xref: /linux/arch/sparc/mm/init_64.c (revision b43ab901d671e3e3cad425ea5e9a3c74e266dcdd)
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/system.h>
32 #include <asm/page.h>
33 #include <asm/pgalloc.h>
34 #include <asm/pgtable.h>
35 #include <asm/oplib.h>
36 #include <asm/iommu.h>
37 #include <asm/io.h>
38 #include <asm/uaccess.h>
39 #include <asm/mmu_context.h>
40 #include <asm/tlbflush.h>
41 #include <asm/dma.h>
42 #include <asm/starfire.h>
43 #include <asm/tlb.h>
44 #include <asm/spitfire.h>
45 #include <asm/sections.h>
46 #include <asm/tsb.h>
47 #include <asm/hypervisor.h>
48 #include <asm/prom.h>
49 #include <asm/mdesc.h>
50 #include <asm/cpudata.h>
51 #include <asm/irq.h>
52 
53 #include "init_64.h"
54 
55 unsigned long kern_linear_pte_xor[2] __read_mostly;
56 
57 /* A bitmap, one bit for every 256MB of physical memory.  If the bit
58  * is clear, we should use a 4MB page (via kern_linear_pte_xor[0]) else
59  * if set we should use a 256MB page (via kern_linear_pte_xor[1]).
60  */
61 unsigned long kpte_linear_bitmap[KPTE_BITMAP_BYTES / sizeof(unsigned long)];
62 
63 #ifndef CONFIG_DEBUG_PAGEALLOC
64 /* A special kernel TSB for 4MB and 256MB linear mappings.
65  * Space is allocated for this right after the trap table
66  * in arch/sparc64/kernel/head.S
67  */
68 extern struct tsb swapper_4m_tsb[KERNEL_TSB4M_NENTRIES];
69 #endif
70 
71 #define MAX_BANKS	32
72 
73 static struct linux_prom64_registers pavail[MAX_BANKS] __devinitdata;
74 static int pavail_ents __devinitdata;
75 
76 static int cmp_p64(const void *a, const void *b)
77 {
78 	const struct linux_prom64_registers *x = a, *y = b;
79 
80 	if (x->phys_addr > y->phys_addr)
81 		return 1;
82 	if (x->phys_addr < y->phys_addr)
83 		return -1;
84 	return 0;
85 }
86 
87 static void __init read_obp_memory(const char *property,
88 				   struct linux_prom64_registers *regs,
89 				   int *num_ents)
90 {
91 	phandle node = prom_finddevice("/memory");
92 	int prop_size = prom_getproplen(node, property);
93 	int ents, ret, i;
94 
95 	ents = prop_size / sizeof(struct linux_prom64_registers);
96 	if (ents > MAX_BANKS) {
97 		prom_printf("The machine has more %s property entries than "
98 			    "this kernel can support (%d).\n",
99 			    property, MAX_BANKS);
100 		prom_halt();
101 	}
102 
103 	ret = prom_getproperty(node, property, (char *) regs, prop_size);
104 	if (ret == -1) {
105 		prom_printf("Couldn't get %s property from /memory.\n");
106 		prom_halt();
107 	}
108 
109 	/* Sanitize what we got from the firmware, by page aligning
110 	 * everything.
111 	 */
112 	for (i = 0; i < ents; i++) {
113 		unsigned long base, size;
114 
115 		base = regs[i].phys_addr;
116 		size = regs[i].reg_size;
117 
118 		size &= PAGE_MASK;
119 		if (base & ~PAGE_MASK) {
120 			unsigned long new_base = PAGE_ALIGN(base);
121 
122 			size -= new_base - base;
123 			if ((long) size < 0L)
124 				size = 0UL;
125 			base = new_base;
126 		}
127 		if (size == 0UL) {
128 			/* If it is empty, simply get rid of it.
129 			 * This simplifies the logic of the other
130 			 * functions that process these arrays.
131 			 */
132 			memmove(&regs[i], &regs[i + 1],
133 				(ents - i - 1) * sizeof(regs[0]));
134 			i--;
135 			ents--;
136 			continue;
137 		}
138 		regs[i].phys_addr = base;
139 		regs[i].reg_size = size;
140 	}
141 
142 	*num_ents = ents;
143 
144 	sort(regs, ents, sizeof(struct linux_prom64_registers),
145 	     cmp_p64, NULL);
146 }
147 
148 unsigned long sparc64_valid_addr_bitmap[VALID_ADDR_BITMAP_BYTES /
149 					sizeof(unsigned long)];
150 EXPORT_SYMBOL(sparc64_valid_addr_bitmap);
151 
152 /* Kernel physical address base and size in bytes.  */
153 unsigned long kern_base __read_mostly;
154 unsigned long kern_size __read_mostly;
155 
156 /* Initial ramdisk setup */
157 extern unsigned long sparc_ramdisk_image64;
158 extern unsigned int sparc_ramdisk_image;
159 extern unsigned int sparc_ramdisk_size;
160 
161 struct page *mem_map_zero __read_mostly;
162 EXPORT_SYMBOL(mem_map_zero);
163 
164 unsigned int sparc64_highest_unlocked_tlb_ent __read_mostly;
165 
166 unsigned long sparc64_kern_pri_context __read_mostly;
167 unsigned long sparc64_kern_pri_nuc_bits __read_mostly;
168 unsigned long sparc64_kern_sec_context __read_mostly;
169 
170 int num_kernel_image_mappings;
171 
172 #ifdef CONFIG_DEBUG_DCFLUSH
173 atomic_t dcpage_flushes = ATOMIC_INIT(0);
174 #ifdef CONFIG_SMP
175 atomic_t dcpage_flushes_xcall = ATOMIC_INIT(0);
176 #endif
177 #endif
178 
179 inline void flush_dcache_page_impl(struct page *page)
180 {
181 	BUG_ON(tlb_type == hypervisor);
182 #ifdef CONFIG_DEBUG_DCFLUSH
183 	atomic_inc(&dcpage_flushes);
184 #endif
185 
186 #ifdef DCACHE_ALIASING_POSSIBLE
187 	__flush_dcache_page(page_address(page),
188 			    ((tlb_type == spitfire) &&
189 			     page_mapping(page) != NULL));
190 #else
191 	if (page_mapping(page) != NULL &&
192 	    tlb_type == spitfire)
193 		__flush_icache_page(__pa(page_address(page)));
194 #endif
195 }
196 
197 #define PG_dcache_dirty		PG_arch_1
198 #define PG_dcache_cpu_shift	32UL
199 #define PG_dcache_cpu_mask	\
200 	((1UL<<ilog2(roundup_pow_of_two(NR_CPUS)))-1UL)
201 
202 #define dcache_dirty_cpu(page) \
203 	(((page)->flags >> PG_dcache_cpu_shift) & PG_dcache_cpu_mask)
204 
205 static inline void set_dcache_dirty(struct page *page, int this_cpu)
206 {
207 	unsigned long mask = this_cpu;
208 	unsigned long non_cpu_bits;
209 
210 	non_cpu_bits = ~(PG_dcache_cpu_mask << PG_dcache_cpu_shift);
211 	mask = (mask << PG_dcache_cpu_shift) | (1UL << PG_dcache_dirty);
212 
213 	__asm__ __volatile__("1:\n\t"
214 			     "ldx	[%2], %%g7\n\t"
215 			     "and	%%g7, %1, %%g1\n\t"
216 			     "or	%%g1, %0, %%g1\n\t"
217 			     "casx	[%2], %%g7, %%g1\n\t"
218 			     "cmp	%%g7, %%g1\n\t"
219 			     "bne,pn	%%xcc, 1b\n\t"
220 			     " nop"
221 			     : /* no outputs */
222 			     : "r" (mask), "r" (non_cpu_bits), "r" (&page->flags)
223 			     : "g1", "g7");
224 }
225 
226 static inline void clear_dcache_dirty_cpu(struct page *page, unsigned long cpu)
227 {
228 	unsigned long mask = (1UL << PG_dcache_dirty);
229 
230 	__asm__ __volatile__("! test_and_clear_dcache_dirty\n"
231 			     "1:\n\t"
232 			     "ldx	[%2], %%g7\n\t"
233 			     "srlx	%%g7, %4, %%g1\n\t"
234 			     "and	%%g1, %3, %%g1\n\t"
235 			     "cmp	%%g1, %0\n\t"
236 			     "bne,pn	%%icc, 2f\n\t"
237 			     " andn	%%g7, %1, %%g1\n\t"
238 			     "casx	[%2], %%g7, %%g1\n\t"
239 			     "cmp	%%g7, %%g1\n\t"
240 			     "bne,pn	%%xcc, 1b\n\t"
241 			     " nop\n"
242 			     "2:"
243 			     : /* no outputs */
244 			     : "r" (cpu), "r" (mask), "r" (&page->flags),
245 			       "i" (PG_dcache_cpu_mask),
246 			       "i" (PG_dcache_cpu_shift)
247 			     : "g1", "g7");
248 }
249 
250 static inline void tsb_insert(struct tsb *ent, unsigned long tag, unsigned long pte)
251 {
252 	unsigned long tsb_addr = (unsigned long) ent;
253 
254 	if (tlb_type == cheetah_plus || tlb_type == hypervisor)
255 		tsb_addr = __pa(tsb_addr);
256 
257 	__tsb_insert(tsb_addr, tag, pte);
258 }
259 
260 unsigned long _PAGE_ALL_SZ_BITS __read_mostly;
261 unsigned long _PAGE_SZBITS __read_mostly;
262 
263 static void flush_dcache(unsigned long pfn)
264 {
265 	struct page *page;
266 
267 	page = pfn_to_page(pfn);
268 	if (page) {
269 		unsigned long pg_flags;
270 
271 		pg_flags = page->flags;
272 		if (pg_flags & (1UL << PG_dcache_dirty)) {
273 			int cpu = ((pg_flags >> PG_dcache_cpu_shift) &
274 				   PG_dcache_cpu_mask);
275 			int this_cpu = get_cpu();
276 
277 			/* This is just to optimize away some function calls
278 			 * in the SMP case.
279 			 */
280 			if (cpu == this_cpu)
281 				flush_dcache_page_impl(page);
282 			else
283 				smp_flush_dcache_page_impl(page, cpu);
284 
285 			clear_dcache_dirty_cpu(page, cpu);
286 
287 			put_cpu();
288 		}
289 	}
290 }
291 
292 void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
293 {
294 	struct mm_struct *mm;
295 	struct tsb *tsb;
296 	unsigned long tag, flags;
297 	unsigned long tsb_index, tsb_hash_shift;
298 	pte_t pte = *ptep;
299 
300 	if (tlb_type != hypervisor) {
301 		unsigned long pfn = pte_pfn(pte);
302 
303 		if (pfn_valid(pfn))
304 			flush_dcache(pfn);
305 	}
306 
307 	mm = vma->vm_mm;
308 
309 	tsb_index = MM_TSB_BASE;
310 	tsb_hash_shift = PAGE_SHIFT;
311 
312 	spin_lock_irqsave(&mm->context.lock, flags);
313 
314 #ifdef CONFIG_HUGETLB_PAGE
315 	if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL) {
316 		if ((tlb_type == hypervisor &&
317 		     (pte_val(pte) & _PAGE_SZALL_4V) == _PAGE_SZHUGE_4V) ||
318 		    (tlb_type != hypervisor &&
319 		     (pte_val(pte) & _PAGE_SZALL_4U) == _PAGE_SZHUGE_4U)) {
320 			tsb_index = MM_TSB_HUGE;
321 			tsb_hash_shift = HPAGE_SHIFT;
322 		}
323 	}
324 #endif
325 
326 	tsb = mm->context.tsb_block[tsb_index].tsb;
327 	tsb += ((address >> tsb_hash_shift) &
328 		(mm->context.tsb_block[tsb_index].tsb_nentries - 1UL));
329 	tag = (address >> 22UL);
330 	tsb_insert(tsb, tag, pte_val(pte));
331 
332 	spin_unlock_irqrestore(&mm->context.lock, flags);
333 }
334 
335 void flush_dcache_page(struct page *page)
336 {
337 	struct address_space *mapping;
338 	int this_cpu;
339 
340 	if (tlb_type == hypervisor)
341 		return;
342 
343 	/* Do not bother with the expensive D-cache flush if it
344 	 * is merely the zero page.  The 'bigcore' testcase in GDB
345 	 * causes this case to run millions of times.
