1/*
2 *  arch/sparc64/mm/init.c
3 *
4 *  Copyright (C) 1996-1999 David S. Miller (davem@caip.rutgers.edu)
5 *  Copyright (C) 1997-1999 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
6 */
7
8#include <linux/module.h>
9#include <linux/kernel.h>
10#include <linux/sched.h>
11#include <linux/string.h>
12#include <linux/init.h>
13#include <linux/bootmem.h>
14#include <linux/mm.h>
15#include <linux/hugetlb.h>
16#include <linux/initrd.h>
17#include <linux/swap.h>
18#include <linux/pagemap.h>
19#include <linux/poison.h>
20#include <linux/fs.h>
21#include <linux/seq_file.h>
22#include <linux/kprobes.h>
23#include <linux/cache.h>
24#include <linux/sort.h>
25#include <linux/ioport.h>
26#include <linux/percpu.h>
27#include <linux/memblock.h>
28#include <linux/mmzone.h>
29#include <linux/gfp.h>
30
31#include <asm/head.h>
32#include <asm/page.h>
33#include <asm/pgalloc.h>
34#include <asm/pgtable.h>
35#include <asm/oplib.h>
36#include <asm/iommu.h>
37#include <asm/io.h>
38#include <asm/uaccess.h>
39#include <asm/mmu_context.h>
40#include <asm/tlbflush.h>
41#include <asm/dma.h>
42#include <asm/starfire.h>
43#include <asm/tlb.h>
44#include <asm/spitfire.h>
45#include <asm/sections.h>
46#include <asm/tsb.h>
47#include <asm/hypervisor.h>
48#include <asm/prom.h>
49#include <asm/mdesc.h>
50#include <asm/cpudata.h>
51#include <asm/setup.h>
52#include <asm/irq.h>
53
54#include "init_64.h"
55
56unsigned long kern_linear_pte_xor[4] __read_mostly;
57static unsigned long page_cache4v_flag;
58
59/* A bitmap, two bits for every 256MB of physical memory.  These two
60 * bits determine what page size we use for kernel linear
61 * translations.  They form an index into kern_linear_pte_xor[].  The
62 * value in the indexed slot is XOR'd with the TLB miss virtual
63 * address to form the resulting TTE.  The mapping is:
64 *
65 *	0	==>	4MB
66 *	1	==>	256MB
67 *	2	==>	2GB
68 *	3	==>	16GB
69 *
70 * All sun4v chips support 256MB pages.  Only SPARC-T4 and later
71 * support 2GB pages, and hopefully future cpus will support the 16GB
72 * pages as well.  For slots 2 and 3, we encode a 256MB TTE xor there
73 * if these larger page sizes are not supported by the cpu.
74 *
75 * It would be nice to determine this from the machine description
76 * 'cpu' properties, but we need to have this table setup before the
77 * MDESC is initialized.
78 */
79
80#ifndef CONFIG_DEBUG_PAGEALLOC
81/* A special kernel TSB for 4MB, 256MB, 2GB and 16GB linear mappings.
82 * Space is allocated for this right after the trap table in
83 * arch/sparc64/kernel/head.S
84 */
85extern struct tsb swapper_4m_tsb[KERNEL_TSB4M_NENTRIES];
86#endif
87extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
88
89static unsigned long cpu_pgsz_mask;
90
91#define MAX_BANKS	1024
92
93static struct linux_prom64_registers pavail[MAX_BANKS];
94static int pavail_ents;
95
96static int cmp_p64(const void *a, const void *b)
97{
98	const struct linux_prom64_registers *x = a, *y = b;
99
100	if (x->phys_addr > y->phys_addr)
101		return 1;
102	if (x->phys_addr < y->phys_addr)
103		return -1;
104	return 0;
105}
106
107static void __init read_obp_memory(const char *property,
108				   struct linux_prom64_registers *regs,
109				   int *num_ents)
110{
111	phandle node = prom_finddevice("/memory");
112	int prop_size = prom_getproplen(node, property);
113	int ents, ret, i;
114
115	ents = prop_size / sizeof(struct linux_prom64_registers);
116	if (ents > MAX_BANKS) {
117		prom_printf("The machine has more %s property entries than "
118			    "this kernel can support (%d).\n",
119			    property, MAX_BANKS);
120		prom_halt();
121	}
122
123	ret = prom_getproperty(node, property, (char *) regs, prop_size);
124	if (ret == -1) {
125		prom_printf("Couldn't get %s property from /memory.\n",
126				property);
127		prom_halt();
128	}
129
130	/* Sanitize what we got from the firmware, by page aligning
131	 * everything.
