xref: /openbmc/linux/arch/sparc/mm/init_64.c (revision 4f6cce39)
1 /*
2  *  arch/sparc64/mm/init.c
3  *
4  *  Copyright (C) 1996-1999 David S. Miller (davem@caip.rutgers.edu)
5  *  Copyright (C) 1997-1999 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
6  */
7 
8 #include <linux/extable.h>
9 #include <linux/kernel.h>
10 #include <linux/sched.h>
11 #include <linux/string.h>
12 #include <linux/init.h>
13 #include <linux/bootmem.h>
14 #include <linux/mm.h>
15 #include <linux/hugetlb.h>
16 #include <linux/initrd.h>
17 #include <linux/swap.h>
18 #include <linux/pagemap.h>
19 #include <linux/poison.h>
20 #include <linux/fs.h>
21 #include <linux/seq_file.h>
22 #include <linux/kprobes.h>
23 #include <linux/cache.h>
24 #include <linux/sort.h>
25 #include <linux/ioport.h>
26 #include <linux/percpu.h>
27 #include <linux/memblock.h>
28 #include <linux/mmzone.h>
29 #include <linux/gfp.h>
30 
31 #include <asm/head.h>
32 #include <asm/page.h>
33 #include <asm/pgalloc.h>
34 #include <asm/pgtable.h>
35 #include <asm/oplib.h>
36 #include <asm/iommu.h>
37 #include <asm/io.h>
38 #include <linux/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 #ifdef CONFIG_HUGETLB_PAGE
328 static int __init setup_hugepagesz(char *string)
329 {
330 	unsigned long long hugepage_size;
331 	unsigned int hugepage_shift;
332 	unsigned short hv_pgsz_idx;
333 	unsigned int hv_pgsz_mask;
334 	int rc = 0;
335 
336 	hugepage_size = memparse(string, &string);
337 	hugepage_shift = ilog2(hugepage_size);
338 
339 	switch (hugepage_shift) {
340 	case HPAGE_256MB_SHIFT:
341 		hv_pgsz_mask = HV_PGSZ_MASK_256MB;
342 		hv_pgsz_idx = HV_PGSZ_IDX_256MB;
343 		break;
344 	case HPAGE_SHIFT:
345 		hv_pgsz_mask = HV_PGSZ_MASK_4MB;
346 		hv_pgsz_idx = HV_PGSZ_IDX_4MB;
347 		break;
348 	case HPAGE_64K_SHIFT:
349 		hv_pgsz_mask = HV_PGSZ_MASK_64K;
350 		hv_pgsz_idx = HV_PGSZ_IDX_64K;
351 		break;
352 	default:
353 		hv_pgsz_mask = 0;
354 	}
355 
356 	if ((hv_pgsz_mask & cpu_pgsz_mask) == 0U) {
357 		pr_warn("hugepagesz=%llu not supported by MMU.\n",
358 			hugepage_size);
359 		goto out;
360 	}
361 
362 	hugetlb_add_hstate(hugepage_shift - PAGE_SHIFT);
363 	rc = 1;
364 
365 out:
366 	return rc;
367 }
368 __setup("hugepagesz=", setup_hugepagesz);
369 #endif	/* CONFIG_HUGETLB_PAGE */
370 
371 void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
372 {
373 	struct mm_struct *mm;
374 	unsigned long flags;
375 	pte_t pte = *ptep;
376 
377 	if (tlb_type != hypervisor) {
378 		unsigned long pfn = pte_pfn(pte);
379 
380 		if (pfn_valid(pfn))
381 			flush_dcache(pfn);
382 	}
383 
384 	mm = vma->vm_mm;
385 
386 	/* Don't insert a non-valid PTE into the TSB, we'll deadlock.  */
387 	if (!pte_accessible(mm, pte))
388 		return;
389 
390 	spin_lock_irqsave(&mm->context.lock, flags);
391 
392 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
393 	if ((mm->context.hugetlb_pte_count || mm->context.thp_pte_count) &&
394 	    is_hugetlb_pmd(__pmd(pte_val(pte)))) {
395 		/* We are fabricating 8MB pages using 4MB real hw pages.  */
396 		pte_val(pte) |= (address & (1UL << REAL_HPAGE_SHIFT));
397 		__update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
398 					address, pte_val(pte));
399 	} else
400 #endif
401 		__update_mmu_tsb_insert(mm, MM_TSB_BASE, PAGE_SHIFT,
402 					address, pte_val(pte));
403 
404 	spin_unlock_irqrestore(&mm->context.lock, flags);
405 }
406 
407 void flush_dcache_page(struct page *page)
408 {
409 	struct address_space *mapping;
410 	int this_cpu;
411 
412 	if (tlb_type == hypervisor)
413 		return;
414 
415 	/* Do not bother with the expensive D-cache flush if it
416 	 * is merely the zero page.  The 'bigcore' testcase in GDB
417 	 * causes this case to run millions of times.
418 	 */
419 	if (page == ZERO_PAGE(0))
420 		return;
421 
422 	this_cpu = get_cpu();
423 
424 	mapping = page_mapping(page);
425 	if (mapping && !mapping_mapped(mapping)) {
426 		int dirty = test_bit(PG_dcache_dirty, &page->flags);
427 		if (dirty) {
428 			int dirty_cpu = dcache_dirty_cpu(page);
429 
430 			if (dirty_cpu == this_cpu)
431 				goto out;
432 			smp_flush_dcache_page_impl(page, dirty_cpu);
433 		}
434 		set_dcache_dirty(page, this_cpu);
435 	} else {
436 		/* We could delay the flush for the !page_mapping
437 		 * case too.  But that case is for exec env/arg
438 		 * pages and those are %99 certainly going to get
439 		 * faulted into the tlb (and thus flushed) anyways.
440 		 */
441 		flush_dcache_page_impl(page);
442 	}
443 
444 out:
445 	put_cpu();
446 }
447 EXPORT_SYMBOL(flush_dcache_page);
448 
449 void __kprobes flush_icache_range(unsigned long start, unsigned long end)
450 {
451 	/* Cheetah and Hypervisor platform cpus have coherent I-cache. */
452 	if (tlb_type == spitfire) {
453 		unsigned long kaddr;
454 
455 		/* This code only runs on Spitfire cpus so this is
456 		 * why we can assume _PAGE_PADDR_4U.
457 		 */
458 		for (kaddr = start; kaddr < end; kaddr += PAGE_SIZE) {
459 			unsigned long paddr, mask = _PAGE_PADDR_4U;
460 
461 			if (kaddr >= PAGE_OFFSET)
462 				paddr = kaddr & mask;
463 			else {
464 				pgd_t *pgdp = pgd_offset_k(kaddr);
465 				pud_t *pudp = pud_offset(pgdp, kaddr);
466 				pmd_t *pmdp = pmd_offset(pudp, kaddr);
467 				pte_t *ptep = pte_offset_kernel(pmdp, kaddr);
468 
469 				paddr = pte_val(*ptep) & mask;
470 			}
471 			__flush_icache_page(paddr);
472 		}
473 	}
474 }
475 EXPORT_SYMBOL(flush_icache_range);
476 
477 void mmu_info(struct seq_file *m)
478 {
479 	static const char *pgsz_strings[] = {
480 		"8K", "64K", "512K", "4MB", "32MB",
481 		"256MB", "2GB", "16GB",
482 	};
483 	int i, printed;
484 
485 	if (tlb_type == cheetah)
486 		seq_printf(m, "MMU Type\t: Cheetah\n");
487 	else if (tlb_type == cheetah_plus)
488 		seq_printf(m, "MMU Type\t: Cheetah+\n");
489 	else if (tlb_type == spitfire)
490 		seq_printf(m, "MMU Type\t: Spitfire\n");
491 	else if (tlb_type == hypervisor)
492 		seq_printf(m, "MMU Type\t: Hypervisor (sun4v)\n");
493 	else
494 		seq_printf(m, "MMU Type\t: ???\n");
495 
496 	seq_printf(m, "MMU PGSZs\t: ");
497 	printed = 0;
498 	for (i = 0; i < ARRAY_SIZE(pgsz_strings); i++) {
499 		if (cpu_pgsz_mask & (1UL << i)) {
500 			seq_printf(m, "%s%s",
501 				   printed ? "," : "", pgsz_strings[i]);
502 			printed++;
503 		}
504 	}
505 	seq_putc(m, '\n');
506 
507 #ifdef CONFIG_DEBUG_DCFLUSH
508 	seq_printf(m, "DCPageFlushes\t: %d\n",
509 		   atomic_read(&dcpage_flushes));
510 #ifdef CONFIG_SMP
511 	seq_printf(m, "DCPageFlushesXC\t: %d\n",
512 		   atomic_read(&dcpage_flushes_xcall));
513 #endif /* CONFIG_SMP */
514 #endif /* CONFIG_DEBUG_DCFLUSH */
515 }
516 
517 struct linux_prom_translation prom_trans[512] __read_mostly;
518 unsigned int prom_trans_ents __read_mostly;
519 
520 unsigned long kern_locked_tte_data;
521 
522 /* The obp translations are saved based on 8k pagesize, since obp can
523  * use a mixture of pagesizes. Misses to the LOW_OBP_ADDRESS ->
524  * HI_OBP_ADDRESS range are handled in ktlb.S.
525  */
526 static inline int in_obp_range(unsigned long vaddr)
527 {
528 	return (vaddr >= LOW_OBP_ADDRESS &&
529 		vaddr < HI_OBP_ADDRESS);
530 }
531 
532 static int cmp_ptrans(const void *a, const void *b)
533 {
534 	const struct linux_prom_translation *x = a, *y = b;
535 
536 	if (x->virt > y->virt)
537 		return 1;
538 	if (x->virt < y->virt)
539 		return -1;
540 	return 0;
541 }
542 
543 /* Read OBP translations property into 'prom_trans[]'.  */
544 static void __init read_obp_translations(void)
545 {
546 	int n, node, ents, first, last, i;
547 
548 	node = prom_finddevice("/virtual-memory");
549 	n = prom_getproplen(node, "translations");
550 	if (unlikely(n == 0 || n == -1)) {
551 		prom_printf("prom_mappings: Couldn't get size.\n");
552 		prom_halt();
553 	}
554 	if (unlikely(n > sizeof(prom_trans))) {
555 		prom_printf("prom_mappings: Size %d is too big.\n", n);
556 		prom_halt();
557 	}
558 
559 	if ((n = prom_getproperty(node, "translations",
560 				  (char *)&prom_trans[0],
561 				  sizeof(prom_trans))) == -1) {
562 		prom_printf("prom_mappings: Couldn't get property.\n");
563 		prom_halt();
564 	}
565 
566 	n = n / sizeof(struct linux_prom_translation);
567 
568 	ents = n;
569 
570 	sort(prom_trans, ents, sizeof(struct linux_prom_translation),
571 	     cmp_ptrans, NULL);
572 
573 	/* Now kick out all the non-OBP entries.  */
574 	for (i = 0; i < ents; i++) {
575 		if (in_obp_range(prom_trans[i].virt))
576 			break;
577 	}
578 	first = i;
579 	for (; i < ents; i++) {
580 		if (!in_obp_range(prom_trans[i].virt))
581 			break;
582 	}
583 	last = i;
584 
585 	for (i = 0; i < (last - first); i++) {
586 		struct linux_prom_translation *src = &prom_trans[i + first];
587 		struct linux_prom_translation *dest = &prom_trans[i];
588 
589 		*dest = *src;
590 	}
591 	for (; i < ents; i++) {
592 		struct linux_prom_translation *dest = &prom_trans[i];
593 		dest->virt = dest->size = dest->data = 0x0UL;
594 	}
595 
596 	prom_trans_ents = last - first;
597 
598 	if (tlb_type == spitfire) {
599 		/* Clear diag TTE bits. */
600 		for (i = 0; i < prom_trans_ents; i++)
601 			prom_trans[i].data &= ~0x0003fe0000000000UL;
602 	}
603 
604 	/* Force execute bit on.  */
605 	for (i = 0; i < prom_trans_ents; i++)
606 		prom_trans[i].data |= (tlb_type == hypervisor ?
