");
243 }
244 
245 static inline void clear_dcache_dirty_cpu(struct page *page, unsigned long cpu)
246 {
247 	unsigned long mask = (1UL << PG_dcache_dirty);
248 
249 	__asm__ __volatile__("! test_and_clear_dcache_dirty\n"
250 			     "1:\n\t"
251 			     "ldx	[%2], %%g7\n\t"
252 			     "srlx	%%g7, %4, %%g1\n\t"
253 			     "and	%%g1, %3, %%g1\n\t"
254 			     "cmp	%%g1, %0\n\t"
255 			     "bne,pn	%%icc, 2f\n\t"
256 			     " andn	%%g7, %1, %%g1\n\t"
257 			     "casx	[%2], %%g7, %%g1\n\t"
258 			     "cmp	%%g7, %%g1\n\t"
259 			     "bne,pn	%%xcc, 1b\n\t"
260 			     " nop\n"
261 			     "2:"
262 			     : /* no outputs */
263 			     : "r" (cpu), "r" (mask), "r" (&page->flags),
264 			       "i" (PG_dcache_cpu_mask),
265 			       "i" (PG_dcache_cpu_shift)
266 			     : "g1", "g7");
267 }
268 
269 static inline void tsb_insert(struct tsb *ent, unsigned long tag, unsigned long pte)
270 {
271 	unsigned long tsb_addr = (unsigned long) ent;
272 
273 	if (tlb_type == cheetah_plus || tlb_type == hypervisor)
274 		tsb_addr = __pa(tsb_addr);
275 
276 	__tsb_insert(tsb_addr, tag, pte);
277 }
278 
279 unsigned long _PAGE_ALL_SZ_BITS __read_mostly;
280 
281 static void flush_dcache(unsigned long pfn)
282 {
283 	struct page *page;
284 
285 	page = pfn_to_page(pfn);
286 	if (page) {
287 		unsigned long pg_flags;
288 
289 		pg_flags = page->flags;
290 		if (pg_flags & (1UL << PG_dcache_dirty)) {
291 			int cpu = ((pg_flags >> PG_dcache_cpu_shift) &
292 				   PG_dcache_cpu_mask);
293 			int this_cpu = get_cpu();
294 
295 			/* This is just to optimize away some function calls
296 			 * in the SMP case.
297 			 */
298 			if (cpu == this_cpu)
299 				flush_dcache_page_impl(page);
300 			else
301 				smp_flush_dcache_page_impl(page, cpu);
302 
303 			clear_dcache_dirty_cpu(page, cpu);
304 
305 			put_cpu();
306 		}
307 	}
308 }
309 
310 /* mm->context.lock must be held */
311 static void __update_mmu_tsb_insert(struct mm_struct *mm, unsigned long tsb_index,
312 				    unsigned long tsb_hash_shift, unsigned long address,
313 				    unsigned long tte)
314 {
315 	struct tsb *tsb = mm->context.tsb_block[tsb_index].tsb;
316 	unsigned long tag;
317 
318 	if (unlikely(!tsb))
319 		return;
320 
321 	tsb += ((address >> tsb_hash_shift) &
322 		(mm->context.tsb_block[tsb_index].tsb_nentries - 1UL));
323 	tag = (address >> 22UL);
324 	tsb_insert(tsb, tag, tte);
325 }
326 
327 #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 	/* Don't insert a non-valid PTE into the TSB, we'll deadlock.  */
355 	if (!pte_accessible(mm, pte))
356 		return;
357 
358 	spin_lock_irqsave(&mm->context.lock, flags);
359 
360 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
361 	if (mm->context.huge_pte_count && is_hugetlb_pte(pte))
362 		__update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
363 					address, pte_val(pte));
364 	else
365 #endif
366 		__update_mmu_tsb_insert(mm, MM_TSB_BASE, PAGE_SHIFT,
367 					address, pte_val(pte));
368 
369 	spin_unlock_irqrestore(&mm->context.lock, flags);
370 }
371 
372 void flush_dcache_page(struct page *page)
373 {
374 	struct address_space *mapping;
375 	int this_cpu;
376 
377 	if (tlb_type == hypervisor)
378 		return;
379 
380 	/* Do not bother with the expensive D-cache flush if it
381 	 * is merely the zero page.  The 'bigcore' testcase in GDB
382 	 * causes this case to run millions of times.
383 	 */
384 	if (page == ZERO_PAGE(0))
385 		return;
386 
387 	this_cpu = get_cpu();
388 
389 	mapping = page_mapping(page);
390 	if (mapping && !mapping_mapped(mapping)) {
391 		int dirty = test_bit(PG_dcache_dirty, &page->flags);
392 		if (dirty) {
393 			int dirty_cpu = dcache_dirty_cpu(page);
394 
395 			if (dirty_cpu == this_cpu)
396 				goto out;
397 			smp_flush_dcache_page_impl(page, dirty_cpu);
398 		}
399 		set_dcache_dirty(page, this_cpu);
400 	} else {
401 		/* We could delay the flush for the !page_mapping
402 		 * case too.  But that case is for exec env/arg
403 		 * pages and those are %99 certainly going to get
404 		 * faulted into the tlb (and thus flushed) anyways.
405 		 */
406 		flush_dcache_page_impl(page);
407 	}
408 
409 out:
410 	put_cpu();
411 }
412 EXPORT_SYMBOL(flush_dcache_page);
413 
414 void __kprobes flush_icache_range(unsigned long start, unsigned long end)
415 {
416 	/* Cheetah and Hypervisor platform cpus have coherent I-cache. */
417 	if (tlb_type == spitfire) {
418 		unsigned long kaddr;
419 
420 		/* This code only runs on Spitfire cpus so this is
421 		 * why we can assume _PAGE_PADDR_4U.
422 		 */
423 		for (kaddr = start; kaddr < end; kaddr += PAGE_SIZE) {
424 			unsigned long paddr, mask = _PAGE_PADDR_4U;
425 
426 			if (kaddr >= PAGE_OFFSET)
427 				paddr = kaddr & mask;
428 			else {
429 				pgd_t *pgdp = pgd_offset_k(kaddr);
430 				pud_t *pudp = pud_offset(pgdp, kaddr);
431 				pmd_t *pmdp = pmd_offset(pudp, kaddr);
432 				pte_t *ptep = pte_offset_kernel(pmdp, kaddr);
433 
434 				paddr = pte_val(*ptep) & mask;
435 			}
436 			__flush_icache_page(paddr);
437 		}
438 	}
439 }
440 EXPORT_SYMBOL(flush_icache_range);
441 
442 void mmu_info(struct seq_file *m)
443 {
444 	static const char *pgsz_strings[] = {
445 		"8K", "64K", "512K", "4MB", "32MB",
446 		"256MB", "2GB", "16GB",
447 	};
448 	int i, printed;
449 
450 	if (tlb_type == cheetah)
451 		seq_printf(m, "MMU Type\t: Cheetah\n");
452 	else if (tlb_type == cheetah_plus)
453 		seq_printf(m, "MMU Type\t: Cheetah+\n");
454 	else if (tlb_type == spitfire)
455 		seq_printf(m, "MMU Type\t: Spitfire\n");
456 	else if (tlb_type == hypervisor)
457 		seq_printf(m, "MMU Type\t: Hypervisor (sun4v)\n");
458 	else
459 		seq_printf(m, "MMU Type\t: ???\n");
460 
461 	seq_printf(m, "MMU PGSZs\t: ");
462 	printed = 0;
463 	for (i = 0; i < ARRAY_SIZE(pgsz_strings); i++) {
464 		if (cpu_pgsz_mask & (1UL << i)) {
465 			seq_printf(m, "%s%s",
466 				   printed ? "," : "", pgsz_strings[i]);
467 			printed++;
468 		}
469 	}
470 	seq_putc(m, '\n');
471 
472 #ifdef CONFIG_DEBUG_DCFLUSH
473 	seq_printf(m, "DCPageFlushes\t: %d\n",
474 		   atomic_read(&dcpage_flushes));
475 #ifdef CONFIG_SMP
476 	seq_printf(m, "DCPageFlushesXC\t: %d\n",
477 		   atomic_read(&dcpage_flushes_xcall));
478 #endif /* CONFIG_SMP */
479 #endif /* CONFIG_DEBUG_DCFLUSH */
480 }
481 
482 struct linux_prom_translation prom_trans[512] __read_mostly;
483 unsigned int prom_trans_ents __read_mostly;
484 
485 unsigned long kern_locked_tte_data;
486 
487 /* The obp translations are saved based on 8k pagesize, since obp can
488  * use a mixture of pagesizes. Misses to the LOW_OBP_ADDRESS ->
489  * HI_OBP_ADDRESS range are handled in ktlb.S.
490  */
491 static inline int in_obp_range(unsigned long vaddr)
492 {
493 	return (vaddr >= LOW_OBP_ADDRESS &&
494 		vaddr < HI_OBP_ADDRESS);
495 }
496 
497 static int cmp_ptrans(const void *a, const void *b)
498 {
499 	const struct linux_prom_translation *x = a, *y = b;
500 
501 	if (x->virt > y->virt)
502 		return 1;
503 	if (x->virt < y->virt)
504 		return -1;
505 	return 0;
506 }
507 
508 /* Read OBP translations property into 'prom_trans[]'.  */
509 static void __init read_obp_translations(void)
510 {
511 	int n, node, ents, first, last, i;
512 
513 	node = prom_finddevice("/virtual-memory");
514 	n = prom_getproplen(node, "translations");
515 	if (unlikely(n == 0 || n == -1)) {
516 		prom_printf("prom_mappings: Couldn't get size.\n");
517 		prom_halt();
518 	}
519 	if (unlikely(n > sizeof(prom_trans))) {
520 		prom_printf("prom_mappings: Size %d is too big.\n", n);
521 		prom_halt();
522 	}
523 
524 	if ((n = prom_getproperty(node, "translations",
525 				  (char *)&prom_trans[0],
526 				  sizeof(prom_trans))) == -1) {
527 		prom_printf("prom_mappings: Couldn't get property.\n");
528 		prom_halt();
529 	}
530 
531 	n = n / sizeof(struct linux_prom_translation);
532 
533 	ents = n;
534 
535 	sort(prom_trans, ents, sizeof(struct linux_prom_translation),
536 	     cmp_ptrans, NULL);
537 
538 	/* Now kick out all the non-OBP entries.  */
539 	for (i = 0; i < ents; i++) {
540 		if (in_obp_range(prom_trans[i].virt))
541 			break;
542 	}
543 	first = i;
544 	for (; i < ents; i++) {
545 		if (!in_obp_range(prom_trans[i].virt))
546 			break;
547 	}
548 	last = i;
549 
550 	for (i = 0; i < (last - first); i++) {
551 		struct linux_prom_translation *src = &prom_trans[i + first];
552 		struct linux_prom_translation *dest = &prom_trans[i];
553 
554 		*dest = *src;
555 	}
556 	for (; i < ents; i++) {
557 		struct linux_prom_translation *dest = &prom_trans[i];
558 		dest->virt = dest->size = dest->data = 0x0UL;
559 	}
560 
561 	prom_trans_ents = last - first;
562 
563 	if (tlb_type == spitfire) {
564 		/* Clear diag TTE bits. */
565 		for (i = 0; i < prom_trans_ents; i++)
566 			prom_trans[i].data &= ~0x0003fe0000000000UL;
567 	}
568 
569 	/* Force execute bit on.  */
570 	for (i = 0; i < prom_trans_ents; i++)
571 		prom_trans[i].data |= (tlb_type == hypervisor ?
