xref: /linux/arch/sparc/mm/init_64.c (revision 95e9fd10f06cb5642028b6b851e32b8c8afb4571)
1 /*
2  *  arch/sparc64/mm/init.c
3  *
4  *  Copyright (C) 1996-1999 David S. Miller (davem@caip.rutgers.edu)
5  *  Copyright (C) 1997-1999 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
6  */
7 
8 #include <linux/module.h>
9 #include <linux/kernel.h>
10 #include <linux/sched.h>
11 #include <linux/string.h>
12 #include <linux/init.h>
13 #include <linux/bootmem.h>
14 #include <linux/mm.h>
15 #include <linux/hugetlb.h>
16 #include <linux/initrd.h>
17 #include <linux/swap.h>
18 #include <linux/pagemap.h>
19 #include <linux/poison.h>
20 #include <linux/fs.h>
21 #include <linux/seq_file.h>
22 #include <linux/kprobes.h>
23 #include <linux/cache.h>
24 #include <linux/sort.h>
25 #include <linux/percpu.h>
26 #include <linux/memblock.h>
27 #include <linux/mmzone.h>
28 #include <linux/gfp.h>
29 
30 #include <asm/head.h>
31 #include <asm/page.h>
32 #include <asm/pgalloc.h>
33 #include <asm/pgtable.h>
34 #include <asm/oplib.h>
35 #include <asm/iommu.h>
36 #include <asm/io.h>
37 #include <asm/uaccess.h>
38 #include <asm/mmu_context.h>
39 #include <asm/tlbflush.h>
40 #include <asm/dma.h>
41 #include <asm/starfire.h>
42 #include <asm/tlb.h>
43 #include <asm/spitfire.h>
44 #include <asm/sections.h>
45 #include <asm/tsb.h>
46 #include <asm/hypervisor.h>
47 #include <asm/prom.h>
48 #include <asm/mdesc.h>
49 #include <asm/cpudata.h>
50 #include <asm/irq.h>
51 
52 #include "init_64.h"
53 
54 unsigned long kern_linear_pte_xor[2] __read_mostly;
55 
56 /* A bitmap, one bit for every 256MB of physical memory.  If the bit
57  * is clear, we should use a 4MB page (via kern_linear_pte_xor[0]) else
58  * if set we should use a 256MB page (via kern_linear_pte_xor[1]).
59  */
60 unsigned long kpte_linear_bitmap[KPTE_BITMAP_BYTES / sizeof(unsigned long)];
61 
62 #ifndef CONFIG_DEBUG_PAGEALLOC
63 /* A special kernel TSB for 4MB and 256MB linear mappings.
64  * Space is allocated for this right after the trap table
65  * in arch/sparc64/kernel/head.S
66  */
67 extern struct tsb swapper_4m_tsb[KERNEL_TSB4M_NENTRIES];
68 #endif
69 
70 #define MAX_BANKS	32
71 
72 static struct linux_prom64_registers pavail[MAX_BANKS] __devinitdata;
73 static int pavail_ents __devinitdata;
74 
75 static int cmp_p64(const void *a, const void *b)
76 {
77 	const struct linux_prom64_registers *x = a, *y = b;
78 
79 	if (x->phys_addr > y->phys_addr)
80 		return 1;
81 	if (x->phys_addr < y->phys_addr)
82 		return -1;
83 	return 0;
84 }
85 
86 static void __init read_obp_memory(const char *property,
87 				   struct linux_prom64_registers *regs,
88 				   int *num_ents)
89 {
90 	phandle node = prom_finddevice("/memory");
91 	int prop_size = prom_getproplen(node, property);
92 	int ents, ret, i;
93 
94 	ents = prop_size / sizeof(struct linux_prom64_registers);
95 	if (ents > MAX_BANKS) {
96 		prom_printf("The machine has more %s property entries than "
97 			    "this kernel can support (%d).\n",
98 			    property, MAX_BANKS);
99 		prom_halt();
100 	}
101 
102 	ret = prom_getproperty(node, property, (char *) regs, prop_size);
103 	if (ret == -1) {
104 		prom_printf("Couldn't get %s property from /memory.\n");
105 		prom_halt();
106 	}
107 
108 	/* Sanitize what we got from the firmware, by page aligning
109 	 * everything.
110 	 */
111 	for (i = 0; i < ents; i++) {
112 		unsigned long base, size;
113 
114 		base = regs[i].phys_addr;
115 		size = regs[i].reg_size;
116 
117 		size &= PAGE_MASK;
118 		if (base & ~PAGE_MASK) {
119 			unsigned long new_base = PAGE_ALIGN(base);
120 
121 			size -= new_base - base;
122 			if ((long) size < 0L)
123 				size = 0UL;
124 			base = new_base;
125 		}
126 		if (size == 0UL) {
127 			/* If it is empty, simply get rid of it.
128 			 * This simplifies the logic of the other
129 			 * functions that process these arrays.
130 			 */
131 			memmove(&regs[i], &regs[i + 1],
132 				(ents - i - 1) * sizeof(regs[0]));
133 			i--;
134 			ents--;
135 			continue;
136 		}
137 		regs[i].phys_addr = base;
138 		regs[i].reg_size = size;
139 	}
140 
141 	*num_ents = ents;
142 
143 	sort(regs, ents, sizeof(struct linux_prom64_registers),
144 	     cmp_p64, NULL);
145 }
146 
147 unsigned long sparc64_valid_addr_bitmap[VALID_ADDR_BITMAP_BYTES /
148 					sizeof(unsigned long)];
149 EXPORT_SYMBOL(sparc64_valid_addr_bitmap);
150 
151 /* Kernel physical address base and size in bytes.  */
152 unsigned long kern_base __read_mostly;
153 unsigned long kern_size __read_mostly;
154 
155 /* Initial ramdisk setup */
156 extern unsigned long sparc_ramdisk_image64;
157 extern unsigned int sparc_ramdisk_image;
158 extern unsigned int sparc_ramdisk_size;
159 
160 struct page *mem_map_zero __read_mostly;
161 EXPORT_SYMBOL(mem_map_zero);
162 
163 unsigned int sparc64_highest_unlocked_tlb_ent __read_mostly;
164 
165 unsigned long sparc64_kern_pri_context __read_mostly;
166 unsigned long sparc64_kern_pri_nuc_bits __read_mostly;
167 unsigned long sparc64_kern_sec_context __read_mostly;
168 
169 int num_kernel_image_mappings;
170 
171 #ifdef CONFIG_DEBUG_DCFLUSH
172 atomic_t dcpage_flushes = ATOMIC_INIT(0);
173 #ifdef CONFIG_SMP
174 atomic_t dcpage_flushes_xcall = ATOMIC_INIT(0);
175 #endif
176 #endif
177 
178 inline void flush_dcache_page_impl(struct page *page)
179 {
180 	BUG_ON(tlb_type == hypervisor);
181 #ifdef CONFIG_DEBUG_DCFLUSH
182 	atomic_inc(&dcpage_flushes);
183 #endif
184 
185 #ifdef DCACHE_ALIASING_POSSIBLE
186 	__flush_dcache_page(page_address(page),
187 			    ((tlb_type == spitfire) &&
188 			     page_mapping(page) != NULL));
189 #else
190 	if (page_mapping(page) != NULL &&
191 	    tlb_type == spitfire)
192 		__flush_icache_page(__pa(page_address(page)));
193 #endif
194 }
195 
196 #define PG_dcache_dirty		PG_arch_1
197 #define PG_dcache_cpu_shift	32UL
198 #define PG_dcache_cpu_mask	\
199 	((1UL<<ilog2(roundup_pow_of_two(NR_CPUS)))-1UL)
200 
201 #define dcache_dirty_cpu(page) \
202 	(((page)->flags >> PG_dcache_cpu_shift) & PG_dcache_cpu_mask)
203 
204 static inline void set_dcache_dirty(struct page *page, int this_cpu)
205 {
206 	unsigned long mask = this_cpu;
207 	unsigned long non_cpu_bits;
208 
209 	non_cpu_bits = ~(PG_dcache_cpu_mask << PG_dcache_cpu_shift);
210 	mask = (mask << PG_dcache_cpu_shift) | (1UL << PG_dcache_dirty);
211 
212 	__asm__ __volatile__("1:\n\t"
213 			     "ldx	[%2], %%g7\n\t"
214 			     "and	%%g7, %1, %%g1\n\t"
215 			     "or	%%g1, %0, %%g1\n\t"
216 			     "casx	[%2], %%g7, %%g1\n\t"
217 			     "cmp	%%g7, %%g1\n\t"
218 			     "bne,pn	%%xcc, 1b\n\t"
219 			     " nop"
220 			     : /* no outputs */
221 			     : "r" (mask), "r" (non_cpu_bits), "r" (&page->flags)
222 			     : "g1", "g7");
223 }
224 
225 static inline void clear_dcache_dirty_cpu(struct page *page, unsigned long cpu)
226 {
227 	unsigned long mask = (1UL << PG_dcache_dirty);
228 
229 	__asm__ __volatile__("! test_and_clear_dcache_dirty\n"
230 			     "1:\n\t"
231 			     "ldx	[%2], %%g7\n\t"
232 			     "srlx	%%g7, %4, %%g1\n\t"
233 			     "and	%%g1, %3, %%g1\n\t"
234 			     "cmp	%%g1, %0\n\t"
235 			     "bne,pn	%%icc, 2f\n\t"
236 			     " andn	%%g7, %1, %%g1\n\t"
237 			     "casx	[%2], %%g7, %%g1\n\t"
238 			     "cmp	%%g7, %%g1\n\t"
239 			     "bne,pn	%%xcc, 1b\n\t"
240 			     " nop\n"
241 			     "2:"
242 			     : /* no outputs */
243 			     : "r" (cpu), "r" (mask), "r" (&page->flags),
244 			       "i" (PG_dcache_cpu_mask),
245 			       "i" (PG_dcache_cpu_shift)
246 			     : "g1", "g7");
247 }
248 
249 static inline void tsb_insert(struct tsb *ent, unsigned long tag, unsigned long pte)
250 {
251 	unsigned long tsb_addr = (unsigned long) ent;
252 
253 	if (tlb_type == cheetah_plus || tlb_type == hypervisor)
254 		tsb_addr = __pa(tsb_addr);
255 
256 	__tsb_insert(tsb_addr, tag, pte);
257 }
258 
259 unsigned long _PAGE_ALL_SZ_BITS __read_mostly;
260 unsigned long _PAGE_SZBITS __read_mostly;
261 
262 static void flush_dcache(unsigned long pfn)
263 {
264 	struct page *page;
265 
266 	page = pfn_to_page(pfn);
267 	if (page) {
268 		unsigned long pg_flags;
269 
270 		pg_flags = page->flags;
271 		if (pg_flags & (1UL << PG_dcache_dirty)) {
272 			int cpu = ((pg_flags >> PG_dcache_cpu_shift) &
273 				   PG_dcache_cpu_mask);
274 			int this_cpu = get_cpu();
275 
276 			/* This is just to optimize away some function calls
277 			 * in the SMP case.