607 				       _PAGE_EXEC_4V : _PAGE_EXEC_4U);
608 }
609 
610 static void __init hypervisor_tlb_lock(unsigned long vaddr,
611 				       unsigned long pte,
612 				       unsigned long mmu)
613 {
614 	unsigned long ret = sun4v_mmu_map_perm_addr(vaddr, 0, pte, mmu);
615 
616 	if (ret != 0) {
617 		prom_printf("hypervisor_tlb_lock[%lx:%x:%lx:%lx]: "
618 			    "errors with %lx\n", vaddr, 0, pte, mmu, ret);
619 		prom_halt();
620 	}
621 }
622 
623 static unsigned long kern_large_tte(unsigned long paddr);
624 
625 static void __init remap_kernel(void)
626 {
627 	unsigned long phys_page, tte_vaddr, tte_data;
628 	int i, tlb_ent = sparc64_highest_locked_tlbent();
629 
630 	tte_vaddr = (unsigned long) KERNBASE;
631 	phys_page = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
632 	tte_data = kern_large_tte(phys_page);
633 
634 	kern_locked_tte_data = tte_data;
635 
636 	/* Now lock us into the TLBs via Hypervisor or OBP. */
637 	if (tlb_type == hypervisor) {
638 		for (i = 0; i < num_kernel_image_mappings; i++) {
639 			hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_DMMU);
640 			hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_IMMU);
641 			tte_vaddr += 0x400000;
642 			tte_data += 0x400000;
643 		}
644 	} else {
645 		for (i = 0; i < num_kernel_image_mappings; i++) {
646 			prom_dtlb_load(tlb_ent - i, tte_data, tte_vaddr);
647 			prom_itlb_load(tlb_ent - i, tte_data, tte_vaddr);
648 			tte_vaddr += 0x400000;
649 			tte_data += 0x400000;
650 		}
651 		sparc64_highest_unlocked_tlb_ent = tlb_ent - i;
652 	}
653 	if (tlb_type == cheetah_plus) {
654 		sparc64_kern_pri_context = (CTX_CHEETAH_PLUS_CTX0 |
655 					    CTX_CHEETAH_PLUS_NUC);
656 		sparc64_kern_pri_nuc_bits = CTX_CHEETAH_PLUS_NUC;
657 		sparc64_kern_sec_context = CTX_CHEETAH_PLUS_CTX0;
658 	}
659 }
660 
661 
662 static void __init inherit_prom_mappings(void)
663 {
664 	/* Now fixup OBP's idea about where we really are mapped. */
665 	printk("Remapping the kernel... ");
666 	remap_kernel();
667 	printk("done.\n");
668 }
669 
670 void prom_world(int enter)
671 {
672 	if (!enter)
673 		set_fs(get_fs());
674 
675 	__asm__ __volatile__("flushw");
676 }
677 
678 void __flush_dcache_range(unsigned long start, unsigned long end)
679 {
680 	unsigned long va;
681 
682 	if (tlb_type == spitfire) {
683 		int n = 0;
684 
685 		for (va = start; va < end; va += 32) {
686 			spitfire_put_dcache_tag(va & 0x3fe0, 0x0);
687 			if (++n >= 512)
688 				break;
689 		}
690 	} else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
691 		start = __pa(start);
692 		end = __pa(end);
693 		for (va = start; va < end; va += 32)
694 			__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
695 					     "membar #Sync"
696 					     : /* no outputs */
697 					     : "r" (va),
698 					       "i" (ASI_DCACHE_INVALIDATE));
699 	}
700 }
701 EXPORT_SYMBOL(__flush_dcache_range);
702 
703 /* get_new_mmu_context() uses "cache + 1".  */
704 DEFINE_SPINLOCK(ctx_alloc_lock);
705 unsigned long tlb_context_cache = CTX_FIRST_VERSION - 1;
706 #define MAX_CTX_NR	(1UL << CTX_NR_BITS)
707 #define CTX_BMAP_SLOTS	BITS_TO_LONGS(MAX_CTX_NR)
708 DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR);
709 
710 /* Caller does TLB context flushing on local CPU if necessary.
711  * The caller also ensures that CTX_VALID(mm->context) is false.
712  *
713  * We must be careful about boundary cases so that we never
714  * let the user have CTX 0 (nucleus) or we ever use a CTX
715  * version of zero (and thus NO_CONTEXT would not be caught
716  * by version mis-match tests in mmu_context.h).
717  *
718  * Always invoked with interrupts disabled.
719  */
720 void get_new_mmu_context(struct mm_struct *mm)
721 {
722 	unsigned long ctx, new_ctx;
723 	unsigned long orig_pgsz_bits;
724 	int new_version;
725 
726 	spin_lock(&ctx_alloc_lock);
727 	orig_pgsz_bits = (mm->context.sparc64_ctx_val & CTX_PGSZ_MASK);
728 	ctx = (tlb_context_cache + 1) & CTX_NR_MASK;
729 	new_ctx = find_next_zero_bit(mmu_context_bmap, 1 << CTX_NR_BITS, ctx);
730 	new_version = 0;
731 	if (new_ctx >= (1 << CTX_NR_BITS)) {
732 		new_ctx = find_next_zero_bit(mmu_context_bmap, ctx, 1);
733 		if (new_ctx >= ctx) {
734 			int i;
735 			new_ctx = (tlb_context_cache & CTX_VERSION_MASK) +
736 				CTX_FIRST_VERSION;
737 			if (new_ctx == 1)
738 				new_ctx = CTX_FIRST_VERSION;
739 
740 			/* Don't call memset, for 16 entries that's just
741 			 * plain silly...
742 			 */
743 			mmu_context_bmap[0] = 3;
744 			mmu_context_bmap[1] = 0;
745 			mmu_context_bmap[2] = 0;
746 			mmu_context_bmap[3] = 0;
747 			for (i = 4; i < CTX_BMAP_SLOTS; i += 4) {
748 				mmu_context_bmap[i + 0] = 0;
749 				mmu_context_bmap[i + 1] = 0;
750 				mmu_context_bmap[i + 2] = 0;
751 				mmu_context_bmap[i + 3] = 0;
752 			}
753 			new_version = 1;
754 			goto out;
755 		}
756 	}
757 	mmu_context_bmap[new_ctx>>6] |= (1UL << (new_ctx & 63));
758 	new_ctx |= (tlb_context_cache & CTX_VERSION_MASK);
759 out:
760 	tlb_context_cache = new_ctx;
761 	mm->context.sparc64_ctx_val = new_ctx | orig_pgsz_bits;
762 	spin_unlock(&ctx_alloc_lock);
763 
764 	if (unlikely(new_version))
765 		smp_new_mmu_context_version();
766 }
767 
768 static int numa_enabled = 1;
769 static int numa_debug;
770 
771 static int __init early_numa(char *p)
772 {
773 	if (!p)
774 		return 0;
775 
776 	if (strstr(p, "off"))
777 		numa_enabled = 0;
778 
779 	if (strstr(p, "debug"))
780 		numa_debug = 1;
781 
782 	return 0;
783 }
784 early_param("numa", early_numa);
785 
786 #define numadbg(f, a...) \
787 do {	if (numa_debug) \
788 		printk(KERN_INFO f, ## a); \
789 } while (0)
790 
791 static void __init find_ramdisk(unsigned long phys_base)
792 {
793 #ifdef CONFIG_BLK_DEV_INITRD
794 	if (sparc_ramdisk_image || sparc_ramdisk_image64) {
795 		unsigned long ramdisk_image;
796 
797 		/* Older versions of the bootloader only supported a
798 		 * 32-bit physical address for the ramdisk image
799 		 * location, stored at sparc_ramdisk_image.  Newer
800 		 * SILO versions set sparc_ramdisk_image to zero and
801 		 * provide a full 64-bit physical address at
802 		 * sparc_ramdisk_image64.
803 		 */
804 		ramdisk_image = sparc_ramdisk_image;
805 		if (!ramdisk_image)
806 			ramdisk_image = sparc_ramdisk_image64;
807 
808 		/* Another bootloader quirk.  The bootloader normalizes
809 		 * the physical address to KERNBASE, so we have to
810 		 * factor that back out and add in the lowest valid
811 		 * physical page address to get the true physical address.
812 		 */
813 		ramdisk_image -= KERNBASE;
814 		ramdisk_image += phys_base;
815 
816 		numadbg("Found ramdisk at physical address 0x%lx, size %u\n",
817 			ramdisk_image, sparc_ramdisk_size);
818 
819 		initrd_start = ramdisk_image;
820 		initrd_end = ramdisk_image + sparc_ramdisk_size;
821 
822 		memblock_reserve(initrd_start, sparc_ramdisk_size);
823 
824 		initrd_start += PAGE_OFFSET;
825 		initrd_end += PAGE_OFFSET;
826 	}
827 #endif
828 }
829 
830 struct node_mem_mask {
831 	unsigned long mask;
832 	unsigned long match;
833 };
834 static struct node_mem_mask node_masks[MAX_NUMNODES];
835 static int num_node_masks;
836 
837 #ifdef CONFIG_NEED_MULTIPLE_NODES
838 
839 struct mdesc_mlgroup {
840 	u64	node;
841 	u64	latency;
842 	u64	match;
843 	u64	mask;
844 };
845 
846 static struct mdesc_mlgroup *mlgroups;
847 static int num_mlgroups;
848 
849 int numa_cpu_lookup_table[NR_CPUS];
850 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
851 
852 struct mdesc_mblock {
853 	u64	base;
854 	u64	size;
855 	u64	offset; /* RA-to-PA */
856 };
857 static struct mdesc_mblock *mblocks;
858 static int num_mblocks;
859 
860 static struct mdesc_mblock * __init addr_to_mblock(unsigned long addr)
861 {
862 	struct mdesc_mblock *m = NULL;
863 	int i;
864 
865 	for (i = 0; i < num_mblocks; i++) {
866 		m = &mblocks[i];
867 
868 		if (addr >= m->base &&
869 		    addr < (m->base + m->size)) {
870 			break;
871 		}
872 	}
873 
874 	return m;
875 }
876 
877 static u64 __init memblock_nid_range_sun4u(u64 start, u64 end, int *nid)
878 {
879 	int prev_nid, new_nid;
880 
881 	prev_nid = -1;
882 	for ( ; start < end; start += PAGE_SIZE) {
883 		for (new_nid = 0; new_nid < num_node_masks; new_nid++) {
884 			struct node_mem_mask *p = &node_masks[new_nid];
885 
886 			if ((start & p->mask) == p->match) {
887 				if (prev_nid == -1)
888 					prev_nid = new_nid;
889 				break;
890 			}
891 		}
892 
893 		if (new_nid == num_node_masks) {
894 			prev_nid = 0;
895 			WARN_ONCE(1, "addr[%Lx] doesn't match a NUMA node rule. Some memory will be owned by node 0.",
896 				  start);
897 			break;
898 		}
899 
900 		if (prev_nid != new_nid)
901 			break;
902 	}
903 	*nid = prev_nid;
904 
905 	return start > end ? end : start;
906 }
907 
908 static u64 __init memblock_nid_range(u64 start, u64 end, int *nid)
909 {
910 	u64 ret_end, pa_start, m_mask, m_match, m_end;
911 	struct mdesc_mblock *mblock;
912 	int _nid, i;
913 
914 	if (tlb_type != hypervisor)
915 		return memblock_nid_range_sun4u(start, end, nid);
916 
917 	mblock = addr_to_mblock(start);
918 	if (!mblock) {
919 		WARN_ONCE(1, "memblock_nid_range: Can't find mblock addr[%Lx]",
920 			  start);
921 
922 		_nid = 0;
923 		ret_end = end;
924 		goto done;
925 	}
926 
927 	pa_start = start + mblock->offset;
928 	m_match = 0;
929 	m_mask = 0;
930 
931 	for (_nid = 0; _nid < num_node_masks; _nid++) {
932 		struct node_mem_mask *const m = &node_masks[_nid];
933 
934 		if ((pa_start & m->mask) == m->match) {
935 			m_match = m->match;
936 			m_mask = m->mask;
937 			break;
938 		}
939 	}
940 
941 	if (num_node_masks == _nid) {
942 		/* We could not find NUMA group, so default to 0, but lets
943 		 * search for latency group, so we could calculate the correct
944 		 * end address that we return
945 		 */
946 		_nid = 0;
947 
948 		for (i = 0; i < num_mlgroups; i++) {
949 			struct mdesc_mlgroup *const m = &mlgroups[i];
950 
951 			if ((pa_start & m->mask) == m->match) {
952 				m_match = m->match;
953 				m_mask = m->mask;
954 				break;
955 			}
956 		}
957 
958 		if (i == num_mlgroups) {
959 			WARN_ONCE(1, "memblock_nid_range: Can't find latency group addr[%Lx]",
960 				  start);
961 
962 			ret_end = end;
963 			goto done;
964 		}
965 	}
966 
967 	/*
968 	 * Each latency group has match and mask, and each memory block has an
969 	 * offset.  An address belongs to a latency group if its address matches
970 	 * the following formula: ((addr + offset) & mask) == match
971 	 * It is, however, slow to check every single page if it matches a
972 	 * particular latency group. As optimization we calculate end value by
973 	 * using bit arithmetics.
974 	 */
975 	m_end = m_match + (1ul << __ffs(m_mask)) - mblock->offset;
976 	m_end += pa_start & ~((1ul << fls64(m_mask)) - 1);
977 	ret_end = m_end > end ? end : m_end;
978 
979 done:
980 	*nid = _nid;
981 	return ret_end;
982 }
983 #endif
984 
985 /* This must be invoked after performing all of the necessary
986  * memblock_set_node() calls for 'nid'.  We need to be able to get
987  * correct data from get_pfn_range_for_nid().