572 				       _PAGE_EXEC_4V : _PAGE_EXEC_4U);
573 }
574 
575 static void __init hypervisor_tlb_lock(unsigned long vaddr,
576 				       unsigned long pte,
577 				       unsigned long mmu)
578 {
579 	unsigned long ret = sun4v_mmu_map_perm_addr(vaddr, 0, pte, mmu);
580 
581 	if (ret != 0) {
582 		prom_printf("hypervisor_tlb_lock[%lx:%x:%lx:%lx]: "
583 			    "errors with %lx\n", vaddr, 0, pte, mmu, ret);
584 		prom_halt();
585 	}
586 }
587 
588 static unsigned long kern_large_tte(unsigned long paddr);
589 
590 static void __init remap_kernel(void)
591 {
592 	unsigned long phys_page, tte_vaddr, tte_data;
593 	int i, tlb_ent = sparc64_highest_locked_tlbent();
594 
595 	tte_vaddr = (unsigned long) KERNBASE;
596 	phys_page = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
597 	tte_data = kern_large_tte(phys_page);
598 
599 	kern_locked_tte_data = tte_data;
600 
601 	/* Now lock us into the TLBs via Hypervisor or OBP. */
602 	if (tlb_type == hypervisor) {
603 		for (i = 0; i < num_kernel_image_mappings; i++) {
604 			hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_DMMU);
605 			hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_IMMU);
606 			tte_vaddr += 0x400000;
607 			tte_data += 0x400000;
608 		}
609 	} else {
610 		for (i = 0; i < num_kernel_image_mappings; i++) {
611 			prom_dtlb_load(tlb_ent - i, tte_data, tte_vaddr);
612 			prom_itlb_load(tlb_ent - i, tte_data, tte_vaddr);
613 			tte_vaddr += 0x400000;
614 			tte_data += 0x400000;
615 		}
616 		sparc64_highest_unlocked_tlb_ent = tlb_ent - i;
617 	}
618 	if (tlb_type == cheetah_plus) {
619 		sparc64_kern_pri_context = (CTX_CHEETAH_PLUS_CTX0 |
620 					    CTX_CHEETAH_PLUS_NUC);
621 		sparc64_kern_pri_nuc_bits = CTX_CHEETAH_PLUS_NUC;
622 		sparc64_kern_sec_context = CTX_CHEETAH_PLUS_CTX0;
623 	}
624 }
625 
626 
627 static void __init inherit_prom_mappings(void)
628 {
629 	/* Now fixup OBP's idea about where we really are mapped. */
630 	printk("Remapping the kernel... ");
631 	remap_kernel();
632 	printk("done.\n");
633 }
634 
635 void prom_world(int enter)
636 {
637 	if (!enter)
638 		set_fs(get_fs());
639 
640 	__asm__ __volatile__("flushw");
641 }
642 
643 void __flush_dcache_range(unsigned long start, unsigned long end)
644 {
645 	unsigned long va;
646 
647 	if (tlb_type == spitfire) {
648 		int n = 0;
649 
650 		for (va = start; va < end; va += 32) {
651 			spitfire_put_dcache_tag(va & 0x3fe0, 0x0);
652 			if (++n >= 512)
653 				break;
654 		}
655 	} else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
656 		start = __pa(start);
657 		end = __pa(end);
658 		for (va = start; va < end; va += 32)
659 			__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
660 					     "membar #Sync"
661 					     : /* no outputs */
662 					     : "r" (va),
663 					       "i" (ASI_DCACHE_INVALIDATE));
664 	}
665 }
666 EXPORT_SYMBOL(__flush_dcache_range);
667 
668 /* get_new_mmu_context() uses "cache + 1".  */
669 DEFINE_SPINLOCK(ctx_alloc_lock);
670 unsigned long tlb_context_cache = CTX_FIRST_VERSION - 1;
671 #define MAX_CTX_NR	(1UL << CTX_NR_BITS)
672 #define CTX_BMAP_SLOTS	BITS_TO_LONGS(MAX_CTX_NR)
673 DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR);
674 
675 /* Caller does TLB context flushing on local CPU if necessary.
676  * The caller also ensures that CTX_VALID(mm->context) is false.
677  *
678  * We must be careful about boundary cases so that we never
679  * let the user have CTX 0 (nucleus) or we ever use a CTX
680  * version of zero (and thus NO_CONTEXT would not be caught
681  * by version mis-match tests in mmu_context.h).
682  *
683  * Always invoked with interrupts disabled.
684  */
685 void get_new_mmu_context(struct mm_struct *mm)
686 {
687 	unsigned long ctx, new_ctx;
688 	unsigned long orig_pgsz_bits;
689 	int new_version;
690 
691 	spin_lock(&ctx_alloc_lock);
692 	orig_pgsz_bits = (mm->context.sparc64_ctx_val & CTX_PGSZ_MASK);
693 	ctx = (tlb_context_cache + 1) & CTX_NR_MASK;
694 	new_ctx = find_next_zero_bit(mmu_context_bmap, 1 << CTX_NR_BITS, ctx);
695 	new_version = 0;
696 	if (new_ctx >= (1 << CTX_NR_BITS)) {
697 		new_ctx = find_next_zero_bit(mmu_context_bmap, ctx, 1);
698 		if (new_ctx >= ctx) {
699 			int i;
700 			new_ctx = (tlb_context_cache & CTX_VERSION_MASK) +
701 				CTX_FIRST_VERSION;
702 			if (new_ctx == 1)
703 				new_ctx = CTX_FIRST_VERSION;
704 
705 			/* Don't call memset, for 16 entries that's just
706 			 * plain silly...
707 			 */
708 			mmu_context_bmap[0] = 3;
709 			mmu_context_bmap[1] = 0;
710 			mmu_context_bmap[2] = 0;
711 			mmu_context_bmap[3] = 0;
712 			for (i = 4; i < CTX_BMAP_SLOTS; i += 4) {
713 				mmu_context_bmap[i + 0] = 0;
714 				mmu_context_bmap[i + 1] = 0;
715 				mmu_context_bmap[i + 2] = 0;
716 				mmu_context_bmap[i + 3] = 0;
717 			}
718 			new_version = 1;
719 			goto out;
720 		}
721 	}
722 	mmu_context_bmap[new_ctx>>6] |= (1UL << (new_ctx & 63));
723 	new_ctx |= (tlb_context_cache & CTX_VERSION_MASK);
724 out:
725 	tlb_context_cache = new_ctx;
726 	mm->context.sparc64_ctx_val = new_ctx | orig_pgsz_bits;
727 	spin_unlock(&ctx_alloc_lock);
728 
729 	if (unlikely(new_version))
730 		smp_new_mmu_context_version();
731 }
732 
733 static int numa_enabled = 1;
734 static int numa_debug;
735 
736 static int __init early_numa(char *p)
737 {
738 	if (!p)
739 		return 0;
740 
741 	if (strstr(p, "off"))
742 		numa_enabled = 0;
743 
744 	if (strstr(p, "debug"))
745 		numa_debug = 1;
746 
747 	return 0;
748 }
749 early_param("numa", early_numa);
750 
751 #define numadbg(f, a...) \
752 do {	if (numa_debug) \
753 		printk(KERN_INFO f, ## a); \
754 } while (0)
755 
756 static void __init find_ramdisk(unsigned long phys_base)
757 {
758 #ifdef CONFIG_BLK_DEV_INITRD
759 	if (sparc_ramdisk_image || sparc_ramdisk_image64) {
760 		unsigned long ramdisk_image;
761 
762 		/* Older versions of the bootloader only supported a
763 		 * 32-bit physical address for the ramdisk image
764 		 * location, stored at sparc_ramdisk_image.  Newer
765 		 * SILO versions set sparc_ramdisk_image to zero and
766 		 * provide a full 64-bit physical address at
767 		 * sparc_ramdisk_image64.
768 		 */
769 		ramdisk_image = sparc_ramdisk_image;
770 		if (!ramdisk_image)
771 			ramdisk_image = sparc_ramdisk_image64;
772 
773 		/* Another bootloader quirk.  The bootloader normalizes
774 		 * the physical address to KERNBASE, so we have to
775 		 * factor that back out and add in the lowest valid
776 		 * physical page address to get the true physical address.
777 		 */
778 		ramdisk_image -= KERNBASE;
779 		ramdisk_image += phys_base;
780 
781 		numadbg("Found ramdisk at physical address 0x%lx, size %u\n",
782 			ramdisk_image, sparc_ramdisk_size);
783 
784 		initrd_start = ramdisk_image;
785 		initrd_end = ramdisk_image + sparc_ramdisk_size;
786 
787 		memblock_reserve(initrd_start, sparc_ramdisk_size);
788 
789 		initrd_start += PAGE_OFFSET;
790 		initrd_end += PAGE_OFFSET;
791 	}
792 #endif
793 }
794 
795 struct node_mem_mask {
796 	unsigned long mask;
797 	unsigned long val;
798 };
799 static struct node_mem_mask node_masks[MAX_NUMNODES];
800 static int num_node_masks;
801 
802 #ifdef CONFIG_NEED_MULTIPLE_NODES
803 
804 int numa_cpu_lookup_table[NR_CPUS];
805 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
806 
807 struct mdesc_mblock {
808 	u64	base;
809 	u64	size;
810 	u64	offset; /* RA-to-PA */
811 };
812 static struct mdesc_mblock *mblocks;
813 static int num_mblocks;
814 
815 static unsigned long ra_to_pa(unsigned long addr)
816 {
817 	int i;
818 
819 	for (i = 0; i < num_mblocks; i++) {
820 		struct mdesc_mblock *m = &mblocks[i];
821 
822 		if (addr >= m->base &&
823 		    addr < (m->base + m->size)) {
824 			addr += m->offset;
825 			break;
826 		}
827 	}
828 	return addr;
829 }
830 
831 static int find_node(unsigned long addr)
832 {
833 	int i;
834 
835 	addr = ra_to_pa(addr);
836 	for (i = 0; i < num_node_masks; i++) {
837 		struct node_mem_mask *p = &node_masks[i];
838 
839 		if ((addr & p->mask) == p->val)
840 			return i;
841 	}
842 	/* The following condition has been observed on LDOM guests.*/
843 	WARN_ONCE(1, "find_node: A physical address doesn't match a NUMA node"
844 		" rule. Some physical memory will be owned by node 0.");
845 	return 0;
846 }
847 
848 static u64 memblock_nid_range(u64 start, u64 end, int *nid)
849 {
850 	*nid = find_node(start);
851 	start += PAGE_SIZE;
852 	while (start < end) {
853 		int n = find_node(start);
854 
855 		if (n != *nid)
856 			break;
857 		start += PAGE_SIZE;
858 	}
859 
860 	if (start > end)
861 		start = end;
862 
863 	return start;
864 }
865 #endif
866 
867 /* This must be invoked after performing all of the necessary
868  * memblock_set_node() calls for 'nid'.  We need to be able to get
869  * correct data from get_pfn_range_for_nid().
870  */
871 static void __init allocate_node_data(int nid)
872 {
873 	struct pglist_data *p;
874 	unsigned long start_pfn, end_pfn;
875 #ifdef CONFIG_NEED_MULTIPLE_NODES
876 	unsigned long paddr;
877 
878 	paddr = memblock_alloc_try_nid(sizeof(struct pglist_data), SMP_CACHE_BYTES, nid);
879 	if (!paddr) {
880 		prom_printf("Cannot allocate pglist_data for nid[%d]\n", nid);
881 		prom_halt();
882 	}
883 	NODE_DATA(nid) = __va(paddr);
884 	memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
885 
886 	NODE_DATA(nid)->node_id = nid;
887 #endif
888 
889 	p = NODE_DATA(nid);
890 
891 	get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
892 	p->node_start_pfn = start_pfn;
893 	p->node_spanned_pages = end_pfn - start_pfn;
894 }
895 
896 static void init_node_masks_nonnuma(void)
897 {
898 #ifdef CONFIG_NEED_MULTIPLE_NODES
899 	int i;
900 #endif
901 
902 	numadbg("Initializing tables for non-numa.\n");
903 
904 	node_masks[0].mask = node_masks[0].val = 0;
905 	num_node_masks = 1;
906 
907 #ifdef CONFIG_NEED_MULTIPLE_NODES
908 	for (i = 0; i < NR_CPUS; i++)
909 		numa_cpu_lookup_table[i] = 0;
910 
911 	cpumask_setall(&numa_cpumask_lookup_table[0]);
912 #endif
913 }
914 
915 #ifdef CONFIG_NEED_MULTIPLE_NODES
916 struct pglist_data *node_data[MAX_NUMNODES];
917 
918 EXPORT_SYMBOL(numa_cpu_lookup_table);
919 EXPORT_SYMBOL(numa_cpumask_lookup_table);
920 EXPORT_SYMBOL(node_data);
921 
922 struct mdesc_mlgroup {
923 	u64	node;
924 	u64	latency;
925 	u64	match;
926 	u64	mask;
927 };
928 static struct mdesc_mlgroup *mlgroups;
929 static int num_mlgroups;
930 
931 static int scan_pio_for_cfg_handle(struct mdesc_handle *md, u64 pio,
932 				   u32 cfg_handle)
933 {
934 	u64 arc;
935 
936 	mdesc_for_each_arc(arc, md, pio, MDESC_ARC_TYPE_FWD) {
937 		u64 target = mdesc_arc_target(md, arc);
938 		const u64 *val;
939 
940 		val = mdesc_get_property(md, target,
941 					 "cfg-handle", NULL);
942 		if (val && *val == cfg_handle)
943 			return 0;
944 	}
945 	return -ENODEV;
946 }
947 
948 static int scan_arcs_for_cfg_handle(struct mdesc_handle *md, u64 grp,
949 				    u32 cfg_handle)
950 {
951 	u64 arc, candidate, best_latency = ~(u64)0;
952 
953 	candidate = MDESC_NODE_NULL;
954 	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
955 		u64 target = mdesc_arc_target(md, arc);
956 		const char *name = mdesc_node_name(md, target);
957 		const u64 *val;
958 
959 		if (strcmp(name, "pio-latency-group"))
960 			continue;
961 
962 		val = mdesc_get_property(md, target, "latency", NULL);
963 		if (!val)
964 			continue;
965 
966 		if (*val < best_latency) {
967 			candidate = target;
968 			best_latency = *val;
969 		}
970 	}
971 
972 	if (candidate == MDESC_NODE_NULL)
973 		return -ENODEV;
974 
975 	return scan_pio_for_cfg_handle(md, candidate, cfg_handle);
976 }
977 
978 int of_node_to_nid(struct device_node *dp)
979 {
980 	const struct linux_prom64_registers *regs;
981 	struct mdesc_handle *md;
982 	u32 cfg_handle;
983 	int count, nid;
984 	u64 grp;
985 
986 	/* This is the right thing to do on currently supported
987 	 * SUN4U NUMA platforms as well, as the PCI controller does
988 	 * not sit behind any particular memory controller.