346 	 */
347 	if (page == ZERO_PAGE(0))
348 		return;
349 
350 	this_cpu = get_cpu();
351 
352 	mapping = page_mapping(page);
353 	if (mapping && !mapping_mapped(mapping)) {
354 		int dirty = test_bit(PG_dcache_dirty, &page->flags);
355 		if (dirty) {
356 			int dirty_cpu = dcache_dirty_cpu(page);
357 
358 			if (dirty_cpu == this_cpu)
359 				goto out;
360 			smp_flush_dcache_page_impl(page, dirty_cpu);
361 		}
362 		set_dcache_dirty(page, this_cpu);
363 	} else {
364 		/* We could delay the flush for the !page_mapping
365 		 * case too.  But that case is for exec env/arg
366 		 * pages and those are %99 certainly going to get
367 		 * faulted into the tlb (and thus flushed) anyways.
368 		 */
369 		flush_dcache_page_impl(page);
370 	}
371 
372 out:
373 	put_cpu();
374 }
375 EXPORT_SYMBOL(flush_dcache_page);
376 
377 void __kprobes flush_icache_range(unsigned long start, unsigned long end)
378 {
379 	/* Cheetah and Hypervisor platform cpus have coherent I-cache. */
380 	if (tlb_type == spitfire) {
381 		unsigned long kaddr;
382 
383 		/* This code only runs on Spitfire cpus so this is
384 		 * why we can assume _PAGE_PADDR_4U.
385 		 */
386 		for (kaddr = start; kaddr < end; kaddr += PAGE_SIZE) {
387 			unsigned long paddr, mask = _PAGE_PADDR_4U;
388 
389 			if (kaddr >= PAGE_OFFSET)
390 				paddr = kaddr & mask;
391 			else {
392 				pgd_t *pgdp = pgd_offset_k(kaddr);
393 				pud_t *pudp = pud_offset(pgdp, kaddr);
394 				pmd_t *pmdp = pmd_offset(pudp, kaddr);
395 				pte_t *ptep = pte_offset_kernel(pmdp, kaddr);
396 
397 				paddr = pte_val(*ptep) & mask;
398 			}
399 			__flush_icache_page(paddr);
400 		}
401 	}
402 }
403 EXPORT_SYMBOL(flush_icache_range);
404 
405 void mmu_info(struct seq_file *m)
406 {
407 	if (tlb_type == cheetah)
408 		seq_printf(m, "MMU Type\t: Cheetah\n");
409 	else if (tlb_type == cheetah_plus)
410 		seq_printf(m, "MMU Type\t: Cheetah+\n");
411 	else if (tlb_type == spitfire)
412 		seq_printf(m, "MMU Type\t: Spitfire\n");
413 	else if (tlb_type == hypervisor)
414 		seq_printf(m, "MMU Type\t: Hypervisor (sun4v)\n");
415 	else
416 		seq_printf(m, "MMU Type\t: ???\n");
417 
418 #ifdef CONFIG_DEBUG_DCFLUSH
419 	seq_printf(m, "DCPageFlushes\t: %d\n",
420 		   atomic_read(&dcpage_flushes));
421 #ifdef CONFIG_SMP
422 	seq_printf(m, "DCPageFlushesXC\t: %d\n",
423 		   atomic_read(&dcpage_flushes_xcall));
424 #endif /* CONFIG_SMP */
425 #endif /* CONFIG_DEBUG_DCFLUSH */
426 }
427 
428 struct linux_prom_translation prom_trans[512] __read_mostly;
429 unsigned int prom_trans_ents __read_mostly;
430 
431 unsigned long kern_locked_tte_data;
432 
433 /* The obp translations are saved based on 8k pagesize, since obp can
434  * use a mixture of pagesizes. Misses to the LOW_OBP_ADDRESS ->
435  * HI_OBP_ADDRESS range are handled in ktlb.S.
436  */
437 static inline int in_obp_range(unsigned long vaddr)
438 {
439 	return (vaddr >= LOW_OBP_ADDRESS &&
440 		vaddr < HI_OBP_ADDRESS);
441 }
442 
443 static int cmp_ptrans(const void *a, const void *b)
444 {
445 	const struct linux_prom_translation *x = a, *y = b;
446 
447 	if (x->virt > y->virt)
448 		return 1;
449 	if (x->virt < y->virt)
450 		return -1;
451 	return 0;
452 }
453 
454 /* Read OBP translations property into 'prom_trans[]'.  */
455 static void __init read_obp_translations(void)
456 {
457 	int n, node, ents, first, last, i;
458 
459 	node = prom_finddevice("/virtual-memory");
460 	n = prom_getproplen(node, "translations");
461 	if (unlikely(n == 0 || n == -1)) {
462 		prom_printf("prom_mappings: Couldn't get size.\n");
463 		prom_halt();
464 	}
465 	if (unlikely(n > sizeof(prom_trans))) {
466 		prom_printf("prom_mappings: Size %Zd is too big.\n", n);
467 		prom_halt();
468 	}
469 
470 	if ((n = prom_getproperty(node, "translations",
471 				  (char *)&prom_trans[0],
472 				  sizeof(prom_trans))) == -1) {
473 		prom_printf("prom_mappings: Couldn't get property.\n");
474 		prom_halt();
475 	}
476 
477 	n = n / sizeof(struct linux_prom_translation);
478 
479 	ents = n;
480 
481 	sort(prom_trans, ents, sizeof(struct linux_prom_translation),
482 	     cmp_ptrans, NULL);
483 
484 	/* Now kick out all the non-OBP entries.  */
485 	for (i = 0; i < ents; i++) {
486 		if (in_obp_range(prom_trans[i].virt))
487 			break;
488 	}
489 	first = i;
490 	for (; i < ents; i++) {
491 		if (!in_obp_range(prom_trans[i].virt))
492 			break;
493 	}
494 	last = i;
495 
496 	for (i = 0; i < (last - first); i++) {
497 		struct linux_prom_translation *src = &prom_trans[i + first];
498 		struct linux_prom_translation *dest = &prom_trans[i];
499 
500 		*dest = *src;
501 	}
502 	for (; i < ents; i++) {
503 		struct linux_prom_translation *dest = &prom_trans[i];
504 		dest->virt = dest->size = dest->data = 0x0UL;
505 	}
506 
507 	prom_trans_ents = last - first;
508 
509 	if (tlb_type == spitfire) {
510 		/* Clear diag TTE bits. */
511 		for (i = 0; i < prom_trans_ents; i++)
512 			prom_trans[i].data &= ~0x0003fe0000000000UL;
513 	}
514 
515 	/* Force execute bit on.  */
516 	for (i = 0; i < prom_trans_ents; i++)
517 		prom_trans[i].data |= (tlb_type == hypervisor ?
518 				       _PAGE_EXEC_4V : _PAGE_EXEC_4U);
519 }
520 
521 static void __init hypervisor_tlb_lock(unsigned long vaddr,
522 				       unsigned long pte,
523 				       unsigned long mmu)
524 {
525 	unsigned long ret = sun4v_mmu_map_perm_addr(vaddr, 0, pte, mmu);
526 
527 	if (ret != 0) {
528 		prom_printf("hypervisor_tlb_lock[%lx:%lx:%lx:%lx]: "
529 			    "errors with %lx\n", vaddr, 0, pte, mmu, ret);
530 		prom_halt();
531 	}
532 }
533 
534 static unsigned long kern_large_tte(unsigned long paddr);
535 
536 static void __init remap_kernel(void)
537 {
538 	unsigned long phys_page, tte_vaddr, tte_data;
539 	int i, tlb_ent = sparc64_highest_locked_tlbent();
540 
541 	tte_vaddr = (unsigned long) KERNBASE;
542 	phys_page = (prom_boot_mapping_phys_low >> 22UL) << 22UL;
543 	tte_data = kern_large_tte(phys_page);
544 
545 	kern_locked_tte_data = tte_data;
546 
547 	/* Now lock us into the TLBs via Hypervisor or OBP. */
548 	if (tlb_type == hypervisor) {
549 		for (i = 0; i < num_kernel_image_mappings; i++) {
550 			hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_DMMU);
551 			hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_IMMU);
552 			tte_vaddr += 0x400000;
553 			tte_data += 0x400000;
554 		}
555 	} else {
556 		for (i = 0; i < num_kernel_image_mappings; i++) {
557 			prom_dtlb_load(tlb_ent - i, tte_data, tte_vaddr);
558 			prom_itlb_load(tlb_ent - i, tte_data, tte_vaddr);
559 			tte_vaddr += 0x400000;
560 			tte_data += 0x400000;
561 		}
562 		sparc64_highest_unlocked_tlb_ent = tlb_ent - i;
563 	}
564 	if (tlb_type == cheetah_plus) {
565 		sparc64_kern_pri_context = (CTX_CHEETAH_PLUS_CTX0 |
566 					    CTX_CHEETAH_PLUS_NUC);
567 		sparc64_kern_pri_nuc_bits = CTX_CHEETAH_PLUS_NUC;
568 		sparc64_kern_sec_context = CTX_CHEETAH_PLUS_CTX0;
569 	}
570 }
571 
572 
573 static void __init inherit_prom_mappings(void)
574 {
575 	/* Now fixup OBP's idea about where we really are mapped. */
576 	printk("Remapping the kernel... ");
577 	remap_kernel();
578 	printk("done.\n");
579 }
580 
581 void prom_world(int enter)
582 {
583 	if (!enter)
584 		set_fs((mm_segment_t) { get_thread_current_ds() });
585 
586 	__asm__ __volatile__("flushw");
587 }
588 
589 void __flush_dcache_range(unsigned long start, unsigned long end)
590 {
591 	unsigned long va;
592 
593 	if (tlb_type == spitfire) {
594 		int n = 0;
595 
596 		for (va = start; va < end; va += 32) {
597 			spitfire_put_dcache_tag(va & 0x3fe0, 0x0);
598 			if (++n >= 512)
599 				break;
600 		}
601 	} else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
602 		start = __pa(start);
603 		end = __pa(end);
604 		for (va = start; va < end; va += 32)
605 			__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
606 					     "membar #Sync"
607 					     : /* no outputs */
608 					     : "r" (va),
609 					       "i" (ASI_DCACHE_INVALIDATE));
610 	}
611 }
612 EXPORT_SYMBOL(__flush_dcache_range);
613 
614 /* get_new_mmu_context() uses "cache + 1".  */
615 DEFINE_SPINLOCK(ctx_alloc_lock);
616 unsigned long tlb_context_cache = CTX_FIRST_VERSION - 1;
617 #define MAX_CTX_NR	(1UL << CTX_NR_BITS)
618 #define CTX_BMAP_SLOTS	BITS_TO_LONGS(MAX_CTX_NR)
619 DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR);
620 
621 /* Caller does TLB context flushing on local CPU if necessary.