278 			 */
279 			if (cpu == this_cpu)
280 				flush_dcache_page_impl(page);
281 			else
282 				smp_flush_dcache_page_impl(page, cpu);
283 
284 			clear_dcache_dirty_cpu(page, cpu);
285 
286 			put_cpu();
287 		}
288 	}
289 }
290 
291 void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
292 {
293 	struct mm_struct *mm;
294 	struct tsb *tsb;
295 	unsigned long tag, flags;
296 	unsigned long tsb_index, tsb_hash_shift;
297 	pte_t pte = *ptep;
298 
299 	if (tlb_type != hypervisor) {
300 		unsigned long pfn = pte_pfn(pte);
301 
302 		if (pfn_valid(pfn))
303 			flush_dcache(pfn);
304 	}
305 
306 	mm = vma->vm_mm;
307 
308 	tsb_index = MM_TSB_BASE;
309 	tsb_hash_shift = PAGE_SHIFT;
310 
311 	spin_lock_irqsave(&mm->context.lock, flags);
312 
313 #ifdef CONFIG_HUGETLB_PAGE
314 	if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL) {
315 		if ((tlb_type == hypervisor &&
316 		     (pte_val(pte) & _PAGE_SZALL_4V) == _PAGE_SZHUGE_4V) ||
317 		    (tlb_type != hypervisor &&
318 		     (pte_val(pte) & _PAGE_SZALL_4U) == _PAGE_SZHUGE_4U)) {
319 			tsb_index = MM_TSB_HUGE;
320 			tsb_hash_shift = HPAGE_SHIFT;
321 		}
322 	}
323 #endif
324 
325 	tsb = mm->context.tsb_block[tsb_index].tsb;
326 	tsb += ((address >> tsb_hash_shift) &
327 		(mm->context.tsb_block[tsb_index].tsb_nentries - 1UL));
328 	tag = (address >> 22UL);
329 	tsb_insert(tsb, tag, pte_val(pte));
330 
331 	spin_unlock_irqrestore(&mm->context.lock, flags);
332 }
333 
334 void flush_dcache_page(struct page *page)
335 {
336 	struct address_space *mapping;
337 	int this_cpu;
338 
339 	if (tlb_type == hypervisor)
340 		return;
341 
342 	/* Do not bother with the expensive D-cache flush if it
343 	 * is merely the zero page.  The 'bigcore' testcase in GDB
344 	 * causes this case to run millions of times.
345 	 */
346 	if (page == ZERO_PAGE(0))
347 		return;
348 
349 	this_cpu = get_cpu();
350 
351 	mapping = page_mapping(page);
352 	if (mapping && !mapping_mapped(mapping)) {
353 		int dirty = test_bit(PG_dcache_dirty, &page->flags);
354 		if (dirty) {
355 			int dirty_cpu = dcache_dirty_cpu(page);
356 
357 			if (dirty_cpu == this_cpu)
358 				goto out;
359 			smp_flush_dcache_page_impl(page, dirty_cpu);
360 		}
361 		set_dcache_dirty(page, this_cpu);
362 	} else {
363 		/* We could delay the flush for the !page_mapping
364 		 * case too.  But that case is for exec env/arg
365 		 * pages and those are %99 certainly going to get
366 		 * faulted into the tlb (and thus flushed) anyways.
367 		 */
368 		flush_dcache_page_impl(page);
369 	}
370 
371 out:
372 	put_cpu();
373 }
374 EXPORT_SYMBOL(flush_dcache_page);
375 
376 void __kprobes flush_icache_range(unsigned long start, unsigned long end)
377 {
378 	/* Cheetah and Hypervisor platform cpus have coherent I-cache. */
379 	if (tlb_type == spitfire) {
380 		unsigned long kaddr;
381 
382 		/* This code only runs on Spitfire cpus so this is
383 		 * why we can assume _PAGE_PADDR_4U.
384 		 */
385 		for (kaddr = start; kaddr < end; kaddr += PAGE_SIZE) {
386 			unsigned long paddr, mask = _PAGE_PADDR_4U;
387 
388 			if (kaddr >= PAGE_OFFSET)
389 				paddr = kaddr & mask;
390 			else {
391 				pgd_t *pgdp = pgd_offset_k(kaddr);
392 				pud_t *pudp = pud_offset(pgdp, kaddr);
393 				pmd_t *pmdp = pmd_offset(pudp, kaddr);
394 				pte_t *ptep = pte_offset_kernel(pmdp, kaddr);
395 
396 				paddr = pte_val(*ptep) & mask;
397 			}
398 			__flush_icache_page(paddr);
399 		}
400 	}
401 }
402 EXPORT_SYMBOL(flush_icache_range);
403 
404 void mmu_info(struct seq_file *m)
405 {
406 	if (tlb_type == cheetah)
407 		seq_printf(m, "MMU Type\t: Cheetah\n");
408 	else if (tlb_type == cheetah_plus)
409 		seq_printf(m, "MMU Type\t: Cheetah+\n");
410 	else if (tlb_type == spitfire)
411 		seq_printf(m, "MMU Type\t: Spitfire\n");
412 	else if (tlb_type == hypervisor)
413 		seq_printf(m, "MMU Type\t: Hypervisor (sun4v)\n");
414 	else
415 		seq_printf(m, "MMU Type\t: ???\n");
416 
417 #ifdef CONFIG_DEBUG_DCFLUSH
418 	seq_printf(m, "DCPageFlushes\t: %d\n",
419 		   atomic_read(&dcpage_flushes));
420 #ifdef CONFIG_SMP
421 	seq_printf(m, "DCPageFlushesXC\t: %d\n",
422 		   atomic_read(&dcpage_flushes_xcall));
423 #endif /* CONFIG_SMP */
424 #endif /* CONFIG_DEBUG_DCFLUSH */
425 }
426 
427 struct linux_prom_translation prom_trans[512] __read_mostly;
428 unsigned int prom_trans_ents __read_mostly;
429 
430 unsigned long kern_locked_tte_data;
431 
432 /* The obp translations are saved based on 8k pagesize, since obp can
433  * use a mixture of pagesizes. Misses to the LOW_OBP_ADDRESS ->
434  * HI_OBP_ADDRESS range are handled in ktlb.S.
435  */
436 static inline int in_obp_range(unsigned long vaddr)
437 {
438 	return (vaddr >= LOW_OBP_ADDRESS &&
439 		vaddr < HI_OBP_ADDRESS);
440 }
441 
442 static int cmp_ptrans(const void *a, const void *b)
443 {
444 	const struct linux_prom_translation *x = a, *y = b;
445 
446 	if (x->virt > y->virt)
447 		return 1;
448 	if (x->virt < y->virt)
449 		return -1;
450 	return 0;
451 }
452 
453 /* Read OBP translations property into 'prom_trans[]'.  */
454 static void __init read_obp_translations(void)
455 {
456 	int n, node, ents, first, last, i;
457 
458 	node = prom_finddevice("/virtual-memory");
459 	n = prom_getproplen(node, "translations");
460 	if (unlikely(n == 0 || n == -1)) {
461 		prom_printf("prom_mappings: Couldn't get size.\n");
462 		prom_halt();
463 	}
464 	if (unlikely(n > sizeof(prom_trans))) {
465 		prom_printf("prom_mappings: Size %Zd is too big.\n", n);
466 		prom_halt();
467 	}
468 
469 	if ((n = prom_getproperty(node, "translations",
470 				  (char *)&prom_trans[0],
471 				  sizeof(prom_trans))) == -1) {
472 		prom_printf("prom_mappings: Couldn't get property.\n");
473 		prom_halt();
474 	}
475 
476 	n = n / sizeof(struct linux_prom_translation);
477 
478 	ents = n;
479 
480 	sort(prom_trans, ents, sizeof(struct linux_prom_translation),
481 	     cmp_ptrans, NULL);
482 
483 	/* Now kick out all the non-OBP entries.  */
484 	for (i = 0; i < ents; i++) {
485 		if (in_obp_range(prom_trans[i].virt))
486 			break;
487 	}
488 	first = i;
489 	for (; i < ents; i++) {
490 		if (!in_obp_range(prom_trans[i].virt))
491 			break;
492 	}
493 	last = i;
494 
495 	for (i = 0; i < (last - first); i++) {
496 		struct linux_prom_translation *src = &prom_trans[i + first];
497 		struct linux_prom_translation *dest = &prom_trans[i];
498 
499 		*dest = *src;
500 	}
501 	for (; i < ents; i++) {
502 		struct linux_prom_translation *dest = &prom_trans[i];
503 		dest->virt = dest->size = dest->data = 0x0UL;
504 	}
505 
506 	prom_trans_ents = last - first;
507 
508 	if (tlb_type == spitfire) {
509 		/* Clear diag TTE bits. */
510 		for (i = 0; i < prom_trans_ents; i++)
511 			prom_trans[i].data &= ~0x0003fe0000000000UL;
512 	}
513 
514 	/* Force execute bit on.  */
515 	for (i = 0; i < prom_trans_ents; i++)
516 		prom_trans[i].data |= (tlb_type == hypervisor ?
517 				       _PAGE_EXEC_4V : _PAGE_EXEC_4U);
518 }
519 
520 static void __init hypervisor_tlb_lock(unsigned long vaddr,
521 				       unsigned long pte,
522 				       unsigned long mmu)
523 {
524 	unsigned long ret = sun4v_mmu_map_perm_addr(vaddr, 0, pte, mmu);
525 
526 	if (ret != 0) {
527 		prom_printf("hypervisor_tlb_lock[%lx:%lx:%lx:%lx]: "
528 			    "errors with %lx\n", vaddr, 0, pte, mmu, ret);
529 		prom_halt();
530 	}
531 }
532 
533 static unsigned long kern_large_tte(unsigned long paddr);
534 
535 static void __init remap_kernel(void)
536 {
537 	unsigned long phys_page, tte_vaddr, tte_data;
538 	int i, tlb_ent = sparc64_highest_locked_tlbent();
539 
540 	tte_vaddr = (unsigned long) KERNBASE;
541 	phys_page = (prom_boot_mapping_phys_low >> 22UL) << 22UL;
542 	tte_data = kern_large_tte(phys_page);
543 
544 	kern_locked_tte_data = tte_data;
545 
546 	/* Now lock us into the TLBs via Hypervisor or OBP. */
547 	if (tlb_type == hypervisor) {
548 		for (i = 0; i < num_kernel_image_mappings; i++) {
549 			hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_DMMU);
550 			hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_IMMU);
551 			tte_vaddr += 0x400000;
552 			tte_data += 0x400000;
553 		}
554 	} else {
555 		for (i = 0; i < num_kernel_image_mappings; i++) {
556 			prom_dtlb_load(tlb_ent - i, tte_data, tte_vaddr);
557 			prom_itlb_load(tlb_ent - i, tte_data, tte_vaddr);
558 			tte_vaddr += 0x400000;
559 			tte_data += 0x400000;
560 		}
561 		sparc64_highest_unlocked_tlb_ent = tlb_ent - i;
562 	}
563 	if (tlb_type == cheetah_plus) {
564 		sparc64_kern_pri_context = (CTX_CHEETAH_PLUS_CTX0 |
565 					    CTX_CHEETAH_PLUS_NUC);
566 		sparc64_kern_pri_nuc_bits = CTX_CHEETAH_PLUS_NUC;
567 		sparc64_kern_sec_context = CTX_CHEETAH_PLUS_CTX0;
568 	}
569 }
570 
571 
572 static void __init inherit_prom_mappings(void)
573 {
574 	/* Now fixup OBP's idea about where we really are mapped. */
575 	printk("Remapping the kernel... ");
576 	remap_kernel();
577 	printk("done.\n");
578 }
579 
580 void prom_world(int enter)
581 {
582 	if (!enter)
583 		set_fs((mm_segment_t) { get_thread_current_ds() });
584 
585 	__asm__ __volatile__("flushw");
586 }
587 
588 void __flush_dcache_range(unsigned long start, unsigned long end)
589 {
590 	unsigned long va;
591 
592 	if (tlb_type == spitfire) {
593 		int n = 0;
594 
595 		for (va = start; va < end; va += 32) {
596 			spitfire_put_dcache_tag(va & 0x3fe0, 0x0);
597 			if (++n >= 512)
598 				break;
599 		}
600 	} else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
601 		start = __pa(start);
602 		end = __pa(end);
603 		for (va = start; va < end; va += 32)
604 			__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
605 					     "membar #Sync"
606 					     : /* no outputs */
607 					     : "r" (va),
608 					       "i" (ASI_DCACHE_INVALIDATE));
609 	}
610 }
611 EXPORT_SYMBOL(__flush_dcache_range);
612 
613 /* get_new_mmu_context() uses "cache + 1".  */
614 DEFINE_SPINLOCK(ctx_alloc_lock);
615 unsigned long tlb_context_cache = CTX_FIRST_VERSION - 1;
616 #define MAX_CTX_NR	(1UL << CTX_NR_BITS)
617 #define CTX_BMAP_SLOTS	BITS_TO_LONGS(MAX_CTX_NR)
618 DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR);
619 
620 /* Caller does TLB context flushing on local CPU if necessary.