989 	 */
990 	if (!mlgroups)
991 		return -1;
992 
993 	regs = of_get_property(dp, "reg", NULL);
994 	if (!regs)
995 		return -1;
996 
997 	cfg_handle = (regs->phys_addr >> 32UL) & 0x0fffffff;
998 
999 	md = mdesc_grab();
1000 
1001 	count = 0;
1002 	nid = -1;
1003 	mdesc_for_each_node_by_name(md, grp, "group") {
1004 		if (!scan_arcs_for_cfg_handle(md, grp, cfg_handle)) {
1005 			nid = count;
1006 			break;
1007 		}
1008 		count++;
1009 	}
1010 
1011 	mdesc_release(md);
1012 
1013 	return nid;
1014 }
1015 
1016 static void __init add_node_ranges(void)
1017 {
1018 	struct memblock_region *reg;
1019 
1020 	for_each_memblock(memory, reg) {
1021 		unsigned long size = reg->size;
1022 		unsigned long start, end;
1023 
1024 		start = reg->base;
1025 		end = start + size;
1026 		while (start < end) {
1027 			unsigned long this_end;
1028 			int nid;
1029 
1030 			this_end = memblock_nid_range(start, end, &nid);
1031 
1032 			numadbg("Setting memblock NUMA node nid[%d] "
1033 				"start[%lx] end[%lx]\n",
1034 				nid, start, this_end);
1035 
1036 			memblock_set_node(start, this_end - start,
1037 					  &memblock.memory, nid);
1038 			start = this_end;
1039 		}
1040 	}
1041 }
1042 
1043 static int __init grab_mlgroups(struct mdesc_handle *md)
1044 {
1045 	unsigned long paddr;
1046 	int count = 0;
1047 	u64 node;
1048 
1049 	mdesc_for_each_node_by_name(md, node, "memory-latency-group")
1050 		count++;
1051 	if (!count)
1052 		return -ENOENT;
1053 
1054 	paddr = memblock_alloc(count * sizeof(struct mdesc_mlgroup),
1055 			  SMP_CACHE_BYTES);
1056 	if (!paddr)
1057 		return -ENOMEM;
1058 
1059 	mlgroups = __va(paddr);
1060 	num_mlgroups = count;
1061 
1062 	count = 0;
1063 	mdesc_for_each_node_by_name(md, node, "memory-latency-group") {
1064 		struct mdesc_mlgroup *m = &mlgroups[count++];
1065 		const u64 *val;
1066 
1067 		m->node = node;
1068 
1069 		val = mdesc_get_property(md, node, "latency", NULL);
1070 		m->latency = *val;
1071 		val = mdesc_get_property(md, node, "address-match", NULL);
1072 		m->match = *val;
1073 		val = mdesc_get_property(md, node, "address-mask", NULL);
1074 		m->mask = *val;
1075 
1076 		numadbg("MLGROUP[%d]: node[%llx] latency[%llx] "
1077 			"match[%llx] mask[%llx]\n",
1078 			count - 1, m->node, m->latency, m->match, m->mask);
1079 	}
1080 
1081 	return 0;
1082 }
1083 
1084 static int __init grab_mblocks(struct mdesc_handle *md)
1085 {
1086 	unsigned long paddr;
1087 	int count = 0;
1088 	u64 node;
1089 
1090 	mdesc_for_each_node_by_name(md, node, "mblock")
1091 		count++;
1092 	if (!count)
1093 		return -ENOENT;
1094 
1095 	paddr = memblock_alloc(count * sizeof(struct mdesc_mblock),
1096 			  SMP_CACHE_BYTES);
1097 	if (!paddr)
1098 		return -ENOMEM;
1099 
1100 	mblocks = __va(paddr);
1101 	num_mblocks = count;
1102 
1103 	count = 0;
1104 	mdesc_for_each_node_by_name(md, node, "mblock") {
1105 		struct mdesc_mblock *m = &mblocks[count++];
1106 		const u64 *val;
1107 
1108 		val = mdesc_get_property(md, node, "base", NULL);
1109 		m->base = *val;
1110 		val = mdesc_get_property(md, node, "size", NULL);
1111 		m->size = *val;
1112 		val = mdesc_get_property(md, node,
1113 					 "address-congruence-offset", NULL);
1114 
1115 		/* The address-congruence-offset property is optional.
1116 		 * Explicity zero it be identifty this.
1117 		 */
1118 		if (val)
1119 			m->offset = *val;
1120 		else
1121 			m->offset = 0UL;
1122 
1123 		numadbg("MBLOCK[%d]: base[%llx] size[%llx] offset[%llx]\n",
1124 			count - 1, m->base, m->size, m->offset);
1125 	}
1126 
1127 	return 0;
1128 }
1129 
1130 static void __init numa_parse_mdesc_group_cpus(struct mdesc_handle *md,
1131 					       u64 grp, cpumask_t *mask)
1132 {
1133 	u64 arc;
1134 
1135 	cpumask_clear(mask);
1136 
1137 	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_BACK) {
1138 		u64 target = mdesc_arc_target(md, arc);
1139 		const char *name = mdesc_node_name(md, target);
1140 		const u64 *id;
1141 
1142 		if (strcmp(name, "cpu"))
1143 			continue;
1144 		id = mdesc_get_property(md, target, "id", NULL);
1145 		if (*id < nr_cpu_ids)
1146 			cpumask_set_cpu(*id, mask);
1147 	}
1148 }
1149 
1150 static struct mdesc_mlgroup * __init find_mlgroup(u64 node)
1151 {
1152 	int i;
1153 
1154 	for (i = 0; i < num_mlgroups; i++) {
1155 		struct mdesc_mlgroup *m = &mlgroups[i];
1156 		if (m->node == node)
1157 			return m;
1158 	}
1159 	return NULL;
1160 }
1161 
1162 int __node_distance(int from, int to)
1163 {
1164 	if ((from >= MAX_NUMNODES) || (to >= MAX_NUMNODES)) {
1165 		pr_warn("Returning default NUMA distance value for %d->%d\n",
1166 			from, to);
1167 		return (from == to) ? LOCAL_DISTANCE : REMOTE_DISTANCE;
1168 	}
1169 	return numa_latency[from][to];
1170 }
1171 
1172 static int find_best_numa_node_for_mlgroup(struct mdesc_mlgroup *grp)
1173 {
1174 	int i;
1175 
1176 	for (i = 0; i < MAX_NUMNODES; i++) {
1177 		struct node_mem_mask *n = &node_masks[i];
1178 
1179 		if ((grp->mask == n->mask) && (grp->match == n->val))
1180 			break;
1181 	}
1182 	return i;
1183 }
1184 
1185 static void find_numa_latencies_for_group(struct mdesc_handle *md, u64 grp,
1186 					  int index)
1187 {
1188 	u64 arc;
1189 
1190 	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1191 		int tnode;
1192 		u64 target = mdesc_arc_target(md, arc);
1193 		struct mdesc_mlgroup *m = find_mlgroup(target);
1194 
1195 		if (!m)
1196 			continue;
1197 		tnode = find_best_numa_node_for_mlgroup(m);
1198 		if (tnode == MAX_NUMNODES)
1199 			continue;
1200 		numa_latency[index][tnode] = m->latency;
1201 	}
1202 }
1203 
1204 static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp,
1205 				      int index)
1206 {
1207 	struct mdesc_mlgroup *candidate = NULL;
1208 	u64 arc, best_latency = ~(u64)0;
1209 	struct node_mem_mask *n;
1210 
1211 	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1212 		u64 target = mdesc_arc_target(md, arc);
1213 		struct mdesc_mlgroup *m = find_mlgroup(target);
1214 		if (!m)
1215 			continue;
1216 		if (m->latency < best_latency) {
1217 			candidate = m;
1218 			best_latency = m->latency;
1219 		}
1220 	}
1221 	if (!candidate)
1222 		return -ENOENT;
1223 
1224 	if (num_node_masks != index) {
1225 		printk(KERN_ERR "Inconsistent NUMA state, "
1226 		       "index[%d] != num_node_masks[%d]\n",
1227 		       index, num_node_masks);
1228 		return -EINVAL;
1229 	}
1230 
1231 	n = &node_masks[num_node_masks++];
1232 
1233 	n->mask = candidate->mask;
1234 	n->val = candidate->match;
1235 
1236 	numadbg("NUMA NODE[%d]: mask[%lx] val[%lx] (latency[%llx])\n",
1237 		index, n->mask, n->val, candidate->latency);
1238 
1239 	return 0;
1240 }
1241 
1242 static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp,
1243 					 int index)
1244 {
1245 	cpumask_t mask;
1246 	int cpu;
1247 
1248 	numa_parse_mdesc_group_cpus(md, grp, &mask);
1249 
1250 	for_each_cpu(cpu, &mask)
1251 		numa_cpu_lookup_table[cpu] = index;
1252 	cpumask_copy(&numa_cpumask_lookup_table[index], &mask);
1253 
1254 	if (numa_debug) {
1255 		printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index);
1256 		for_each_cpu(cpu, &mask)
1257 			printk("%d ", cpu);
1258 		printk("]\n");
1259 	}
1260 
1261 	return numa_attach_mlgroup(md, grp, index);
1262 }
1263 
1264 static int __init numa_parse_mdesc(void)
1265 {
1266 	struct mdesc_handle *md = mdesc_grab();
1267 	int i, j, err, count;
1268 	u64 node;
1269 
1270 	node = mdesc_node_by_name(md, MDESC_NODE_NULL, "latency-groups");
1271 	if (node == MDESC_NODE_NULL) {
1272 		mdesc_release(md);
1273 		return -ENOENT;
1274 	}
1275 
1276 	err = grab_mblocks(md);
1277 	if (err < 0)
1278 		goto out;
1279 
1280 	err = grab_mlgroups(md);
1281 	if (err < 0)
1282 		goto out;
1283 
1284 	count = 0;
1285 	mdesc_for_each_node_by_name(md, node, "group") {
1286 		err = numa_parse_mdesc_group(md, node, count);
1287 		if (err < 0)
1288 			break;
1289 		count++;
1290 	}
1291 
1292 	count = 0;
1293 	mdesc_for_each_node_by_name(md, node, "group") {
1294 		find_numa_latencies_for_group(md, node, count);
1295 		count++;
1296 	}
1297 
1298 	/* Normalize numa latency matrix according to ACPI SLIT spec. */
1299 	for (i = 0; i < MAX_NUMNODES; i++) {
1300 		u64 self_latency = numa_latency[i][i];
1301 
1302 		for (j = 0; j < MAX_NUMNODES; j++) {
1303 			numa_latency[i][j] =
1304 				(numa_latency[i][j] * LOCAL_DISTANCE) /
1305 				self_latency;
1306 		}
1307 	}
1308 
1309 	add_node_ranges();
1310 
1311 	for (i = 0; i < num_node_masks; i++) {
1312 		allocate_node_data(i);
1313 		node_set_online(i);
1314 	}
1315 
1316 	err = 0;
1317 out:
1318 	mdesc_release(md);
1319 	return err;
1320 }
1321 
1322 static int __init numa_parse_jbus(void)
1323 {
1324 	unsigned long cpu, index;
1325 
1326 	/* NUMA node id is encoded in bits 36 and higher, and there is
1327 	 * a 1-to-1 mapping from CPU ID to NUMA node ID.