622  * The caller also ensures that CTX_VALID(mm->context) is false.
623  *
624  * We must be careful about boundary cases so that we never
625  * let the user have CTX 0 (nucleus) or we ever use a CTX
626  * version of zero (and thus NO_CONTEXT would not be caught
627  * by version mis-match tests in mmu_context.h).
628  *
629  * Always invoked with interrupts disabled.
630  */
631 void get_new_mmu_context(struct mm_struct *mm)
632 {
633 	unsigned long ctx, new_ctx;
634 	unsigned long orig_pgsz_bits;
635 	unsigned long flags;
636 	int new_version;
637 
638 	spin_lock_irqsave(&ctx_alloc_lock, flags);
639 	orig_pgsz_bits = (mm->context.sparc64_ctx_val & CTX_PGSZ_MASK);
640 	ctx = (tlb_context_cache + 1) & CTX_NR_MASK;
641 	new_ctx = find_next_zero_bit(mmu_context_bmap, 1 << CTX_NR_BITS, ctx);
642 	new_version = 0;
643 	if (new_ctx >= (1 << CTX_NR_BITS)) {
644 		new_ctx = find_next_zero_bit(mmu_context_bmap, ctx, 1);
645 		if (new_ctx >= ctx) {
646 			int i;
647 			new_ctx = (tlb_context_cache & CTX_VERSION_MASK) +
648 				CTX_FIRST_VERSION;
649 			if (new_ctx == 1)
650 				new_ctx = CTX_FIRST_VERSION;
651 
652 			/* Don't call memset, for 16 entries that's just
653 			 * plain silly...
654 			 */
655 			mmu_context_bmap[0] = 3;
656 			mmu_context_bmap[1] = 0;
657 			mmu_context_bmap[2] = 0;
658 			mmu_context_bmap[3] = 0;
659 			for (i = 4; i < CTX_BMAP_SLOTS; i += 4) {
660 				mmu_context_bmap[i + 0] = 0;
661 				mmu_context_bmap[i + 1] = 0;
662 				mmu_context_bmap[i + 2] = 0;
663 				mmu_context_bmap[i + 3] = 0;
664 			}
665 			new_version = 1;
666 			goto out;
667 		}
668 	}
669 	mmu_context_bmap[new_ctx>>6] |= (1UL << (new_ctx & 63));
670 	new_ctx |= (tlb_context_cache & CTX_VERSION_MASK);
671 out:
672 	tlb_context_cache = new_ctx;
673 	mm->context.sparc64_ctx_val = new_ctx | orig_pgsz_bits;
674 	spin_unlock_irqrestore(&ctx_alloc_lock, flags);
675 
676 	if (unlikely(new_version))
677 		smp_new_mmu_context_version();
678 }
679 
680 static int numa_enabled = 1;
681 static int numa_debug;
682 
683 static int __init early_numa(char *p)
684 {
685 	if (!p)
686 		return 0;
687 
688 	if (strstr(p, "off"))
689 		numa_enabled = 0;
690 
691 	if (strstr(p, "debug"))
692 		numa_debug = 1;
693 
694 	return 0;
695 }
696 early_param("numa", early_numa);
697 
698 #define numadbg(f, a...) \
699 do {	if (numa_debug) \
700 		printk(KERN_INFO f, ## a); \
701 } while (0)
702 
703 static void __init find_ramdisk(unsigned long phys_base)
704 {
705 #ifdef CONFIG_BLK_DEV_INITRD
706 	if (sparc_ramdisk_image || sparc_ramdisk_image64) {
707 		unsigned long ramdisk_image;
708 
709 		/* Older versions of the bootloader only supported a
710 		 * 32-bit physical address for the ramdisk image
711 		 * location, stored at sparc_ramdisk_image.  Newer
712 		 * SILO versions set sparc_ramdisk_image to zero and
713 		 * provide a full 64-bit physical address at
714 		 * sparc_ramdisk_image64.
715 		 */
716 		ramdisk_image = sparc_ramdisk_image;
717 		if (!ramdisk_image)
718 			ramdisk_image = sparc_ramdisk_image64;
719 
720 		/* Another bootloader quirk.  The bootloader normalizes
721 		 * the physical address to KERNBASE, so we have to
722 		 * factor that back out and add in the lowest valid
723 		 * physical page address to get the true physical address.
724 		 */
725 		ramdisk_image -= KERNBASE;
726 		ramdisk_image += phys_base;
727 
728 		numadbg("Found ramdisk at physical address 0x%lx, size %u\n",
729 			ramdisk_image, sparc_ramdisk_size);
730 
731 		initrd_start = ramdisk_image;
732 		initrd_end = ramdisk_image + sparc_ramdisk_size;
733 
734 		memblock_reserve(initrd_start, sparc_ramdisk_size);
735 
736 		initrd_start += PAGE_OFFSET;
737 		initrd_end += PAGE_OFFSET;
738 	}
739 #endif
740 }
741 
742 struct node_mem_mask {
743 	unsigned long mask;
744 	unsigned long val;
745 	unsigned long bootmem_paddr;
746 };
747 static struct node_mem_mask node_masks[MAX_NUMNODES];
748 static int num_node_masks;
749 
750 int numa_cpu_lookup_table[NR_CPUS];
751 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
752 
753 #ifdef CONFIG_NEED_MULTIPLE_NODES
754 
755 struct mdesc_mblock {
756 	u64	base;
757 	u64	size;
758 	u64	offset; /* RA-to-PA */
759 };
760 static struct mdesc_mblock *mblocks;
761 static int num_mblocks;
762 
763 static unsigned long ra_to_pa(unsigned long addr)
764 {
765 	int i;
766 
767 	for (i = 0; i < num_mblocks; i++) {
768 		struct mdesc_mblock *m = &mblocks[i];
769 
770 		if (addr >= m->base &&
771 		    addr < (m->base + m->size)) {
772 			addr += m->offset;
773 			break;
774 		}
775 	}
776 	return addr;
777 }
778 
779 static int find_node(unsigned long addr)
780 {
781 	int i;
782 
783 	addr = ra_to_pa(addr);
784 	for (i = 0; i < num_node_masks; i++) {
785 		struct node_mem_mask *p = &node_masks[i];
786 
787 		if ((addr & p->mask) == p->val)
788 			return i;
789 	}
790 	return -1;
791 }
792 
793 static u64 memblock_nid_range(u64 start, u64 end, int *nid)
794 {
795 	*nid = find_node(start);
796 	start += PAGE_SIZE;
797 	while (start < end) {
798 		int n = find_node(start);
799 
800 		if (n != *nid)
801 			break;
802 		start += PAGE_SIZE;
803 	}
804 
805 	if (start > end)
806 		start = end;
807 
808 	return start;
809 }
810 #else
811 static u64 memblock_nid_range(u64 start, u64 end, int *nid)
812 {
813 	*nid = 0;
814 	return end;
815 }
816 #endif
817 
818 /* This must be invoked after performing all of the necessary
819  * memblock_set_node() calls for 'nid'.  We need to be able to get
820  * correct data from get_pfn_range_for_nid().
821  */
822 static void __init allocate_node_data(int nid)
823 {
824 	unsigned long paddr, num_pages, start_pfn, end_pfn;
825 	struct pglist_data *p;
826 
827 #ifdef CONFIG_NEED_MULTIPLE_NODES
828 	paddr = memblock_alloc_try_nid(sizeof(struct pglist_data), SMP_CACHE_BYTES, nid);
829 	if (!paddr) {
830 		prom_printf("Cannot allocate pglist_data for nid[%d]\n", nid);
831 		prom_halt();
832 	}
833 	NODE_DATA(nid) = __va(paddr);
834 	memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
835 
836 	NODE_DATA(nid)->bdata = &bootmem_node_data[nid];
837 #endif
838 
839 	p = NODE_DATA(nid);
840 
841 	get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
842 	p->node_start_pfn = start_pfn;
843 	p->node_spanned_pages = end_pfn - start_pfn;
844 
845 	if (p->node_spanned_pages) {
846 		num_pages = bootmem_bootmap_pages(p->node_spanned_pages);
847 
848 		paddr = memblock_alloc_try_nid(num_pages << PAGE_SHIFT, PAGE_SIZE, nid);
849 		if (!paddr) {
850 			prom_printf("Cannot allocate bootmap for nid[%d]\n",
851 				  nid);
852 			prom_halt();
853 		}
854 		node_masks[nid].bootmem_paddr = paddr;
855 	}
856 }
857 
858 static void init_node_masks_nonnuma(void)
859 {
860 	int i;
861 
862 	numadbg("Initializing tables for non-numa.\n");
863 
864 	node_masks[0].mask = node_masks[0].val = 0;
865 	num_node_masks = 1;
866 
867 	for (i = 0; i < NR_CPUS; i++)
868 		numa_cpu_lookup_table[i] = 0;
869 
870 	cpumask_setall(&numa_cpumask_lookup_table[0]);
871 }
872 
873 #ifdef CONFIG_NEED_MULTIPLE_NODES
874 struct pglist_data *node_data[MAX_NUMNODES];
875 
876 EXPORT_SYMBOL(numa_cpu_lookup_table);
877 EXPORT_SYMBOL(numa_cpumask_lookup_table);
878 EXPORT_SYMBOL(node_data);
879 
880 struct mdesc_mlgroup {
881 	u64	node;
882 	u64	latency;
883 	u64	match;
884 	u64	mask;
885 };
886 static struct mdesc_mlgroup *mlgroups;
887 static int num_mlgroups;
888 
889 static int scan_pio_for_cfg_handle(struct mdesc_handle *md, u64 pio,
890 				   u32 cfg_handle)
891 {
892 	u64 arc;
893 
894 	mdesc_for_each_arc(arc, md, pio, MDESC_ARC_TYPE_FWD) {
895 		u64 target = mdesc_arc_target(md, arc);
896 		const u64 *val;
897 
898 		val = mdesc_get_property(md, target,
899 					 "cfg-handle", NULL);
900 		if (val && *val == cfg_handle)
901 			return 0;
902 	}
903 	return -ENODEV;
904 }
905 
906 static int scan_arcs_for_cfg_handle(struct mdesc_handle *md, u64 grp,
907 				    u32 cfg_handle)
908 {
909 	u64 arc, candidate, best_latency = ~(u64)0;
910 
911 	candidate = MDESC_NODE_NULL;
912 	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
913 		u64 target = mdesc_arc_target(md, arc);
914 		const char *name = mdesc_node_name(md, target);
915 		const u64 *val;
916 
917 		if (strcmp(name, "pio-latency-group"))
918 			continue;
919 
920 		val = mdesc_get_property(md, target, "latency", NULL);
921 		if (!val)
922 			continue;
923 
924 		if (*val < best_latency) {
925 			candidate = target;
926 			best_latency = *val;
927 		}
928 	}
929 
930 	if (candidate == MDESC_NODE_NULL)
931 		return -ENODEV;
932 
933 	return scan_pio_for_cfg_handle(md, candidate, cfg_handle);
934 }
935 
936 int of_node_to_nid(struct device_node *dp)
937 {
938 	const struct linux_prom64_registers *regs;
939 	struct mdesc_handle *md;
940 	u32 cfg_handle;
941 	int count, nid;
942 	u64 grp;
943 
944 	/* This is the right thing to do on currently supported
945 	 * SUN4U NUMA platforms as well, as the PCI controller does
946 	 * not sit behind any particular memory controller.