621  * The caller also ensures that CTX_VALID(mm->context) is false.
622  *
623  * We must be careful about boundary cases so that we never
624  * let the user have CTX 0 (nucleus) or we ever use a CTX
625  * version of zero (and thus NO_CONTEXT would not be caught
626  * by version mis-match tests in mmu_context.h).
627  *
628  * Always invoked with interrupts disabled.
629  */
630 void get_new_mmu_context(struct mm_struct *mm)
631 {
632 	unsigned long ctx, new_ctx;
633 	unsigned long orig_pgsz_bits;
634 	unsigned long flags;
635 	int new_version;
636 
637 	spin_lock_irqsave(&ctx_alloc_lock, flags);
638 	orig_pgsz_bits = (mm->context.sparc64_ctx_val & CTX_PGSZ_MASK);
639 	ctx = (tlb_context_cache + 1) & CTX_NR_MASK;
640 	new_ctx = find_next_zero_bit(mmu_context_bmap, 1 << CTX_NR_BITS, ctx);
641 	new_version = 0;
642 	if (new_ctx >= (1 << CTX_NR_BITS)) {
643 		new_ctx = find_next_zero_bit(mmu_context_bmap, ctx, 1);
644 		if (new_ctx >= ctx) {
645 			int i;
646 			new_ctx = (tlb_context_cache & CTX_VERSION_MASK) +
647 				CTX_FIRST_VERSION;
648 			if (new_ctx == 1)
649 				new_ctx = CTX_FIRST_VERSION;
650 
651 			/* Don't call memset, for 16 entries that's just
652 			 * plain silly...
653 			 */
654 			mmu_context_bmap[0] = 3;
655 			mmu_context_bmap[1] = 0;
656 			mmu_context_bmap[2] = 0;
657 			mmu_context_bmap[3] = 0;
658 			for (i = 4; i < CTX_BMAP_SLOTS; i += 4) {
659 				mmu_context_bmap[i + 0] = 0;
660 				mmu_context_bmap[i + 1] = 0;
661 				mmu_context_bmap[i + 2] = 0;
662 				mmu_context_bmap[i + 3] = 0;
663 			}
664 			new_version = 1;
665 			goto out;
666 		}
667 	}
668 	mmu_context_bmap[new_ctx>>6] |= (1UL << (new_ctx & 63));
669 	new_ctx |= (tlb_context_cache & CTX_VERSION_MASK);
670 out:
671 	tlb_context_cache = new_ctx;
672 	mm->context.sparc64_ctx_val = new_ctx | orig_pgsz_bits;
673 	spin_unlock_irqrestore(&ctx_alloc_lock, flags);
674 
675 	if (unlikely(new_version))
676 		smp_new_mmu_context_version();
677 }
678 
679 static int numa_enabled = 1;
680 static int numa_debug;
681 
682 static int __init early_numa(char *p)
683 {
684 	if (!p)
685 		return 0;
686 
687 	if (strstr(p, "off"))
688 		numa_enabled = 0;
689 
690 	if (strstr(p, "debug"))
691 		numa_debug = 1;
692 
693 	return 0;
694 }
695 early_param("numa", early_numa);
696 
697 #define numadbg(f, a...) \
698 do {	if (numa_debug) \
699 		printk(KERN_INFO f, ## a); \
700 } while (0)
701 
702 static void __init find_ramdisk(unsigned long phys_base)
703 {
704 #ifdef CONFIG_BLK_DEV_INITRD
705 	if (sparc_ramdisk_image || sparc_ramdisk_image64) {
706 		unsigned long ramdisk_image;
707 
708 		/* Older versions of the bootloader only supported a
709 		 * 32-bit physical address for the ramdisk image
710 		 * location, stored at sparc_ramdisk_image.  Newer
711 		 * SILO versions set sparc_ramdisk_image to zero and
712 		 * provide a full 64-bit physical address at
713 		 * sparc_ramdisk_image64.
714 		 */
715 		ramdisk_image = sparc_ramdisk_image;
716 		if (!ramdisk_image)
717 			ramdisk_image = sparc_ramdisk_image64;
718 
719 		/* Another bootloader quirk.  The bootloader normalizes
720 		 * the physical address to KERNBASE, so we have to
721 		 * factor that back out and add in the lowest valid
722 		 * physical page address to get the true physical address.
723 		 */
724 		ramdisk_image -= KERNBASE;
725 		ramdisk_image += phys_base;
726 
727 		numadbg("Found ramdisk at physical address 0x%lx, size %u\n",
728 			ramdisk_image, sparc_ramdisk_size);
729 
730 		initrd_start = ramdisk_image;
731 		initrd_end = ramdisk_image + sparc_ramdisk_size;
732 
733 		memblock_reserve(initrd_start, sparc_ramdisk_size);
734 
735 		initrd_start += PAGE_OFFSET;
736 		initrd_end += PAGE_OFFSET;
737 	}
738 #endif
739 }
740 
741 struct node_mem_mask {
742 	unsigned long mask;
743 	unsigned long val;
744 };
745 static struct node_mem_mask node_masks[MAX_NUMNODES];
746 static int num_node_masks;
747 
748 int numa_cpu_lookup_table[NR_CPUS];
749 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
750 
751 #ifdef CONFIG_NEED_MULTIPLE_NODES
752 
753 struct mdesc_mblock {
754 	u64	base;
755 	u64	size;
756 	u64	offset; /* RA-to-PA */
757 };
758 static struct mdesc_mblock *mblocks;
759 static int num_mblocks;
760 
761 static unsigned long ra_to_pa(unsigned long addr)
762 {
763 	int i;
764 
765 	for (i = 0; i < num_mblocks; i++) {
766 		struct mdesc_mblock *m = &mblocks[i];
767 
768 		if (addr >= m->base &&
769 		    addr < (m->base + m->size)) {
770 			addr += m->offset;
771 			break;
772 		}
773 	}
774 	return addr;
775 }
776 
777 static int find_node(unsigned long addr)
778 {
779 	int i;
780 
781 	addr = ra_to_pa(addr);
782 	for (i = 0; i < num_node_masks; i++) {
783 		struct node_mem_mask *p = &node_masks[i];
784 
785 		if ((addr & p->mask) == p->val)
786 			return i;
787 	}
788 	return -1;
789 }
790 
791 static u64 memblock_nid_range(u64 start, u64 end, int *nid)
792 {
793 	*nid = find_node(start);
794 	start += PAGE_SIZE;
795 	while (start < end) {
796 		int n = find_node(start);
797 
798 		if (n != *nid)
799 			break;
800 		start += PAGE_SIZE;
801 	}
802 
803 	if (start > end)
804 		start = end;
805 
806 	return start;
807 }
808 #endif
809 
810 /* This must be invoked after performing all of the necessary
811  * memblock_set_node() calls for 'nid'.  We need to be able to get
812  * correct data from get_pfn_range_for_nid().
813  */
814 static void __init allocate_node_data(int nid)
815 {
816 	struct pglist_data *p;
817 	unsigned long start_pfn, end_pfn;
818 #ifdef CONFIG_NEED_MULTIPLE_NODES
819 	unsigned long paddr;
820 
821 	paddr = memblock_alloc_try_nid(sizeof(struct pglist_data), SMP_CACHE_BYTES, nid);
822 	if (!paddr) {
823 		prom_printf("Cannot allocate pglist_data for nid[%d]\n", nid);
824 		prom_halt();
825 	}
826 	NODE_DATA(nid) = __va(paddr);
827 	memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
828 
829 	NODE_DATA(nid)->node_id = nid;
830 #endif
831 
832 	p = NODE_DATA(nid);
833 
834 	get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
835 	p->node_start_pfn = start_pfn;
836 	p->node_spanned_pages = end_pfn - start_pfn;
837 }
838 
839 static void init_node_masks_nonnuma(void)
840 {
841 	int i;
842 
843 	numadbg("Initializing tables for non-numa.\n");
844 
845 	node_masks[0].mask = node_masks[0].val = 0;
846 	num_node_masks = 1;
847 
848 	for (i = 0; i < NR_CPUS; i++)
849 		numa_cpu_lookup_table[i] = 0;
850 
851 	cpumask_setall(&numa_cpumask_lookup_table[0]);
852 }
853 
854 #ifdef CONFIG_NEED_MULTIPLE_NODES
855 struct pglist_data *node_data[MAX_NUMNODES];
856 
857 EXPORT_SYMBOL(numa_cpu_lookup_table);
858 EXPORT_SYMBOL(numa_cpumask_lookup_table);
859 EXPORT_SYMBOL(node_data);
860 
861 struct mdesc_mlgroup {
862 	u64	node;
863 	u64	latency;
864 	u64	match;
865 	u64	mask;
866 };
867 static struct mdesc_mlgroup *mlgroups;
868 static int num_mlgroups;
869 
870 static int scan_pio_for_cfg_handle(struct mdesc_handle *md, u64 pio,
871 				   u32 cfg_handle)
872 {
873 	u64 arc;
874 
875 	mdesc_for_each_arc(arc, md, pio, MDESC_ARC_TYPE_FWD) {
876 		u64 target = mdesc_arc_target(md, arc);
877 		const u64 *val;
878 
879 		val = mdesc_get_property(md, target,
880 					 "cfg-handle", NULL);
881 		if (val && *val == cfg_handle)
882 			return 0;
883 	}
884 	return -ENODEV;
885 }
886 
887 static int scan_arcs_for_cfg_handle(struct mdesc_handle *md, u64 grp,
888 				    u32 cfg_handle)
889 {
890 	u64 arc, candidate, best_latency = ~(u64)0;
891 
892 	candidate = MDESC_NODE_NULL;
893 	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
894 		u64 target = mdesc_arc_target(md, arc);
895 		const char *name = mdesc_node_name(md, target);
896 		const u64 *val;
897 
898 		if (strcmp(name, "pio-latency-group"))
899 			continue;
900 
901 		val = mdesc_get_property(md, target, "latency", NULL);
902 		if (!val)
903 			continue;
904 
905 		if (*val < best_latency) {
906 			candidate = target;
907 			best_latency = *val;
908 		}
909 	}
910 
911 	if (candidate == MDESC_NODE_NULL)
912 		return -ENODEV;
913 
914 	return scan_pio_for_cfg_handle(md, candidate, cfg_handle);
915 }
916 
917 int of_node_to_nid(struct device_node *dp)
918 {
919 	const struct linux_prom64_registers *regs;
920 	struct mdesc_handle *md;
921 	u32 cfg_handle;
922 	int count, nid;
923 	u64 grp;
924 
925 	/* This is the right thing to do on currently supported
926 	 * SUN4U NUMA platforms as well, as the PCI controller does
927 	 * not sit behind any particular memory controller.