988  */
989 static void __init allocate_node_data(int nid)
990 {
991 	struct pglist_data *p;
992 	unsigned long start_pfn, end_pfn;
993 #ifdef CONFIG_NEED_MULTIPLE_NODES
994 	unsigned long paddr;
995 
996 	paddr = memblock_alloc_try_nid(sizeof(struct pglist_data), SMP_CACHE_BYTES, nid);
997 	if (!paddr) {
998 		prom_printf("Cannot allocate pglist_data for nid[%d]\n", nid);
999 		prom_halt();
1000 	}
1001 	NODE_DATA(nid) = __va(paddr);
1002 	memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
1003 
1004 	NODE_DATA(nid)->node_id = nid;
1005 #endif
1006 
1007 	p = NODE_DATA(nid);
1008 
1009 	get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1010 	p->node_start_pfn = start_pfn;
1011 	p->node_spanned_pages = end_pfn - start_pfn;
1012 }
1013 
1014 static void init_node_masks_nonnuma(void)
1015 {
1016 #ifdef CONFIG_NEED_MULTIPLE_NODES
1017 	int i;
1018 #endif
1019 
1020 	numadbg("Initializing tables for non-numa.\n");
1021 
1022 	node_masks[0].mask = 0;
1023 	node_masks[0].match = 0;
1024 	num_node_masks = 1;
1025 
1026 #ifdef CONFIG_NEED_MULTIPLE_NODES
1027 	for (i = 0; i < NR_CPUS; i++)
1028 		numa_cpu_lookup_table[i] = 0;
1029 
1030 	cpumask_setall(&numa_cpumask_lookup_table[0]);
1031 #endif
1032 }
1033 
1034 #ifdef CONFIG_NEED_MULTIPLE_NODES
1035 struct pglist_data *node_data[MAX_NUMNODES];
1036 
1037 EXPORT_SYMBOL(numa_cpu_lookup_table);
1038 EXPORT_SYMBOL(numa_cpumask_lookup_table);
1039 EXPORT_SYMBOL(node_data);
1040 
1041 static int scan_pio_for_cfg_handle(struct mdesc_handle *md, u64 pio,
1042 				   u32 cfg_handle)
1043 {
1044 	u64 arc;
1045 
1046 	mdesc_for_each_arc(arc, md, pio, MDESC_ARC_TYPE_FWD) {
1047 		u64 target = mdesc_arc_target(md, arc);
1048 		const u64 *val;
1049 
1050 		val = mdesc_get_property(md, target,
1051 					 "cfg-handle", NULL);
1052 		if (val && *val == cfg_handle)
1053 			return 0;
1054 	}
1055 	return -ENODEV;
1056 }
1057 
1058 static int scan_arcs_for_cfg_handle(struct mdesc_handle *md, u64 grp,
1059 				    u32 cfg_handle)
1060 {
1061 	u64 arc, candidate, best_latency = ~(u64)0;
1062 
1063 	candidate = MDESC_NODE_NULL;
1064 	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1065 		u64 target = mdesc_arc_target(md, arc);
1066 		const char *name = mdesc_node_name(md, target);
1067 		const u64 *val;
1068 
1069 		if (strcmp(name, "pio-latency-group"))
1070 			continue;
1071 
1072 		val = mdesc_get_property(md, target, "latency", NULL);
1073 		if (!val)
1074 			continue;
1075 
1076 		if (*val < best_latency) {
1077 			candidate = target;
1078 			best_latency = *val;
1079 		}
1080 	}
1081 
1082 	if (candidate == MDESC_NODE_NULL)
1083 		return -ENODEV;
1084 
1085 	return scan_pio_for_cfg_handle(md, candidate, cfg_handle);
1086 }
1087 
1088 int of_node_to_nid(struct device_node *dp)
1089 {
1090 	const struct linux_prom64_registers *regs;
1091 	struct mdesc_handle *md;
1092 	u32 cfg_handle;
1093 	int count, nid;
1094 	u64 grp;
1095 
1096 	/* This is the right thing to do on currently supported
1097 	 * SUN4U NUMA platforms as well, as the PCI controller does
1098 	 * not sit behind any particular memory controller.
1099 	 */
1100 	if (!mlgroups)
1101 		return -1;
1102 
1103 	regs = of_get_property(dp, "reg", NULL);
1104 	if (!regs)
1105 		return -1;
1106 
1107 	cfg_handle = (regs->phys_addr >> 32UL) & 0x0fffffff;
1108 
1109 	md = mdesc_grab();
1110 
1111 	count = 0;
1112 	nid = -1;
1113 	mdesc_for_each_node_by_name(md, grp, "group") {
1114 		if (!scan_arcs_for_cfg_handle(md, grp, cfg_handle)) {
1115 			nid = count;
1116 			break;
1117 		}
1118 		count++;
1119 	}
1120 
1121 	mdesc_release(md);
1122 
1123 	return nid;
1124 }
1125 
1126 static void __init add_node_ranges(void)
1127 {
1128 	struct memblock_region *reg;
1129 	unsigned long prev_max;
1130 
1131 memblock_resized:
1132 	prev_max = memblock.memory.max;
1133 
1134 	for_each_memblock(memory, reg) {
1135 		unsigned long size = reg->size;
1136 		unsigned long start, end;
1137 
1138 		start = reg->base;
1139 		end = start + size;
1140 		while (start < end) {
1141 			unsigned long this_end;
1142 			int nid;
1143 
1144 			this_end = memblock_nid_range(start, end, &nid);
1145 
1146 			numadbg("Setting memblock NUMA node nid[%d] "
1147 				"start[%lx] end[%lx]\n",
1148 				nid, start, this_end);
1149 
1150 			memblock_set_node(start, this_end - start,
1151 					  &memblock.memory, nid);
1152 			if (memblock.memory.max != prev_max)
1153 				goto memblock_resized;
1154 			start = this_end;
1155 		}
1156 	}
1157 }
1158 
1159 static int __init grab_mlgroups(struct mdesc_handle *md)
1160 {
1161 	unsigned long paddr;
1162 	int count = 0;
1163 	u64 node;
1164 
1165 	mdesc_for_each_node_by_name(md, node, "memory-latency-group")
1166 		count++;
1167 	if (!count)
1168 		return -ENOENT;
1169 
1170 	paddr = memblock_alloc(count * sizeof(struct mdesc_mlgroup),
1171 			  SMP_CACHE_BYTES);
1172 	if (!paddr)
1173 		return -ENOMEM;
1174 
1175 	mlgroups = __va(paddr);
1176 	num_mlgroups = count;
1177 
1178 	count = 0;
1179 	mdesc_for_each_node_by_name(md, node, "memory-latency-group") {
1180 		struct mdesc_mlgroup *m = &mlgroups[count++];
1181 		const u64 *val;
1182 
1183 		m->node = node;
1184 
1185 		val = mdesc_get_property(md, node, "latency", NULL);
1186 		m->latency = *val;
1187 		val = mdesc_get_property(md, node, "address-match", NULL);
1188 		m->match = *val;
1189 		val = mdesc_get_property(md, node, "address-mask", NULL);
1190 		m->mask = *val;
1191 
1192 		numadbg("MLGROUP[%d]: node[%llx] latency[%llx] "
1193 			"match[%llx] mask[%llx]\n",
1194 			count - 1, m->node, m->latency, m->match, m->mask);
1195 	}
1196 
1197 	return 0;
1198 }
1199 
1200 static int __init grab_mblocks(struct mdesc_handle *md)
1201 {
1202 	unsigned long paddr;
1203 	int count = 0;
1204 	u64 node;
1205 
1206 	mdesc_for_each_node_by_name(md, node, "mblock")
1207 		count++;
1208 	if (!count)
1209 		return -ENOENT;
1210 
1211 	paddr = memblock_alloc(count * sizeof(struct mdesc_mblock),
1212 			  SMP_CACHE_BYTES);
1213 	if (!paddr)
1214 		return -ENOMEM;
1215 
1216 	mblocks = __va(paddr);
1217 	num_mblocks = count;
1218 
1219 	count = 0;
1220 	mdesc_for_each_node_by_name(md, node, "mblock") {
1221 		struct mdesc_mblock *m = &mblocks[count++];
1222 		const u64 *val;
1223 
1224 		val = mdesc_get_property(md, node, "base", NULL);
1225 		m->base = *val;
1226 		val = mdesc_get_property(md, node, "size", NULL);
1227 		m->size = *val;
1228 		val = mdesc_get_property(md, node,
1229 					 "address-congruence-offset", NULL);
1230 
1231 		/* The address-congruence-offset property is optional.
1232 		 * Explicity zero it be identifty this.
1233 		 */
1234 		if (val)
1235 			m->offset = *val;
1236 		else
1237 			m->offset = 0UL;
1238 
1239 		numadbg("MBLOCK[%d]: base[%llx] size[%llx] offset[%llx]\n",
1240 			count - 1, m->base, m->size, m->offset);
1241 	}
1242 
1243 	return 0;
1244 }
1245 
1246 static void __init numa_parse_mdesc_group_cpus(struct mdesc_handle *md,
1247 					       u64 grp, cpumask_t *mask)
1248 {
1249 	u64 arc;
1250 
1251 	cpumask_clear(mask);
1252 
1253 	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_BACK) {
1254 		u64 target = mdesc_arc_target(md, arc);
1255 		const char *name = mdesc_node_name(md, target);
1256 		const u64 *id;
1257 
1258 		if (strcmp(name, "cpu"))
1259 			continue;
1260 		id = mdesc_get_property(md, target, "id", NULL);
1261 		if (*id < nr_cpu_ids)
1262 			cpumask_set_cpu(*id, mask);
1263 	}
1264 }
1265 
1266 static struct mdesc_mlgroup * __init find_mlgroup(u64 node)
1267 {
1268 	int i;
1269 
1270 	for (i = 0; i < num_mlgroups; i++) {
1271 		struct mdesc_mlgroup *m = &mlgroups[i];
1272 		if (m->node == node)
1273 			return m;
1274 	}
1275 	return NULL;
1276 }
1277 
1278 int __node_distance(int from, int to)
1279 {
1280 	if ((from >= MAX_NUMNODES) || (to >= MAX_NUMNODES)) {
1281 		pr_warn("Returning default NUMA distance value for %d->%d\n",
1282 			from, to);
1283 		return (from == to) ? LOCAL_DISTANCE : REMOTE_DISTANCE;
1284 	}
1285 	return numa_latency[from][to];
1286 }
1287 
1288 static int __init find_best_numa_node_for_mlgroup(struct mdesc_mlgroup *grp)
1289 {
1290 	int i;
1291 
1292 	for (i = 0; i < MAX_NUMNODES; i++) {
1293 		struct node_mem_mask *n = &node_masks[i];
1294 
1295 		if ((grp->mask == n->mask) && (grp->match == n->match))
1296 			break;
1297 	}
1298 	return i;
1299 }
1300 
1301 static void __init find_numa_latencies_for_group(struct mdesc_handle *md,
1302 						 u64 grp, int index)
1303 {
1304 	u64 arc;
1305 
1306 	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1307 		int tnode;
1308 		u64 target = mdesc_arc_target(md, arc);
1309 		struct mdesc_mlgroup *m = find_mlgroup(target);
1310 
1311 		if (!m)
1312 			continue;
1313 		tnode = find_best_numa_node_for_mlgroup(m);
1314 		if (tnode == MAX_NUMNODES)
1315 			continue;
1316 		numa_latency[index][tnode] = m->latency;
1317 	}
1318 }
1319 
1320 static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp,
1321 				      int index)
1322 {
1323 	struct mdesc_mlgroup *candidate = NULL;
1324 	u64 arc, best_latency = ~(u64)0;
1325 	struct node_mem_mask *n;
1326 
1327 	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1328 		u64 target = mdesc_arc_target(md, arc);
1329 		struct mdesc_mlgroup *m = find_mlgroup(target);
1330 		if (!m)
1331 			continue;
1332 		if (m->latency < best_latency) {
1333 			candidate = m;
1334 			best_latency = m->latency;
1335 		}
1336 	}
1337 	if (!candidate)
1338 		return -ENOENT;
1339 
1340 	if (num_node_masks != index) {
1341 		printk(KERN_ERR "Inconsistent NUMA state, "
1342 		       "index[%d] != num_node_masks[%d]\n",
1343 		       index, num_node_masks);
1344 		return -EINVAL;
1345 	}
1346 
1347 	n = &node_masks[num_node_masks++];
1348 
1349 	n->mask = candidate->mask;
1350 	n->match = candidate->match;
1351 
1352 	numadbg("NUMA NODE[%d]: mask[%lx] match[%lx] (latency[%llx])\n",
1353 		index, n->mask, n->match, candidate->latency);
1354 
1355 	return 0;
1356 }
1357 
1358 static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp,
1359 					 int index)
1360 {
1361 	cpumask_t mask;
1362 	int cpu;
1363 
1364 	numa_parse_mdesc_group_cpus(md, grp, &mask);
1365 
1366 	for_each_cpu(cpu, &mask)
1367 		numa_cpu_lookup_table[cpu] = index;
1368 	cpumask_copy(&numa_cpumask_lookup_table[index], &mask);
1369 
1370 	if (numa_debug) {
1371 		printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index);
1372 		for_each_cpu(cpu, &mask)
1373 			printk("%d ", cpu);
1374 		printk("]\n");
1375 	}
1376 
1377 	return numa_attach_mlgroup(md, grp, index);
1378 }
1379 
1380 static int __init numa_parse_mdesc(void)
1381 {
1382 	struct mdesc_handle *md = mdesc_grab();
1383 	int i, j, err, count;
1384 	u64 node;
1385 
1386 	node = mdesc_node_by_name(md, MDESC_NODE_NULL, "latency-groups");
1387 	if (node == MDESC_NODE_NULL) {
1388 		mdesc_release(md);
1389 		return -ENOENT;
1390 	}
1391 
1392 	err = grab_mblocks(md);
1393 	if (err < 0)
1394 		goto out;
1395 
1396 	err = grab_mlgroups(md);
1397 	if (err < 0)
1398 		goto out;
1399 
1400 	count = 0;
1401 	mdesc_for_each_node_by_name(md, node, "group") {
1402 		err = numa_parse_mdesc_group(md, node, count);
1403 		if (err < 0)
1404 			break;
1405 		count++;
1406 	}
1407 
1408 	count = 0;
1409 	mdesc_for_each_node_by_name(md, node, "group") {
1410 		find_numa_latencies_for_group(md, node, count);
1411 		count++;
1412 	}
1413 
1414 	/* Normalize numa latency matrix according to ACPI SLIT spec. */
1415 	for (i = 0; i < MAX_NUMNODES; i++) {
1416 		u64 self_latency = numa_latency[i][i];
1417 
1418 		for (j = 0; j < MAX_NUMNODES; j++) {
1419 			numa_latency[i][j] =
1420 				(numa_latency[i][j] * LOCAL_DISTANCE) /
1421 				self_latency;
1422 		}
1423 	}
1424 
1425 	add_node_ranges();
1426 
1427 	for (i = 0; i < num_node_masks; i++) {
1428 		allocate_node_data(i);
1429 		node_set_online(i);
1430 	}
1431 
1432 	err = 0;
1433 out:
1434 	mdesc_release(md);
1435 	return err;
1436 }
1437 
1438 static int __init numa_parse_jbus(void)
1439 {
1440 	unsigned long cpu, index;
1441 
1442 	/* NUMA node id is encoded in bits 36 and higher, and there is
1443 	 * a 1-to-1 mapping from CPU ID to NUMA node ID.