947 	 */
948 	if (!mlgroups)
949 		return -1;
950 
951 	regs = of_get_property(dp, "reg", NULL);
952 	if (!regs)
953 		return -1;
954 
955 	cfg_handle = (regs->phys_addr >> 32UL) & 0x0fffffff;
956 
957 	md = mdesc_grab();
958 
959 	count = 0;
960 	nid = -1;
961 	mdesc_for_each_node_by_name(md, grp, "group") {
962 		if (!scan_arcs_for_cfg_handle(md, grp, cfg_handle)) {
963 			nid = count;
964 			break;
965 		}
966 		count++;
967 	}
968 
969 	mdesc_release(md);
970 
971 	return nid;
972 }
973 
974 static void __init add_node_ranges(void)
975 {
976 	struct memblock_region *reg;
977 
978 	for_each_memblock(memory, reg) {
979 		unsigned long size = reg->size;
980 		unsigned long start, end;
981 
982 		start = reg->base;
983 		end = start + size;
984 		while (start < end) {
985 			unsigned long this_end;
986 			int nid;
987 
988 			this_end = memblock_nid_range(start, end, &nid);
989 
990 			numadbg("Setting memblock NUMA node nid[%d] "
991 				"start[%lx] end[%lx]\n",
992 				nid, start, this_end);
993 
994 			memblock_set_node(start, this_end - start, nid);
995 			start = this_end;
996 		}
997 	}
998 }
999 
1000 static int __init grab_mlgroups(struct mdesc_handle *md)
1001 {
1002 	unsigned long paddr;
1003 	int count = 0;
1004 	u64 node;
1005 
1006 	mdesc_for_each_node_by_name(md, node, "memory-latency-group")
1007 		count++;
1008 	if (!count)
1009 		return -ENOENT;
1010 
1011 	paddr = memblock_alloc(count * sizeof(struct mdesc_mlgroup),
1012 			  SMP_CACHE_BYTES);
1013 	if (!paddr)
1014 		return -ENOMEM;
1015 
1016 	mlgroups = __va(paddr);
1017 	num_mlgroups = count;
1018 
1019 	count = 0;
1020 	mdesc_for_each_node_by_name(md, node, "memory-latency-group") {
1021 		struct mdesc_mlgroup *m = &mlgroups[count++];
1022 		const u64 *val;
1023 
1024 		m->node = node;
1025 
1026 		val = mdesc_get_property(md, node, "latency", NULL);
1027 		m->latency = *val;
1028 		val = mdesc_get_property(md, node, "address-match", NULL);
1029 		m->match = *val;
1030 		val = mdesc_get_property(md, node, "address-mask", NULL);
1031 		m->mask = *val;
1032 
1033 		numadbg("MLGROUP[%d]: node[%llx] latency[%llx] "
1034 			"match[%llx] mask[%llx]\n",
1035 			count - 1, m->node, m->latency, m->match, m->mask);
1036 	}
1037 
1038 	return 0;
1039 }
1040 
1041 static int __init grab_mblocks(struct mdesc_handle *md)
1042 {
1043 	unsigned long paddr;
1044 	int count = 0;
1045 	u64 node;
1046 
1047 	mdesc_for_each_node_by_name(md, node, "mblock")
1048 		count++;
1049 	if (!count)
1050 		return -ENOENT;
1051 
1052 	paddr = memblock_alloc(count * sizeof(struct mdesc_mblock),
1053 			  SMP_CACHE_BYTES);
1054 	if (!paddr)
1055 		return -ENOMEM;
1056 
1057 	mblocks = __va(paddr);
1058 	num_mblocks = count;
1059 
1060 	count = 0;
1061 	mdesc_for_each_node_by_name(md, node, "mblock") {
1062 		struct mdesc_mblock *m = &mblocks[count++];
1063 		const u64 *val;
1064 
1065 		val = mdesc_get_property(md, node, "base", NULL);
1066 		m->base = *val;
1067 		val = mdesc_get_property(md, node, "size", NULL);
1068 		m->size = *val;
1069 		val = mdesc_get_property(md, node,
1070 					 "address-congruence-offset", NULL);
1071 		m->offset = *val;
1072 
1073 		numadbg("MBLOCK[%d]: base[%llx] size[%llx] offset[%llx]\n",
1074 			count - 1, m->base, m->size, m->offset);
1075 	}
1076 
1077 	return 0;
1078 }
1079 
1080 static void __init numa_parse_mdesc_group_cpus(struct mdesc_handle *md,
1081 					       u64 grp, cpumask_t *mask)
1082 {
1083 	u64 arc;
1084 
1085 	cpumask_clear(mask);
1086 
1087 	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_BACK) {
1088 		u64 target = mdesc_arc_target(md, arc);
1089 		const char *name = mdesc_node_name(md, target);
1090 		const u64 *id;
1091 
1092 		if (strcmp(name, "cpu"))
1093 			continue;
1094 		id = mdesc_get_property(md, target, "id", NULL);
1095 		if (*id < nr_cpu_ids)
1096 			cpumask_set_cpu(*id, mask);
1097 	}
1098 }
1099 
1100 static struct mdesc_mlgroup * __init find_mlgroup(u64 node)
1101 {
1102 	int i;
1103 
1104 	for (i = 0; i < num_mlgroups; i++) {
1105 		struct mdesc_mlgroup *m = &mlgroups[i];
1106 		if (m->node == node)
1107 			return m;
1108 	}
1109 	return NULL;
1110 }
1111 
1112 static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp,
1113 				      int index)
1114 {
1115 	struct mdesc_mlgroup *candidate = NULL;
1116 	u64 arc, best_latency = ~(u64)0;
1117 	struct node_mem_mask *n;
1118 
1119 	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1120 		u64 target = mdesc_arc_target(md, arc);
1121 		struct mdesc_mlgroup *m = find_mlgroup(target);
1122 		if (!m)
1123 			continue;
1124 		if (m->latency < best_latency) {
1125 			candidate = m;
1126 			best_latency = m->latency;
1127 		}
1128 	}
1129 	if (!candidate)
1130 		return -ENOENT;
1131 
1132 	if (num_node_masks != index) {
1133 		printk(KERN_ERR "Inconsistent NUMA state, "
1134 		       "index[%d] != num_node_masks[%d]\n",
1135 		       index, num_node_masks);
1136 		return -EINVAL;
1137 	}
1138 
1139 	n = &node_masks[num_node_masks++];
1140 
1141 	n->mask = candidate->mask;
1142 	n->val = candidate->match;
1143 
1144 	numadbg("NUMA NODE[%d]: mask[%lx] val[%lx] (latency[%llx])\n",
1145 		index, n->mask, n->val, candidate->latency);
1146 
1147 	return 0;
1148 }
1149 
1150 static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp,
1151 					 int index)
1152 {
1153 	cpumask_t mask;
1154 	int cpu;
1155 
1156 	numa_parse_mdesc_group_cpus(md, grp, &mask);
1157 
1158 	for_each_cpu(cpu, &mask)
1159 		numa_cpu_lookup_table[cpu] = index;
1160 	cpumask_copy(&numa_cpumask_lookup_table[index], &mask);
1161 
1162 	if (numa_debug) {
1163 		printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index);
1164 		for_each_cpu(cpu, &mask)
1165 			printk("%d ", cpu);
1166 		printk("]\n");
1167 	}
1168 
1169 	return numa_attach_mlgroup(md, grp, index);
1170 }
1171 
1172 static int __init numa_parse_mdesc(void)
1173 {
1174 	struct mdesc_handle *md = mdesc_grab();
1175 	int i, err, count;
1176 	u64 node;
1177 
1178 	node = mdesc_node_by_name(md, MDESC_NODE_NULL, "latency-groups");
1179 	if (node == MDESC_NODE_NULL) {
1180 		mdesc_release(md);
1181 		return -ENOENT;
1182 	}
1183 
1184 	err = grab_mblocks(md);
1185 	if (err < 0)
1186 		goto out;
1187 
1188 	err = grab_mlgroups(md);
1189 	if (err < 0)
1190 		goto out;
1191 
1192 	count = 0;
1193 	mdesc_for_each_node_by_name(md, node, "group") {
1194 		err = numa_parse_mdesc_group(md, node, count);
1195 		if (err < 0)
1196 			break;
1197 		count++;
1198 	}
1199 
1200 	add_node_ranges();
1201 
1202 	for (i = 0; i < num_node_masks; i++) {
1203 		allocate_node_data(i);
1204 		node_set_online(i);
1205 	}
1206 
1207 	err = 0;
1208 out:
1209 	mdesc_release(md);
1210 	return err;
1211 }
1212 
1213 static int __init numa_parse_jbus(void)
1214 {
1215 	unsigned long cpu, index;
1216 
1217 	/* NUMA node id is encoded in bits 36 and higher, and there is
1218 	 * a 1-to-1 mapping from CPU ID to NUMA node ID.