928 	 */
929 	if (!mlgroups)
930 		return -1;
931 
932 	regs = of_get_property(dp, "reg", NULL);
933 	if (!regs)
934 		return -1;
935 
936 	cfg_handle = (regs->phys_addr >> 32UL) & 0x0fffffff;
937 
938 	md = mdesc_grab();
939 
940 	count = 0;
941 	nid = -1;
942 	mdesc_for_each_node_by_name(md, grp, "group") {
943 		if (!scan_arcs_for_cfg_handle(md, grp, cfg_handle)) {
944 			nid = count;
945 			break;
946 		}
947 		count++;
948 	}
949 
950 	mdesc_release(md);
951 
952 	return nid;
953 }
954 
955 static void __init add_node_ranges(void)
956 {
957 	struct memblock_region *reg;
958 
959 	for_each_memblock(memory, reg) {
960 		unsigned long size = reg->size;
961 		unsigned long start, end;
962 
963 		start = reg->base;
964 		end = start + size;
965 		while (start < end) {
966 			unsigned long this_end;
967 			int nid;
968 
969 			this_end = memblock_nid_range(start, end, &nid);
970 
971 			numadbg("Setting memblock NUMA node nid[%d] "
972 				"start[%lx] end[%lx]\n",
973 				nid, start, this_end);
974 
975 			memblock_set_node(start, this_end - start, nid);
976 			start = this_end;
977 		}
978 	}
979 }
980 
981 static int __init grab_mlgroups(struct mdesc_handle *md)
982 {
983 	unsigned long paddr;
984 	int count = 0;
985 	u64 node;
986 
987 	mdesc_for_each_node_by_name(md, node, "memory-latency-group")
988 		count++;
989 	if (!count)
990 		return -ENOENT;
991 
992 	paddr = memblock_alloc(count * sizeof(struct mdesc_mlgroup),
993 			  SMP_CACHE_BYTES);
994 	if (!paddr)
995 		return -ENOMEM;
996 
997 	mlgroups = __va(paddr);
998 	num_mlgroups = count;
999 
1000 	count = 0;
1001 	mdesc_for_each_node_by_name(md, node, "memory-latency-group") {
1002 		struct mdesc_mlgroup *m = &mlgroups[count++];
1003 		const u64 *val;
1004 
1005 		m->node = node;
1006 
1007 		val = mdesc_get_property(md, node, "latency", NULL);
1008 		m->latency = *val;
1009 		val = mdesc_get_property(md, node, "address-match", NULL);
1010 		m->match = *val;
1011 		val = mdesc_get_property(md, node, "address-mask", NULL);
1012 		m->mask = *val;
1013 
1014 		numadbg("MLGROUP[%d]: node[%llx] latency[%llx] "
1015 			"match[%llx] mask[%llx]\n",
1016 			count - 1, m->node, m->latency, m->match, m->mask);
1017 	}
1018 
1019 	return 0;
1020 }
1021 
1022 static int __init grab_mblocks(struct mdesc_handle *md)
1023 {
1024 	unsigned long paddr;
1025 	int count = 0;
1026 	u64 node;
1027 
1028 	mdesc_for_each_node_by_name(md, node, "mblock")
1029 		count++;
1030 	if (!count)
1031 		return -ENOENT;
1032 
1033 	paddr = memblock_alloc(count * sizeof(struct mdesc_mblock),
1034 			  SMP_CACHE_BYTES);
1035 	if (!paddr)
1036 		return -ENOMEM;
1037 
1038 	mblocks = __va(paddr);
1039 	num_mblocks = count;
1040 
1041 	count = 0;
1042 	mdesc_for_each_node_by_name(md, node, "mblock") {
1043 		struct mdesc_mblock *m = &mblocks[count++];
1044 		const u64 *val;
1045 
1046 		val = mdesc_get_property(md, node, "base", NULL);
1047 		m->base = *val;
1048 		val = mdesc_get_property(md, node, "size", NULL);
1049 		m->size = *val;
1050 		val = mdesc_get_property(md, node,
1051 					 "address-congruence-offset", NULL);
1052 		m->offset = *val;
1053 
1054 		numadbg("MBLOCK[%d]: base[%llx] size[%llx] offset[%llx]\n",
1055 			count - 1, m->base, m->size, m->offset);
1056 	}
1057 
1058 	return 0;
1059 }
1060 
1061 static void __init numa_parse_mdesc_group_cpus(struct mdesc_handle *md,
1062 					       u64 grp, cpumask_t *mask)
1063 {
1064 	u64 arc;
1065 
1066 	cpumask_clear(mask);
1067 
1068 	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_BACK) {
1069 		u64 target = mdesc_arc_target(md, arc);
1070 		const char *name = mdesc_node_name(md, target);
1071 		const u64 *id;
1072 
1073 		if (strcmp(name, "cpu"))
1074 			continue;
1075 		id = mdesc_get_property(md, target, "id", NULL);
1076 		if (*id < nr_cpu_ids)
1077 			cpumask_set_cpu(*id, mask);
1078 	}
1079 }
1080 
1081 static struct mdesc_mlgroup * __init find_mlgroup(u64 node)
1082 {
1083 	int i;
1084 
1085 	for (i = 0; i < num_mlgroups; i++) {
1086 		struct mdesc_mlgroup *m = &mlgroups[i];
1087 		if (m->node == node)
1088 			return m;
1089 	}
1090 	return NULL;
1091 }
1092 
1093 static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp,
1094 				      int index)
1095 {
1096 	struct mdesc_mlgroup *candidate = NULL;
1097 	u64 arc, best_latency = ~(u64)0;
1098 	struct node_mem_mask *n;
1099 
1100 	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1101 		u64 target = mdesc_arc_target(md, arc);
1102 		struct mdesc_mlgroup *m = find_mlgroup(target);
1103 		if (!m)
1104 			continue;
1105 		if (m->latency < best_latency) {
1106 			candidate = m;
1107 			best_latency = m->latency;
1108 		}
1109 	}
1110 	if (!candidate)
1111 		return -ENOENT;
1112 
1113 	if (num_node_masks != index) {
1114 		printk(KERN_ERR "Inconsistent NUMA state, "
1115 		       "index[%d] != num_node_masks[%d]\n",
1116 		       index, num_node_masks);
1117 		return -EINVAL;
1118 	}
1119 
1120 	n = &node_masks[num_node_masks++];
1121 
1122 	n->mask = candidate->mask;
1123 	n->val = candidate->match;
1124 
1125 	numadbg("NUMA NODE[%d]: mask[%lx] val[%lx] (latency[%llx])\n",
1126 		index, n->mask, n->val, candidate->latency);
1127 
1128 	return 0;
1129 }
1130 
1131 static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp,
1132 					 int index)
1133 {
1134 	cpumask_t mask;
1135 	int cpu;
1136 
1137 	numa_parse_mdesc_group_cpus(md, grp, &mask);
1138 
1139 	for_each_cpu(cpu, &mask)
1140 		numa_cpu_lookup_table[cpu] = index;
1141 	cpumask_copy(&numa_cpumask_lookup_table[index], &mask);
1142 
1143 	if (numa_debug) {
1144 		printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index);
1145 		for_each_cpu(cpu, &mask)
1146 			printk("%d ", cpu);
1147 		printk("]\n");
1148 	}
1149 
1150 	return numa_attach_mlgroup(md, grp, index);
1151 }
1152 
1153 static int __init numa_parse_mdesc(void)
1154 {
1155 	struct mdesc_handle *md = mdesc_grab();
1156 	int i, err, count;
1157 	u64 node;
1158 
1159 	node = mdesc_node_by_name(md, MDESC_NODE_NULL, "latency-groups");
1160 	if (node == MDESC_NODE_NULL) {
1161 		mdesc_release(md);
1162 		return -ENOENT;
1163 	}
1164 
1165 	err = grab_mblocks(md);
1166 	if (err < 0)
1167 		goto out;
1168 
1169 	err = grab_mlgroups(md);
1170 	if (err < 0)
1171 		goto out;
1172 
1173 	count = 0;
1174 	mdesc_for_each_node_by_name(md, node, "group") {
1175 		err = numa_parse_mdesc_group(md, node, count);
1176 		if (err < 0)
1177 			break;
1178 		count++;
1179 	}
1180 
1181 	add_node_ranges();
1182 
1183 	for (i = 0; i < num_node_masks; i++) {
1184 		allocate_node_data(i);
1185 		node_set_online(i);
1186 	}
1187 
1188 	err = 0;
1189 out:
1190 	mdesc_release(md);
1191 	return err;
1192 }
1193 
1194 static int __init numa_parse_jbus(void)
1195 {
1196 	unsigned long cpu, index;
1197 
1198 	/* NUMA node id is encoded in bits 36 and higher, and there is
1199 	 * a 1-to-1 mapping from CPU ID to NUMA node ID.