1328 	 */
1329 	index = 0;
1330 	for_each_present_cpu(cpu) {
1331 		numa_cpu_lookup_table[cpu] = index;
1332 		cpumask_copy(&numa_cpumask_lookup_table[index], cpumask_of(cpu));
1333 		node_masks[index].mask = ~((1UL << 36UL) - 1UL);
1334 		node_masks[index].val = cpu << 36UL;
1335 
1336 		index++;
1337 	}
1338 	num_node_masks = index;
1339 
1340 	add_node_ranges();
1341 
1342 	for (index = 0; index < num_node_masks; index++) {
1343 		allocate_node_data(index);
1344 		node_set_online(index);
1345 	}
1346 
1347 	return 0;
1348 }
1349 
1350 static int __init numa_parse_sun4u(void)
1351 {
1352 	if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1353 		unsigned long ver;
1354 
1355 		__asm__ ("rdpr %%ver, %0" : "=r" (ver));
1356 		if ((ver >> 32UL) == __JALAPENO_ID ||
1357 		    (ver >> 32UL) == __SERRANO_ID)
1358 			return numa_parse_jbus();
1359 	}
1360 	return -1;
1361 }
1362 
1363 static int __init bootmem_init_numa(void)
1364 {
1365 	int i, j;
1366 	int err = -1;
1367 
1368 	numadbg("bootmem_init_numa()\n");
1369 
1370 	/* Some sane defaults for numa latency values */
1371 	for (i = 0; i < MAX_NUMNODES; i++) {
1372 		for (j = 0; j < MAX_NUMNODES; j++)
1373 			numa_latency[i][j] = (i == j) ?
1374 				LOCAL_DISTANCE : REMOTE_DISTANCE;
1375 	}
1376 
1377 	if (numa_enabled) {
1378 		if (tlb_type == hypervisor)
1379 			err = numa_parse_mdesc();
1380 		else
1381 			err = numa_parse_sun4u();
1382 	}
1383 	return err;
1384 }
1385 
1386 #else
1387 
1388 static int bootmem_init_numa(void)
1389 {
1390 	return -1;
1391 }
1392 
1393 #endif
1394 
1395 static void __init bootmem_init_nonnuma(void)
1396 {
1397 	unsigned long top_of_ram = memblock_end_of_DRAM();
1398 	unsigned long total_ram = memblock_phys_mem_size();
1399 
1400 	numadbg("bootmem_init_nonnuma()\n");
1401 
1402 	printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
1403 	       top_of_ram, total_ram);
1404 	printk(KERN_INFO "Memory hole size: %ldMB\n",
1405 	       (top_of_ram - total_ram) >> 20);
1406 
1407 	init_node_masks_nonnuma();
1408 	memblock_set_node(0, (phys_addr_t)ULLONG_MAX, &memblock.memory, 0);
1409 	allocate_node_data(0);
1410 	node_set_online(0);
1411 }
1412 
1413 static unsigned long __init bootmem_init(unsigned long phys_base)
1414 {
1415 	unsigned long end_pfn;
1416 
1417 	end_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1418 	max_pfn = max_low_pfn = end_pfn;
1419 	min_low_pfn = (phys_base >> PAGE_SHIFT);
1420 
1421 	if (bootmem_init_numa() < 0)
1422 		bootmem_init_nonnuma();
1423 
1424 	/* Dump memblock with node info. */
1425 	memblock_dump_all();
1426 
1427 	/* XXX cpu notifier XXX */
1428 
1429 	sparse_memory_present_with_active_regions(MAX_NUMNODES);
1430 	sparse_init();
1431 
1432 	return end_pfn;
1433 }
1434 
1435 static struct linux_prom64_registers pall[MAX_BANKS] __initdata;
1436 static int pall_ents __initdata;
1437 
1438 static unsigned long max_phys_bits = 40;
1439 
1440 bool kern_addr_valid(unsigned long addr)
1441 {
1442 	pgd_t *pgd;
1443 	pud_t *pud;
1444 	pmd_t *pmd;
1445 	pte_t *pte;
1446 
1447 	if ((long)addr < 0L) {
1448 		unsigned long pa = __pa(addr);
1449 
1450 		if ((addr >> max_phys_bits) != 0UL)
1451 			return false;
1452 
1453 		return pfn_valid(pa >> PAGE_SHIFT);
1454 	}
1455 
1456 	if (addr >= (unsigned long) KERNBASE &&
1457 	    addr < (unsigned long)&_end)
1458 		return true;
1459 
1460 	pgd = pgd_offset_k(addr);
1461 	if (pgd_none(*pgd))
1462 		return 0;
1463 
1464 	pud = pud_offset(pgd, addr);
1465 	if (pud_none(*pud))
1466 		return 0;
1467 
1468 	if (pud_large(*pud))
1469 		return pfn_valid(pud_pfn(*pud));
1470 
1471 	pmd = pmd_offset(pud, addr);
1472 	if (pmd_none(*pmd))
1473 		return 0;
1474 
1475 	if (pmd_large(*pmd))
1476 		return pfn_valid(pmd_pfn(*pmd));
1477 
1478 	pte = pte_offset_kernel(pmd, addr);
1479 	if (pte_none(*pte))
1480 		return 0;
1481 
1482 	return pfn_valid(pte_pfn(*pte));
1483 }
1484 EXPORT_SYMBOL(kern_addr_valid);
1485 
1486 static unsigned long __ref kernel_map_hugepud(unsigned long vstart,
1487 					      unsigned long vend,
1488 					      pud_t *pud)
1489 {
1490 	const unsigned long mask16gb = (1UL << 34) - 1UL;
1491 	u64 pte_val = vstart;
1492 
1493 	/* Each PUD is 8GB */
1494 	if ((vstart & mask16gb) ||
1495 	    (vend - vstart <= mask16gb)) {
1496 		pte_val ^= kern_linear_pte_xor[2];
1497 		pud_val(*pud) = pte_val | _PAGE_PUD_HUGE;
1498 
1499 		return vstart + PUD_SIZE;
1500 	}
1501 
1502 	pte_val ^= kern_linear_pte_xor[3];
1503 	pte_val |= _PAGE_PUD_HUGE;
1504 
1505 	vend = vstart + mask16gb + 1UL;
1506 	while (vstart < vend) {
1507 		pud_val(*pud) = pte_val;
1508 
1509 		pte_val += PUD_SIZE;
1510 		vstart += PUD_SIZE;
1511 		pud++;
1512 	}
1513 	return vstart;
1514 }
1515 
1516 static bool kernel_can_map_hugepud(unsigned long vstart, unsigned long vend,
1517 				   bool guard)
1518 {
1519 	if (guard && !(vstart & ~PUD_MASK) && (vend - vstart) >= PUD_SIZE)
1520 		return true;
1521 
1522 	return false;
1523 }
1524 
1525 static unsigned long __ref kernel_map_hugepmd(unsigned long vstart,
1526 					      unsigned long vend,
1527 					      pmd_t *pmd)
1528 {
1529 	const unsigned long mask256mb = (1UL << 28) - 1UL;
1530 	const unsigned long mask2gb = (1UL << 31) - 1UL;
1531 	u64 pte_val = vstart;
1532 
1533 	/* Each PMD is 8MB */
1534 	if ((vstart & mask256mb) ||
1535 	    (vend - vstart <= mask256mb)) {
1536 		pte_val ^= kern_linear_pte_xor[0];
1537 		pmd_val(*pmd) = pte_val | _PAGE_PMD_HUGE;
1538 
1539 		return vstart + PMD_SIZE;
1540 	}
1541 
1542 	if ((vstart & mask2gb) ||
1543 	    (vend - vstart <= mask2gb)) {
1544 		pte_val ^= kern_linear_pte_xor[1];
1545 		pte_val |= _PAGE_PMD_HUGE;
1546 		vend = vstart + mask256mb + 1UL;
1547 	} else {
1548 		pte_val ^= kern_linear_pte_xor[2];
1549 		pte_val |= _PAGE_PMD_HUGE;
1550 		vend = vstart + mask2gb + 1UL;
1551 	}
1552 
1553 	while (vstart < vend) {
1554 		pmd_val(*pmd) = pte_val;
1555 
1556 		pte_val += PMD_SIZE;
1557 		vstart += PMD_SIZE;
1558 		pmd++;
1559 	}
1560 
1561 	return vstart;
1562 }
1563 
1564 static bool kernel_can_map_hugepmd(unsigned long vstart, unsigned long vend,
1565 				   bool guard)
1566 {
1567 	if (guard && !(vstart & ~PMD_MASK) && (vend - vstart) >= PMD_SIZE)
1568 		return true;
1569 
1570 	return false;
1571 }
1572 
1573 static unsigned long __ref kernel_map_range(unsigned long pstart,
1574 					    unsigned long pend, pgprot_t prot,
1575 					    bool use_huge)
1576 {
1577 	unsigned long vstart = PAGE_OFFSET + pstart;
1578 	unsigned long vend = PAGE_OFFSET + pend;
1579 	unsigned long alloc_bytes = 0UL;
1580 
1581 	if ((vstart & ~PAGE_MASK) || (vend & ~PAGE_MASK)) {
1582 		prom_printf("kernel_map: Unaligned physmem[%lx:%lx]\n",
1583 			    vstart, vend);
1584 		prom_halt();
1585 	}
1586 
1587 	while (vstart < vend) {
1588 		unsigned long this_end, paddr = __pa(vstart);
1589 		pgd_t *pgd = pgd_offset_k(vstart);
1590 		pud_t *pud;
1591 		pmd_t *pmd;
1592 		pte_t *pte;
1593 
1594 		if (pgd_none(*pgd)) {
1595 			pud_t *new;
1596 
1597 			new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1598 			alloc_bytes += PAGE_SIZE;
1599 			pgd_populate(&init_mm, pgd, new);
1600 		}
1601 		pud = pud_offset(pgd, vstart);
1602 		if (pud_none(*pud)) {
1603 			pmd_t *new;
1604 
1605 			if (kernel_can_map_hugepud(vstart, vend, use_huge)) {
1606 				vstart = kernel_map_hugepud(vstart, vend, pud);
1607 				continue;
1608 			}
1609 			new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1610 			alloc_bytes += PAGE_SIZE;
1611 			pud_populate(&init_mm, pud, new);
1612 		}
1613 
1614 		pmd = pmd_offset(pud, vstart);
1615 		if (pmd_none(*pmd)) {
1616 			pte_t *new;
1617 
1618 			if (kernel_can_map_hugepmd(vstart, vend, use_huge)) {
1619 				vstart = kernel_map_hugepmd(vstart, vend, pmd);
1620 				continue;
1621 			}
1622 			new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1623 			alloc_bytes += PAGE_SIZE;
1624 			pmd_populate_kernel(&init_mm, pmd, new);
1625 		}
1626 
1627 		pte = pte_offset_kernel(pmd, vstart);
1628 		this_end = (vstart + PMD_SIZE) & PMD_MASK;
1629 		if (this_end > vend)
1630 			this_end = vend;
1631 
1632 		while (vstart < this_end) {
1633 			pte_val(*pte) = (paddr | pgprot_val(prot));
1634 
1635 			vstart += PAGE_SIZE;
1636 			paddr += PAGE_SIZE;
1637 			pte++;
1638 		}
1639 	}
1640 
1641 	return alloc_bytes;
1642 }
1643 
1644 static void __init flush_all_kernel_tsbs(void)
1645 {
1646 	int i;
1647 
1648 	for (i = 0; i < KERNEL_TSB_NENTRIES; i++) {
1649 		struct tsb *ent = &swapper_tsb[i];
1650 
1651 		ent->tag = (1UL << TSB_TAG_INVALID_BIT);
1652 	}
1653 #ifndef CONFIG_DEBUG_PAGEALLOC
1654 	for (i = 0; i < KERNEL_TSB4M_NENTRIES; i++) {
1655 		struct tsb *ent = &swapper_4m_tsb[i];
1656 
1657 		ent->tag = (1UL << TSB_TAG_INVALID_BIT);
1658 	}
1659 #endif
1660 }
1661 
1662 extern unsigned int kvmap_linear_patch[1];
1663 
1664 static void __init kernel_physical_mapping_init(void)
1665 {
1666 	unsigned long i, mem_alloced = 0UL;
1667 	bool use_huge = true;
1668 
1669 #ifdef CONFIG_DEBUG_PAGEALLOC
1670 	use_huge = false;
1671 #endif
1672 	for (i = 0; i < pall_ents; i++) {
1673 		unsigned long phys_start, phys_end;
1674 
1675 		phys_start = pall[i].phys_addr;
1676 		phys_end = phys_start + pall[i].reg_size;
1677 
1678 		mem_alloced += kernel_map_range(phys_start, phys_end,
1679 						PAGE_KERNEL, use_huge);
1680 	}
1681 
1682 	printk("Allocated %ld bytes for kernel page tables.\n",
1683 	       mem_alloced);
1684 
1685 	kvmap_linear_patch[0] = 0x01000000; /* nop */
1686 	flushi(&kvmap_linear_patch[0]);
1687 
1688 	flush_all_kernel_tsbs();
1689 
1690 	__flush_tlb_all();
1691 }
1692 
1693 #ifdef CONFIG_DEBUG_PAGEALLOC
1694 void __kernel_map_pages(struct page *page, int numpages, int enable)
1695 {
1696 	unsigned long phys_start = page_to_pfn(page) << PAGE_SHIFT;
1697 	unsigned long phys_end = phys_start + (numpages * PAGE_SIZE);
1698 
1699 	kernel_map_range(phys_start, phys_end,
1700 			 (enable ? PAGE_KERNEL : __pgprot(0)), false);
1701 
1702 	flush_tsb_kernel_range(PAGE_OFFSET + phys_start,
1703 			       PAGE_OFFSET + phys_end);
1704 
1705 	/* we should perform an IPI and flush all tlbs,
1706 	 * but that can deadlock->flush only current cpu.