1219 	 */
1220 	index = 0;
1221 	for_each_present_cpu(cpu) {
1222 		numa_cpu_lookup_table[cpu] = index;
1223 		cpumask_copy(&numa_cpumask_lookup_table[index], cpumask_of(cpu));
1224 		node_masks[index].mask = ~((1UL << 36UL) - 1UL);
1225 		node_masks[index].val = cpu << 36UL;
1226 
1227 		index++;
1228 	}
1229 	num_node_masks = index;
1230 
1231 	add_node_ranges();
1232 
1233 	for (index = 0; index < num_node_masks; index++) {
1234 		allocate_node_data(index);
1235 		node_set_online(index);
1236 	}
1237 
1238 	return 0;
1239 }
1240 
1241 static int __init numa_parse_sun4u(void)
1242 {
1243 	if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1244 		unsigned long ver;
1245 
1246 		__asm__ ("rdpr %%ver, %0" : "=r" (ver));
1247 		if ((ver >> 32UL) == __JALAPENO_ID ||
1248 		    (ver >> 32UL) == __SERRANO_ID)
1249 			return numa_parse_jbus();
1250 	}
1251 	return -1;
1252 }
1253 
1254 static int __init bootmem_init_numa(void)
1255 {
1256 	int err = -1;
1257 
1258 	numadbg("bootmem_init_numa()\n");
1259 
1260 	if (numa_enabled) {
1261 		if (tlb_type == hypervisor)
1262 			err = numa_parse_mdesc();
1263 		else
1264 			err = numa_parse_sun4u();
1265 	}
1266 	return err;
1267 }
1268 
1269 #else
1270 
1271 static int bootmem_init_numa(void)
1272 {
1273 	return -1;
1274 }
1275 
1276 #endif
1277 
1278 static void __init bootmem_init_nonnuma(void)
1279 {
1280 	unsigned long top_of_ram = memblock_end_of_DRAM();
1281 	unsigned long total_ram = memblock_phys_mem_size();
1282 
1283 	numadbg("bootmem_init_nonnuma()\n");
1284 
1285 	printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
1286 	       top_of_ram, total_ram);
1287 	printk(KERN_INFO "Memory hole size: %ldMB\n",
1288 	       (top_of_ram - total_ram) >> 20);
1289 
1290 	init_node_masks_nonnuma();
1291 	memblock_set_node(0, (phys_addr_t)ULLONG_MAX, 0);
1292 	allocate_node_data(0);
1293 	node_set_online(0);
1294 }
1295 
1296 static void __init reserve_range_in_node(int nid, unsigned long start,
1297 					 unsigned long end)
1298 {
1299 	numadbg("    reserve_range_in_node(nid[%d],start[%lx],end[%lx]\n",
1300 		nid, start, end);
1301 	while (start < end) {
1302 		unsigned long this_end;
1303 		int n;
1304 
1305 		this_end = memblock_nid_range(start, end, &n);
1306 		if (n == nid) {
1307 			numadbg("      MATCH reserving range [%lx:%lx]\n",
1308 				start, this_end);
1309 			reserve_bootmem_node(NODE_DATA(nid), start,
1310 					     (this_end - start), BOOTMEM_DEFAULT);
1311 		} else
1312 			numadbg("      NO MATCH, advancing start to %lx\n",
1313 				this_end);
1314 
1315 		start = this_end;
1316 	}
1317 }
1318 
1319 static void __init trim_reserved_in_node(int nid)
1320 {
1321 	struct memblock_region *reg;
1322 
1323 	numadbg("  trim_reserved_in_node(%d)\n", nid);
1324 
1325 	for_each_memblock(reserved, reg)
1326 		reserve_range_in_node(nid, reg->base, reg->base + reg->size);
1327 }
1328 
1329 static void __init bootmem_init_one_node(int nid)
1330 {
1331 	struct pglist_data *p;
1332 
1333 	numadbg("bootmem_init_one_node(%d)\n", nid);
1334 
1335 	p = NODE_DATA(nid);
1336 
1337 	if (p->node_spanned_pages) {
1338 		unsigned long paddr = node_masks[nid].bootmem_paddr;
1339 		unsigned long end_pfn;
1340 
1341 		end_pfn = p->node_start_pfn + p->node_spanned_pages;
1342 
1343 		numadbg("  init_bootmem_node(%d, %lx, %lx, %lx)\n",
1344 			nid, paddr >> PAGE_SHIFT, p->node_start_pfn, end_pfn);
1345 
1346 		init_bootmem_node(p, paddr >> PAGE_SHIFT,
1347 				  p->node_start_pfn, end_pfn);
1348 
1349 		numadbg("  free_bootmem_with_active_regions(%d, %lx)\n",
1350 			nid, end_pfn);
1351 		free_bootmem_with_active_regions(nid, end_pfn);
1352 
1353 		trim_reserved_in_node(nid);
1354 
1355 		numadbg("  sparse_memory_present_with_active_regions(%d)\n",
1356 			nid);
1357 		sparse_memory_present_with_active_regions(nid);
1358 	}
1359 }
1360 
1361 static unsigned long __init bootmem_init(unsigned long phys_base)
1362 {
1363 	unsigned long end_pfn;
1364 	int nid;
1365 
1366 	end_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1367 	max_pfn = max_low_pfn = end_pfn;
1368 	min_low_pfn = (phys_base >> PAGE_SHIFT);
1369 
1370 	if (bootmem_init_numa() < 0)
1371 		bootmem_init_nonnuma();
1372 
1373 	/* XXX cpu notifier XXX */
1374 
1375 	for_each_online_node(nid)
1376 		bootmem_init_one_node(nid);
1377 
1378 	sparse_init();
1379 
1380 	return end_pfn;
1381 }
1382 
1383 static struct linux_prom64_registers pall[MAX_BANKS] __initdata;
1384 static int pall_ents __initdata;
1385 
1386 #ifdef CONFIG_DEBUG_PAGEALLOC
1387 static unsigned long __ref kernel_map_range(unsigned long pstart,
1388 					    unsigned long pend, pgprot_t prot)
1389 {
1390 	unsigned long vstart = PAGE_OFFSET + pstart;
1391 	unsigned long vend = PAGE_OFFSET + pend;
1392 	unsigned long alloc_bytes = 0UL;
1393 
1394 	if ((vstart & ~PAGE_MASK) || (vend & ~PAGE_MASK)) {
1395 		prom_printf("kernel_map: Unaligned physmem[%lx:%lx]\n",
1396 			    vstart, vend);
1397 		prom_halt();
1398 	}
1399 
1400 	while (vstart < vend) {
1401 		unsigned long this_end, paddr = __pa(vstart);
1402 		pgd_t *pgd = pgd_offset_k(vstart);
1403 		pud_t *pud;
1404 		pmd_t *pmd;
1405 		pte_t *pte;
1406 
1407 		pud = pud_offset(pgd, vstart);
1408 		if (pud_none(*pud)) {
1409 			pmd_t *new;
1410 
1411 			new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1412 			alloc_bytes += PAGE_SIZE;
1413 			pud_populate(&init_mm, pud, new);
1414 		}
1415 
1416 		pmd = pmd_offset(pud, vstart);
1417 		if (!pmd_present(*pmd)) {
1418 			pte_t *new;
1419 
1420 			new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1421 			alloc_bytes += PAGE_SIZE;
1422 			pmd_populate_kernel(&init_mm, pmd, new);
1423 		}
1424 
1425 		pte = pte_offset_kernel(pmd, vstart);
1426 		this_end = (vstart + PMD_SIZE) & PMD_MASK;
1427 		if (this_end > vend)
1428 			this_end = vend;
1429 
1430 		while (vstart < this_end) {
1431 			pte_val(*pte) = (paddr | pgprot_val(prot));
1432 
1433 			vstart += PAGE_SIZE;
1434 			paddr += PAGE_SIZE;
1435 			pte++;
1436 		}
1437 	}
1438 
1439 	return alloc_bytes;
1440 }
1441 
1442 extern unsigned int kvmap_linear_patch[1];
1443 #endif /* CONFIG_DEBUG_PAGEALLOC */
1444 
1445 static void __init mark_kpte_bitmap(unsigned long start, unsigned long end)
1446 {
1447 	const unsigned long shift_256MB = 28;
1448 	const unsigned long mask_256MB = ((1UL << shift_256MB) - 1UL);
1449 	const unsigned long size_256MB = (1UL << shift_256MB);
1450 
1451 	while (start < end) {
1452 		long remains;
1453 
1454 		remains = end - start;
1455 		if (remains < size_256MB)
1456 			break;
1457 
1458 		if (start & mask_256MB) {
1459 			start = (start + size_256MB) & ~mask_256MB;
1460 			continue;
1461 		}
1462 
1463 		while (remains >= size_256MB) {
1464 			unsigned long index = start >> shift_256MB;
1465 
1466 			__set_bit(index, kpte_linear_bitmap);
1467 
1468 			start += size_256MB;
1469 			remains -= size_256MB;
1470 		}
1471 	}
1472 }
1473 
1474 static void __init init_kpte_bitmap(void)
1475 {
1476 	unsigned long i;
1477 
1478 	for (i = 0; i < pall_ents; i++) {
1479 		unsigned long phys_start, phys_end;
1480 
1481 		phys_start = pall[i].phys_addr;
1482 		phys_end = phys_start + pall[i].reg_size;
1483 
1484 		mark_kpte_bitmap(phys_start, phys_end);
1485 	}
1486 }
1487 
1488 static void __init kernel_physical_mapping_init(void)
1489 {
1490 #ifdef CONFIG_DEBUG_PAGEALLOC
1491 	unsigned long i, mem_alloced = 0UL;
1492 
1493 	for (i = 0; i < pall_ents; i++) {
1494 		unsigned long phys_start, phys_end;
1495 
1496 		phys_start = pall[i].phys_addr;
1497 		phys_end = phys_start + pall[i].reg_size;
1498 
1499 		mem_alloced += kernel_map_range(phys_start, phys_end,
1500 						PAGE_KERNEL);
1501 	}
1502 
1503 	printk("Allocated %ld bytes for kernel page tables.\n",
1504 	       mem_alloced);
1505 
1506 	kvmap_linear_patch[0] = 0x01000000; /* nop */
1507 	flushi(&kvmap_linear_patch[0]);
1508 
1509 	__flush_tlb_all();
1510 #endif
1511 }
1512 
1513 #ifdef CONFIG_DEBUG_PAGEALLOC
1514 void kernel_map_pages(struct page *page, int numpages, int enable)
1515 {
1516 	unsigned long phys_start = page_to_pfn(page) << PAGE_SHIFT;
1517 	unsigned long phys_end = phys_start + (numpages * PAGE_SIZE);
1518 
1519 	kernel_map_range(phys_start, phys_end,
1520 			 (enable ? PAGE_KERNEL : __pgprot(0)));
1521 
1522 	flush_tsb_kernel_range(PAGE_OFFSET + phys_start,
1523 			       PAGE_OFFSET + phys_end);
1524 
1525 	/* we should perform an IPI and flush all tlbs,
1526 	 * but that can deadlock->flush only current cpu.