1444 	 */
1445 	index = 0;
1446 	for_each_present_cpu(cpu) {
1447 		numa_cpu_lookup_table[cpu] = index;
1448 		cpumask_copy(&numa_cpumask_lookup_table[index], cpumask_of(cpu));
1449 		node_masks[index].mask = ~((1UL << 36UL) - 1UL);
1450 		node_masks[index].match = cpu << 36UL;
1451 
1452 		index++;
1453 	}
1454 	num_node_masks = index;
1455 
1456 	add_node_ranges();
1457 
1458 	for (index = 0; index < num_node_masks; index++) {
1459 		allocate_node_data(index);
1460 		node_set_online(index);
1461 	}
1462 
1463 	return 0;
1464 }
1465 
1466 static int __init numa_parse_sun4u(void)
1467 {
1468 	if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1469 		unsigned long ver;
1470 
1471 		__asm__ ("rdpr %%ver, %0" : "=r" (ver));
1472 		if ((ver >> 32UL) == __JALAPENO_ID ||
1473 		    (ver >> 32UL) == __SERRANO_ID)
1474 			return numa_parse_jbus();
1475 	}
1476 	return -1;
1477 }
1478 
1479 static int __init bootmem_init_numa(void)
1480 {
1481 	int i, j;
1482 	int err = -1;
1483 
1484 	numadbg("bootmem_init_numa()\n");
1485 
1486 	/* Some sane defaults for numa latency values */
1487 	for (i = 0; i < MAX_NUMNODES; i++) {
1488 		for (j = 0; j < MAX_NUMNODES; j++)
1489 			numa_latency[i][j] = (i == j) ?
1490 				LOCAL_DISTANCE : REMOTE_DISTANCE;
1491 	}
1492 
1493 	if (numa_enabled) {
1494 		if (tlb_type == hypervisor)
1495 			err = numa_parse_mdesc();
1496 		else
1497 			err = numa_parse_sun4u();
1498 	}
1499 	return err;
1500 }
1501 
1502 #else
1503 
1504 static int bootmem_init_numa(void)
1505 {
1506 	return -1;
1507 }
1508 
1509 #endif
1510 
1511 static void __init bootmem_init_nonnuma(void)
1512 {
1513 	unsigned long top_of_ram = memblock_end_of_DRAM();
1514 	unsigned long total_ram = memblock_phys_mem_size();
1515 
1516 	numadbg("bootmem_init_nonnuma()\n");
1517 
1518 	printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
1519 	       top_of_ram, total_ram);
1520 	printk(KERN_INFO "Memory hole size: %ldMB\n",
1521 	       (top_of_ram - total_ram) >> 20);
1522 
1523 	init_node_masks_nonnuma();
1524 	memblock_set_node(0, (phys_addr_t)ULLONG_MAX, &memblock.memory, 0);
1525 	allocate_node_data(0);
1526 	node_set_online(0);
1527 }
1528 
1529 static unsigned long __init bootmem_init(unsigned long phys_base)
1530 {
1531 	unsigned long end_pfn;
1532 
1533 	end_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1534 	max_pfn = max_low_pfn = end_pfn;
1535 	min_low_pfn = (phys_base >> PAGE_SHIFT);
1536 
1537 	if (bootmem_init_numa() < 0)
1538 		bootmem_init_nonnuma();
1539 
1540 	/* Dump memblock with node info. */
1541 	memblock_dump_all();
1542 
1543 	/* XXX cpu notifier XXX */
1544 
1545 	sparse_memory_present_with_active_regions(MAX_NUMNODES);
1546 	sparse_init();
1547 
1548 	return end_pfn;
1549 }
1550 
1551 static struct linux_prom64_registers pall[MAX_BANKS] __initdata;
1552 static int pall_ents __initdata;
1553 
1554 static unsigned long max_phys_bits = 40;
1555 
1556 bool kern_addr_valid(unsigned long addr)
1557 {
1558 	pgd_t *pgd;
1559 	pud_t *pud;
1560 	pmd_t *pmd;
1561 	pte_t *pte;
1562 
1563 	if ((long)addr < 0L) {
1564 		unsigned long pa = __pa(addr);
1565 
1566 		if ((addr >> max_phys_bits) != 0UL)
1567 			return false;
1568 
1569 		return pfn_valid(pa >> PAGE_SHIFT);
1570 	}
1571 
1572 	if (addr >= (unsigned long) KERNBASE &&
1573 	    addr < (unsigned long)&_end)
1574 		return true;
1575 
1576 	pgd = pgd_offset_k(addr);
1577 	if (pgd_none(*pgd))
1578 		return 0;
1579 
1580 	pud = pud_offset(pgd, addr);
1581 	if (pud_none(*pud))
1582 		return 0;
1583 
1584 	if (pud_large(*pud))
1585 		return pfn_valid(pud_pfn(*pud));
1586 
1587 	pmd = pmd_offset(pud, addr);
1588 	if (pmd_none(*pmd))
1589 		return 0;
1590 
1591 	if (pmd_large(*pmd))
1592 		return pfn_valid(pmd_pfn(*pmd));
1593 
1594 	pte = pte_offset_kernel(pmd, addr);
1595 	if (pte_none(*pte))
1596 		return 0;
1597 
1598 	return pfn_valid(pte_pfn(*pte));
1599 }
1600 EXPORT_SYMBOL(kern_addr_valid);
1601 
1602 static unsigned long __ref kernel_map_hugepud(unsigned long vstart,
1603 					      unsigned long vend,
1604 					      pud_t *pud)
1605 {
1606 	const unsigned long mask16gb = (1UL << 34) - 1UL;
1607 	u64 pte_val = vstart;
1608 
1609 	/* Each PUD is 8GB */
1610 	if ((vstart & mask16gb) ||
1611 	    (vend - vstart <= mask16gb)) {
1612 		pte_val ^= kern_linear_pte_xor[2];
1613 		pud_val(*pud) = pte_val | _PAGE_PUD_HUGE;
1614 
1615 		return vstart + PUD_SIZE;
1616 	}
1617 
1618 	pte_val ^= kern_linear_pte_xor[3];
1619 	pte_val |= _PAGE_PUD_HUGE;
1620 
1621 	vend = vstart + mask16gb + 1UL;
1622 	while (vstart < vend) {
1623 		pud_val(*pud) = pte_val;
1624 
1625 		pte_val += PUD_SIZE;
1626 		vstart += PUD_SIZE;
1627 		pud++;
1628 	}
1629 	return vstart;
1630 }
1631 
1632 static bool kernel_can_map_hugepud(unsigned long vstart, unsigned long vend,
1633 				   bool guard)
1634 {
1635 	if (guard && !(vstart & ~PUD_MASK) && (vend - vstart) >= PUD_SIZE)
1636 		return true;
1637 
1638 	return false;
1639 }
1640 
1641 static unsigned long __ref kernel_map_hugepmd(unsigned long vstart,
1642 					      unsigned long vend,
1643 					      pmd_t *pmd)
1644 {
1645 	const unsigned long mask256mb = (1UL << 28) - 1UL;
1646 	const unsigned long mask2gb = (1UL << 31) - 1UL;
1647 	u64 pte_val = vstart;
1648 
1649 	/* Each PMD is 8MB */
1650 	if ((vstart & mask256mb) ||
1651 	    (vend - vstart <= mask256mb)) {
1652 		pte_val ^= kern_linear_pte_xor[0];
1653 		pmd_val(*pmd) = pte_val | _PAGE_PMD_HUGE;
1654 
1655 		return vstart + PMD_SIZE;
1656 	}
1657 
1658 	if ((vstart & mask2gb) ||
1659 	    (vend - vstart <= mask2gb)) {
1660 		pte_val ^= kern_linear_pte_xor[1];
1661 		pte_val |= _PAGE_PMD_HUGE;
1662 		vend = vstart + mask256mb + 1UL;
1663 	} else {
1664 		pte_val ^= kern_linear_pte_xor[2];
1665 		pte_val |= _PAGE_PMD_HUGE;
1666 		vend = vstart + mask2gb + 1UL;
1667 	}
1668 
1669 	while (vstart < vend) {
1670 		pmd_val(*pmd) = pte_val;
1671 
1672 		pte_val += PMD_SIZE;
1673 		vstart += PMD_SIZE;
1674 		pmd++;
1675 	}
1676 
1677 	return vstart;
1678 }
1679 
1680 static bool kernel_can_map_hugepmd(unsigned long vstart, unsigned long vend,
1681 				   bool guard)
1682 {
1683 	if (guard && !(vstart & ~PMD_MASK) && (vend - vstart) >= PMD_SIZE)
1684 		return true;
1685 
1686 	return false;
1687 }
1688 
1689 static unsigned long __ref kernel_map_range(unsigned long pstart,
1690 					    unsigned long pend, pgprot_t prot,
1691 					    bool use_huge)
1692 {
1693 	unsigned long vstart = PAGE_OFFSET + pstart;
1694 	unsigned long vend = PAGE_OFFSET + pend;
1695 	unsigned long alloc_bytes = 0UL;
1696 
1697 	if ((vstart & ~PAGE_MASK) || (vend & ~PAGE_MASK)) {
1698 		prom_printf("kernel_map: Unaligned physmem[%lx:%lx]\n",
1699 			    vstart, vend);
1700 		prom_halt();
1701 	}
1702 
1703 	while (vstart < vend) {
1704 		unsigned long this_end, paddr = __pa(vstart);
1705 		pgd_t *pgd = pgd_offset_k(vstart);
1706 		pud_t *pud;
1707 		pmd_t *pmd;
1708 		pte_t *pte;
1709 
1710 		if (pgd_none(*pgd)) {
1711 			pud_t *new;
1712 
1713 			new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1714 			alloc_bytes += PAGE_SIZE;
1715 			pgd_populate(&init_mm, pgd, new);
1716 		}
1717 		pud = pud_offset(pgd, vstart);
1718 		if (pud_none(*pud)) {
1719 			pmd_t *new;
1720 
1721 			if (kernel_can_map_hugepud(vstart, vend, use_huge)) {
1722 				vstart = kernel_map_hugepud(vstart, vend, pud);
1723 				continue;
1724 			}
1725 			new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1726 			alloc_bytes += PAGE_SIZE;
1727 			pud_populate(&init_mm, pud, new);
1728 		}
1729 
1730 		pmd = pmd_offset(pud, vstart);
1731 		if (pmd_none(*pmd)) {
1732 			pte_t *new;
1733 
1734 			if (kernel_can_map_hugepmd(vstart, vend, use_huge)) {
1735 				vstart = kernel_map_hugepmd(vstart, vend, pmd);
1736 				continue;
1737 			}
1738 			new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1739 			alloc_bytes += PAGE_SIZE;
1740 			pmd_populate_kernel(&init_mm, pmd, new);
1741 		}
1742 
1743 		pte = pte_offset_kernel(pmd, vstart);
1744 		this_end = (vstart + PMD_SIZE) & PMD_MASK;
1745 		if (this_end > vend)
1746 			this_end = vend;
1747 
1748 		while (vstart < this_end) {
1749 			pte_val(*pte) = (paddr | pgprot_val(prot));
1750 
1751 			vstart += PAGE_SIZE;
1752 			paddr += PAGE_SIZE;
1753 			pte++;
1754 		}
1755 	}
1756 
1757 	return alloc_bytes;
1758 }
1759 
1760 static void __init flush_all_kernel_tsbs(void)
1761 {
1762 	int i;
1763 
1764 	for (i = 0; i < KERNEL_TSB_NENTRIES; i++) {
1765 		struct tsb *ent = &swapper_tsb[i];
1766 
1767 		ent->tag = (1UL << TSB_TAG_INVALID_BIT);
1768 	}
1769 #ifndef CONFIG_DEBUG_PAGEALLOC
1770 	for (i = 0; i < KERNEL_TSB4M_NENTRIES; i++) {
1771 		struct tsb *ent = &swapper_4m_tsb[i];
1772 
1773 		ent->tag = (1UL << TSB_TAG_INVALID_BIT);
1774 	}
1775 #endif
1776 }
1777 
1778 extern unsigned int kvmap_linear_patch[1];
1779 
1780 static void __init kernel_physical_mapping_init(void)
1781 {
1782 	unsigned long i, mem_alloced = 0UL;
1783 	bool use_huge = true;
1784 
1785 #ifdef CONFIG_DEBUG_PAGEALLOC
1786 	use_huge = false;
1787 #endif
1788 	for (i = 0; i < pall_ents; i++) {
1789 		unsigned long phys_start, phys_end;
1790 
1791 		phys_start = pall[i].phys_addr;
1792 		phys_end = phys_start + pall[i].reg_size;
1793 
1794 		mem_alloced += kernel_map_range(phys_start, phys_end,
1795 						PAGE_KERNEL, use_huge);
1796 	}
1797 
1798 	printk("Allocated %ld bytes for kernel page tables.\n",
1799 	       mem_alloced);
1800 
1801 	kvmap_linear_patch[0] = 0x01000000; /* nop */
1802 	flushi(&kvmap_linear_patch[0]);
1803 
1804 	flush_all_kernel_tsbs();
1805 
1806 	__flush_tlb_all();
1807 }
1808 
1809 #ifdef CONFIG_DEBUG_PAGEALLOC
1810 void __kernel_map_pages(struct page *page, int numpages, int enable)
1811 {
1812 	unsigned long phys_start = page_to_pfn(page) << PAGE_SHIFT;
1813 	unsigned long phys_end = phys_start + (numpages * PAGE_SIZE);
1814 
1815 	kernel_map_range(phys_start, phys_end,
1816 			 (enable ? PAGE_KERNEL : __pgprot(0)), false);
1817 
1818 	flush_tsb_kernel_range(PAGE_OFFSET + phys_start,
1819 			       PAGE_OFFSET + phys_end);
1820 
1821 	/* we should perform an IPI and flush all tlbs,
1822 	 * but that can deadlock->flush only current cpu.