1200 	 */
1201 	index = 0;
1202 	for_each_present_cpu(cpu) {
1203 		numa_cpu_lookup_table[cpu] = index;
1204 		cpumask_copy(&numa_cpumask_lookup_table[index], cpumask_of(cpu));
1205 		node_masks[index].mask = ~((1UL << 36UL) - 1UL);
1206 		node_masks[index].val = cpu << 36UL;
1207 
1208 		index++;
1209 	}
1210 	num_node_masks = index;
1211 
1212 	add_node_ranges();
1213 
1214 	for (index = 0; index < num_node_masks; index++) {
1215 		allocate_node_data(index);
1216 		node_set_online(index);
1217 	}
1218 
1219 	return 0;
1220 }
1221 
1222 static int __init numa_parse_sun4u(void)
1223 {
1224 	if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1225 		unsigned long ver;
1226 
1227 		__asm__ ("rdpr %%ver, %0" : "=r" (ver));
1228 		if ((ver >> 32UL) == __JALAPENO_ID ||
1229 		    (ver >> 32UL) == __SERRANO_ID)
1230 			return numa_parse_jbus();
1231 	}
1232 	return -1;
1233 }
1234 
1235 static int __init bootmem_init_numa(void)
1236 {
1237 	int err = -1;
1238 
1239 	numadbg("bootmem_init_numa()\n");
1240 
1241 	if (numa_enabled) {
1242 		if (tlb_type == hypervisor)
1243 			err = numa_parse_mdesc();
1244 		else
1245 			err = numa_parse_sun4u();
1246 	}
1247 	return err;
1248 }
1249 
1250 #else
1251 
1252 static int bootmem_init_numa(void)
1253 {
1254 	return -1;
1255 }
1256 
1257 #endif
1258 
1259 static void __init bootmem_init_nonnuma(void)
1260 {
1261 	unsigned long top_of_ram = memblock_end_of_DRAM();
1262 	unsigned long total_ram = memblock_phys_mem_size();
1263 
1264 	numadbg("bootmem_init_nonnuma()\n");
1265 
1266 	printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
1267 	       top_of_ram, total_ram);
1268 	printk(KERN_INFO "Memory hole size: %ldMB\n",
1269 	       (top_of_ram - total_ram) >> 20);
1270 
1271 	init_node_masks_nonnuma();
1272 	memblock_set_node(0, (phys_addr_t)ULLONG_MAX, 0);
1273 	allocate_node_data(0);
1274 	node_set_online(0);
1275 }
1276 
1277 static unsigned long __init bootmem_init(unsigned long phys_base)
1278 {
1279 	unsigned long end_pfn;
1280 
1281 	end_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1282 	max_pfn = max_low_pfn = end_pfn;
1283 	min_low_pfn = (phys_base >> PAGE_SHIFT);
1284 
1285 	if (bootmem_init_numa() < 0)
1286 		bootmem_init_nonnuma();
1287 
1288 	/* Dump memblock with node info. */
1289 	memblock_dump_all();
1290 
1291 	/* XXX cpu notifier XXX */
1292 
1293 	sparse_memory_present_with_active_regions(MAX_NUMNODES);
1294 	sparse_init();
1295 
1296 	return end_pfn;
1297 }
1298 
1299 static struct linux_prom64_registers pall[MAX_BANKS] __initdata;
1300 static int pall_ents __initdata;
1301 
1302 #ifdef CONFIG_DEBUG_PAGEALLOC
1303 static unsigned long __ref kernel_map_range(unsigned long pstart,
1304 					    unsigned long pend, pgprot_t prot)
1305 {
1306 	unsigned long vstart = PAGE_OFFSET + pstart;
1307 	unsigned long vend = PAGE_OFFSET + pend;
1308 	unsigned long alloc_bytes = 0UL;
1309 
1310 	if ((vstart & ~PAGE_MASK) || (vend & ~PAGE_MASK)) {
1311 		prom_printf("kernel_map: Unaligned physmem[%lx:%lx]\n",
1312 			    vstart, vend);
1313 		prom_halt();
1314 	}
1315 
1316 	while (vstart < vend) {
1317 		unsigned long this_end, paddr = __pa(vstart);
1318 		pgd_t *pgd = pgd_offset_k(vstart);
1319 		pud_t *pud;
1320 		pmd_t *pmd;
1321 		pte_t *pte;
1322 
1323 		pud = pud_offset(pgd, vstart);
1324 		if (pud_none(*pud)) {
1325 			pmd_t *new;
1326 
1327 			new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1328 			alloc_bytes += PAGE_SIZE;
1329 			pud_populate(&init_mm, pud, new);
1330 		}
1331 
1332 		pmd = pmd_offset(pud, vstart);
1333 		if (!pmd_present(*pmd)) {
1334 			pte_t *new;
1335 
1336 			new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1337 			alloc_bytes += PAGE_SIZE;
1338 			pmd_populate_kernel(&init_mm, pmd, new);
1339 		}
1340 
1341 		pte = pte_offset_kernel(pmd, vstart);
1342 		this_end = (vstart + PMD_SIZE) & PMD_MASK;
1343 		if (this_end > vend)
1344 			this_end = vend;
1345 
1346 		while (vstart < this_end) {
1347 			pte_val(*pte) = (paddr | pgprot_val(prot));
1348 
1349 			vstart += PAGE_SIZE;
1350 			paddr += PAGE_SIZE;
1351 			pte++;
1352 		}
1353 	}
1354 
1355 	return alloc_bytes;
1356 }
1357 
1358 extern unsigned int kvmap_linear_patch[1];
1359 #endif /* CONFIG_DEBUG_PAGEALLOC */
1360 
1361 static void __init mark_kpte_bitmap(unsigned long start, unsigned long end)
1362 {
1363 	const unsigned long shift_256MB = 28;
1364 	const unsigned long mask_256MB = ((1UL << shift_256MB) - 1UL);
1365 	const unsigned long size_256MB = (1UL << shift_256MB);
1366 
1367 	while (start < end) {
1368 		long remains;
1369 
1370 		remains = end - start;
1371 		if (remains < size_256MB)
1372 			break;
1373 
1374 		if (start & mask_256MB) {
1375 			start = (start + size_256MB) & ~mask_256MB;
1376 			continue;
1377 		}
1378 
1379 		while (remains >= size_256MB) {
1380 			unsigned long index = start >> shift_256MB;
1381 
1382 			__set_bit(index, kpte_linear_bitmap);
1383 
1384 			start += size_256MB;
1385 			remains -= size_256MB;
1386 		}
1387 	}
1388 }
1389 
1390 static void __init init_kpte_bitmap(void)
1391 {
1392 	unsigned long i;
1393 
1394 	for (i = 0; i < pall_ents; i++) {
1395 		unsigned long phys_start, phys_end;
1396 
1397 		phys_start = pall[i].phys_addr;
1398 		phys_end = phys_start + pall[i].reg_size;
1399 
1400 		mark_kpte_bitmap(phys_start, phys_end);
1401 	}
1402 }
1403 
1404 static void __init kernel_physical_mapping_init(void)
1405 {
1406 #ifdef CONFIG_DEBUG_PAGEALLOC
1407 	unsigned long i, mem_alloced = 0UL;
1408 
1409 	for (i = 0; i < pall_ents; i++) {
1410 		unsigned long phys_start, phys_end;
1411 
1412 		phys_start = pall[i].phys_addr;
1413 		phys_end = phys_start + pall[i].reg_size;
1414 
1415 		mem_alloced += kernel_map_range(phys_start, phys_end,
1416 						PAGE_KERNEL);
1417 	}
1418 
1419 	printk("Allocated %ld bytes for kernel page tables.\n",
1420 	       mem_alloced);
1421 
1422 	kvmap_linear_patch[0] = 0x01000000; /* nop */
1423 	flushi(&kvmap_linear_patch[0]);
1424 
1425 	__flush_tlb_all();
1426 #endif
1427 }
1428 
1429 #ifdef CONFIG_DEBUG_PAGEALLOC
1430 void kernel_map_pages(struct page *page, int numpages, int enable)
1431 {
1432 	unsigned long phys_start = page_to_pfn(page) << PAGE_SHIFT;
1433 	unsigned long phys_end = phys_start + (numpages * PAGE_SIZE);
1434 
1435 	kernel_map_range(phys_start, phys_end,
1436 			 (enable ? PAGE_KERNEL : __pgprot(0)));
1437 
1438 	flush_tsb_kernel_range(PAGE_OFFSET + phys_start,
1439 			       PAGE_OFFSET + phys_end);
1440 
1441 	/* we should perform an IPI and flush all tlbs,
1442 	 * but that can deadlock->flush only current cpu.
1443 	 */
1444 	__flush_tlb_kernel_range(PAGE_OFFSET + phys_start,
1445 				 PAGE_OFFSET + phys_end);
1446 }
1447 #endif
1448 
1449 unsigned long __init find_ecache_flush_span(unsigned long size)
1450 {
1451 	int i;
1452 
1453 	for (i = 0; i < pavail_ents; i++) {
1454 		if (pavail[i].reg_size >= size)
1455 			return pavail[i].phys_addr;
1456 	}
1457 
1458 	return ~0UL;
1459 }
1460 
1461 static void __init tsb_phys_patch(void)
1462 {
1463 	struct tsb_ldquad_phys_patch_entry *pquad;
1464 	struct tsb_phys_patch_entry *p;
1465 
1466 	pquad = &__tsb_ldquad_phys_patch;
1467 	while (pquad < &__tsb_ldquad_phys_patch_end) {
1468 		unsigned long addr = pquad->addr;
1469 
1470 		if (tlb_type == hypervisor)
1471 			*(unsigned int *) addr = pquad->sun4v_insn;
1472 		else
1473 			*(unsigned int *) addr = pquad->sun4u_insn;
1474 		wmb();
1475 		__asm__ __volatile__("flush	%0"
1476 				     : /* no outputs */
1477 				     : "r" (addr));
1478 
1479 		pquad++;
1480 	}
1481 
1482 	p = &__tsb_phys_patch;
1483 	while (p < &__tsb_phys_patch_end) {
1484 		unsigned long addr = p->addr;
1485 
1486 		*(unsigned int *) addr = p->insn;
1487 		wmb();
1488 		__asm__ __volatile__("flush	%0"
1489 				     : /* no outputs */
1490 				     : "r" (addr));
1491 
1492 		p++;
1493 	}
1494 }
1495 
1496 /* Don't mark as init, we give this to the Hypervisor.  */
1497 #ifndef CONFIG_DEBUG_PAGEALLOC
1498 #define NUM_KTSB_DESCR	2
1499 #else
1500 #define NUM_KTSB_DESCR	1
1501 #endif
1502 static struct hv_tsb_descr ktsb_descr[NUM_KTSB_DESCR];
1503 extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
1504 
1505 static void patch_one_ktsb_phys(unsigned int *start, unsigned int *end, unsigned long pa)
1506 {
1507 	pa >>= KTSB_PHYS_SHIFT;
1508 
1509 	while (start < end) {
1510 		unsigned int *ia = (unsigned int *)(unsigned long)*start;
1511 
1512 		ia[0] = (ia[0] & ~0x3fffff) | (pa >> 10);
1513 		__asm__ __volatile__("flush	%0" : : "r" (ia));
1514 
1515 		ia[1] = (ia[1] & ~0x3ff) | (pa & 0x3ff);
1516 		__asm__ __volatile__("flush	%0" : : "r" (ia + 1));
1517 
1518 		start++;
1519 	}
1520 }
1521 
1522 static void ktsb_phys_patch(void)
1523 {
1524 	extern unsigned int __swapper_tsb_phys_patch;
1525 	extern unsigned int __swapper_tsb_phys_patch_end;
1526 	unsigned long ktsb_pa;
1527 
1528 	ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
1529 	patch_one_ktsb_phys(&__swapper_tsb_phys_patch,
1530 			    &__swapper_tsb_phys_patch_end, ktsb_pa);
1531 #ifndef CONFIG_DEBUG_PAGEALLOC
1532 	{
1533 	extern unsigned int __swapper_4m_tsb_phys_patch;
1534 	extern unsigned int __swapper_4m_tsb_phys_patch_end;
1535 	ktsb_pa = (kern_base +
1536 		   ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
1537 	patch_one_ktsb_phys(&__swapper_4m_tsb_phys_patch,
1538 			    &__swapper_4m_tsb_phys_patch_end, ktsb_pa);
1539 	}
1540 #endif
1541 }
1542 
1543 static void __init sun4v_ktsb_init(void)
1544 {
1545 	unsigned long ktsb_pa;
1546 
1547 	/* First KTSB for PAGE_SIZE mappings.  */
1548 	ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
1549 
1550 	switch (PAGE_SIZE) {
1551 	case 8 * 1024:
1552 	default:
1553 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_8K;
1554 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_8K;
1555 		break;
1556 
1557 	case 64 * 1024:
1558 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_64K;
1559 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_64K;
1560 		break;
1561 
1562 	case 512 * 1024:
1563 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_512K;
1564 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_512K;
1565 		break;
1566 
1567 	case 4 * 1024 * 1024:
1568 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_4MB;
1569 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_4MB;
1570 		break;
1571 	}
1572 
1573 	ktsb_descr[0].assoc = 1;
1574 	ktsb_descr[0].num_ttes = KERNEL_TSB_NENTRIES;
1575 	ktsb_descr[0].ctx_idx = 0;
1576 	ktsb_descr[0].tsb_base = ktsb_pa;
1577 	ktsb_descr[0].resv = 0;
1578 
1579 #ifndef CONFIG_DEBUG_PAGEALLOC
1580 	/* Second KTSB for 4MB/256MB mappings.  */
1581 	ktsb_pa = (kern_base +
1582 		   ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
1583 
1584 	ktsb_descr[1].pgsz_idx = HV_PGSZ_IDX_4MB;
1585 	ktsb_descr[1].pgsz_mask = (HV_PGSZ_MASK_4MB |
1586 				   HV_PGSZ_MASK_256MB);
1587 	ktsb_descr[1].assoc = 1;
1588 	ktsb_descr[1].num_ttes = KERNEL_TSB4M_NENTRIES;
1589 	ktsb_descr[1].ctx_idx = 0;
1590 	ktsb_descr[1].tsb_base = ktsb_pa;
1591 	ktsb_descr[1].resv = 0;
1592 #endif
1593 }
1594 
1595 void __cpuinit sun4v_ktsb_register(void)
1596 {
1597 	unsigned long pa, ret;
1598 
1599 	pa = kern_base + ((unsigned long)&ktsb_descr[0] - KERNBASE);
1600 
1601 	ret = sun4v_mmu_tsb_ctx0(NUM_KTSB_DESCR, pa);
1602 	if (ret != 0) {
1603 		prom_printf("hypervisor_mmu_tsb_ctx0[%lx]: "
1604 			    "errors with %lx\n", pa, ret);
1605 		prom_halt();
1606 	}
1607 }
1608 
1609 /* paging_init() sets up the page tables */
1610 
1611 static unsigned long last_valid_pfn;
1612 pgd_t swapper_pg_dir[2048];
1613 
1614 static void sun4u_pgprot_init(void);
1615 static void sun4v_pgprot_init(void);
1616 
1617 void __init paging_init(void)
1618 {
1619 	unsigned long end_pfn, shift, phys_base;
1620 	unsigned long real_end, i;
1621 	int node;
1622 
1623 	/* These build time checkes make sure that the dcache_dirty_cpu()
1624 	 * page->flags usage will work.