1707 	 */
1708 	__flush_tlb_kernel_range(PAGE_OFFSET + phys_start,
1709 				 PAGE_OFFSET + phys_end);
1710 }
1711 #endif
1712 
1713 unsigned long __init find_ecache_flush_span(unsigned long size)
1714 {
1715 	int i;
1716 
1717 	for (i = 0; i < pavail_ents; i++) {
1718 		if (pavail[i].reg_size >= size)
1719 			return pavail[i].phys_addr;
1720 	}
1721 
1722 	return ~0UL;
1723 }
1724 
1725 unsigned long PAGE_OFFSET;
1726 EXPORT_SYMBOL(PAGE_OFFSET);
1727 
1728 unsigned long VMALLOC_END   = 0x0000010000000000UL;
1729 EXPORT_SYMBOL(VMALLOC_END);
1730 
1731 unsigned long sparc64_va_hole_top =    0xfffff80000000000UL;
1732 unsigned long sparc64_va_hole_bottom = 0x0000080000000000UL;
1733 
1734 static void __init setup_page_offset(void)
1735 {
1736 	if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1737 		/* Cheetah/Panther support a full 64-bit virtual
1738 		 * address, so we can use all that our page tables
1739 		 * support.
1740 		 */
1741 		sparc64_va_hole_top =    0xfff0000000000000UL;
1742 		sparc64_va_hole_bottom = 0x0010000000000000UL;
1743 
1744 		max_phys_bits = 42;
1745 	} else if (tlb_type == hypervisor) {
1746 		switch (sun4v_chip_type) {
1747 		case SUN4V_CHIP_NIAGARA1:
1748 		case SUN4V_CHIP_NIAGARA2:
1749 			/* T1 and T2 support 48-bit virtual addresses.  */
1750 			sparc64_va_hole_top =    0xffff800000000000UL;
1751 			sparc64_va_hole_bottom = 0x0000800000000000UL;
1752 
1753 			max_phys_bits = 39;
1754 			break;
1755 		case SUN4V_CHIP_NIAGARA3:
1756 			/* T3 supports 48-bit virtual addresses.  */
1757 			sparc64_va_hole_top =    0xffff800000000000UL;
1758 			sparc64_va_hole_bottom = 0x0000800000000000UL;
1759 
1760 			max_phys_bits = 43;
1761 			break;
1762 		case SUN4V_CHIP_NIAGARA4:
1763 		case SUN4V_CHIP_NIAGARA5:
1764 		case SUN4V_CHIP_SPARC64X:
1765 		case SUN4V_CHIP_SPARC_M6:
1766 			/* T4 and later support 52-bit virtual addresses.  */
1767 			sparc64_va_hole_top =    0xfff8000000000000UL;
1768 			sparc64_va_hole_bottom = 0x0008000000000000UL;
1769 			max_phys_bits = 47;
1770 			break;
1771 		case SUN4V_CHIP_SPARC_M7:
1772 		default:
1773 			/* M7 and later support 52-bit virtual addresses.  */
1774 			sparc64_va_hole_top =    0xfff8000000000000UL;
1775 			sparc64_va_hole_bottom = 0x0008000000000000UL;
1776 			max_phys_bits = 49;
1777 			break;
1778 		}
1779 	}
1780 
1781 	if (max_phys_bits > MAX_PHYS_ADDRESS_BITS) {
1782 		prom_printf("MAX_PHYS_ADDRESS_BITS is too small, need %lu\n",
1783 			    max_phys_bits);
1784 		prom_halt();
1785 	}
1786 
1787 	PAGE_OFFSET = sparc64_va_hole_top;
1788 	VMALLOC_END = ((sparc64_va_hole_bottom >> 1) +
1789 		       (sparc64_va_hole_bottom >> 2));
1790 
1791 	pr_info("MM: PAGE_OFFSET is 0x%016lx (max_phys_bits == %lu)\n",
1792 		PAGE_OFFSET, max_phys_bits);
1793 	pr_info("MM: VMALLOC [0x%016lx --> 0x%016lx]\n",
1794 		VMALLOC_START, VMALLOC_END);
1795 	pr_info("MM: VMEMMAP [0x%016lx --> 0x%016lx]\n",
1796 		VMEMMAP_BASE, VMEMMAP_BASE << 1);
1797 }
1798 
1799 static void __init tsb_phys_patch(void)
1800 {
1801 	struct tsb_ldquad_phys_patch_entry *pquad;
1802 	struct tsb_phys_patch_entry *p;
1803 
1804 	pquad = &__tsb_ldquad_phys_patch;
1805 	while (pquad < &__tsb_ldquad_phys_patch_end) {
1806 		unsigned long addr = pquad->addr;
1807 
1808 		if (tlb_type == hypervisor)
1809 			*(unsigned int *) addr = pquad->sun4v_insn;
1810 		else
1811 			*(unsigned int *) addr = pquad->sun4u_insn;
1812 		wmb();
1813 		__asm__ __volatile__("flush	%0"
1814 				     : /* no outputs */
1815 				     : "r" (addr));
1816 
1817 		pquad++;
1818 	}
1819 
1820 	p = &__tsb_phys_patch;
1821 	while (p < &__tsb_phys_patch_end) {
1822 		unsigned long addr = p->addr;
1823 
1824 		*(unsigned int *) addr = p->insn;
1825 		wmb();
1826 		__asm__ __volatile__("flush	%0"
1827 				     : /* no outputs */
1828 				     : "r" (addr));
1829 
1830 		p++;
1831 	}
1832 }
1833 
1834 /* Don't mark as init, we give this to the Hypervisor.  */
1835 #ifndef CONFIG_DEBUG_PAGEALLOC
1836 #define NUM_KTSB_DESCR	2
1837 #else
1838 #define NUM_KTSB_DESCR	1
1839 #endif
1840 static struct hv_tsb_descr ktsb_descr[NUM_KTSB_DESCR];
1841 
1842 /* The swapper TSBs are loaded with a base sequence of:
1843  *
1844  *	sethi	%uhi(SYMBOL), REG1
1845  *	sethi	%hi(SYMBOL), REG2
1846  *	or	REG1, %ulo(SYMBOL), REG1
1847  *	or	REG2, %lo(SYMBOL), REG2
1848  *	sllx	REG1, 32, REG1
1849  *	or	REG1, REG2, REG1
1850  *
1851  * When we use physical addressing for the TSB accesses, we patch the
1852  * first four instructions in the above sequence.
1853  */
1854 
1855 static void patch_one_ktsb_phys(unsigned int *start, unsigned int *end, unsigned long pa)
1856 {
1857 	unsigned long high_bits, low_bits;
1858 
1859 	high_bits = (pa >> 32) & 0xffffffff;
1860 	low_bits = (pa >> 0) & 0xffffffff;
1861 
1862 	while (start < end) {
1863 		unsigned int *ia = (unsigned int *)(unsigned long)*start;
1864 
1865 		ia[0] = (ia[0] & ~0x3fffff) | (high_bits >> 10);
1866 		__asm__ __volatile__("flush	%0" : : "r" (ia));
1867 
1868 		ia[1] = (ia[1] & ~0x3fffff) | (low_bits >> 10);
1869 		__asm__ __volatile__("flush	%0" : : "r" (ia + 1));
1870 
1871 		ia[2] = (ia[2] & ~0x1fff) | (high_bits & 0x3ff);
1872 		__asm__ __volatile__("flush	%0" : : "r" (ia + 2));
1873 
1874 		ia[3] = (ia[3] & ~0x1fff) | (low_bits & 0x3ff);
1875 		__asm__ __volatile__("flush	%0" : : "r" (ia + 3));
1876 
1877 		start++;
1878 	}
1879 }
1880 
1881 static void ktsb_phys_patch(void)
1882 {
1883 	extern unsigned int __swapper_tsb_phys_patch;
1884 	extern unsigned int __swapper_tsb_phys_patch_end;
1885 	unsigned long ktsb_pa;
1886 
1887 	ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
1888 	patch_one_ktsb_phys(&__swapper_tsb_phys_patch,
1889 			    &__swapper_tsb_phys_patch_end, ktsb_pa);
1890 #ifndef CONFIG_DEBUG_PAGEALLOC
1891 	{
1892 	extern unsigned int __swapper_4m_tsb_phys_patch;
1893 	extern unsigned int __swapper_4m_tsb_phys_patch_end;
1894 	ktsb_pa = (kern_base +
1895 		   ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
1896 	patch_one_ktsb_phys(&__swapper_4m_tsb_phys_patch,
1897 			    &__swapper_4m_tsb_phys_patch_end, ktsb_pa);
1898 	}
1899 #endif
1900 }
1901 
1902 static void __init sun4v_ktsb_init(void)
1903 {
1904 	unsigned long ktsb_pa;
1905 
1906 	/* First KTSB for PAGE_SIZE mappings.  */
1907 	ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
1908 
1909 	switch (PAGE_SIZE) {
1910 	case 8 * 1024:
1911 	default:
1912 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_8K;
1913 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_8K;
1914 		break;
1915 
1916 	case 64 * 1024:
1917 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_64K;
1918 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_64K;
1919 		break;
1920 
1921 	case 512 * 1024:
1922 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_512K;
1923 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_512K;
1924 		break;
1925 
1926 	case 4 * 1024 * 1024:
1927 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_4MB;
1928 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_4MB;
1929 		break;
1930 	}
1931 
1932 	ktsb_descr[0].assoc = 1;
1933 	ktsb_descr[0].num_ttes = KERNEL_TSB_NENTRIES;
1934 	ktsb_descr[0].ctx_idx = 0;
1935 	ktsb_descr[0].tsb_base = ktsb_pa;
1936 	ktsb_descr[0].resv = 0;
1937 
1938 #ifndef CONFIG_DEBUG_PAGEALLOC
1939 	/* Second KTSB for 4MB/256MB/2GB/16GB mappings.  */
1940 	ktsb_pa = (kern_base +
1941 		   ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
1942 
1943 	ktsb_descr[1].pgsz_idx = HV_PGSZ_IDX_4MB;
1944 	ktsb_descr[1].pgsz_mask = ((HV_PGSZ_MASK_4MB |
1945 				    HV_PGSZ_MASK_256MB |
1946 				    HV_PGSZ_MASK_2GB |
1947 				    HV_PGSZ_MASK_16GB) &
1948 				   cpu_pgsz_mask);
1949 	ktsb_descr[1].assoc = 1;
1950 	ktsb_descr[1].num_ttes = KERNEL_TSB4M_NENTRIES;
1951 	ktsb_descr[1].ctx_idx = 0;
1952 	ktsb_descr[1].tsb_base = ktsb_pa;
1953 	ktsb_descr[1].resv = 0;
1954 #endif
1955 }
1956 
1957 void sun4v_ktsb_register(void)
1958 {
1959 	unsigned long pa, ret;
1960 
1961 	pa = kern_base + ((unsigned long)&ktsb_descr[0] - KERNBASE);
1962 
1963 	ret = sun4v_mmu_tsb_ctx0(NUM_KTSB_DESCR, pa);
1964 	if (ret != 0) {
1965 		prom_printf("hypervisor_mmu_tsb_ctx0[%lx]: "
1966 			    "errors with %lx\n", pa, ret);
1967 		prom_halt();
1968 	}
1969 }
1970 
1971 static void __init sun4u_linear_pte_xor_finalize(void)
1972 {
1973 #ifndef CONFIG_DEBUG_PAGEALLOC
1974 	/* This is where we would add Panther support for
1975 	 * 32MB and 256MB pages.
1976 	 */
1977 #endif
1978 }
1979 
1980 static void __init sun4v_linear_pte_xor_finalize(void)
1981 {
1982 	unsigned long pagecv_flag;
1983 
1984 	/* Bit 9 of TTE is no longer CV bit on M7 processor and it instead
1985 	 * enables MCD error. Do not set bit 9 on M7 processor.