1823 	 */
1824 	__flush_tlb_kernel_range(PAGE_OFFSET + phys_start,
1825 				 PAGE_OFFSET + phys_end);
1826 }
1827 #endif
1828 
1829 unsigned long __init find_ecache_flush_span(unsigned long size)
1830 {
1831 	int i;
1832 
1833 	for (i = 0; i < pavail_ents; i++) {
1834 		if (pavail[i].reg_size >= size)
1835 			return pavail[i].phys_addr;
1836 	}
1837 
1838 	return ~0UL;
1839 }
1840 
1841 unsigned long PAGE_OFFSET;
1842 EXPORT_SYMBOL(PAGE_OFFSET);
1843 
1844 unsigned long VMALLOC_END   = 0x0000010000000000UL;
1845 EXPORT_SYMBOL(VMALLOC_END);
1846 
1847 unsigned long sparc64_va_hole_top =    0xfffff80000000000UL;
1848 unsigned long sparc64_va_hole_bottom = 0x0000080000000000UL;
1849 
1850 static void __init setup_page_offset(void)
1851 {
1852 	if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1853 		/* Cheetah/Panther support a full 64-bit virtual
1854 		 * address, so we can use all that our page tables
1855 		 * support.
1856 		 */
1857 		sparc64_va_hole_top =    0xfff0000000000000UL;
1858 		sparc64_va_hole_bottom = 0x0010000000000000UL;
1859 
1860 		max_phys_bits = 42;
1861 	} else if (tlb_type == hypervisor) {
1862 		switch (sun4v_chip_type) {
1863 		case SUN4V_CHIP_NIAGARA1:
1864 		case SUN4V_CHIP_NIAGARA2:
1865 			/* T1 and T2 support 48-bit virtual addresses.  */
1866 			sparc64_va_hole_top =    0xffff800000000000UL;
1867 			sparc64_va_hole_bottom = 0x0000800000000000UL;
1868 
1869 			max_phys_bits = 39;
1870 			break;
1871 		case SUN4V_CHIP_NIAGARA3:
1872 			/* T3 supports 48-bit virtual addresses.  */
1873 			sparc64_va_hole_top =    0xffff800000000000UL;
1874 			sparc64_va_hole_bottom = 0x0000800000000000UL;
1875 
1876 			max_phys_bits = 43;
1877 			break;
1878 		case SUN4V_CHIP_NIAGARA4:
1879 		case SUN4V_CHIP_NIAGARA5:
1880 		case SUN4V_CHIP_SPARC64X:
1881 		case SUN4V_CHIP_SPARC_M6:
1882 			/* T4 and later support 52-bit virtual addresses.  */
1883 			sparc64_va_hole_top =    0xfff8000000000000UL;
1884 			sparc64_va_hole_bottom = 0x0008000000000000UL;
1885 			max_phys_bits = 47;
1886 			break;
1887 		case SUN4V_CHIP_SPARC_M7:
1888 		case SUN4V_CHIP_SPARC_SN:
1889 		default:
1890 			/* M7 and later support 52-bit virtual addresses.  */
1891 			sparc64_va_hole_top =    0xfff8000000000000UL;
1892 			sparc64_va_hole_bottom = 0x0008000000000000UL;
1893 			max_phys_bits = 49;
1894 			break;
1895 		}
1896 	}
1897 
1898 	if (max_phys_bits > MAX_PHYS_ADDRESS_BITS) {
1899 		prom_printf("MAX_PHYS_ADDRESS_BITS is too small, need %lu\n",
1900 			    max_phys_bits);
1901 		prom_halt();
1902 	}
1903 
1904 	PAGE_OFFSET = sparc64_va_hole_top;
1905 	VMALLOC_END = ((sparc64_va_hole_bottom >> 1) +
1906 		       (sparc64_va_hole_bottom >> 2));
1907 
1908 	pr_info("MM: PAGE_OFFSET is 0x%016lx (max_phys_bits == %lu)\n",
1909 		PAGE_OFFSET, max_phys_bits);
1910 	pr_info("MM: VMALLOC [0x%016lx --> 0x%016lx]\n",
1911 		VMALLOC_START, VMALLOC_END);
1912 	pr_info("MM: VMEMMAP [0x%016lx --> 0x%016lx]\n",
1913 		VMEMMAP_BASE, VMEMMAP_BASE << 1);
1914 }
1915 
1916 static void __init tsb_phys_patch(void)
1917 {
1918 	struct tsb_ldquad_phys_patch_entry *pquad;
1919 	struct tsb_phys_patch_entry *p;
1920 
1921 	pquad = &__tsb_ldquad_phys_patch;
1922 	while (pquad < &__tsb_ldquad_phys_patch_end) {
1923 		unsigned long addr = pquad->addr;
1924 
1925 		if (tlb_type == hypervisor)
1926 			*(unsigned int *) addr = pquad->sun4v_insn;
1927 		else
1928 			*(unsigned int *) addr = pquad->sun4u_insn;
1929 		wmb();
1930 		__asm__ __volatile__("flush	%0"
1931 				     : /* no outputs */
1932 				     : "r" (addr));
1933 
1934 		pquad++;
1935 	}
1936 
1937 	p = &__tsb_phys_patch;
1938 	while (p < &__tsb_phys_patch_end) {
1939 		unsigned long addr = p->addr;
1940 
1941 		*(unsigned int *) addr = p->insn;
1942 		wmb();
1943 		__asm__ __volatile__("flush	%0"
1944 				     : /* no outputs */
1945 				     : "r" (addr));
1946 
1947 		p++;
1948 	}
1949 }
1950 
1951 /* Don't mark as init, we give this to the Hypervisor.  */
1952 #ifndef CONFIG_DEBUG_PAGEALLOC
1953 #define NUM_KTSB_DESCR	2
1954 #else
1955 #define NUM_KTSB_DESCR	1
1956 #endif
1957 static struct hv_tsb_descr ktsb_descr[NUM_KTSB_DESCR];
1958 
1959 /* The swapper TSBs are loaded with a base sequence of:
1960  *
1961  *	sethi	%uhi(SYMBOL), REG1
1962  *	sethi	%hi(SYMBOL), REG2
1963  *	or	REG1, %ulo(SYMBOL), REG1
1964  *	or	REG2, %lo(SYMBOL), REG2
1965  *	sllx	REG1, 32, REG1
1966  *	or	REG1, REG2, REG1
1967  *
1968  * When we use physical addressing for the TSB accesses, we patch the
1969  * first four instructions in the above sequence.
1970  */
1971 
1972 static void patch_one_ktsb_phys(unsigned int *start, unsigned int *end, unsigned long pa)
1973 {
1974 	unsigned long high_bits, low_bits;
1975 
1976 	high_bits = (pa >> 32) & 0xffffffff;
1977 	low_bits = (pa >> 0) & 0xffffffff;
1978 
1979 	while (start < end) {
1980 		unsigned int *ia = (unsigned int *)(unsigned long)*start;
1981 
1982 		ia[0] = (ia[0] & ~0x3fffff) | (high_bits >> 10);
1983 		__asm__ __volatile__("flush	%0" : : "r" (ia));
1984 
1985 		ia[1] = (ia[1] & ~0x3fffff) | (low_bits >> 10);
1986 		__asm__ __volatile__("flush	%0" : : "r" (ia + 1));
1987 
1988 		ia[2] = (ia[2] & ~0x1fff) | (high_bits & 0x3ff);
1989 		__asm__ __volatile__("flush	%0" : : "r" (ia + 2));
1990 
1991 		ia[3] = (ia[3] & ~0x1fff) | (low_bits & 0x3ff);
1992 		__asm__ __volatile__("flush	%0" : : "r" (ia + 3));
1993 
1994 		start++;
1995 	}
1996 }
1997 
1998 static void ktsb_phys_patch(void)
1999 {
2000 	extern unsigned int __swapper_tsb_phys_patch;
2001 	extern unsigned int __swapper_tsb_phys_patch_end;
2002 	unsigned long ktsb_pa;
2003 
2004 	ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
2005 	patch_one_ktsb_phys(&__swapper_tsb_phys_patch,
2006 			    &__swapper_tsb_phys_patch_end, ktsb_pa);
2007 #ifndef CONFIG_DEBUG_PAGEALLOC
2008 	{
2009 	extern unsigned int __swapper_4m_tsb_phys_patch;
2010 	extern unsigned int __swapper_4m_tsb_phys_patch_end;
2011 	ktsb_pa = (kern_base +
2012 		   ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
2013 	patch_one_ktsb_phys(&__swapper_4m_tsb_phys_patch,
2014 			    &__swapper_4m_tsb_phys_patch_end, ktsb_pa);
2015 	}
2016 #endif
2017 }
2018 
2019 static void __init sun4v_ktsb_init(void)
2020 {
2021 	unsigned long ktsb_pa;
2022 
2023 	/* First KTSB for PAGE_SIZE mappings.  */
2024 	ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
2025 
2026 	switch (PAGE_SIZE) {
2027 	case 8 * 1024:
2028 	default:
2029 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_8K;
2030 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_8K;
2031 		break;
2032 
2033 	case 64 * 1024:
2034 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_64K;
2035 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_64K;
2036 		break;
2037 
2038 	case 512 * 1024:
2039 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_512K;
2040 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_512K;
2041 		break;
2042 
2043 	case 4 * 1024 * 1024:
2044 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_4MB;
2045 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_4MB;
2046 		break;
2047 	}
2048 
2049 	ktsb_descr[0].assoc = 1;
2050 	ktsb_descr[0].num_ttes = KERNEL_TSB_NENTRIES;
2051 	ktsb_descr[0].ctx_idx = 0;
2052 	ktsb_descr[0].tsb_base = ktsb_pa;
2053 	ktsb_descr[0].resv = 0;
2054 
2055 #ifndef CONFIG_DEBUG_PAGEALLOC
2056 	/* Second KTSB for 4MB/256MB/2GB/16GB mappings.  */
2057 	ktsb_pa = (kern_base +
2058 		   ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
2059 
2060 	ktsb_descr[1].pgsz_idx = HV_PGSZ_IDX_4MB;
2061 	ktsb_descr[1].pgsz_mask = ((HV_PGSZ_MASK_4MB |
2062 				    HV_PGSZ_MASK_256MB |
2063 				    HV_PGSZ_MASK_2GB |
2064 				    HV_PGSZ_MASK_16GB) &
2065 				   cpu_pgsz_mask);
2066 	ktsb_descr[1].assoc = 1;
2067 	ktsb_descr[1].num_ttes = KERNEL_TSB4M_NENTRIES;
2068 	ktsb_descr[1].ctx_idx = 0;
2069 	ktsb_descr[1].tsb_base = ktsb_pa;
2070 	ktsb_descr[1].resv = 0;
2071 #endif
2072 }
2073 
2074 void sun4v_ktsb_register(void)
2075 {
2076 	unsigned long pa, ret;
2077 
2078 	pa = kern_base + ((unsigned long)&ktsb_descr[0] - KERNBASE);
2079 
2080 	ret = sun4v_mmu_tsb_ctx0(NUM_KTSB_DESCR, pa);
2081 	if (ret != 0) {
2082 		prom_printf("hypervisor_mmu_tsb_ctx0[%lx]: "
2083 			    "errors with %lx\n", pa, ret);
2084 		prom_halt();
2085 	}
2086 }
2087 
2088 static void __init sun4u_linear_pte_xor_finalize(void)
2089 {
2090 #ifndef CONFIG_DEBUG_PAGEALLOC
2091 	/* This is where we would add Panther support for
2092 	 * 32MB and 256MB pages.