1625 	 *
1626 	 * When a page gets marked as dcache-dirty, we store the
1627 	 * cpu number starting at bit 32 in the page->flags.  Also,
1628 	 * functions like clear_dcache_dirty_cpu use the cpu mask
1629 	 * in 13-bit signed-immediate instruction fields.
1630 	 */
1631 
1632 	/*
1633 	 * Page flags must not reach into upper 32 bits that are used
1634 	 * for the cpu number
1635 	 */
1636 	BUILD_BUG_ON(NR_PAGEFLAGS > 32);
1637 
1638 	/*
1639 	 * The bit fields placed in the high range must not reach below
1640 	 * the 32 bit boundary. Otherwise we cannot place the cpu field
1641 	 * at the 32 bit boundary.
1642 	 */
1643 	BUILD_BUG_ON(SECTIONS_WIDTH + NODES_WIDTH + ZONES_WIDTH +
1644 		ilog2(roundup_pow_of_two(NR_CPUS)) > 32);
1645 
1646 	BUILD_BUG_ON(NR_CPUS > 4096);
1647 
1648 	kern_base = (prom_boot_mapping_phys_low >> 22UL) << 22UL;
1649 	kern_size = (unsigned long)&_end - (unsigned long)KERNBASE;
1650 
1651 	/* Invalidate both kernel TSBs.  */
1652 	memset(swapper_tsb, 0x40, sizeof(swapper_tsb));
1653 #ifndef CONFIG_DEBUG_PAGEALLOC
1654 	memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
1655 #endif
1656 
1657 	if (tlb_type == hypervisor)
1658 		sun4v_pgprot_init();
1659 	else
1660 		sun4u_pgprot_init();
1661 
1662 	if (tlb_type == cheetah_plus ||
1663 	    tlb_type == hypervisor) {
1664 		tsb_phys_patch();
1665 		ktsb_phys_patch();
1666 	}
1667 
1668 	if (tlb_type == hypervisor) {
1669 		sun4v_patch_tlb_handlers();
1670 		sun4v_ktsb_init();
1671 	}
1672 
1673 	/* Find available physical memory...
1674 	 *
1675 	 * Read it twice in order to work around a bug in openfirmware.
1676 	 * The call to grab this table itself can cause openfirmware to
1677 	 * allocate memory, which in turn can take away some space from
1678 	 * the list of available memory.  Reading it twice makes sure
1679 	 * we really do get the final value.
1680 	 */
1681 	read_obp_translations();
1682 	read_obp_memory("reg", &pall[0], &pall_ents);
1683 	read_obp_memory("available", &pavail[0], &pavail_ents);
1684 	read_obp_memory("available", &pavail[0], &pavail_ents);
1685 
1686 	phys_base = 0xffffffffffffffffUL;
1687 	for (i = 0; i < pavail_ents; i++) {
1688 		phys_base = min(phys_base, pavail[i].phys_addr);
1689 		memblock_add(pavail[i].phys_addr, pavail[i].reg_size);
1690 	}
1691 
1692 	memblock_reserve(kern_base, kern_size);
1693 
1694 	find_ramdisk(phys_base);
1695 
1696 	memblock_enforce_memory_limit(cmdline_memory_size);
1697 
1698 	memblock_allow_resize();
1699 	memblock_dump_all();
1700 
1701 	set_bit(0, mmu_context_bmap);
1702 
1703 	shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE);
1704 
1705 	real_end = (unsigned long)_end;
1706 	num_kernel_image_mappings = DIV_ROUND_UP(real_end - KERNBASE, 1 << 22);
1707 	printk("Kernel: Using %d locked TLB entries for main kernel image.\n",
1708 	       num_kernel_image_mappings);
1709 
1710 	/* Set kernel pgd to upper alias so physical page computations
1711 	 * work.
1712 	 */
1713 	init_mm.pgd += ((shift) / (sizeof(pgd_t)));
1714 
1715 	memset(swapper_low_pmd_dir, 0, sizeof(swapper_low_pmd_dir));
1716 
1717 	/* Now can init the kernel/bad page tables. */
1718 	pud_set(pud_offset(&swapper_pg_dir[0], 0),
1719 		swapper_low_pmd_dir + (shift / sizeof(pgd_t)));
1720 
1721 	inherit_prom_mappings();
1722 
1723 	init_kpte_bitmap();
1724 
1725 	/* Ok, we can use our TLB miss and window trap handlers safely.  */
1726 	setup_tba();
1727 
1728 	__flush_tlb_all();
1729 
1730 	if (tlb_type == hypervisor)
1731 		sun4v_ktsb_register();
1732 
1733 	prom_build_devicetree();
1734 	of_populate_present_mask();
1735 #ifndef CONFIG_SMP
1736 	of_fill_in_cpu_data();
1737 #endif
1738 
1739 	if (tlb_type == hypervisor) {
1740 		sun4v_mdesc_init();
1741 		mdesc_populate_present_mask(cpu_all_mask);
1742 #ifndef CONFIG_SMP
1743 		mdesc_fill_in_cpu_data(cpu_all_mask);
1744 #endif
1745 	}
1746 
1747 	/* Setup bootmem... */
1748 	last_valid_pfn = end_pfn = bootmem_init(phys_base);
1749 
1750 	/* Once the OF device tree and MDESC have been setup, we know
1751 	 * the list of possible cpus.  Therefore we can allocate the
1752 	 * IRQ stacks.
1753 	 */
1754 	for_each_possible_cpu(i) {
1755 		node = cpu_to_node(i);
1756 
1757 		softirq_stack[i] = __alloc_bootmem_node(NODE_DATA(node),
1758 							THREAD_SIZE,
1759 							THREAD_SIZE, 0);
1760 		hardirq_stack[i] = __alloc_bootmem_node(NODE_DATA(node),
1761 							THREAD_SIZE,
1762 							THREAD_SIZE, 0);
1763 	}
1764 
1765 	kernel_physical_mapping_init();
1766 
1767 	{
1768 		unsigned long max_zone_pfns[MAX_NR_ZONES];
1769 
1770 		memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1771 
1772 		max_zone_pfns[ZONE_NORMAL] = end_pfn;
1773 
1774 		free_area_init_nodes(max_zone_pfns);
1775 	}
1776 
1777 	printk("Booting Linux...\n");
1778 }
1779 
1780 int __devinit page_in_phys_avail(unsigned long paddr)
1781 {
1782 	int i;
1783 
1784 	paddr &= PAGE_MASK;
1785 
1786 	for (i = 0; i < pavail_ents; i++) {
1787 		unsigned long start, end;
1788 
1789 		start = pavail[i].phys_addr;
1790 		end = start + pavail[i].reg_size;
1791 
1792 		if (paddr >= start && paddr < end)
1793 			return 1;
1794 	}
1795 	if (paddr >= kern_base && paddr < (kern_base + kern_size))
1796 		return 1;
1797 #ifdef CONFIG_BLK_DEV_INITRD
1798 	if (paddr >= __pa(initrd_start) &&
1799 	    paddr < __pa(PAGE_ALIGN(initrd_end)))
1800 		return 1;
1801 #endif
1802 
1803 	return 0;
1804 }
1805 
1806 static struct linux_prom64_registers pavail_rescan[MAX_BANKS] __initdata;
1807 static int pavail_rescan_ents __initdata;
1808 
1809 /* Certain OBP calls, such as fetching "available" properties, can
1810  * claim physical memory.  So, along with initializing the valid
1811  * address bitmap, what we do here is refetch the physical available
1812  * memory list again, and make sure it provides at least as much
1813  * memory as 'pavail' does.