1527 	 */
1528 	__flush_tlb_kernel_range(PAGE_OFFSET + phys_start,
1529 				 PAGE_OFFSET + phys_end);
1530 }
1531 #endif
1532 
1533 unsigned long __init find_ecache_flush_span(unsigned long size)
1534 {
1535 	int i;
1536 
1537 	for (i = 0; i < pavail_ents; i++) {
1538 		if (pavail[i].reg_size >= size)
1539 			return pavail[i].phys_addr;
1540 	}
1541 
1542 	return ~0UL;
1543 }
1544 
1545 static void __init tsb_phys_patch(void)
1546 {
1547 	struct tsb_ldquad_phys_patch_entry *pquad;
1548 	struct tsb_phys_patch_entry *p;
1549 
1550 	pquad = &__tsb_ldquad_phys_patch;
1551 	while (pquad < &__tsb_ldquad_phys_patch_end) {
1552 		unsigned long addr = pquad->addr;
1553 
1554 		if (tlb_type == hypervisor)
1555 			*(unsigned int *) addr = pquad->sun4v_insn;
1556 		else
1557 			*(unsigned int *) addr = pquad->sun4u_insn;
1558 		wmb();
1559 		__asm__ __volatile__("flush	%0"
1560 				     : /* no outputs */
1561 				     : "r" (addr));
1562 
1563 		pquad++;
1564 	}
1565 
1566 	p = &__tsb_phys_patch;
1567 	while (p < &__tsb_phys_patch_end) {
1568 		unsigned long addr = p->addr;
1569 
1570 		*(unsigned int *) addr = p->insn;
1571 		wmb();
1572 		__asm__ __volatile__("flush	%0"
1573 				     : /* no outputs */
1574 				     : "r" (addr));
1575 
1576 		p++;
1577 	}
1578 }
1579 
1580 /* Don't mark as init, we give this to the Hypervisor.  */
1581 #ifndef CONFIG_DEBUG_PAGEALLOC
1582 #define NUM_KTSB_DESCR	2
1583 #else
1584 #define NUM_KTSB_DESCR	1
1585 #endif
1586 static struct hv_tsb_descr ktsb_descr[NUM_KTSB_DESCR];
1587 extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
1588 
1589 static void patch_one_ktsb_phys(unsigned int *start, unsigned int *end, unsigned long pa)
1590 {
1591 	pa >>= KTSB_PHYS_SHIFT;
1592 
1593 	while (start < end) {
1594 		unsigned int *ia = (unsigned int *)(unsigned long)*start;
1595 
1596 		ia[0] = (ia[0] & ~0x3fffff) | (pa >> 10);
1597 		__asm__ __volatile__("flush	%0" : : "r" (ia));
1598 
1599 		ia[1] = (ia[1] & ~0x3ff) | (pa & 0x3ff);
1600 		__asm__ __volatile__("flush	%0" : : "r" (ia + 1));
1601 
1602 		start++;
1603 	}
1604 }
1605 
1606 static void ktsb_phys_patch(void)
1607 {
1608 	extern unsigned int __swapper_tsb_phys_patch;
1609 	extern unsigned int __swapper_tsb_phys_patch_end;
1610 	unsigned long ktsb_pa;
1611 
1612 	ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
1613 	patch_one_ktsb_phys(&__swapper_tsb_phys_patch,
1614 			    &__swapper_tsb_phys_patch_end, ktsb_pa);
1615 #ifndef CONFIG_DEBUG_PAGEALLOC
1616 	{
1617 	extern unsigned int __swapper_4m_tsb_phys_patch;
1618 	extern unsigned int __swapper_4m_tsb_phys_patch_end;
1619 	ktsb_pa = (kern_base +
1620 		   ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
1621 	patch_one_ktsb_phys(&__swapper_4m_tsb_phys_patch,
1622 			    &__swapper_4m_tsb_phys_patch_end, ktsb_pa);
1623 	}
1624 #endif
1625 }
1626 
1627 static void __init sun4v_ktsb_init(void)
1628 {
1629 	unsigned long ktsb_pa;
1630 
1631 	/* First KTSB for PAGE_SIZE mappings.  */
1632 	ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
1633 
1634 	switch (PAGE_SIZE) {
1635 	case 8 * 1024:
1636 	default:
1637 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_8K;
1638 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_8K;
1639 		break;
1640 
1641 	case 64 * 1024:
1642 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_64K;
1643 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_64K;
1644 		break;
1645 
1646 	case 512 * 1024:
1647 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_512K;
1648 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_512K;
1649 		break;
1650 
1651 	case 4 * 1024 * 1024:
1652 		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_4MB;
1653 		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_4MB;
1654 		break;
1655 	}
1656 
1657 	ktsb_descr[0].assoc = 1;
1658 	ktsb_descr[0].num_ttes = KERNEL_TSB_NENTRIES;
1659 	ktsb_descr[0].ctx_idx = 0;
1660 	ktsb_descr[0].tsb_base = ktsb_pa;
1661 	ktsb_descr[0].resv = 0;
1662 
1663 #ifndef CONFIG_DEBUG_PAGEALLOC
1664 	/* Second KTSB for 4MB/256MB mappings.  */
1665 	ktsb_pa = (kern_base +
1666 		   ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
1667 
1668 	ktsb_descr[1].pgsz_idx = HV_PGSZ_IDX_4MB;
1669 	ktsb_descr[1].pgsz_mask = (HV_PGSZ_MASK_4MB |
1670 				   HV_PGSZ_MASK_256MB);
1671 	ktsb_descr[1].assoc = 1;
1672 	ktsb_descr[1].num_ttes = KERNEL_TSB4M_NENTRIES;
1673 	ktsb_descr[1].ctx_idx = 0;
1674 	ktsb_descr[1].tsb_base = ktsb_pa;
1675 	ktsb_descr[1].resv = 0;
1676 #endif
1677 }
1678 
1679 void __cpuinit sun4v_ktsb_register(void)
1680 {
1681 	unsigned long pa, ret;
1682 
1683 	pa = kern_base + ((unsigned long)&ktsb_descr[0] - KERNBASE);
1684 
1685 	ret = sun4v_mmu_tsb_ctx0(NUM_KTSB_DESCR, pa);
1686 	if (ret != 0) {
1687 		prom_printf("hypervisor_mmu_tsb_ctx0[%lx]: "
1688 			    "errors with %lx\n", pa, ret);
1689 		prom_halt();
1690 	}
1691 }
1692 
1693 /* paging_init() sets up the page tables */
1694 
1695 static unsigned long last_valid_pfn;
1696 pgd_t swapper_pg_dir[2048];
1697 
1698 static void sun4u_pgprot_init(void);
1699 static void sun4v_pgprot_init(void);
1700 
1701 void __init paging_init(void)
1702 {
1703 	unsigned long end_pfn, shift, phys_base;
1704 	unsigned long real_end, i;
1705 
1706 	/* These build time checkes make sure that the dcache_dirty_cpu()
1707 	 * page->flags usage will work.
1708 	 *
1709 	 * When a page gets marked as dcache-dirty, we store the
1710 	 * cpu number starting at bit 32 in the page->flags.  Also,
1711 	 * functions like clear_dcache_dirty_cpu use the cpu mask
1712 	 * in 13-bit signed-immediate instruction fields.
1713 	 */
1714 
1715 	/*
1716 	 * Page flags must not reach into upper 32 bits that are used
1717 	 * for the cpu number
1718 	 */
1719 	BUILD_BUG_ON(NR_PAGEFLAGS > 32);
1720 
1721 	/*
1722 	 * The bit fields placed in the high range must not reach below
1723 	 * the 32 bit boundary. Otherwise we cannot place the cpu field
1724 	 * at the 32 bit boundary.
1725 	 */
1726 	BUILD_BUG_ON(SECTIONS_WIDTH + NODES_WIDTH + ZONES_WIDTH +
1727 		ilog2(roundup_pow_of_two(NR_CPUS)) > 32);
1728 
1729 	BUILD_BUG_ON(NR_CPUS > 4096);
1730 
1731 	kern_base = (prom_boot_mapping_phys_low >> 22UL) << 22UL;
1732 	kern_size = (unsigned long)&_end - (unsigned long)KERNBASE;
1733 
1734 	/* Invalidate both kernel TSBs.  */
1735 	memset(swapper_tsb, 0x40, sizeof(swapper_tsb));
1736 #ifndef CONFIG_DEBUG_PAGEALLOC
1737 	memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
1738 #endif
1739 
1740 	if (tlb_type == hypervisor)
1741 		sun4v_pgprot_init();
1742 	else
1743 		sun4u_pgprot_init();
1744 
1745 	if (tlb_type == cheetah_plus ||
1746 	    tlb_type == hypervisor) {
1747 		tsb_phys_patch();
1748 		ktsb_phys_patch();
1749 	}
1750 
1751 	if (tlb_type == hypervisor) {
1752 		sun4v_patch_tlb_handlers();
1753 		sun4v_ktsb_init();
1754 	}
1755 
1756 	/* Find available physical memory...
1757 	 *
1758 	 * Read it twice in order to work around a bug in openfirmware.
1759 	 * The call to grab this table itself can cause openfirmware to
1760 	 * allocate memory, which in turn can take away some space from
1761 	 * the list of available memory.  Reading it twice makes sure
1762 	 * we really do get the final value.
1763 	 */
1764 	read_obp_translations();
1765 	read_obp_memory("reg", &pall[0], &pall_ents);
1766 	read_obp_memory("available", &pavail[0], &pavail_ents);
1767 	read_obp_memory("available", &pavail[0], &pavail_ents);
1768 
1769 	phys_base = 0xffffffffffffffffUL;
1770 	for (i = 0; i < pavail_ents; i++) {
1771 		phys_base = min(phys_base, pavail[i].phys_addr);
1772 		memblock_add(pavail[i].phys_addr, pavail[i].reg_size);
1773 	}
1774 
1775 	memblock_reserve(kern_base, kern_size);
1776 
1777 	find_ramdisk(phys_base);
1778 
1779 	memblock_enforce_memory_limit(cmdline_memory_size);
1780 
1781 	memblock_allow_resize();
1782 	memblock_dump_all();
1783 
1784 	set_bit(0, mmu_context_bmap);
1785 
1786 	shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE);
1787 
1788 	real_end = (unsigned long)_end;
1789 	num_kernel_image_mappings = DIV_ROUND_UP(real_end - KERNBASE, 1 << 22);
1790 	printk("Kernel: Using %d locked TLB entries for main kernel image.\n",
1791 	       num_kernel_image_mappings);
1792 
1793 	/* Set kernel pgd to upper alias so physical page computations
1794 	 * work.
1795 	 */
1796 	init_mm.pgd += ((shift) / (sizeof(pgd_t)));
1797 
1798 	memset(swapper_low_pmd_dir, 0, sizeof(swapper_low_pmd_dir));
1799 
1800 	/* Now can init the kernel/bad page tables. */
1801 	pud_set(pud_offset(&swapper_pg_dir[0], 0),
1802 		swapper_low_pmd_dir + (shift / sizeof(pgd_t)));
1803 
1804 	inherit_prom_mappings();
1805 
1806 	init_kpte_bitmap();
1807 
1808 	/* Ok, we can use our TLB miss and window trap handlers safely.  */
1809 	setup_tba();
1810 
1811 	__flush_tlb_all();
1812 
1813 	if (tlb_type == hypervisor)
1814 		sun4v_ktsb_register();
1815 
1816 	prom_build_devicetree();
1817 	of_populate_present_mask();
1818 #ifndef CONFIG_SMP
1819 	of_fill_in_cpu_data();
1820 #endif
1821 
1822 	if (tlb_type == hypervisor) {
1823 		sun4v_mdesc_init();
1824 		mdesc_populate_present_mask(cpu_all_mask);
1825 #ifndef CONFIG_SMP
1826 		mdesc_fill_in_cpu_data(cpu_all_mask);
1827 #endif
1828 	}
1829 
1830 	/* Once the OF device tree and MDESC have been setup, we know
1831 	 * the list of possible cpus.  Therefore we can allocate the
1832 	 * IRQ stacks.
1833 	 */
1834 	for_each_possible_cpu(i) {
1835 		/* XXX Use node local allocations... XXX */
1836 		softirq_stack[i] = __va(memblock_alloc(THREAD_SIZE, THREAD_SIZE));
1837 		hardirq_stack[i] = __va(memblock_alloc(THREAD_SIZE, THREAD_SIZE));
1838 	}
1839 
1840 	/* Setup bootmem... */
1841 	last_valid_pfn = end_pfn = bootmem_init(phys_base);
1842 
1843 #ifndef CONFIG_NEED_MULTIPLE_NODES
1844 	max_mapnr = last_valid_pfn;
1845 #endif
1846 	kernel_physical_mapping_init();
1847 
1848 	{
1849 		unsigned long max_zone_pfns[MAX_NR_ZONES];
1850 
1851 		memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1852 
1853 		max_zone_pfns[ZONE_NORMAL] = end_pfn;
1854 
1855 		free_area_init_nodes(max_zone_pfns);
1856 	}
1857 
1858 	printk("Booting Linux...\n");
1859 }
1860 
1861 int __devinit page_in_phys_avail(unsigned long paddr)
1862 {
1863 	int i;
1864 
1865 	paddr &= PAGE_MASK;
1866 
1867 	for (i = 0; i < pavail_ents; i++) {
1868 		unsigned long start, end;
1869 
1870 		start = pavail[i].phys_addr;
1871 		end = start + pavail[i].reg_size;
1872 
1873 		if (paddr >= start && paddr < end)
1874 			return 1;
1875 	}
1876 	if (paddr >= kern_base && paddr < (kern_base + kern_size))
1877 		return 1;
1878 #ifdef CONFIG_BLK_DEV_INITRD
1879 	if (paddr >= __pa(initrd_start) &&
1880 	    paddr < __pa(PAGE_ALIGN(initrd_end)))
1881 		return 1;
1882 #endif
1883 
1884 	return 0;
1885 }
1886 
1887 static struct linux_prom64_registers pavail_rescan[MAX_BANKS] __initdata;
1888 static int pavail_rescan_ents __initdata;
1889 
1890 /* Certain OBP calls, such as fetching "available" properties, can
1891  * claim physical memory.  So, along with initializing the valid
1892  * address bitmap, what we do here is refetch the physical available
1893  * memory list again, and make sure it provides at least as much
1894  * memory as 'pavail' does.