2093 	 */
2094 #endif
2095 }
2096 
2097 static void __init sun4v_linear_pte_xor_finalize(void)
2098 {
2099 	unsigned long pagecv_flag;
2100 
2101 	/* Bit 9 of TTE is no longer CV bit on M7 processor and it instead
2102 	 * enables MCD error. Do not set bit 9 on M7 processor.
2103 	 */
2104 	switch (sun4v_chip_type) {
2105 	case SUN4V_CHIP_SPARC_M7:
2106 	case SUN4V_CHIP_SPARC_SN:
2107 		pagecv_flag = 0x00;
2108 		break;
2109 	default:
2110 		pagecv_flag = _PAGE_CV_4V;
2111 		break;
2112 	}
2113 #ifndef CONFIG_DEBUG_PAGEALLOC
2114 	if (cpu_pgsz_mask & HV_PGSZ_MASK_256MB) {
2115 		kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZ256MB_4V) ^
2116 			PAGE_OFFSET;
2117 		kern_linear_pte_xor[1] |= (_PAGE_CP_4V | pagecv_flag |
2118 					   _PAGE_P_4V | _PAGE_W_4V);
2119 	} else {
2120 		kern_linear_pte_xor[1] = kern_linear_pte_xor[0];
2121 	}
2122 
2123 	if (cpu_pgsz_mask & HV_PGSZ_MASK_2GB) {
2124 		kern_linear_pte_xor[2] = (_PAGE_VALID | _PAGE_SZ2GB_4V) ^
2125 			PAGE_OFFSET;
2126 		kern_linear_pte_xor[2] |= (_PAGE_CP_4V | pagecv_flag |
2127 					   _PAGE_P_4V | _PAGE_W_4V);
2128 	} else {
2129 		kern_linear_pte_xor[2] = kern_linear_pte_xor[1];
2130 	}
2131 
2132 	if (cpu_pgsz_mask & HV_PGSZ_MASK_16GB) {
2133 		kern_linear_pte_xor[3] = (_PAGE_VALID | _PAGE_SZ16GB_4V) ^
2134 			PAGE_OFFSET;
2135 		kern_linear_pte_xor[3] |= (_PAGE_CP_4V | pagecv_flag |
2136 					   _PAGE_P_4V | _PAGE_W_4V);
2137 	} else {
2138 		kern_linear_pte_xor[3] = kern_linear_pte_xor[2];
2139 	}
2140 #endif
2141 }
2142 
2143 /* paging_init() sets up the page tables */
2144 
2145 static unsigned long last_valid_pfn;
2146 
2147 static void sun4u_pgprot_init(void);
2148 static void sun4v_pgprot_init(void);
2149 
2150 static phys_addr_t __init available_memory(void)
2151 {
2152 	phys_addr_t available = 0ULL;
2153 	phys_addr_t pa_start, pa_end;
2154 	u64 i;
2155 
2156 	for_each_free_mem_range(i, NUMA_NO_NODE, MEMBLOCK_NONE, &pa_start,
2157 				&pa_end, NULL)
2158 		available = available + (pa_end  - pa_start);
2159 
2160 	return available;
2161 }
2162 
2163 #define _PAGE_CACHE_4U	(_PAGE_CP_4U | _PAGE_CV_4U)
2164 #define _PAGE_CACHE_4V	(_PAGE_CP_4V | _PAGE_CV_4V)
2165 #define __DIRTY_BITS_4U	 (_PAGE_MODIFIED_4U | _PAGE_WRITE_4U | _PAGE_W_4U)
2166 #define __DIRTY_BITS_4V	 (_PAGE_MODIFIED_4V | _PAGE_WRITE_4V | _PAGE_W_4V)
2167 #define __ACCESS_BITS_4U (_PAGE_ACCESSED_4U | _PAGE_READ_4U | _PAGE_R)
2168 #define __ACCESS_BITS_4V (_PAGE_ACCESSED_4V | _PAGE_READ_4V | _PAGE_R)
2169 
2170 /* We need to exclude reserved regions. This exclusion will include
2171  * vmlinux and initrd. To be more precise the initrd size could be used to
2172  * compute a new lower limit because it is freed later during initialization.
2173  */
2174 static void __init reduce_memory(phys_addr_t limit_ram)
2175 {
2176 	phys_addr_t avail_ram = available_memory();
2177 	phys_addr_t pa_start, pa_end;
2178 	u64 i;
2179 
2180 	if (limit_ram >= avail_ram)
2181 		return;
2182 
2183 	for_each_free_mem_range(i, NUMA_NO_NODE, MEMBLOCK_NONE, &pa_start,
2184 				&pa_end, NULL) {
2185 		phys_addr_t region_size = pa_end - pa_start;
2186 		phys_addr_t clip_start = pa_start;
2187 
2188 		avail_ram = avail_ram - region_size;
2189 		/* Are we consuming too much? */
2190 		if (avail_ram < limit_ram) {
2191 			phys_addr_t give_back = limit_ram - avail_ram;
2192 
2193 			region_size = region_size - give_back;
2194 			clip_start = clip_start + give_back;
2195 		}
2196 
2197 		memblock_remove(clip_start, region_size);
2198 
2199 		if (avail_ram <= limit_ram)
2200 			break;
2201 		i = 0UL;
2202 	}
2203 }
2204 
2205 void __init paging_init(void)
2206 {
2207 	unsigned long end_pfn, shift, phys_base;
2208 	unsigned long real_end, i;
2209 
2210 	setup_page_offset();
2211 
2212 	/* These build time checkes make sure that the dcache_dirty_cpu()
2213 	 * page->flags usage will work.
2214 	 *
2215 	 * When a page gets marked as dcache-dirty, we store the
2216 	 * cpu number starting at bit 32 in the page->flags.  Also,
2217 	 * functions like clear_dcache_dirty_cpu use the cpu mask
2218 	 * in 13-bit signed-immediate instruction fields.
2219 	 */
2220 
2221 	/*
2222 	 * Page flags must not reach into upper 32 bits that are used
2223 	 * for the cpu number
2224 	 */
2225 	BUILD_BUG_ON(NR_PAGEFLAGS > 32);
2226 
2227 	/*
2228 	 * The bit fields placed in the high range must not reach below
2229 	 * the 32 bit boundary. Otherwise we cannot place the cpu field
2230 	 * at the 32 bit boundary.
2231 	 */
2232 	BUILD_BUG_ON(SECTIONS_WIDTH + NODES_WIDTH + ZONES_WIDTH +
2233 		ilog2(roundup_pow_of_two(NR_CPUS)) > 32);
2234 
2235 	BUILD_BUG_ON(NR_CPUS > 4096);
2236 
2237 	kern_base = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
2238 	kern_size = (unsigned long)&_end - (unsigned long)KERNBASE;
2239 
2240 	/* Invalidate both kernel TSBs.  */
2241 	memset(swapper_tsb, 0x40, sizeof(swapper_tsb));
2242 #ifndef CONFIG_DEBUG_PAGEALLOC
2243 	memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2244 #endif
2245 
2246 	/* TTE.cv bit on sparc v9 occupies the same position as TTE.mcde
2247 	 * bit on M7 processor. This is a conflicting usage of the same
2248 	 * bit. Enabling TTE.cv on M7 would turn on Memory Corruption
2249 	 * Detection error on all pages and this will lead to problems
2250 	 * later. Kernel does not run with MCD enabled and hence rest
2251 	 * of the required steps to fully configure memory corruption
2252 	 * detection are not taken. We need to ensure TTE.mcde is not
2253 	 * set on M7 processor. Compute the value of cacheability
2254 	 * flag for use later taking this into consideration.
2255 	 */
2256 	switch (sun4v_chip_type) {
2257 	case SUN4V_CHIP_SPARC_M7:
2258 	case SUN4V_CHIP_SPARC_SN:
2259 		page_cache4v_flag = _PAGE_CP_4V;
2260 		break;
2261 	default:
2262 		page_cache4v_flag = _PAGE_CACHE_4V;
2263 		break;
2264 	}
2265 
2266 	if (tlb_type == hypervisor)
2267 		sun4v_pgprot_init();
2268 	else
2269 		sun4u_pgprot_init();
2270 
2271 	if (tlb_type == cheetah_plus ||
2272 	    tlb_type == hypervisor) {
2273 		tsb_phys_patch();
2274 		ktsb_phys_patch();
2275 	}
2276 
2277 	if (tlb_type == hypervisor)
2278 		sun4v_patch_tlb_handlers();
2279 
2280 	/* Find available physical memory...
2281 	 *
2282 	 * Read it twice in order to work around a bug in openfirmware.
2283 	 * The call to grab this table itself can cause openfirmware to
2284 	 * allocate memory, which in turn can take away some space from
2285 	 * the list of available memory.  Reading it twice makes sure
2286 	 * we really do get the final value.
2287 	 */
2288 	read_obp_translations();
2289 	read_obp_memory("reg", &pall[0], &pall_ents);
2290 	read_obp_memory("available", &pavail[0], &pavail_ents);
2291 	read_obp_memory("available", &pavail[0], &pavail_ents);
2292 
2293 	phys_base = 0xffffffffffffffffUL;
2294 	for (i = 0; i < pavail_ents; i++) {
2295 		phys_base = min(phys_base, pavail[i].phys_addr);
2296 		memblock_add(pavail[i].phys_addr, pavail[i].reg_size);
2297 	}
2298 
2299 	memblock_reserve(kern_base, kern_size);
2300 
2301 	find_ramdisk(phys_base);
2302 
2303 	if (cmdline_memory_size)
2304 		reduce_memory(cmdline_memory_size);
2305 
2306 	memblock_allow_resize();
2307 	memblock_dump_all();
2308 
2309 	set_bit(0, mmu_context_bmap);
2310 
2311 	shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE);
2312 
2313 	real_end = (unsigned long)_end;
2314 	num_kernel_image_mappings = DIV_ROUND_UP(real_end - KERNBASE, 1 << ILOG2_4MB);
2315 	printk("Kernel: Using %d locked TLB entries for main kernel image.\n",
2316 	       num_kernel_image_mappings);
2317 
2318 	/* Set kernel pgd to upper alias so physical page computations
2319 	 * work.
2320 	 */
2321 	init_mm.pgd += ((shift) / (sizeof(pgd_t)));
2322 
2323 	memset(swapper_pg_dir, 0, sizeof(swapper_pg_dir));
2324 
2325 	inherit_prom_mappings();
2326 
2327 	/* Ok, we can use our TLB miss and window trap handlers safely.  */
2328 	setup_tba();
2329 
2330 	__flush_tlb_all();
2331 
2332 	prom_build_devicetree();
2333 	of_populate_present_mask();
2334 #ifndef CONFIG_SMP
2335 	of_fill_in_cpu_data();
2336 #endif
2337 
2338 	if (tlb_type == hypervisor) {
2339 		sun4v_mdesc_init();
2340 		mdesc_populate_present_mask(cpu_all_mask);
2341 #ifndef CONFIG_SMP
2342 		mdesc_fill_in_cpu_data(cpu_all_mask);
2343 #endif
2344 		mdesc_get_page_sizes(cpu_all_mask, &cpu_pgsz_mask);
2345 
2346 		sun4v_linear_pte_xor_finalize();
2347 
2348 		sun4v_ktsb_init();
2349 		sun4v_ktsb_register();
2350 	} else {
2351 		unsigned long impl, ver;
2352 
2353 		cpu_pgsz_mask = (HV_PGSZ_MASK_8K | HV_PGSZ_MASK_64K |
2354 				 HV_PGSZ_MASK_512K | HV_PGSZ_MASK_4MB);
2355 
2356 		__asm__ __volatile__("rdpr %%ver, %0" : "=r" (ver));
2357 		impl = ((ver >> 32) & 0xffff);
2358 		if (impl == PANTHER_IMPL)
2359 			cpu_pgsz_mask |= (HV_PGSZ_MASK_32MB |
2360 					  HV_PGSZ_MASK_256MB);
2361 
2362 		sun4u_linear_pte_xor_finalize();
2363 	}
2364 
2365 	/* Flush the TLBs and the 4M TSB so that the updated linear
2366 	 * pte XOR settings are realized for all mappings.