1814  */
1815 static void __init setup_valid_addr_bitmap_from_pavail(unsigned long *bitmap)
1816 {
1817 	int i;
1818 
1819 	read_obp_memory("available", &pavail_rescan[0], &pavail_rescan_ents);
1820 
1821 	for (i = 0; i < pavail_ents; i++) {
1822 		unsigned long old_start, old_end;
1823 
1824 		old_start = pavail[i].phys_addr;
1825 		old_end = old_start + pavail[i].reg_size;
1826 		while (old_start < old_end) {
1827 			int n;
1828 
1829 			for (n = 0; n < pavail_rescan_ents; n++) {
1830 				unsigned long new_start, new_end;
1831 
1832 				new_start = pavail_rescan[n].phys_addr;
1833 				new_end = new_start +
1834 					pavail_rescan[n].reg_size;
1835 
1836 				if (new_start <= old_start &&
1837 				    new_end >= (old_start + PAGE_SIZE)) {
1838 					set_bit(old_start >> 22, bitmap);
1839 					goto do_next_page;
1840 				}
1841 			}
1842 
1843 			prom_printf("mem_init: Lost memory in pavail\n");
1844 			prom_printf("mem_init: OLD start[%lx] size[%lx]\n",
1845 				    pavail[i].phys_addr,
1846 				    pavail[i].reg_size);
1847 			prom_printf("mem_init: NEW start[%lx] size[%lx]\n",
1848 				    pavail_rescan[i].phys_addr,
1849 				    pavail_rescan[i].reg_size);
1850 			prom_printf("mem_init: Cannot continue, aborting.\n");
1851 			prom_halt();
1852 
1853 		do_next_page:
1854 			old_start += PAGE_SIZE;
1855 		}
1856 	}
1857 }
1858 
1859 static void __init patch_tlb_miss_handler_bitmap(void)
1860 {
1861 	extern unsigned int valid_addr_bitmap_insn[];
1862 	extern unsigned int valid_addr_bitmap_patch[];
1863 
1864 	valid_addr_bitmap_insn[1] = valid_addr_bitmap_patch[1];
1865 	mb();
1866 	valid_addr_bitmap_insn[0] = valid_addr_bitmap_patch[0];
1867 	flushi(&valid_addr_bitmap_insn[0]);
1868 }
1869 
1870 void __init mem_init(void)
1871 {
1872 	unsigned long codepages, datapages, initpages;
1873 	unsigned long addr, last;
1874 
1875 	addr = PAGE_OFFSET + kern_base;
1876 	last = PAGE_ALIGN(kern_size) + addr;
1877 	while (addr < last) {
1878 		set_bit(__pa(addr) >> 22, sparc64_valid_addr_bitmap);
1879 		addr += PAGE_SIZE;
1880 	}
1881 
1882 	setup_valid_addr_bitmap_from_pavail(sparc64_valid_addr_bitmap);
1883 	patch_tlb_miss_handler_bitmap();
1884 
1885 	high_memory = __va(last_valid_pfn << PAGE_SHIFT);
1886 
1887 #ifdef CONFIG_NEED_MULTIPLE_NODES
1888 	{
1889 		int i;
1890 		for_each_online_node(i) {
1891 			if (NODE_DATA(i)->node_spanned_pages != 0) {
1892 				totalram_pages +=
1893 					free_all_bootmem_node(NODE_DATA(i));
1894 			}
1895 		}
1896 		totalram_pages += free_low_memory_core_early(MAX_NUMNODES);
1897 	}
1898 #else
1899 	totalram_pages = free_all_bootmem();
1900 #endif
1901 
1902 	/* We subtract one to account for the mem_map_zero page
1903 	 * allocated below.
1904 	 */
1905 	totalram_pages -= 1;
1906 	num_physpages = totalram_pages;
1907 
1908 	/*
1909 	 * Set up the zero page, mark it reserved, so that page count
1910 	 * is not manipulated when freeing the page from user ptes.
1911 	 */
1912 	mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0);
1913 	if (mem_map_zero == NULL) {
1914 		prom_printf("paging_init: Cannot alloc zero page.\n");
1915 		prom_halt();
1916 	}
1917 	SetPageReserved(mem_map_zero);
1918 
1919 	codepages = (((unsigned long) _etext) - ((unsigned long) _start));
1920 	codepages = PAGE_ALIGN(codepages) >> PAGE_SHIFT;
1921 	datapages = (((unsigned long) _edata) - ((unsigned long) _etext));
1922 	datapages = PAGE_ALIGN(datapages) >> PAGE_SHIFT;
1923 	initpages = (((unsigned long) __init_end) - ((unsigned long) __init_begin));
1924 	initpages = PAGE_ALIGN(initpages) >> PAGE_SHIFT;
1925 
1926 	printk("Memory: %luk available (%ldk kernel code, %ldk data, %ldk init) [%016lx,%016lx]\n",
1927 	       nr_free_pages() << (PAGE_SHIFT-10),
1928 	       codepages << (PAGE_SHIFT-10),
1929 	       datapages << (PAGE_SHIFT-10),
1930 	       initpages << (PAGE_SHIFT-10),
1931 	       PAGE_OFFSET, (last_valid_pfn << PAGE_SHIFT));
1932 
1933 	if (tlb_type == cheetah || tlb_type == cheetah_plus)
1934 		cheetah_ecache_flush_init();
1935 }
1936 
1937 void free_initmem(void)
1938 {
1939 	unsigned long addr, initend;
1940 	int do_free = 1;
1941 
1942 	/* If the physical memory maps were trimmed by kernel command
1943 	 * line options, don't even try freeing this initmem stuff up.
1944 	 * The kernel image could have been in the trimmed out region
1945 	 * and if so the freeing below will free invalid page structs.
1946 	 */
1947 	if (cmdline_memory_size)
1948 		do_free = 0;
1949 
1950 	/*
1951 	 * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes.
1952 	 */
1953 	addr = PAGE_ALIGN((unsigned long)(__init_begin));
1954 	initend = (unsigned long)(__init_end) & PAGE_MASK;
1955 	for (; addr < initend; addr += PAGE_SIZE) {
1956 		unsigned long page;
1957 		struct page *p;
1958 
1959 		page = (addr +
1960 			((unsigned long) __va(kern_base)) -
1961 			((unsigned long) KERNBASE));
1962 		memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
1963 
1964 		if (do_free) {
1965 			p = virt_to_page(page);
1966 
1967 			ClearPageReserved(p);
1968 			init_page_count(p);
1969 			__free_page(p);
1970 			num_physpages++;
1971 			totalram_pages++;
1972 		}
1973 	}
1974 }
1975 
1976 #ifdef CONFIG_BLK_DEV_INITRD
1977 void free_initrd_mem(unsigned long start, unsigned long end)
1978 {
1979 	if (start < end)
1980 		printk ("Freeing initrd memory: %ldk freed\n", (end - start) >> 10);
1981 	for (; start < end; start += PAGE_SIZE) {
1982 		struct page *p = virt_to_page(start);
1983 
1984 		ClearPageReserved(p);
1985 		init_page_count(p);
1986 		__free_page(p);
1987 		num_physpages++;
1988 		totalram_pages++;
1989 	}
1990 }
1991 #endif
1992 
1993 #define _PAGE_CACHE_4U	(_PAGE_CP_4U | _PAGE_CV_4U)
1994 #define _PAGE_CACHE_4V	(_PAGE_CP_4V | _PAGE_CV_4V)
1995 #define __DIRTY_BITS_4U	 (_PAGE_MODIFIED_4U | _PAGE_WRITE_4U | _PAGE_W_4U)
1996 #define __DIRTY_BITS_4V	 (_PAGE_MODIFIED_4V | _PAGE_WRITE_4V | _PAGE_W_4V)
1997 #define __ACCESS_BITS_4U (_PAGE_ACCESSED_4U | _PAGE_READ_4U | _PAGE_R)
1998 #define __ACCESS_BITS_4V (_PAGE_ACCESSED_4V | _PAGE_READ_4V | _PAGE_R)
1999 
2000 pgprot_t PAGE_KERNEL __read_mostly;
2001 EXPORT_SYMBOL(PAGE_KERNEL);
2002 
2003 pgprot_t PAGE_KERNEL_LOCKED __read_mostly;
2004 pgprot_t PAGE_COPY __read_mostly;
2005 
2006 pgprot_t PAGE_SHARED __read_mostly;
2007 EXPORT_SYMBOL(PAGE_SHARED);
2008 
2009 unsigned long pg_iobits __read_mostly;
2010 
2011 unsigned long _PAGE_IE __read_mostly;
2012 EXPORT_SYMBOL(_PAGE_IE);
2013 
2014 unsigned long _PAGE_E __read_mostly;
2015 EXPORT_SYMBOL(_PAGE_E);
2016 
2017 unsigned long _PAGE_CACHE __read_mostly;
2018 EXPORT_SYMBOL(_PAGE_CACHE);
2019 
2020 #ifdef CONFIG_SPARSEMEM_VMEMMAP
2021 unsigned long vmemmap_table[VMEMMAP_SIZE];
2022 
2023 static long __meminitdata addr_start, addr_end;
2024 static int __meminitdata node_start;
2025 
2026 int __meminit vmemmap_populate(struct page *start, unsigned long nr, int node)
2027 {
2028 	unsigned long vstart = (unsigned long) start;
2029 	unsigned long vend = (unsigned long) (start + nr);
2030 	unsigned long phys_start = (vstart - VMEMMAP_BASE);
2031 	unsigned long phys_end = (vend - VMEMMAP_BASE);
2032 	unsigned long addr = phys_start & VMEMMAP_CHUNK_MASK;
2033 	unsigned long end = VMEMMAP_ALIGN(phys_end);
2034 	unsigned long pte_base;
2035 
2036 	pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2037 		    _PAGE_CP_4U | _PAGE_CV_4U |
2038 		    _PAGE_P_4U | _PAGE_W_4U);
2039 	if (tlb_type == hypervisor)
2040 		pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2041 			    _PAGE_CP_4V | _PAGE_CV_4V |
2042 			    _PAGE_P_4V | _PAGE_W_4V);
2043 
2044 	for (; addr < end; addr += VMEMMAP_CHUNK) {
2045 		unsigned long *vmem_pp =
2046 			vmemmap_table + (addr >> VMEMMAP_CHUNK_SHIFT);
2047 		void *block;
2048 
2049 		if (!(*vmem_pp & _PAGE_VALID)) {
2050 			block = vmemmap_alloc_block(1UL << 22, node);
2051 			if (!block)
2052 				return -ENOMEM;
2053 
2054 			*vmem_pp = pte_base | __pa(block);
2055 
2056 			/* check to see if we have contiguous blocks */
2057 			if (addr_end != addr || node_start != node) {
2058 				if (addr_start)
2059 					printk(KERN_DEBUG " [%lx-%lx] on node %d\n",
2060 					       addr_start, addr_end-1, node_start);
2061 				addr_start = addr;
2062 				node_start = node;
2063 			}
2064 			addr_end = addr + VMEMMAP_CHUNK;
2065 		}
2066 	}
2067 	return 0;
2068 }
2069 
2070 void __meminit vmemmap_populate_print_last(void)
2071 {
2072 	if (addr_start) {
2073 		printk(KERN_DEBUG " [%lx-%lx] on node %d\n",
2074 		       addr_start, addr_end-1, node_start);
2075 		addr_start = 0;
2076 		addr_end = 0;
2077 		node_start = 0;
2078 	}
2079 }
2080 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
2081 
2082 static void prot_init_common(unsigned long page_none,
2083 			     unsigned long page_shared,
2084 			     unsigned long page_copy,
2085 			     unsigned long page_readonly,
2086 			     unsigned long page_exec_bit)
2087 {
2088 	PAGE_COPY = __pgprot(page_copy);