1986 	 */
1987 	switch (sun4v_chip_type) {
1988 	case SUN4V_CHIP_SPARC_M7:
1989 		pagecv_flag = 0x00;
1990 		break;
1991 	default:
1992 		pagecv_flag = _PAGE_CV_4V;
1993 		break;
1994 	}
1995 #ifndef CONFIG_DEBUG_PAGEALLOC
1996 	if (cpu_pgsz_mask & HV_PGSZ_MASK_256MB) {
1997 		kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZ256MB_4V) ^
1998 			PAGE_OFFSET;
1999 		kern_linear_pte_xor[1] |= (_PAGE_CP_4V | pagecv_flag |
2000 					   _PAGE_P_4V | _PAGE_W_4V);
2001 	} else {
2002 		kern_linear_pte_xor[1] = kern_linear_pte_xor[0];
2003 	}
2004 
2005 	if (cpu_pgsz_mask & HV_PGSZ_MASK_2GB) {
2006 		kern_linear_pte_xor[2] = (_PAGE_VALID | _PAGE_SZ2GB_4V) ^
2007 			PAGE_OFFSET;
2008 		kern_linear_pte_xor[2] |= (_PAGE_CP_4V | pagecv_flag |
2009 					   _PAGE_P_4V | _PAGE_W_4V);
2010 	} else {
2011 		kern_linear_pte_xor[2] = kern_linear_pte_xor[1];
2012 	}
2013 
2014 	if (cpu_pgsz_mask & HV_PGSZ_MASK_16GB) {
2015 		kern_linear_pte_xor[3] = (_PAGE_VALID | _PAGE_SZ16GB_4V) ^
2016 			PAGE_OFFSET;
2017 		kern_linear_pte_xor[3] |= (_PAGE_CP_4V | pagecv_flag |
2018 					   _PAGE_P_4V | _PAGE_W_4V);
2019 	} else {
2020 		kern_linear_pte_xor[3] = kern_linear_pte_xor[2];
2021 	}
2022 #endif
2023 }
2024 
2025 /* paging_init() sets up the page tables */
2026 
2027 static unsigned long last_valid_pfn;
2028 
2029 static void sun4u_pgprot_init(void);
2030 static void sun4v_pgprot_init(void);
2031 
2032 static phys_addr_t __init available_memory(void)
2033 {
2034 	phys_addr_t available = 0ULL;
2035 	phys_addr_t pa_start, pa_end;
2036 	u64 i;
2037 
2038 	for_each_free_mem_range(i, NUMA_NO_NODE, MEMBLOCK_NONE, &pa_start,
2039 				&pa_end, NULL)
2040 		available = available + (pa_end  - pa_start);
2041 
2042 	return available;
2043 }
2044 
2045 #define _PAGE_CACHE_4U	(_PAGE_CP_4U | _PAGE_CV_4U)
2046 #define _PAGE_CACHE_4V	(_PAGE_CP_4V | _PAGE_CV_4V)
2047 #define __DIRTY_BITS_4U	 (_PAGE_MODIFIED_4U | _PAGE_WRITE_4U | _PAGE_W_4U)
2048 #define __DIRTY_BITS_4V	 (_PAGE_MODIFIED_4V | _PAGE_WRITE_4V | _PAGE_W_4V)
2049 #define __ACCESS_BITS_4U (_PAGE_ACCESSED_4U | _PAGE_READ_4U | _PAGE_R)
2050 #define __ACCESS_BITS_4V (_PAGE_ACCESSED_4V | _PAGE_READ_4V | _PAGE_R)
2051 
2052 /* We need to exclude reserved regions. This exclusion will include
2053  * vmlinux and initrd. To be more precise the initrd size could be used to
2054  * compute a new lower limit because it is freed later during initialization.
2055  */
2056 static void __init reduce_memory(phys_addr_t limit_ram)
2057 {
2058 	phys_addr_t avail_ram = available_memory();
2059 	phys_addr_t pa_start, pa_end;
2060 	u64 i;
2061 
2062 	if (limit_ram >= avail_ram)
2063 		return;
2064 
2065 	for_each_free_mem_range(i, NUMA_NO_NODE, MEMBLOCK_NONE, &pa_start,
2066 				&pa_end, NULL) {
2067 		phys_addr_t region_size = pa_end - pa_start;
2068 		phys_addr_t clip_start = pa_start;
2069 
2070 		avail_ram = avail_ram - region_size;
2071 		/* Are we consuming too much? */
2072 		if (avail_ram < limit_ram) {
2073 			phys_addr_t give_back = limit_ram - avail_ram;
2074 
2075 			region_size = region_size - give_back;
2076 			clip_start = clip_start + give_back;
2077 		}
2078 
2079 		memblock_remove(clip_start, region_size);
2080 
2081 		if (avail_ram <= limit_ram)
2082 			break;
2083 		i = 0UL;
2084 	}
2085 }
2086 
2087 void __init paging_init(void)
2088 {
2089 	unsigned long end_pfn, shift, phys_base;
2090 	unsigned long real_end, i;
2091 	int node;
2092 
2093 	setup_page_offset();
2094 
2095 	/* These build time checkes make sure that the dcache_dirty_cpu()
2096 	 * page->flags usage will work.
2097 	 *
2098 	 * When a page gets marked as dcache-dirty, we store the
2099 	 * cpu number starting at bit 32 in the page->flags.  Also,
2100 	 * functions like clear_dcache_dirty_cpu use the cpu mask
2101 	 * in 13-bit signed-immediate instruction fields.
2102 	 */
2103 
2104 	/*
2105 	 * Page flags must not reach into upper 32 bits that are used
2106 	 * for the cpu number
2107 	 */
2108 	BUILD_BUG_ON(NR_PAGEFLAGS > 32);
2109 
2110 	/*
2111 	 * The bit fields placed in the high range must not reach below
2112 	 * the 32 bit boundary. Otherwise we cannot place the cpu field
2113 	 * at the 32 bit boundary.
2114 	 */
2115 	BUILD_BUG_ON(SECTIONS_WIDTH + NODES_WIDTH + ZONES_WIDTH +
2116 		ilog2(roundup_pow_of_two(NR_CPUS)) > 32);
2117 
2118 	BUILD_BUG_ON(NR_CPUS > 4096);
2119 
2120 	kern_base = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
2121 	kern_size = (unsigned long)&_end - (unsigned long)KERNBASE;
2122 
2123 	/* Invalidate both kernel TSBs.  */
2124 	memset(swapper_tsb, 0x40, sizeof(swapper_tsb));
2125 #ifndef CONFIG_DEBUG_PAGEALLOC
2126 	memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2127 #endif
2128 
2129 	/* TTE.cv bit on sparc v9 occupies the same position as TTE.mcde
2130 	 * bit on M7 processor. This is a conflicting usage of the same
2131 	 * bit. Enabling TTE.cv on M7 would turn on Memory Corruption
2132 	 * Detection error on all pages and this will lead to problems
2133 	 * later. Kernel does not run with MCD enabled and hence rest
2134 	 * of the required steps to fully configure memory corruption
2135 	 * detection are not taken. We need to ensure TTE.mcde is not
2136 	 * set on M7 processor. Compute the value of cacheability
2137 	 * flag for use later taking this into consideration.
2138 	 */
2139 	switch (sun4v_chip_type) {
2140 	case SUN4V_CHIP_SPARC_M7:
2141 		page_cache4v_flag = _PAGE_CP_4V;
2142 		break;
2143 	default:
2144 		page_cache4v_flag = _PAGE_CACHE_4V;
2145 		break;
2146 	}
2147 
2148 	if (tlb_type == hypervisor)
2149 		sun4v_pgprot_init();
2150 	else
2151 		sun4u_pgprot_init();
2152 
2153 	if (tlb_type == cheetah_plus ||
2154 	    tlb_type == hypervisor) {
2155 		tsb_phys_patch();
2156 		ktsb_phys_patch();
2157 	}
2158 
2159 	if (tlb_type == hypervisor)
2160 		sun4v_patch_tlb_handlers();
2161 
2162 	/* Find available physical memory...
2163 	 *
2164 	 * Read it twice in order to work around a bug in openfirmware.
2165 	 * The call to grab this table itself can cause openfirmware to
2166 	 * allocate memory, which in turn can take away some space from
2167 	 * the list of available memory.  Reading it twice makes sure
2168 	 * we really do get the final value.
2169 	 */
2170 	read_obp_translations();
2171 	read_obp_memory("reg", &pall[0], &pall_ents);
2172 	read_obp_memory("available", &pavail[0], &pavail_ents);
2173 	read_obp_memory("available", &pavail[0], &pavail_ents);
2174 
2175 	phys_base = 0xffffffffffffffffUL;
2176 	for (i = 0; i < pavail_ents; i++) {
2177 		phys_base = min(phys_base, pavail[i].phys_addr);
2178 		memblock_add(pavail[i].phys_addr, pavail[i].reg_size);
2179 	}
2180 
2181 	memblock_reserve(kern_base, kern_size);
2182 
2183 	find_ramdisk(phys_base);
2184 
2185 	if (cmdline_memory_size)
2186 		reduce_memory(cmdline_memory_size);
2187 
2188 	memblock_allow_resize();
2189 	memblock_dump_all();
2190 
2191 	set_bit(0, mmu_context_bmap);
2192 
2193 	shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE);
2194 
2195 	real_end = (unsigned long)_end;
2196 	num_kernel_image_mappings = DIV_ROUND_UP(real_end - KERNBASE, 1 << ILOG2_4MB);
2197 	printk("Kernel: Using %d locked TLB entries for main kernel image.\n",
2198 	       num_kernel_image_mappings);
2199 
2200 	/* Set kernel pgd to upper alias so physical page computations
2201 	 * work.
2202 	 */
2203 	init_mm.pgd += ((shift) / (sizeof(pgd_t)));
2204 
2205 	memset(swapper_pg_dir, 0, sizeof(swapper_pg_dir));
2206 
2207 	inherit_prom_mappings();
2208 
2209 	/* Ok, we can use our TLB miss and window trap handlers safely.  */
2210 	setup_tba();
2211 
2212 	__flush_tlb_all();
2213 
2214 	prom_build_devicetree();
2215 	of_populate_present_mask();
2216 #ifndef CONFIG_SMP
2217 	of_fill_in_cpu_data();
2218 #endif
2219 
2220 	if (tlb_type == hypervisor) {
2221 		sun4v_mdesc_init();
2222 		mdesc_populate_present_mask(cpu_all_mask);
2223 #ifndef CONFIG_SMP
2224 		mdesc_fill_in_cpu_data(cpu_all_mask);
2225 #endif
2226 		mdesc_get_page_sizes(cpu_all_mask, &cpu_pgsz_mask);
2227 
2228 		sun4v_linear_pte_xor_finalize();
2229 
2230 		sun4v_ktsb_init();
2231 		sun4v_ktsb_register();
2232 	} else {
2233 		unsigned long impl, ver;
2234 
2235 		cpu_pgsz_mask = (HV_PGSZ_MASK_8K | HV_PGSZ_MASK_64K |
2236 				 HV_PGSZ_MASK_512K | HV_PGSZ_MASK_4MB);
2237 
2238 		__asm__ __volatile__("rdpr %%ver, %0" : "=r" (ver));
2239 		impl = ((ver >> 32) & 0xffff);
2240 		if (impl == PANTHER_IMPL)
2241 			cpu_pgsz_mask |= (HV_PGSZ_MASK_32MB |
2242 					  HV_PGSZ_MASK_256MB);
2243 
2244 		sun4u_linear_pte_xor_finalize();
2245 	}
2246 
2247 	/* Flush the TLBs and the 4M TSB so that the updated linear
2248 	 * pte XOR settings are realized for all mappings.
2249 	 */
2250 	__flush_tlb_all();
2251 #ifndef CONFIG_DEBUG_PAGEALLOC
2252 	memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2253 #endif
2254 	__flush_tlb_all();
2255 
2256 	/* Setup bootmem... */
2257 	last_valid_pfn = end_pfn = bootmem_init(phys_base);
2258 
2259 	/* Once the OF device tree and MDESC have been setup, we know
2260 	 * the list of possible cpus.  Therefore we can allocate the
2261 	 * IRQ stacks.