2367 	 */
2368 	__flush_tlb_all();
2369 #ifndef CONFIG_DEBUG_PAGEALLOC
2370 	memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2371 #endif
2372 	__flush_tlb_all();
2373 
2374 	/* Setup bootmem... */
2375 	last_valid_pfn = end_pfn = bootmem_init(phys_base);
2376 
2377 	kernel_physical_mapping_init();
2378 
2379 	{
2380 		unsigned long max_zone_pfns[MAX_NR_ZONES];
2381 
2382 		memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
2383 
2384 		max_zone_pfns[ZONE_NORMAL] = end_pfn;
2385 
2386 		free_area_init_nodes(max_zone_pfns);
2387 	}
2388 
2389 	printk("Booting Linux...\n");
2390 }
2391 
2392 int page_in_phys_avail(unsigned long paddr)
2393 {
2394 	int i;
2395 
2396 	paddr &= PAGE_MASK;
2397 
2398 	for (i = 0; i < pavail_ents; i++) {
2399 		unsigned long start, end;
2400 
2401 		start = pavail[i].phys_addr;
2402 		end = start + pavail[i].reg_size;
2403 
2404 		if (paddr >= start && paddr < end)
2405 			return 1;
2406 	}
2407 	if (paddr >= kern_base && paddr < (kern_base + kern_size))
2408 		return 1;
2409 #ifdef CONFIG_BLK_DEV_INITRD
2410 	if (paddr >= __pa(initrd_start) &&
2411 	    paddr < __pa(PAGE_ALIGN(initrd_end)))
2412 		return 1;
2413 #endif
2414 
2415 	return 0;
2416 }
2417 
2418 static void __init register_page_bootmem_info(void)
2419 {
2420 #ifdef CONFIG_NEED_MULTIPLE_NODES
2421 	int i;
2422 
2423 	for_each_online_node(i)
2424 		if (NODE_DATA(i)->node_spanned_pages)
2425 			register_page_bootmem_info_node(NODE_DATA(i));
2426 #endif
2427 }
2428 void __init mem_init(void)
2429 {
2430 	high_memory = __va(last_valid_pfn << PAGE_SHIFT);
2431 
2432 	register_page_bootmem_info();
2433 	free_all_bootmem();
2434 
2435 	/*
2436 	 * Set up the zero page, mark it reserved, so that page count
2437 	 * is not manipulated when freeing the page from user ptes.
2438 	 */
2439 	mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0);
2440 	if (mem_map_zero == NULL) {
2441 		prom_printf("paging_init: Cannot alloc zero page.\n");
2442 		prom_halt();
2443 	}
2444 	mark_page_reserved(mem_map_zero);
2445 
2446 	mem_init_print_info(NULL);
2447 
2448 	if (tlb_type == cheetah || tlb_type == cheetah_plus)
2449 		cheetah_ecache_flush_init();
2450 }
2451 
2452 void free_initmem(void)
2453 {
2454 	unsigned long addr, initend;
2455 	int do_free = 1;
2456 
2457 	/* If the physical memory maps were trimmed by kernel command
2458 	 * line options, don't even try freeing this initmem stuff up.
2459 	 * The kernel image could have been in the trimmed out region
2460 	 * and if so the freeing below will free invalid page structs.
2461 	 */
2462 	if (cmdline_memory_size)
2463 		do_free = 0;
2464 
2465 	/*
2466 	 * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes.
2467 	 */
2468 	addr = PAGE_ALIGN((unsigned long)(__init_begin));
2469 	initend = (unsigned long)(__init_end) & PAGE_MASK;
2470 	for (; addr < initend; addr += PAGE_SIZE) {
2471 		unsigned long page;
2472 
2473 		page = (addr +
2474 			((unsigned long) __va(kern_base)) -
2475 			((unsigned long) KERNBASE));
2476 		memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
2477 
2478 		if (do_free)
2479 			free_reserved_page(virt_to_page(page));
2480 	}
2481 }
2482 
2483 #ifdef CONFIG_BLK_DEV_INITRD
2484 void free_initrd_mem(unsigned long start, unsigned long end)
2485 {
2486 	free_reserved_area((void *)start, (void *)end, POISON_FREE_INITMEM,
2487 			   "initrd");
2488 }
2489 #endif
2490 
2491 pgprot_t PAGE_KERNEL __read_mostly;
2492 EXPORT_SYMBOL(PAGE_KERNEL);
2493 
2494 pgprot_t PAGE_KERNEL_LOCKED __read_mostly;
2495 pgprot_t PAGE_COPY __read_mostly;
2496 
2497 pgprot_t PAGE_SHARED __read_mostly;
2498 EXPORT_SYMBOL(PAGE_SHARED);
2499 
2500 unsigned long pg_iobits __read_mostly;
2501 
2502 unsigned long _PAGE_IE __read_mostly;
2503 EXPORT_SYMBOL(_PAGE_IE);
2504 
2505 unsigned long _PAGE_E __read_mostly;
2506 EXPORT_SYMBOL(_PAGE_E);
2507 
2508 unsigned long _PAGE_CACHE __read_mostly;
2509 EXPORT_SYMBOL(_PAGE_CACHE);
2510 
2511 #ifdef CONFIG_SPARSEMEM_VMEMMAP
2512 int __meminit vmemmap_populate(unsigned long vstart, unsigned long vend,
2513 			       int node)
2514 {
2515 	unsigned long pte_base;
2516 
2517 	pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2518 		    _PAGE_CP_4U | _PAGE_CV_4U |
2519 		    _PAGE_P_4U | _PAGE_W_4U);
2520 	if (tlb_type == hypervisor)
2521 		pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2522 			    page_cache4v_flag | _PAGE_P_4V | _PAGE_W_4V);
2523 
2524 	pte_base |= _PAGE_PMD_HUGE;
2525 
2526 	vstart = vstart & PMD_MASK;
2527 	vend = ALIGN(vend, PMD_SIZE);
2528 	for (; vstart < vend; vstart += PMD_SIZE) {
2529 		pgd_t *pgd = pgd_offset_k(vstart);
2530 		unsigned long pte;
2531 		pud_t *pud;
2532 		pmd_t *pmd;
2533 
2534 		if (pgd_none(*pgd)) {
2535 			pud_t *new = vmemmap_alloc_block(PAGE_SIZE, node);
2536 
2537 			if (!new)
2538 				return -ENOMEM;
2539 			pgd_populate(&init_mm, pgd, new);
2540 		}
2541 
2542 		pud = pud_offset(pgd, vstart);
2543 		if (pud_none(*pud)) {
2544 			pmd_t *new = vmemmap_alloc_block(PAGE_SIZE, node);
2545 
2546 			if (!new)
2547 				return -ENOMEM;
2548 			pud_populate(&init_mm, pud, new);
2549 		}
2550 
2551 		pmd = pmd_offset(pud, vstart);
2552 
2553 		pte = pmd_val(*pmd);
2554 		if (!(pte & _PAGE_VALID)) {
2555 			void *block = vmemmap_alloc_block(PMD_SIZE, node);
2556 
2557 			if (!block)
2558 				return -ENOMEM;
2559 
2560 			pmd_val(*pmd) = pte_base | __pa(block);
2561 		}
2562 	}
2563 
2564 	return 0;
2565 }
2566 
2567 void vmemmap_free(unsigned long start, unsigned long end)
2568 {
2569 }
2570 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
2571 
2572 static void prot_init_common(unsigned long page_none,
2573 			     unsigned long page_shared,
2574 			     unsigned long page_copy,
2575 			     unsigned long page_readonly,
2576 			     unsigned long page_exec_bit)
2577 {
2578 	PAGE_COPY = __pgprot(page_copy);
2579 	PAGE_SHARED = __pgprot(page_shared);
2580 
2581 	protection_map[0x0] = __pgprot(page_none);
2582 	protection_map[0x1] = __pgprot(page_readonly & ~page_exec_bit);
2583 	protection_map[0x2] = __pgprot(page_copy & ~page_exec_bit);
2584 	protection_map[0x3] = __pgprot(page_copy & ~page_exec_bit);
2585 	protection_map[0x4] = __pgprot(page_readonly);
2586 	protection_map[0x5] = __pgprot(page_readonly);
2587 	protection_map[0x6] = __pgprot(page_copy);
2588 	protection_map[0x7] = __pgprot(page_copy);
2589 	protection_map[0x8] = __pgprot(page_none);
2590 	protection_map[0x9] = __pgprot(page_readonly & ~page_exec_bit);
2591 	protection_map[0xa] = __pgprot(page_shared & ~page_exec_bit);
2592 	protection_map[0xb] = __pgprot(page_shared & ~page_exec_bit);
2593 	protection_map[0xc] = __pgprot(page_readonly);
2594 	protection_map[0xd] = __pgprot(page_readonly);
2595 	protection_map[0xe] = __pgprot(page_shared);
2596 	protection_map[0xf] = __pgprot(page_shared);
2597 }
2598 
2599 static void __init sun4u_pgprot_init(void)
2600 {
2601 	unsigned long page_none, page_shared, page_copy, page_readonly;
2602 	unsigned long page_exec_bit;
2603 	int i;
2604 
2605 	PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2606 				_PAGE_CACHE_4U | _PAGE_P_4U |
2607 				__ACCESS_BITS_4U | __DIRTY_BITS_4U |
2608 				_PAGE_EXEC_4U);
2609 	PAGE_KERNEL_LOCKED = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2610 				       _PAGE_CACHE_4U | _PAGE_P_4U |
2611 				       __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2612 				       _PAGE_EXEC_4U | _PAGE_L_4U);
2613 
2614 	_PAGE_IE = _PAGE_IE_4U;
2615 	_PAGE_E = _PAGE_E_4U;
2616 	_PAGE_CACHE = _PAGE_CACHE_4U;
2617 
2618 	pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4U | __DIRTY_BITS_4U |
2619 		     __ACCESS_BITS_4U | _PAGE_E_4U);
2620 
2621 #ifdef CONFIG_DEBUG_PAGEALLOC
2622 	kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2623 #else
2624 	kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4U) ^
2625 		PAGE_OFFSET;
2626 #endif
2627 	kern_linear_pte_xor[0] |= (_PAGE_CP_4U | _PAGE_CV_4U |
2628 				   _PAGE_P_4U | _PAGE_W_4U);
2629 
2630 	for (i = 1; i < 4; i++)
2631 		kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2632 
2633 	_PAGE_ALL_SZ_BITS =  (_PAGE_SZ4MB_4U | _PAGE_SZ512K_4U |
2634 			      _PAGE_SZ64K_4U | _PAGE_SZ8K_4U |
2635 			      _PAGE_SZ32MB_4U | _PAGE_SZ256MB_4U);
2636 
2637 
2638 	page_none = _PAGE_PRESENT_4U | _PAGE_ACCESSED_4U | _PAGE_CACHE_4U;
2639 	page_shared = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2640 		       __ACCESS_BITS_4U | _PAGE_WRITE_4U | _PAGE_EXEC_4U);
2641 	page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2642 		       __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2643 	page_readonly   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2644 			   __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2645 
2646 	page_exec_bit = _PAGE_EXEC_4U;
2647 
2648 	prot_init_common(page_none, page_shared, page_copy, page_readonly,
2649 			 page_exec_bit);
2650 }
2651 
2652 static void __init sun4v_pgprot_init(void)
2653 {
2654 	unsigned long page_none, page_shared, page_copy, page_readonly;
2655 	unsigned long page_exec_bit;
2656 	int i;
2657 
2658 	PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4V | _PAGE_VALID |
2659 				page_cache4v_flag | _PAGE_P_4V |
2660 				__ACCESS_BITS_4V | __DIRTY_BITS_4V |
2661 				_PAGE_EXEC_4V);
2662 	PAGE_KERNEL_LOCKED = PAGE_KERNEL;
2663 
2664 	_PAGE_IE = _PAGE_IE_4V;
2665 	_PAGE_E = _PAGE_E_4V;
2666 	_PAGE_CACHE = page_cache4v_flag;
2667 
2668 #ifdef CONFIG_DEBUG_PAGEALLOC
2669 	kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2670 #else
2671 	kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4V) ^
2672 		PAGE_OFFSET;
2673 #endif
2674 	kern_linear_pte_xor[0] |= (page_cache4v_flag | _PAGE_P_4V |
2675 				   _PAGE_W_4V);
2676 
2677 	for (i = 1; i < 4; i++)
2678 		kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2679 
2680 	pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4V | __DIRTY_BITS_4V |
2681 		     __ACCESS_BITS_4V | _PAGE_E_4V);
2682 
2683 	_PAGE_ALL_SZ_BITS = (_PAGE_SZ16GB_4V | _PAGE_SZ2GB_4V |
2684 			     _PAGE_SZ256MB_4V | _PAGE_SZ32MB_4V |
2685 			     _PAGE_SZ4MB_4V | _PAGE_SZ512K_4V |
2686 			     _PAGE_SZ64K_4V | _PAGE_SZ8K_4V);
2687 
2688 	page_none = _PAGE_PRESENT_4V | _PAGE_ACCESSED_4V | page_cache4v_flag;
2689 	page_shared = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2690 		       __ACCESS_BITS_4V | _PAGE_WRITE_4V | _PAGE_EXEC_4V);
2691 	page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2692 		       __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2693 	page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2694 			 __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2695 
2696 	page_exec_bit = _PAGE_EXEC_4V;
2697 
2698 	prot_init_common(page_none, page_shared, page_copy, page_readonly,
2699 			 page_exec_bit);
2700 }
2701 
2702 unsigned long pte_sz_bits(unsigned long sz)
2703 {
2704 	if (tlb_type == hypervisor) {
2705 		switch (sz) {
2706 		case 8 * 1024:
2707 		default:
2708 			return _PAGE_SZ8K_4V;
2709 		case 64 * 1024:
2710 			return _PAGE_SZ64K_4V;
2711 		case 512 * 1024:
2712 			return _PAGE_SZ512K_4V;
2713 		case 4 * 1024 * 1024:
2714 			return _PAGE_SZ4MB_4V;
2715 		}
2716 	} else {
2717 		switch (sz) {
2718 		case 8 * 1024:
2719 		default:
2720 			return _PAGE_SZ8K_4U;
2721 		case 64 * 1024:
2722 			return _PAGE_SZ64K_4U;
2723 		case 512 * 1024:
2724 			return _PAGE_SZ512K_4U;
2725 		case 4 * 1024 * 1024:
2726 			return _PAGE_SZ4MB_4U;
2727 		}
2728 	}
2729 }
2730 
2731 pte_t mk_pte_io(unsigned long page, pgprot_t prot, int space, unsigned long page_size)
2732 {
2733 	pte_t pte;
2734 
2735 	pte_val(pte)  = page | pgprot_val(pgprot_noncached(prot));
2736 	pte_val(pte) |= (((unsigned long)space) << 32);
2737 	pte_val(pte) |= pte_sz_bits(page_size);
2738 
2739 	return pte;
2740 }
2741 
2742 static unsigned long kern_large_tte(unsigned long paddr)
2743 {
2744 	unsigned long val;
2745 
2746 	val = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2747 	       _PAGE_CP_4U | _PAGE_CV_4U | _PAGE_P_4U |
2748 	       _PAGE_EXEC_4U | _PAGE_L_4U | _PAGE_W_4U);
2749 	if (tlb_type == hypervisor)
2750 		val = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2751 		       page_cache4v_flag | _PAGE_P_4V |
2752 		       _PAGE_EXEC_4V | _PAGE_W_4V);
2753 
2754 	return val | paddr;
2755 }
2756 
2757 /* If not locked, zap it. */
2758 void __flush_tlb_all(void)
2759 {
2760 	unsigned long pstate;
2761 	int i;
2762 
2763 	__asm__ __volatile__("flushw\n\t"
2764 			     "rdpr	%%pstate, %0\n\t"
2765 			     "wrpr	%0, %1, %%pstate"
2766 			     : "=r" (pstate)
2767 			     : "i" (PSTATE_IE));
2768 	if (tlb_type == hypervisor) {
2769 		sun4v_mmu_demap_all();
2770 	} else if (tlb_type == spitfire) {
2771 		for (i = 0; i < 64; i++) {
2772 			/* Spitfire Errata #32 workaround */
2773 			/* NOTE: Always runs on spitfire, so no
2774 			 *       cheetah+ page size encodings.