1895  */
1896 static void __init setup_valid_addr_bitmap_from_pavail(unsigned long *bitmap)
1897 {
1898 	int i;
1899 
1900 	read_obp_memory("available", &pavail_rescan[0], &pavail_rescan_ents);
1901 
1902 	for (i = 0; i < pavail_ents; i++) {
1903 		unsigned long old_start, old_end;
1904 
1905 		old_start = pavail[i].phys_addr;
1906 		old_end = old_start + pavail[i].reg_size;
1907 		while (old_start < old_end) {
1908 			int n;
1909 
1910 			for (n = 0; n < pavail_rescan_ents; n++) {
1911 				unsigned long new_start, new_end;
1912 
1913 				new_start = pavail_rescan[n].phys_addr;
1914 				new_end = new_start +
1915 					pavail_rescan[n].reg_size;
1916 
1917 				if (new_start <= old_start &&
1918 				    new_end >= (old_start + PAGE_SIZE)) {
1919 					set_bit(old_start >> 22, bitmap);
1920 					goto do_next_page;
1921 				}
1922 			}
1923 
1924 			prom_printf("mem_init: Lost memory in pavail\n");
1925 			prom_printf("mem_init: OLD start[%lx] size[%lx]\n",
1926 				    pavail[i].phys_addr,
1927 				    pavail[i].reg_size);
1928 			prom_printf("mem_init: NEW start[%lx] size[%lx]\n",
1929 				    pavail_rescan[i].phys_addr,
1930 				    pavail_rescan[i].reg_size);
1931 			prom_printf("mem_init: Cannot continue, aborting.\n");
1932 			prom_halt();
1933 
1934 		do_next_page:
1935 			old_start += PAGE_SIZE;
1936 		}
1937 	}
1938 }
1939 
1940 static void __init patch_tlb_miss_handler_bitmap(void)
1941 {
1942 	extern unsigned int valid_addr_bitmap_insn[];
1943 	extern unsigned int valid_addr_bitmap_patch[];
1944 
1945 	valid_addr_bitmap_insn[1] = valid_addr_bitmap_patch[1];
1946 	mb();
1947 	valid_addr_bitmap_insn[0] = valid_addr_bitmap_patch[0];
1948 	flushi(&valid_addr_bitmap_insn[0]);
1949 }
1950 
1951 void __init mem_init(void)
1952 {
1953 	unsigned long codepages, datapages, initpages;
1954 	unsigned long addr, last;
1955 
1956 	addr = PAGE_OFFSET + kern_base;
1957 	last = PAGE_ALIGN(kern_size) + addr;
1958 	while (addr < last) {
1959 		set_bit(__pa(addr) >> 22, sparc64_valid_addr_bitmap);
1960 		addr += PAGE_SIZE;
1961 	}
1962 
1963 	setup_valid_addr_bitmap_from_pavail(sparc64_valid_addr_bitmap);
1964 	patch_tlb_miss_handler_bitmap();
1965 
1966 	high_memory = __va(last_valid_pfn << PAGE_SHIFT);
1967 
1968 #ifdef CONFIG_NEED_MULTIPLE_NODES
1969 	{
1970 		int i;
1971 		for_each_online_node(i) {
1972 			if (NODE_DATA(i)->node_spanned_pages != 0) {
1973 				totalram_pages +=
1974 					free_all_bootmem_node(NODE_DATA(i));
1975 			}
1976 		}
1977 	}
1978 #else
1979 	totalram_pages = free_all_bootmem();
1980 #endif
1981 
1982 	/* We subtract one to account for the mem_map_zero page
1983 	 * allocated below.
1984 	 */
1985 	totalram_pages -= 1;
1986 	num_physpages = totalram_pages;
1987 
1988 	/*
1989 	 * Set up the zero page, mark it reserved, so that page count
1990 	 * is not manipulated when freeing the page from user ptes.
1991 	 */
1992 	mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0);
1993 	if (mem_map_zero == NULL) {
1994 		prom_printf("paging_init: Cannot alloc zero page.\n");
1995 		prom_halt();
1996 	}
1997 	SetPageReserved(mem_map_zero);
1998 
1999 	codepages = (((unsigned long) _etext) - ((unsigned long) _start));
2000 	codepages = PAGE_ALIGN(codepages) >> PAGE_SHIFT;
2001 	datapages = (((unsigned long) _edata) - ((unsigned long) _etext));
2002 	datapages = PAGE_ALIGN(datapages) >> PAGE_SHIFT;
2003 	initpages = (((unsigned long) __init_end) - ((unsigned long) __init_begin));
2004 	initpages = PAGE_ALIGN(initpages) >> PAGE_SHIFT;
2005 
2006 	printk("Memory: %luk available (%ldk kernel code, %ldk data, %ldk init) [%016lx,%016lx]\n",
2007 	       nr_free_pages() << (PAGE_SHIFT-10),
2008 	       codepages << (PAGE_SHIFT-10),
2009 	       datapages << (PAGE_SHIFT-10),
2010 	       initpages << (PAGE_SHIFT-10),
2011 	       PAGE_OFFSET, (last_valid_pfn << PAGE_SHIFT));
2012 
2013 	if (tlb_type == cheetah || tlb_type == cheetah_plus)
2014 		cheetah_ecache_flush_init();
2015 }
2016 
2017 void free_initmem(void)
2018 {
2019 	unsigned long addr, initend;
2020 	int do_free = 1;
2021 
2022 	/* If the physical memory maps were trimmed by kernel command
2023 	 * line options, don't even try freeing this initmem stuff up.
2024 	 * The kernel image could have been in the trimmed out region
2025 	 * and if so the freeing below will free invalid page structs.
2026 	 */
2027 	if (cmdline_memory_size)
2028 		do_free = 0;
2029 
2030 	/*
2031 	 * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes.
2032 	 */
2033 	addr = PAGE_ALIGN((unsigned long)(__init_begin));
2034 	initend = (unsigned long)(__init_end) & PAGE_MASK;
2035 	for (; addr < initend; addr += PAGE_SIZE) {
2036 		unsigned long page;
2037 		struct page *p;
2038 
2039 		page = (addr +
2040 			((unsigned long) __va(kern_base)) -
2041 			((unsigned long) KERNBASE));
2042 		memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
2043 
2044 		if (do_free) {
2045 			p = virt_to_page(page);
2046 
2047 			ClearPageReserved(p);
2048 			init_page_count(p);
2049 			__free_page(p);
2050 			num_physpages++;
2051 			totalram_pages++;
2052 		}
2053 	}
2054 }
2055 
2056 #ifdef CONFIG_BLK_DEV_INITRD
2057 void free_initrd_mem(unsigned long start, unsigned long end)
2058 {
2059 	if (start < end)
2060 		printk ("Freeing initrd memory: %ldk freed\n", (end - start) >> 10);
2061 	for (; start < end; start += PAGE_SIZE) {
2062 		struct page *p = virt_to_page(start);
2063 
2064 		ClearPageReserved(p);
2065 		init_page_count(p);
2066 		__free_page(p);
2067 		num_physpages++;
2068 		totalram_pages++;
2069 	}
2070 }
2071 #endif
2072 
2073 #define _PAGE_CACHE_4U	(_PAGE_CP_4U | _PAGE_CV_4U)
2074 #define _PAGE_CACHE_4V	(_PAGE_CP_4V | _PAGE_CV_4V)
2075 #define __DIRTY_BITS_4U	 (_PAGE_MODIFIED_4U | _PAGE_WRITE_4U | _PAGE_W_4U)
2076 #define __DIRTY_BITS_4V	 (_PAGE_MODIFIED_4V | _PAGE_WRITE_4V | _PAGE_W_4V)
2077 #define __ACCESS_BITS_4U (_PAGE_ACCESSED_4U | _PAGE_READ_4U | _PAGE_R)
2078 #define __ACCESS_BITS_4V (_PAGE_ACCESSED_4V | _PAGE_READ_4V | _PAGE_R)
2079 
2080 pgprot_t PAGE_KERNEL __read_mostly;
2081 EXPORT_SYMBOL(PAGE_KERNEL);
2082 
2083 pgprot_t PAGE_KERNEL_LOCKED __read_mostly;
2084 pgprot_t PAGE_COPY __read_mostly;
2085 
2086 pgprot_t PAGE_SHARED __read_mostly;
2087 EXPORT_SYMBOL(PAGE_SHARED);
2088 
2089 unsigned long pg_iobits __read_mostly;
2090 
2091 unsigned long _PAGE_IE __read_mostly;
2092 EXPORT_SYMBOL(_PAGE_IE);
2093 
2094 unsigned long _PAGE_E __read_mostly;
2095 EXPORT_SYMBOL(_PAGE_E);
2096 
2097 unsigned long _PAGE_CACHE __read_mostly;
2098 EXPORT_SYMBOL(_PAGE_CACHE);
2099 
2100 #ifdef CONFIG_SPARSEMEM_VMEMMAP
2101 unsigned long vmemmap_table[VMEMMAP_SIZE];
2102 
2103 int __meminit vmemmap_populate(struct page *start, unsigned long nr, int node)
2104 {
2105 	unsigned long vstart = (unsigned long) start;
2106 	unsigned long vend = (unsigned long) (start + nr);
2107 	unsigned long phys_start = (vstart - VMEMMAP_BASE);
2108 	unsigned long phys_end = (vend - VMEMMAP_BASE);
2109 	unsigned long addr = phys_start & VMEMMAP_CHUNK_MASK;
2110 	unsigned long end = VMEMMAP_ALIGN(phys_end);
2111 	unsigned long pte_base;
2112 
2113 	pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2114 		    _PAGE_CP_4U | _PAGE_CV_4U |
2115 		    _PAGE_P_4U | _PAGE_W_4U);
2116 	if (tlb_type == hypervisor)
2117 		pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2118 			    _PAGE_CP_4V | _PAGE_CV_4V |
2119 			    _PAGE_P_4V | _PAGE_W_4V);
2120 
2121 	for (; addr < end; addr += VMEMMAP_CHUNK) {
2122 		unsigned long *vmem_pp =
2123 			vmemmap_table + (addr >> VMEMMAP_CHUNK_SHIFT);
2124 		void *block;
2125 
2126 		if (!(*vmem_pp & _PAGE_VALID)) {
2127 			block = vmemmap_alloc_block(1UL << 22, node);
2128 			if (!block)
2129 				return -ENOMEM;
2130 
2131 			*vmem_pp = pte_base | __pa(block);
2132 
2133 			printk(KERN_INFO "[%p-%p] page_structs=%lu "
2134 			       "node=%d entry=%lu/%lu\n", start, block, nr,
2135 			       node,
2136 			       addr >> VMEMMAP_CHUNK_SHIFT,
2137 			       VMEMMAP_SIZE);
2138 		}
2139 	}
2140 	return 0;
2141 }
2142 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
2143 
2144 static void prot_init_common(unsigned long page_none,
2145 			     unsigned long page_shared,
2146 			     unsigned long page_copy,
2147 			     unsigned long page_readonly,
2148 			     unsigned long page_exec_bit)
2149 {
2150 	PAGE_COPY = __pgprot(page_copy);
2151 	PAGE_SHARED = __pgprot(page_shared);
2152 
2153 	protection_map[0x0] = __pgprot(page_none);
2154 	protection_map[0x1] = __pgprot(page_readonly & ~page_exec_bit);
2155 	protection_map[0x2] = __pgprot(page_copy & ~page_exec_bit);
2156 	protection_map[0x3] = __pgprot(page_copy & ~page_exec_bit);