2262 	 */
2263 	for_each_possible_cpu(i) {
2264 		node = cpu_to_node(i);
2265 
2266 		softirq_stack[i] = __alloc_bootmem_node(NODE_DATA(node),
2267 							THREAD_SIZE,
2268 							THREAD_SIZE, 0);
2269 		hardirq_stack[i] = __alloc_bootmem_node(NODE_DATA(node),
2270 							THREAD_SIZE,
2271 							THREAD_SIZE, 0);
2272 	}
2273 
2274 	kernel_physical_mapping_init();
2275 
2276 	{
2277 		unsigned long max_zone_pfns[MAX_NR_ZONES];
2278 
2279 		memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
2280 
2281 		max_zone_pfns[ZONE_NORMAL] = end_pfn;
2282 
2283 		free_area_init_nodes(max_zone_pfns);
2284 	}
2285 
2286 	printk("Booting Linux...\n");
2287 }
2288 
2289 int page_in_phys_avail(unsigned long paddr)
2290 {
2291 	int i;
2292 
2293 	paddr &= PAGE_MASK;
2294 
2295 	for (i = 0; i < pavail_ents; i++) {
2296 		unsigned long start, end;
2297 
2298 		start = pavail[i].phys_addr;
2299 		end = start + pavail[i].reg_size;
2300 
2301 		if (paddr >= start && paddr < end)
2302 			return 1;
2303 	}
2304 	if (paddr >= kern_base && paddr < (kern_base + kern_size))
2305 		return 1;
2306 #ifdef CONFIG_BLK_DEV_INITRD
2307 	if (paddr >= __pa(initrd_start) &&
2308 	    paddr < __pa(PAGE_ALIGN(initrd_end)))
2309 		return 1;
2310 #endif
2311 
2312 	return 0;
2313 }
2314 
2315 static void __init register_page_bootmem_info(void)
2316 {
2317 #ifdef CONFIG_NEED_MULTIPLE_NODES
2318 	int i;
2319 
2320 	for_each_online_node(i)
2321 		if (NODE_DATA(i)->node_spanned_pages)
2322 			register_page_bootmem_info_node(NODE_DATA(i));
2323 #endif
2324 }
2325 void __init mem_init(void)
2326 {
2327 	high_memory = __va(last_valid_pfn << PAGE_SHIFT);
2328 
2329 	register_page_bootmem_info();
2330 	free_all_bootmem();
2331 
2332 	/*
2333 	 * Set up the zero page, mark it reserved, so that page count
2334 	 * is not manipulated when freeing the page from user ptes.
2335 	 */
2336 	mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0);
2337 	if (mem_map_zero == NULL) {
2338 		prom_printf("paging_init: Cannot alloc zero page.\n");
2339 		prom_halt();
2340 	}
2341 	mark_page_reserved(mem_map_zero);
2342 
2343 	mem_init_print_info(NULL);
2344 
2345 	if (tlb_type == cheetah || tlb_type == cheetah_plus)
2346 		cheetah_ecache_flush_init();
2347 }
2348 
2349 void free_initmem(void)
2350 {
2351 	unsigned long addr, initend;
2352 	int do_free = 1;
2353 
2354 	/* If the physical memory maps were trimmed by kernel command
2355 	 * line options, don't even try freeing this initmem stuff up.
2356 	 * The kernel image could have been in the trimmed out region
2357 	 * and if so the freeing below will free invalid page structs.
2358 	 */
2359 	if (cmdline_memory_size)
2360 		do_free = 0;
2361 
2362 	/*
2363 	 * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes.
2364 	 */
2365 	addr = PAGE_ALIGN((unsigned long)(__init_begin));
2366 	initend = (unsigned long)(__init_end) & PAGE_MASK;
2367 	for (; addr < initend; addr += PAGE_SIZE) {
2368 		unsigned long page;
2369 
2370 		page = (addr +
2371 			((unsigned long) __va(kern_base)) -
2372 			((unsigned long) KERNBASE));
2373 		memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
2374 
2375 		if (do_free)
2376 			free_reserved_page(virt_to_page(page));
2377 	}
2378 }
2379 
2380 #ifdef CONFIG_BLK_DEV_INITRD
2381 void free_initrd_mem(unsigned long start, unsigned long end)
2382 {
2383 	free_reserved_area((void *)start, (void *)end, POISON_FREE_INITMEM,
2384 			   "initrd");
2385 }
2386 #endif
2387 
2388 pgprot_t PAGE_KERNEL __read_mostly;
2389 EXPORT_SYMBOL(PAGE_KERNEL);
2390 
2391 pgprot_t PAGE_KERNEL_LOCKED __read_mostly;
2392 pgprot_t PAGE_COPY __read_mostly;
2393 
2394 pgprot_t PAGE_SHARED __read_mostly;
2395 EXPORT_SYMBOL(PAGE_SHARED);
2396 
2397 unsigned long pg_iobits __read_mostly;
2398 
2399 unsigned long _PAGE_IE __read_mostly;
2400 EXPORT_SYMBOL(_PAGE_IE);
2401 
2402 unsigned long _PAGE_E __read_mostly;
2403 EXPORT_SYMBOL(_PAGE_E);
2404 
2405 unsigned long _PAGE_CACHE __read_mostly;
2406 EXPORT_SYMBOL(_PAGE_CACHE);
2407 
2408 #ifdef CONFIG_SPARSEMEM_VMEMMAP
2409 int __meminit vmemmap_populate(unsigned long vstart, unsigned long vend,
2410 			       int node)
2411 {
2412 	unsigned long pte_base;
2413 
2414 	pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2415 		    _PAGE_CP_4U | _PAGE_CV_4U |
2416 		    _PAGE_P_4U | _PAGE_W_4U);
2417 	if (tlb_type == hypervisor)
2418 		pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2419 			    page_cache4v_flag | _PAGE_P_4V | _PAGE_W_4V);
2420 
2421 	pte_base |= _PAGE_PMD_HUGE;
2422 
2423 	vstart = vstart & PMD_MASK;
2424 	vend = ALIGN(vend, PMD_SIZE);
2425 	for (; vstart < vend; vstart += PMD_SIZE) {
2426 		pgd_t *pgd = pgd_offset_k(vstart);
2427 		unsigned long pte;
2428 		pud_t *pud;
2429 		pmd_t *pmd;
2430 
2431 		if (pgd_none(*pgd)) {
2432 			pud_t *new = vmemmap_alloc_block(PAGE_SIZE, node);
2433 
2434 			if (!new)
2435 				return -ENOMEM;
2436 			pgd_populate(&init_mm, pgd, new);
2437 		}
2438 
2439 		pud = pud_offset(pgd, vstart);
2440 		if (pud_none(*pud)) {
2441 			pmd_t *new = vmemmap_alloc_block(PAGE_SIZE, node);
2442 
2443 			if (!new)
2444 				return -ENOMEM;
2445 			pud_populate(&init_mm, pud, new);
2446 		}
2447 
2448 		pmd = pmd_offset(pud, vstart);
2449 
2450 		pte = pmd_val(*pmd);
2451 		if (!(pte & _PAGE_VALID)) {
2452 			void *block = vmemmap_alloc_block(PMD_SIZE, node);
2453 
2454 			if (!block)
2455 				return -ENOMEM;
2456 
2457 			pmd_val(*pmd) = pte_base | __pa(block);
2458 		}
2459 	}
2460 
2461 	return 0;
2462 }
2463 
2464 void vmemmap_free(unsigned long start, unsigned long end)
2465 {
2466 }
2467 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
2468 
2469 static void prot_init_common(unsigned long page_none,
2470 			     unsigned long page_shared,
2471 			     unsigned long page_copy,
2472 			     unsigned long page_readonly,
2473 			     unsigned long page_exec_bit)
2474 {
2475 	PAGE_COPY = __pgprot(page_copy);
2476 	PAGE_SHARED = __pgprot(page_shared);
2477 
2478 	protection_map[0x0] = __pgprot(page_none);
2479 	protection_map[0x1] = __pgprot(page_readonly & ~page_exec_bit);
2480 	protection_map[0x2] = __pgprot(page_copy & ~page_exec_bit);
2481 	protection_map[0x3] = __pgprot(page_copy & ~page_exec_bit);
2482 	protection_map[0x4] = __pgprot(page_readonly);
2483 	protection_map[0x5] = __pgprot(page_readonly);
2484 	protection_map[0x6] = __pgprot(page_copy);
2485 	protection_map[0x7] = __pgprot(page_copy);
2486 	protection_map[0x8] = __pgprot(page_none);
2487 	protection_map[0x9] = __pgprot(page_readonly & ~page_exec_bit);
2488 	protection_map[0xa] = __pgprot(page_shared & ~page_exec_bit);
2489 	protection_map[0xb] = __pgprot(page_shared & ~page_exec_bit);
2490 	protection_map[0xc] = __pgprot(page_readonly);
2491 	protection_map[0xd] = __pgprot(page_readonly);
2492 	protection_map[0xe] = __pgprot(page_shared);
2493 	protection_map[0xf] = __pgprot(page_shared);
2494 }
2495 
2496 static void __init sun4u_pgprot_init(void)
2497 {
2498 	unsigned long page_none, page_shared, page_copy, page_readonly;
2499 	unsigned long page_exec_bit;
2500 	int i;
2501 
2502 	PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2503 				_PAGE_CACHE_4U | _PAGE_P_4U |
2504 				__ACCESS_BITS_4U | __DIRTY_BITS_4U |
2505 				_PAGE_EXEC_4U);
2506 	PAGE_KERNEL_LOCKED = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2507 				       _PAGE_CACHE_4U | _PAGE_P_4U |
2508 				       __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2509 				       _PAGE_EXEC_4U | _PAGE_L_4U);
2510 
2511 	_PAGE_IE = _PAGE_IE_4U;
2512 	_PAGE_E = _PAGE_E_4U;
2513 	_PAGE_CACHE = _PAGE_CACHE_4U;
2514 
2515 	pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4U | __DIRTY_BITS_4U |
2516 		     __ACCESS_BITS_4U | _PAGE_E_4U);
2517 
2518 #ifdef CONFIG_DEBUG_PAGEALLOC
2519 	kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2520 #else
2521 	kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4U) ^
2522 		PAGE_OFFSET;
2523 #endif
2524 	kern_linear_pte_xor[0] |= (_PAGE_CP_4U | _PAGE_CV_4U |
2525 				   _PAGE_P_4U | _PAGE_W_4U);
2526 
2527 	for (i = 1; i < 4; i++)
2528 		kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2529 
2530 	_PAGE_ALL_SZ_BITS =  (_PAGE_SZ4MB_4U | _PAGE_SZ512K_4U |
2531 			      _PAGE_SZ64K_4U | _PAGE_SZ8K_4U |
2532 			      _PAGE_SZ32MB_4U | _PAGE_SZ256MB_4U);
2533 
2534 
2535 	page_none = _PAGE_PRESENT_4U | _PAGE_ACCESSED_4U | _PAGE_CACHE_4U;
2536 	page_shared = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2537 		       __ACCESS_BITS_4U | _PAGE_WRITE_4U | _PAGE_EXEC_4U);
2538 	page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2539 		       __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2540 	page_readonly   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2541 			   __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2542 
2543 	page_exec_bit = _PAGE_EXEC_4U;
2544 
2545 	prot_init_common(page_none, page_shared, page_copy, page_readonly,
2546 			 page_exec_bit);
2547 }
2548 
2549 static void __init sun4v_pgprot_init(void)
2550 {
2551 	unsigned long page_none, page_shared, page_copy, page_readonly;
2552 	unsigned long page_exec_bit;
2553 	int i;
2554 
2555 	PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4V | _PAGE_VALID |
2556 				page_cache4v_flag | _PAGE_P_4V |
2557 				__ACCESS_BITS_4V | __DIRTY_BITS_4V |
2558 				_PAGE_EXEC_4V);
2559 	PAGE_KERNEL_LOCKED = PAGE_KERNEL;
2560 
2561 	_PAGE_IE = _PAGE_IE_4V;
2562 	_PAGE_E = _PAGE_E_4V;
2563 	_PAGE_CACHE = page_cache4v_flag;
2564 
2565 #ifdef CONFIG_DEBUG_PAGEALLOC
2566 	kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2567 #else
2568 	kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4V) ^
2569 		PAGE_OFFSET;
2570 #endif
2571 	kern_linear_pte_xor[0] |= (page_cache4v_flag | _PAGE_P_4V |
2572 				   _PAGE_W_4V);
2573 
2574 	for (i = 1; i < 4; i++)
2575 		kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2576 
2577 	pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4V | __DIRTY_BITS_4V |
2578 		     __ACCESS_BITS_4V | _PAGE_E_4V);
2579 
2580 	_PAGE_ALL_SZ_BITS = (_PAGE_SZ16GB_4V | _PAGE_SZ2GB_4V |
2581 			     _PAGE_SZ256MB_4V | _PAGE_SZ32MB_4V |
2582 			     _PAGE_SZ4MB_4V | _PAGE_SZ512K_4V |
2583 			     _PAGE_SZ64K_4V | _PAGE_SZ8K_4V);
2584 
2585 	page_none = _PAGE_PRESENT_4V | _PAGE_ACCESSED_4V | page_cache4v_flag;
2586 	page_shared = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2587 		       __ACCESS_BITS_4V | _PAGE_WRITE_4V | _PAGE_EXEC_4V);
2588 	page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2589 		       __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2590 	page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2591 			 __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2592 
2593 	page_exec_bit = _PAGE_EXEC_4V;
2594 
2595 	prot_init_common(page_none, page_shared, page_copy, page_readonly,
2596 			 page_exec_bit);
2597 }
2598 
2599 unsigned long pte_sz_bits(unsigned long sz)
2600 {
2601 	if (tlb_type == hypervisor) {
2602 		switch (sz) {
2603 		case 8 * 1024:
2604 		default:
2605 			return _PAGE_SZ8K_4V;
2606 		case 64 * 1024:
2607 			return _PAGE_SZ64K_4V;
2608 		case 512 * 1024:
2609 			return _PAGE_SZ512K_4V;
2610 		case 4 * 1024 * 1024:
2611 			return _PAGE_SZ4MB_4V;
2612 		}
2613 	} else {
2614 		switch (sz) {
2615 		case 8 * 1024:
2616 		default:
2617 			return _PAGE_SZ8K_4U;
2618 		case 64 * 1024:
2619 			return _PAGE_SZ64K_4U;
2620 		case 512 * 1024:
2621 			return _PAGE_SZ512K_4U;
2622 		case 4 * 1024 * 1024:
2623 			return _PAGE_SZ4MB_4U;
2624 		}
2625 	}
2626 }
2627 
2628 pte_t mk_pte_io(unsigned long page, pgprot_t prot, int space, unsigned long page_size)
2629 {
2630 	pte_t pte;
2631 
2632 	pte_val(pte)  = page | pgprot_val(pgprot_noncached(prot));
2633 	pte_val(pte) |= (((unsigned long)space) << 32);
2634 	pte_val(pte) |= pte_sz_bits(page_size);
2635 
2636 	return pte;
2637 }
2638 
2639 static unsigned long kern_large_tte(unsigned long paddr)
2640 {
2641 	unsigned long val;
2642 
2643 	val = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2644 	       _PAGE_CP_4U | _PAGE_CV_4U | _PAGE_P_4U |
2645 	       _PAGE_EXEC_4U | _PAGE_L_4U | _PAGE_W_4U);
2646 	if (tlb_type == hypervisor)
2647 		val = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2648 		       page_cache4v_flag | _PAGE_P_4V |
2649 		       _PAGE_EXEC_4V | _PAGE_W_4V);
2650 
2651 	return val | paddr;
2652 }
2653 
2654 /* If not locked, zap it. */
2655 void __flush_tlb_all(void)
2656 {
2657 	unsigned long pstate;
2658 	int i;
2659 
2660 	__asm__ __volatile__("flushw\n\t"
2661 			     "rdpr	%%pstate, %0\n\t"
2662 			     "wrpr	%0, %1, %%pstate"
2663 			     : "=r" (pstate)
2664 			     : "i" (PSTATE_IE));
2665 	if (tlb_type == hypervisor) {
2666 		sun4v_mmu_demap_all();
2667 	} else if (tlb_type == spitfire) {
2668 		for (i = 0; i < 64; i++) {
2669 			/* Spitfire Errata #32 workaround */
2670 			/* NOTE: Always runs on spitfire, so no
2671 			 *       cheetah+ page size encodings.