2089 	PAGE_SHARED = __pgprot(page_shared);
2090 
2091 	protection_map[0x0] = __pgprot(page_none);
2092 	protection_map[0x1] = __pgprot(page_readonly & ~page_exec_bit);
2093 	protection_map[0x2] = __pgprot(page_copy & ~page_exec_bit);
2094 	protection_map[0x3] = __pgprot(page_copy & ~page_exec_bit);
2095 	protection_map[0x4] = __pgprot(page_readonly);
2096 	protection_map[0x5] = __pgprot(page_readonly);
2097 	protection_map[0x6] = __pgprot(page_copy);
2098 	protection_map[0x7] = __pgprot(page_copy);
2099 	protection_map[0x8] = __pgprot(page_none);
2100 	protection_map[0x9] = __pgprot(page_readonly & ~page_exec_bit);
2101 	protection_map[0xa] = __pgprot(page_shared & ~page_exec_bit);
2102 	protection_map[0xb] = __pgprot(page_shared & ~page_exec_bit);
2103 	protection_map[0xc] = __pgprot(page_readonly);
2104 	protection_map[0xd] = __pgprot(page_readonly);
2105 	protection_map[0xe] = __pgprot(page_shared);
2106 	protection_map[0xf] = __pgprot(page_shared);
2107 }
2108 
2109 static void __init sun4u_pgprot_init(void)
2110 {
2111 	unsigned long page_none, page_shared, page_copy, page_readonly;
2112 	unsigned long page_exec_bit;
2113 
2114 	PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2115 				_PAGE_CACHE_4U | _PAGE_P_4U |
2116 				__ACCESS_BITS_4U | __DIRTY_BITS_4U |
2117 				_PAGE_EXEC_4U);
2118 	PAGE_KERNEL_LOCKED = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2119 				       _PAGE_CACHE_4U | _PAGE_P_4U |
2120 				       __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2121 				       _PAGE_EXEC_4U | _PAGE_L_4U);
2122 
2123 	_PAGE_IE = _PAGE_IE_4U;
2124 	_PAGE_E = _PAGE_E_4U;
2125 	_PAGE_CACHE = _PAGE_CACHE_4U;
2126 
2127 	pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4U | __DIRTY_BITS_4U |
2128 		     __ACCESS_BITS_4U | _PAGE_E_4U);
2129 
2130 #ifdef CONFIG_DEBUG_PAGEALLOC
2131 	kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZBITS_4U) ^
2132 		0xfffff80000000000UL;
2133 #else
2134 	kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4U) ^
2135 		0xfffff80000000000UL;
2136 #endif
2137 	kern_linear_pte_xor[0] |= (_PAGE_CP_4U | _PAGE_CV_4U |
2138 				   _PAGE_P_4U | _PAGE_W_4U);
2139 
2140 	/* XXX Should use 256MB on Panther. XXX */
2141 	kern_linear_pte_xor[1] = kern_linear_pte_xor[0];
2142 
2143 	_PAGE_SZBITS = _PAGE_SZBITS_4U;
2144 	_PAGE_ALL_SZ_BITS =  (_PAGE_SZ4MB_4U | _PAGE_SZ512K_4U |
2145 			      _PAGE_SZ64K_4U | _PAGE_SZ8K_4U |
2146 			      _PAGE_SZ32MB_4U | _PAGE_SZ256MB_4U);
2147 
2148 
2149 	page_none = _PAGE_PRESENT_4U | _PAGE_ACCESSED_4U | _PAGE_CACHE_4U;
2150 	page_shared = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2151 		       __ACCESS_BITS_4U | _PAGE_WRITE_4U | _PAGE_EXEC_4U);
2152 	page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2153 		       __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2154 	page_readonly   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2155 			   __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2156 
2157 	page_exec_bit = _PAGE_EXEC_4U;
2158 
2159 	prot_init_common(page_none, page_shared, page_copy, page_readonly,
2160 			 page_exec_bit);
2161 }
2162 
2163 static void __init sun4v_pgprot_init(void)
2164 {
2165 	unsigned long page_none, page_shared, page_copy, page_readonly;
2166 	unsigned long page_exec_bit;
2167 
2168 	PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4V | _PAGE_VALID |
2169 				_PAGE_CACHE_4V | _PAGE_P_4V |
2170 				__ACCESS_BITS_4V | __DIRTY_BITS_4V |
2171 				_PAGE_EXEC_4V);
2172 	PAGE_KERNEL_LOCKED = PAGE_KERNEL;
2173 
2174 	_PAGE_IE = _PAGE_IE_4V;
2175 	_PAGE_E = _PAGE_E_4V;
2176 	_PAGE_CACHE = _PAGE_CACHE_4V;
2177 
2178 #ifdef CONFIG_DEBUG_PAGEALLOC
2179 	kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZBITS_4V) ^
2180 		0xfffff80000000000UL;
2181 #else
2182 	kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4V) ^
2183 		0xfffff80000000000UL;
2184 #endif
2185 	kern_linear_pte_xor[0] |= (_PAGE_CP_4V | _PAGE_CV_4V |
2186 				   _PAGE_P_4V | _PAGE_W_4V);
2187 
2188 #ifdef CONFIG_DEBUG_PAGEALLOC
2189 	kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZBITS_4V) ^
2190 		0xfffff80000000000UL;
2191 #else
2192 	kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZ256MB_4V) ^
2193 		0xfffff80000000000UL;
2194 #endif
2195 	kern_linear_pte_xor[1] |= (_PAGE_CP_4V | _PAGE_CV_4V |
2196 				   _PAGE_P_4V | _PAGE_W_4V);
2197 
2198 	pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4V | __DIRTY_BITS_4V |
2199 		     __ACCESS_BITS_4V | _PAGE_E_4V);
2200 
2201 	_PAGE_SZBITS = _PAGE_SZBITS_4V;
2202 	_PAGE_ALL_SZ_BITS = (_PAGE_SZ16GB_4V | _PAGE_SZ2GB_4V |
2203 			     _PAGE_SZ256MB_4V | _PAGE_SZ32MB_4V |
2204 			     _PAGE_SZ4MB_4V | _PAGE_SZ512K_4V |
2205 			     _PAGE_SZ64K_4V | _PAGE_SZ8K_4V);
2206 
2207 	page_none = _PAGE_PRESENT_4V | _PAGE_ACCESSED_4V | _PAGE_CACHE_4V;
2208 	page_shared = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2209 		       __ACCESS_BITS_4V | _PAGE_WRITE_4V | _PAGE_EXEC_4V);
2210 	page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2211 		       __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2212 	page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2213 			 __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2214 
2215 	page_exec_bit = _PAGE_EXEC_4V;
2216 
2217 	prot_init_common(page_none, page_shared, page_copy, page_readonly,
2218 			 page_exec_bit);
2219 }
2220 
2221 unsigned long pte_sz_bits(unsigned long sz)
2222 {
2223 	if (tlb_type == hypervisor) {
2224 		switch (sz) {
2225 		case 8 * 1024:
2226 		default:
2227 			return _PAGE_SZ8K_4V;
2228 		case 64 * 1024:
2229 			return _PAGE_SZ64K_4V;
2230 		case 512 * 1024:
2231 			return _PAGE_SZ512K_4V;
2232 		case 4 * 1024 * 1024:
2233 			return _PAGE_SZ4MB_4V;
2234 		}
2235 	} else {
2236 		switch (sz) {
2237 		case 8 * 1024:
2238 		default:
2239 			return _PAGE_SZ8K_4U;
2240 		case 64 * 1024:
2241 			return _PAGE_SZ64K_4U;
2242 		case 512 * 1024:
2243 			return _PAGE_SZ512K_4U;
2244 		case 4 * 1024 * 1024:
2245 			return _PAGE_SZ4MB_4U;
2246 		}
2247 	}
2248 }
2249 
2250 pte_t mk_pte_io(unsigned long page, pgprot_t prot, int space, unsigned long page_size)
2251 {
2252 	pte_t pte;
2253 
2254 	pte_val(pte)  = page | pgprot_val(pgprot_noncached(prot));
2255 	pte_val(pte) |= (((unsigned long)space) << 32);
2256 	pte_val(pte) |= pte_sz_bits(page_size);
2257 
2258 	return pte;
2259 }
2260 
2261 static unsigned long kern_large_tte(unsigned long paddr)
2262 {
2263 	unsigned long val;
2264 
2265 	val = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2266 	       _PAGE_CP_4U | _PAGE_CV_4U | _PAGE_P_4U |
2267 	       _PAGE_EXEC_4U | _PAGE_L_4U | _PAGE_W_4U);
2268 	if (tlb_type == hypervisor)
2269 		val = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2270 		       _PAGE_CP_4V | _PAGE_CV_4V | _PAGE_P_4V |
2271 		       _PAGE_EXEC_4V | _PAGE_W_4V);
2272 
2273 	return val | paddr;
2274 }
2275 
2276 /* If not locked, zap it. */
2277 void __flush_tlb_all(void)
2278 {
2279 	unsigned long pstate;
2280 	int i;
2281 
2282 	__asm__ __volatile__("flushw\n\t"
2283 			     "rdpr	%%pstate, %0\n\t"
2284 			     "wrpr	%0, %1, %%pstate"
2285 			     : "=r" (pstate)
2286 			     : "i" (PSTATE_IE));
2287 	if (tlb_type == hypervisor) {
2288 		sun4v_mmu_demap_all();
2289 	} else if (tlb_type == spitfire) {
2290 		for (i = 0; i < 64; i++) {
2291 			/* Spitfire Errata #32 workaround */
2292 			/* NOTE: Always runs on spitfire, so no
2293 			 *       cheetah+ page size encodings.
2294 			 */
2295 			__asm__ __volatile__("stxa	%0, [%1] %2\n\t"
2296 					     "flush	%%g6"
2297 					     : /* No outputs */
2298 					     : "r" (0),
2299 					     "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2300 
2301 			if (!(spitfire_get_dtlb_data(i) & _PAGE_L_4U)) {
2302 				__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2303 						     "membar #Sync"
2304 						     : /* no outputs */
2305 						     : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
2306 				spitfire_put_dtlb_data(i, 0x0UL);
2307 			}
2308 
2309 			/* Spitfire Errata #32 workaround */
2310 			/* NOTE: Always runs on spitfire, so no
2311 			 *       cheetah+ page size encodings.
2312 			 */
2313 			__asm__ __volatile__("stxa	%0, [%1] %2\n\t"
2314 					     "flush	%%g6"
2315 					     : /* No outputs */
2316 					     : "r" (0),
2317 					     "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2318 
2319 			if (!(spitfire_get_itlb_data(i) & _PAGE_L_4U)) {
2320 				__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2321 						     "membar #Sync"
2322 						     : /* no outputs */
2323 						     : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
2324 				spitfire_put_itlb_data(i, 0x0UL);
2325 			}
2326 		}
2327 	} else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
2328 		cheetah_flush_dtlb_all();
2329 		cheetah_flush_itlb_all();
2330 	}
2331 	__asm__ __volatile__("wrpr	%0, 0, %%pstate"
2332 			     : : "r" (pstate));
2333 }
2334