2157 	protection_map[0x4] = __pgprot(page_readonly);
2158 	protection_map[0x5] = __pgprot(page_readonly);
2159 	protection_map[0x6] = __pgprot(page_copy);
2160 	protection_map[0x7] = __pgprot(page_copy);
2161 	protection_map[0x8] = __pgprot(page_none);
2162 	protection_map[0x9] = __pgprot(page_readonly & ~page_exec_bit);
2163 	protection_map[0xa] = __pgprot(page_shared & ~page_exec_bit);
2164 	protection_map[0xb] = __pgprot(page_shared & ~page_exec_bit);
2165 	protection_map[0xc] = __pgprot(page_readonly);
2166 	protection_map[0xd] = __pgprot(page_readonly);
2167 	protection_map[0xe] = __pgprot(page_shared);
2168 	protection_map[0xf] = __pgprot(page_shared);
2169 }
2170 
2171 static void __init sun4u_pgprot_init(void)
2172 {
2173 	unsigned long page_none, page_shared, page_copy, page_readonly;
2174 	unsigned long page_exec_bit;
2175 
2176 	PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2177 				_PAGE_CACHE_4U | _PAGE_P_4U |
2178 				__ACCESS_BITS_4U | __DIRTY_BITS_4U |
2179 				_PAGE_EXEC_4U);
2180 	PAGE_KERNEL_LOCKED = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2181 				       _PAGE_CACHE_4U | _PAGE_P_4U |
2182 				       __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2183 				       _PAGE_EXEC_4U | _PAGE_L_4U);
2184 
2185 	_PAGE_IE = _PAGE_IE_4U;
2186 	_PAGE_E = _PAGE_E_4U;
2187 	_PAGE_CACHE = _PAGE_CACHE_4U;
2188 
2189 	pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4U | __DIRTY_BITS_4U |
2190 		     __ACCESS_BITS_4U | _PAGE_E_4U);
2191 
2192 #ifdef CONFIG_DEBUG_PAGEALLOC
2193 	kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZBITS_4U) ^
2194 		0xfffff80000000000UL;
2195 #else
2196 	kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4U) ^
2197 		0xfffff80000000000UL;
2198 #endif
2199 	kern_linear_pte_xor[0] |= (_PAGE_CP_4U | _PAGE_CV_4U |
2200 				   _PAGE_P_4U | _PAGE_W_4U);
2201 
2202 	/* XXX Should use 256MB on Panther. XXX */
2203 	kern_linear_pte_xor[1] = kern_linear_pte_xor[0];
2204 
2205 	_PAGE_SZBITS = _PAGE_SZBITS_4U;
2206 	_PAGE_ALL_SZ_BITS =  (_PAGE_SZ4MB_4U | _PAGE_SZ512K_4U |
2207 			      _PAGE_SZ64K_4U | _PAGE_SZ8K_4U |
2208 			      _PAGE_SZ32MB_4U | _PAGE_SZ256MB_4U);
2209 
2210 
2211 	page_none = _PAGE_PRESENT_4U | _PAGE_ACCESSED_4U | _PAGE_CACHE_4U;
2212 	page_shared = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2213 		       __ACCESS_BITS_4U | _PAGE_WRITE_4U | _PAGE_EXEC_4U);
2214 	page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2215 		       __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2216 	page_readonly   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2217 			   __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2218 
2219 	page_exec_bit = _PAGE_EXEC_4U;
2220 
2221 	prot_init_common(page_none, page_shared, page_copy, page_readonly,
2222 			 page_exec_bit);
2223 }
2224 
2225 static void __init sun4v_pgprot_init(void)
2226 {
2227 	unsigned long page_none, page_shared, page_copy, page_readonly;
2228 	unsigned long page_exec_bit;
2229 
2230 	PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4V | _PAGE_VALID |
2231 				_PAGE_CACHE_4V | _PAGE_P_4V |
2232 				__ACCESS_BITS_4V | __DIRTY_BITS_4V |
2233 				_PAGE_EXEC_4V);
2234 	PAGE_KERNEL_LOCKED = PAGE_KERNEL;
2235 
2236 	_PAGE_IE = _PAGE_IE_4V;
2237 	_PAGE_E = _PAGE_E_4V;
2238 	_PAGE_CACHE = _PAGE_CACHE_4V;
2239 
2240 #ifdef CONFIG_DEBUG_PAGEALLOC
2241 	kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZBITS_4V) ^
2242 		0xfffff80000000000UL;
2243 #else
2244 	kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4V) ^
2245 		0xfffff80000000000UL;
2246 #endif
2247 	kern_linear_pte_xor[0] |= (_PAGE_CP_4V | _PAGE_CV_4V |
2248 				   _PAGE_P_4V | _PAGE_W_4V);
2249 
2250 #ifdef CONFIG_DEBUG_PAGEALLOC
2251 	kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZBITS_4V) ^
2252 		0xfffff80000000000UL;
2253 #else
2254 	kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZ256MB_4V) ^
2255 		0xfffff80000000000UL;
2256 #endif
2257 	kern_linear_pte_xor[1] |= (_PAGE_CP_4V | _PAGE_CV_4V |
2258 				   _PAGE_P_4V | _PAGE_W_4V);
2259 
2260 	pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4V | __DIRTY_BITS_4V |
2261 		     __ACCESS_BITS_4V | _PAGE_E_4V);
2262 
2263 	_PAGE_SZBITS = _PAGE_SZBITS_4V;
2264 	_PAGE_ALL_SZ_BITS = (_PAGE_SZ16GB_4V | _PAGE_SZ2GB_4V |
2265 			     _PAGE_SZ256MB_4V | _PAGE_SZ32MB_4V |
2266 			     _PAGE_SZ4MB_4V | _PAGE_SZ512K_4V |
2267 			     _PAGE_SZ64K_4V | _PAGE_SZ8K_4V);
2268 
2269 	page_none = _PAGE_PRESENT_4V | _PAGE_ACCESSED_4V | _PAGE_CACHE_4V;
2270 	page_shared = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2271 		       __ACCESS_BITS_4V | _PAGE_WRITE_4V | _PAGE_EXEC_4V);
2272 	page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2273 		       __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2274 	page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2275 			 __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2276 
2277 	page_exec_bit = _PAGE_EXEC_4V;
2278 
2279 	prot_init_common(page_none, page_shared, page_copy, page_readonly,
2280 			 page_exec_bit);
2281 }
2282 
2283 unsigned long pte_sz_bits(unsigned long sz)
2284 {
2285 	if (tlb_type == hypervisor) {
2286 		switch (sz) {
2287 		case 8 * 1024:
2288 		default:
2289 			return _PAGE_SZ8K_4V;
2290 		case 64 * 1024:
2291 			return _PAGE_SZ64K_4V;
2292 		case 512 * 1024:
2293 			return _PAGE_SZ512K_4V;
2294 		case 4 * 1024 * 1024:
2295 			return _PAGE_SZ4MB_4V;
2296 		}
2297 	} else {
2298 		switch (sz) {
2299 		case 8 * 1024:
2300 		default:
2301 			return _PAGE_SZ8K_4U;
2302 		case 64 * 1024:
2303 			return _PAGE_SZ64K_4U;
2304 		case 512 * 1024:
2305 			return _PAGE_SZ512K_4U;
2306 		case 4 * 1024 * 1024:
2307 			return _PAGE_SZ4MB_4U;
2308 		}
2309 	}
2310 }
2311 
2312 pte_t mk_pte_io(unsigned long page, pgprot_t prot, int space, unsigned long page_size)
2313 {
2314 	pte_t pte;
2315 
2316 	pte_val(pte)  = page | pgprot_val(pgprot_noncached(prot));
2317 	pte_val(pte) |= (((unsigned long)space) << 32);
2318 	pte_val(pte) |= pte_sz_bits(page_size);
2319 
2320 	return pte;
2321 }
2322 
2323 static unsigned long kern_large_tte(unsigned long paddr)
2324 {
2325 	unsigned long val;
2326 
2327 	val = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2328 	       _PAGE_CP_4U | _PAGE_CV_4U | _PAGE_P_4U |
2329 	       _PAGE_EXEC_4U | _PAGE_L_4U | _PAGE_W_4U);
2330 	if (tlb_type == hypervisor)
2331 		val = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2332 		       _PAGE_CP_4V | _PAGE_CV_4V | _PAGE_P_4V |
2333 		       _PAGE_EXEC_4V | _PAGE_W_4V);
2334 
2335 	return val | paddr;
2336 }
2337 
2338 /* If not locked, zap it. */
2339 void __flush_tlb_all(void)
2340 {
2341 	unsigned long pstate;
2342 	int i;
2343 
2344 	__asm__ __volatile__("flushw\n\t"
2345 			     "rdpr	%%pstate, %0\n\t"
2346 			     "wrpr	%0, %1, %%pstate"
2347 			     : "=r" (pstate)
2348 			     : "i" (PSTATE_IE));
2349 	if (tlb_type == hypervisor) {
2350 		sun4v_mmu_demap_all();
2351 	} else if (tlb_type == spitfire) {
2352 		for (i = 0; i < 64; i++) {
2353 			/* Spitfire Errata #32 workaround */
2354 			/* NOTE: Always runs on spitfire, so no
2355 			 *       cheetah+ page size encodings.
2356 			 */
2357 			__asm__ __volatile__("stxa	%0, [%1] %2\n\t"
2358 					     "flush	%%g6"
2359 					     : /* No outputs */
2360 					     : "r" (0),
2361 					     "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2362 
2363 			if (!(spitfire_get_dtlb_data(i) & _PAGE_L_4U)) {
2364 				__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2365 						     "membar #Sync"
2366 						     : /* no outputs */
2367 						     : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
2368 				spitfire_put_dtlb_data(i, 0x0UL);
2369 			}
2370 
2371 			/* Spitfire Errata #32 workaround */
2372 			/* NOTE: Always runs on spitfire, so no
2373 			 *       cheetah+ page size encodings.
2374 			 */
2375 			__asm__ __volatile__("stxa	%0, [%1] %2\n\t"
2376 					     "flush	%%g6"
2377 					     : /* No outputs */
2378 					     : "r" (0),
2379 					     "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2380 
2381 			if (!(spitfire_get_itlb_data(i) & _PAGE_L_4U)) {
2382 				__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2383 						     "membar #Sync"
2384 						     : /* no outputs */
2385 						     : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
2386 				spitfire_put_itlb_data(i, 0x0UL);
2387 			}
2388 		}
2389 	} else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
2390 		cheetah_flush_dtlb_all();
2391 		cheetah_flush_itlb_all();
2392 	}
2393 	__asm__ __volatile__("wrpr	%0, 0, %%pstate"
2394 			     : : "r" (pstate));
2395 }
2396