2672 			 */
2673 			__asm__ __volatile__("stxa	%0, [%1] %2\n\t"
2674 					     "flush	%%g6"
2675 					     : /* No outputs */
2676 					     : "r" (0),
2677 					     "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2678 
2679 			if (!(spitfire_get_dtlb_data(i) & _PAGE_L_4U)) {
2680 				__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2681 						     "membar #Sync"
2682 						     : /* no outputs */
2683 						     : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
2684 				spitfire_put_dtlb_data(i, 0x0UL);
2685 			}
2686 
2687 			/* Spitfire Errata #32 workaround */
2688 			/* NOTE: Always runs on spitfire, so no
2689 			 *       cheetah+ page size encodings.
2690 			 */
2691 			__asm__ __volatile__("stxa	%0, [%1] %2\n\t"
2692 					     "flush	%%g6"
2693 					     : /* No outputs */
2694 					     : "r" (0),
2695 					     "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2696 
2697 			if (!(spitfire_get_itlb_data(i) & _PAGE_L_4U)) {
2698 				__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2699 						     "membar #Sync"
2700 						     : /* no outputs */
2701 						     : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
2702 				spitfire_put_itlb_data(i, 0x0UL);
2703 			}
2704 		}
2705 	} else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
2706 		cheetah_flush_dtlb_all();
2707 		cheetah_flush_itlb_all();
2708 	}
2709 	__asm__ __volatile__("wrpr	%0, 0, %%pstate"
2710 			     : : "r" (pstate));
2711 }
2712 
2713 pte_t *pte_alloc_one_kernel(struct mm_struct *mm,
2714 			    unsigned long address)
2715 {
2716 	struct page *page = alloc_page(GFP_KERNEL | __GFP_NOTRACK |
2717 				       __GFP_REPEAT | __GFP_ZERO);
2718 	pte_t *pte = NULL;
2719 
2720 	if (page)
2721 		pte = (pte_t *) page_address(page);
2722 
2723 	return pte;
2724 }
2725 
2726 pgtable_t pte_alloc_one(struct mm_struct *mm,
2727 			unsigned long address)
2728 {
2729 	struct page *page = alloc_page(GFP_KERNEL | __GFP_NOTRACK |
2730 				       __GFP_REPEAT | __GFP_ZERO);
2731 	if (!page)
2732 		return NULL;
2733 	if (!pgtable_page_ctor(page)) {
2734 		free_hot_cold_page(page, 0);
2735 		return NULL;
2736 	}
2737 	return (pte_t *) page_address(page);
2738 }
2739 
2740 void pte_free_kernel(struct mm_struct *mm, pte_t *pte)
2741 {
2742 	free_page((unsigned long)pte);
2743 }
2744 
2745 static void __pte_free(pgtable_t pte)
2746 {
2747 	struct page *page = virt_to_page(pte);
2748 
2749 	pgtable_page_dtor(page);
2750 	__free_page(page);
2751 }
2752 
2753 void pte_free(struct mm_struct *mm, pgtable_t pte)
2754 {
2755 	__pte_free(pte);
2756 }
2757 
2758 void pgtable_free(void *table, bool is_page)
2759 {
2760 	if (is_page)
2761 		__pte_free(table);
2762 	else
2763 		kmem_cache_free(pgtable_cache, table);
2764 }
2765 
2766 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2767 void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
2768 			  pmd_t *pmd)
2769 {
2770 	unsigned long pte, flags;
2771 	struct mm_struct *mm;
2772 	pmd_t entry = *pmd;
2773 
2774 	if (!pmd_large(entry) || !pmd_young(entry))
2775 		return;
2776 
2777 	pte = pmd_val(entry);
2778 
2779 	/* Don't insert a non-valid PMD into the TSB, we'll deadlock.  */
2780 	if (!(pte & _PAGE_VALID))
2781 		return;
2782 
2783 	/* We are fabricating 8MB pages using 4MB real hw pages.  */
2784 	pte |= (addr & (1UL << REAL_HPAGE_SHIFT));
2785 
2786 	mm = vma->vm_mm;
2787 
2788 	spin_lock_irqsave(&mm->context.lock, flags);
2789 
2790 	if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL)
2791 		__update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
2792 					addr, pte);
2793 
2794 	spin_unlock_irqrestore(&mm->context.lock, flags);
2795 }
2796 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2797 
2798 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
2799 static void context_reload(void *__data)
2800 {
2801 	struct mm_struct *mm = __data;
2802 
2803 	if (mm == current->mm)
2804 		load_secondary_context(mm);
2805 }
2806 
2807 void hugetlb_setup(struct pt_regs *regs)
2808 {
2809 	struct mm_struct *mm = current->mm;
2810 	struct tsb_config *tp;
2811 
2812 	if (faulthandler_disabled() || !mm) {
2813 		const struct exception_table_entry *entry;
2814 
2815 		entry = search_exception_tables(regs->tpc);
2816 		if (entry) {
2817 			regs->tpc = entry->fixup;
2818 			regs->tnpc = regs->tpc + 4;
2819 			return;
2820 		}
2821 		pr_alert("Unexpected HugeTLB setup in atomic context.\n");
2822 		die_if_kernel("HugeTSB in atomic", regs);
2823 	}
2824 
2825 	tp = &mm->context.tsb_block[MM_TSB_HUGE];
2826 	if (likely(tp->tsb == NULL))
2827 		tsb_grow(mm, MM_TSB_HUGE, 0);
2828 
2829 	tsb_context_switch(mm);
2830 	smp_tsb_sync(mm);
2831 
2832 	/* On UltraSPARC-III+ and later, configure the second half of
2833 	 * the Data-TLB for huge pages.
2834 	 */
2835 	if (tlb_type == cheetah_plus) {
2836 		unsigned long ctx;
2837 
2838 		spin_lock(&ctx_alloc_lock);
2839 		ctx = mm->context.sparc64_ctx_val;
2840 		ctx &= ~CTX_PGSZ_MASK;
2841 		ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT;
2842 		ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT;
2843 
2844 		if (ctx != mm->context.sparc64_ctx_val) {
2845 			/* When changing the page size fields, we
2846 			 * must perform a context flush so that no
2847 			 * stale entries match.  This flush must
2848 			 * occur with the original context register
2849 			 * settings.
2850 			 */
2851 			do_flush_tlb_mm(mm);
2852 
2853 			/* Reload the context register of all processors
2854 			 * also executing in this address space.
2855 			 */
2856 			mm->context.sparc64_ctx_val = ctx;
2857 			on_each_cpu(context_reload, mm, 0);
2858 		}
2859 		spin_unlock(&ctx_alloc_lock);
2860 	}
2861 }
2862 #endif
2863 
2864 static struct resource code_resource = {
2865 	.name	= "Kernel code",
2866 	.flags	= IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
2867 };
2868 
2869 static struct resource data_resource = {
2870 	.name	= "Kernel data",
2871 	.flags	= IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
2872 };
2873 
2874 static struct resource bss_resource = {
2875 	.name	= "Kernel bss",
2876 	.flags	= IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
2877 };
2878 
2879 static inline resource_size_t compute_kern_paddr(void *addr)
2880 {
2881 	return (resource_size_t) (addr - KERNBASE + kern_base);
2882 }
2883 
2884 static void __init kernel_lds_init(void)
2885 {
2886 	code_resource.start = compute_kern_paddr(_text);
2887 	code_resource.end   = compute_kern_paddr(_etext - 1);
2888 	data_resource.start = compute_kern_paddr(_etext);
2889 	data_resource.end   = compute_kern_paddr(_edata - 1);
2890 	bss_resource.start  = compute_kern_paddr(__bss_start);
2891 	bss_resource.end    = compute_kern_paddr(_end - 1);
2892 }
2893 
2894 static int __init report_memory(void)
2895 {
2896 	int i;
2897 	struct resource *res;
2898 
2899 	kernel_lds_init();
2900 
2901 	for (i = 0; i < pavail_ents; i++) {
2902 		res = kzalloc(sizeof(struct resource), GFP_KERNEL);
2903 
2904 		if (!res) {
2905 			pr_warn("Failed to allocate source.\n");
2906 			break;
2907 		}
2908 
2909 		res->name = "System RAM";
2910 		res->start = pavail[i].phys_addr;
2911 		res->end = pavail[i].phys_addr + pavail[i].reg_size - 1;
2912 		res->flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM;
2913 
2914 		if (insert_resource(&iomem_resource, res) < 0) {
2915 			pr_warn("Resource insertion failed.\n");
2916 			break;
2917 		}
2918 
2919 		insert_resource(res, &code_resource);
2920 		insert_resource(res, &data_resource);
2921 		insert_resource(res, &bss_resource);
2922 	}
2923 
2924 	return 0;
2925 }
2926 arch_initcall(report_memory);
2927 
2928 #ifdef CONFIG_SMP
2929 #define do_flush_tlb_kernel_range	smp_flush_tlb_kernel_range
2930 #else
2931 #define do_flush_tlb_kernel_range	__flush_tlb_kernel_range
2932 #endif
2933 
2934 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
2935 {
2936 	if (start < HI_OBP_ADDRESS && end > LOW_OBP_ADDRESS) {
2937 		if (start < LOW_OBP_ADDRESS) {
2938 			flush_tsb_kernel_range(start, LOW_OBP_ADDRESS);
2939 			do_flush_tlb_kernel_range(start, LOW_OBP_ADDRESS);
2940 		}
2941 		if (end > HI_OBP_ADDRESS) {
2942 			flush_tsb_kernel_range(HI_OBP_ADDRESS, end);
2943 			do_flush_tlb_kernel_range(HI_OBP_ADDRESS, end);
2944 		}
2945 	} else {
2946 		flush_tsb_kernel_range(start, end);
2947 		do_flush_tlb_kernel_range(start, end);
2948 	}
2949 }
2950