2775 			 */
2776 			__asm__ __volatile__("stxa	%0, [%1] %2\n\t"
2777 					     "flush	%%g6"
2778 					     : /* No outputs */
2779 					     : "r" (0),
2780 					     "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2781 
2782 			if (!(spitfire_get_dtlb_data(i) & _PAGE_L_4U)) {
2783 				__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2784 						     "membar #Sync"
2785 						     : /* no outputs */
2786 						     : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
2787 				spitfire_put_dtlb_data(i, 0x0UL);
2788 			}
2789 
2790 			/* Spitfire Errata #32 workaround */
2791 			/* NOTE: Always runs on spitfire, so no
2792 			 *       cheetah+ page size encodings.
2793 			 */
2794 			__asm__ __volatile__("stxa	%0, [%1] %2\n\t"
2795 					     "flush	%%g6"
2796 					     : /* No outputs */
2797 					     : "r" (0),
2798 					     "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2799 
2800 			if (!(spitfire_get_itlb_data(i) & _PAGE_L_4U)) {
2801 				__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2802 						     "membar #Sync"
2803 						     : /* no outputs */
2804 						     : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
2805 				spitfire_put_itlb_data(i, 0x0UL);
2806 			}
2807 		}
2808 	} else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
2809 		cheetah_flush_dtlb_all();
2810 		cheetah_flush_itlb_all();
2811 	}
2812 	__asm__ __volatile__("wrpr	%0, 0, %%pstate"
2813 			     : : "r" (pstate));
2814 }
2815 
2816 pte_t *pte_alloc_one_kernel(struct mm_struct *mm,
2817 			    unsigned long address)
2818 {
2819 	struct page *page = alloc_page(GFP_KERNEL | __GFP_NOTRACK | __GFP_ZERO);
2820 	pte_t *pte = NULL;
2821 
2822 	if (page)
2823 		pte = (pte_t *) page_address(page);
2824 
2825 	return pte;
2826 }
2827 
2828 pgtable_t pte_alloc_one(struct mm_struct *mm,
2829 			unsigned long address)
2830 {
2831 	struct page *page = alloc_page(GFP_KERNEL | __GFP_NOTRACK | __GFP_ZERO);
2832 	if (!page)
2833 		return NULL;
2834 	if (!pgtable_page_ctor(page)) {
2835 		free_hot_cold_page(page, 0);
2836 		return NULL;
2837 	}
2838 	return (pte_t *) page_address(page);
2839 }
2840 
2841 void pte_free_kernel(struct mm_struct *mm, pte_t *pte)
2842 {
2843 	free_page((unsigned long)pte);
2844 }
2845 
2846 static void __pte_free(pgtable_t pte)
2847 {
2848 	struct page *page = virt_to_page(pte);
2849 
2850 	pgtable_page_dtor(page);
2851 	__free_page(page);
2852 }
2853 
2854 void pte_free(struct mm_struct *mm, pgtable_t pte)
2855 {
2856 	__pte_free(pte);
2857 }
2858 
2859 void pgtable_free(void *table, bool is_page)
2860 {
2861 	if (is_page)
2862 		__pte_free(table);
2863 	else
2864 		kmem_cache_free(pgtable_cache, table);
2865 }
2866 
2867 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2868 void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
2869 			  pmd_t *pmd)
2870 {
2871 	unsigned long pte, flags;
2872 	struct mm_struct *mm;
2873 	pmd_t entry = *pmd;
2874 
2875 	if (!pmd_large(entry) || !pmd_young(entry))
2876 		return;
2877 
2878 	pte = pmd_val(entry);
2879 
2880 	/* Don't insert a non-valid PMD into the TSB, we'll deadlock.  */
2881 	if (!(pte & _PAGE_VALID))
2882 		return;
2883 
2884 	/* We are fabricating 8MB pages using 4MB real hw pages.  */
2885 	pte |= (addr & (1UL << REAL_HPAGE_SHIFT));
2886 
2887 	mm = vma->vm_mm;
2888 
2889 	spin_lock_irqsave(&mm->context.lock, flags);
2890 
2891 	if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL)
2892 		__update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
2893 					addr, pte);
2894 
2895 	spin_unlock_irqrestore(&mm->context.lock, flags);
2896 }
2897 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2898 
2899 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
2900 static void context_reload(void *__data)
2901 {
2902 	struct mm_struct *mm = __data;
2903 
2904 	if (mm == current->mm)
2905 		load_secondary_context(mm);
2906 }
2907 
2908 void hugetlb_setup(struct pt_regs *regs)
2909 {
2910 	struct mm_struct *mm = current->mm;
2911 	struct tsb_config *tp;
2912 
2913 	if (faulthandler_disabled() || !mm) {
2914 		const struct exception_table_entry *entry;
2915 
2916 		entry = search_exception_tables(regs->tpc);
2917 		if (entry) {
2918 			regs->tpc = entry->fixup;
2919 			regs->tnpc = regs->tpc + 4;
2920 			return;
2921 		}
2922 		pr_alert("Unexpected HugeTLB setup in atomic context.\n");
2923 		die_if_kernel("HugeTSB in atomic", regs);
2924 	}
2925 
2926 	tp = &mm->context.tsb_block[MM_TSB_HUGE];
2927 	if (likely(tp->tsb == NULL))
2928 		tsb_grow(mm, MM_TSB_HUGE, 0);
2929 
2930 	tsb_context_switch(mm);
2931 	smp_tsb_sync(mm);
2932 
2933 	/* On UltraSPARC-III+ and later, configure the second half of
2934 	 * the Data-TLB for huge pages.
2935 	 */
2936 	if (tlb_type == cheetah_plus) {
2937 		bool need_context_reload = false;
2938 		unsigned long ctx;
2939 
2940 		spin_lock_irq(&ctx_alloc_lock);
2941 		ctx = mm->context.sparc64_ctx_val;
2942 		ctx &= ~CTX_PGSZ_MASK;
2943 		ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT;
2944 		ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT;
2945 
2946 		if (ctx != mm->context.sparc64_ctx_val) {
2947 			/* When changing the page size fields, we
2948 			 * must perform a context flush so that no
2949 			 * stale entries match.  This flush must
2950 			 * occur with the original context register
2951 			 * settings.
2952 			 */
2953 			do_flush_tlb_mm(mm);
2954 
2955 			/* Reload the context register of all processors
2956 			 * also executing in this address space.
2957 			 */
2958 			mm->context.sparc64_ctx_val = ctx;
2959 			need_context_reload = true;
2960 		}
2961 		spin_unlock_irq(&ctx_alloc_lock);
2962 
2963 		if (need_context_reload)
2964 			on_each_cpu(context_reload, mm, 0);
2965 	}
2966 }
2967 #endif
2968 
2969 static struct resource code_resource = {
2970 	.name	= "Kernel code",
2971 	.flags	= IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
2972 };
2973 
2974 static struct resource data_resource = {
2975 	.name	= "Kernel data",
2976 	.flags	= IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
2977 };
2978 
2979 static struct resource bss_resource = {
2980 	.name	= "Kernel bss",
2981 	.flags	= IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
2982 };
2983 
2984 static inline resource_size_t compute_kern_paddr(void *addr)
2985 {
2986 	return (resource_size_t) (addr - KERNBASE + kern_base);
2987 }
2988 
2989 static void __init kernel_lds_init(void)
2990 {
2991 	code_resource.start = compute_kern_paddr(_text);
2992 	code_resource.end   = compute_kern_paddr(_etext - 1);
2993 	data_resource.start = compute_kern_paddr(_etext);
2994 	data_resource.end   = compute_kern_paddr(_edata - 1);
2995 	bss_resource.start  = compute_kern_paddr(__bss_start);
2996 	bss_resource.end    = compute_kern_paddr(_end - 1);
2997 }
2998 
2999 static int __init report_memory(void)
3000 {
3001 	int i;
3002 	struct resource *res;
3003 
3004 	kernel_lds_init();
3005 
3006 	for (i = 0; i < pavail_ents; i++) {
3007 		res = kzalloc(sizeof(struct resource), GFP_KERNEL);
3008 
3009 		if (!res) {
3010 			pr_warn("Failed to allocate source.\n");
3011 			break;
3012 		}
3013 
3014 		res->name = "System RAM";
3015 		res->start = pavail[i].phys_addr;
3016 		res->end = pavail[i].phys_addr + pavail[i].reg_size - 1;
3017 		res->flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM;
3018 
3019 		if (insert_resource(&iomem_resource, res) < 0) {
3020 			pr_warn("Resource insertion failed.\n");
3021 			break;
3022 		}
3023 
3024 		insert_resource(res, &code_resource);
3025 		insert_resource(res, &data_resource);
3026 		insert_resource(res, &bss_resource);
3027 	}
3028 
3029 	return 0;
3030 }
3031 arch_initcall(report_memory);
3032 
3033 #ifdef CONFIG_SMP
3034 #define do_flush_tlb_kernel_range	smp_flush_tlb_kernel_range
3035 #else
3036 #define do_flush_tlb_kernel_range	__flush_tlb_kernel_range
3037 #endif
3038 
3039 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
3040 {
3041 	if (start < HI_OBP_ADDRESS && end > LOW_OBP_ADDRESS) {
3042 		if (start < LOW_OBP_ADDRESS) {
3043 			flush_tsb_kernel_range(start, LOW_OBP_ADDRESS);
3044 			do_flush_tlb_kernel_range(start, LOW_OBP_ADDRESS);
3045 		}
3046 		if (end > HI_OBP_ADDRESS) {
3047 			flush_tsb_kernel_range(HI_OBP_ADDRESS, end);
3048 			do_flush_tlb_kernel_range(HI_OBP_ADDRESS, end);
3049 		}
3050 	} else {
3051 		flush_tsb_kernel_range(start, end);
3052 		do_flush_tlb_kernel_range(start, end);
3053 	}
3054 }
3055