132	 */
133	for (i = 0; i < ents; i++) {
134		unsigned long base, size;
135
136		base = regs[i].phys_addr;
137		size = regs[i].reg_size;
138
139		size &= PAGE_MASK;
140		if (base & ~PAGE_MASK) {
141			unsigned long new_base = PAGE_ALIGN(base);
142
143			size -= new_base - base;
144			if ((long) size < 0L)
145				size = 0UL;
146			base = new_base;
147		}
148		if (size == 0UL) {
149			/* If it is empty, simply get rid of it.
150			 * This simplifies the logic of the other
151			 * functions that process these arrays.
152			 */
153			memmove(&regs[i], &regs[i + 1],
154				(ents - i - 1) * sizeof(regs[0]));
155			i--;
156			ents--;
157			continue;
158		}
159		regs[i].phys_addr = base;
160		regs[i].reg_size = size;
161	}
162
163	*num_ents = ents;
164
165	sort(regs, ents, sizeof(struct linux_prom64_registers),
166	     cmp_p64, NULL);
167}
168
169/* Kernel physical address base and size in bytes.  */
170unsigned long kern_base __read_mostly;
171unsigned long kern_size __read_mostly;
172
173/* Initial ramdisk setup */
174extern unsigned long sparc_ramdisk_image64;
175extern unsigned int sparc_ramdisk_image;
176extern unsigned int sparc_ramdisk_size;
177
178struct page *mem_map_zero __read_mostly;
179EXPORT_SYMBOL(mem_map_zero);
180
181unsigned int sparc64_highest_unlocked_tlb_ent __read_mostly;
182
183unsigned long sparc64_kern_pri_context __read_mostly;
184unsigned long sparc64_kern_pri_nuc_bits __read_mostly;
185unsigned long sparc64_kern_sec_context __read_mostly;
186
187int num_kernel_image_mappings;
188
189#ifdef CONFIG_DEBUG_DCFLUSH
190atomic_t dcpage_flushes = ATOMIC_INIT(0);
191#ifdef CONFIG_SMP
192atomic_t dcpage_flushes_xcall = ATOMIC_INIT(0);
193#endif
194#endif
195
196inline void flush_dcache_page_impl(struct page *page)
197{
198	BUG_ON(tlb_type == hypervisor);
199#ifdef CONFIG_DEBUG_DCFLUSH
200	atomic_inc(&dcpage_flushes);
201#endif
202
203#ifdef DCACHE_ALIASING_POSSIBLE
204	__flush_dcache_page(page_address(page),
205			    ((tlb_type == spitfire) &&
206			     page_mapping(page) != NULL));
207#else
208	if (page_mapping(page) != NULL &&
209	    tlb_type == spitfire)
210		__flush_icache_page(__pa(page_address(page)));
211#endif
212}
213
214#define PG_dcache_dirty		PG_arch_1
215#define PG_dcache_cpu_shift	32UL
216#define PG_dcache_cpu_mask	\
217	((1UL<<ilog2(roundup_pow_of_two(NR_CPUS)))-1UL)
218
219#define dcache_dirty_cpu(page) \
220	(((page)->flags >> PG_dcache_cpu_shift) & PG_dcache_cpu_mask)
221
222static inline void set_dcache_dirty(struct page *page, int this_cpu)
223{
224	unsigned long mask = this_cpu;
225	unsigned long non_cpu_bits;
226
227	non_cpu_bits = ~(PG_dcache_cpu_mask << PG_dcache_cpu_shift);
228	mask = (mask << PG_dcache_cpu_shift) | (1UL << PG_dcache_dirty);
229
230	__asm__ __volatile__("1:\n\t"
231			     "ldx	[%2], %%g7\n\t"
232			     "and	%%g7, %1, %%g1\n\t"
233			     "or	%%g1, %0, %%g1\n\t"
234			     "casx	[%2], %%g7, %%g1\n\t"
235			     "cmp	%%g7, %%g1\n\t"
236			     "bne,pn	%%xcc, 1b\n\t"
237			     " nop"
238			     : /* no outputs */
239			     : "r" (mask), "r" (non_cpu_bits), "r" (&page->flags)
240			     : "g1", "g7");
241}
242
243static inline void clear_dcache_dirty_cpu(struct page *page, unsigned long cpu)
244{
245	unsigned long mask = (1UL << PG_dcache_dirty);
246
247	__asm__ __volatile__("! test_and_clear_dcache_dirty\n"
248			     "1:\n\t"
249			     "ldx	[%2], %%g7\n\t"
250			     "srlx	%%g7, %4, %%g1\n\t"
251			     "and	%%g1, %3, %%g1\n\t"
252			     "cmp	%%g1, %0\n\t"
253			     "bne,pn	%%icc, 2f\n\t"
254			     " andn	%%g7, %1, %%g1\n\t"
255			     "casx	[%2], %%g7, %%g1\n\t"
256			     "cmp	%%g7, %%g1\n\t"
257			     "bne,pn	%%xcc, 1b\n\t"
258			     " nop\n"
259			     "2:"
260			     : /* no outputs */
261			     : "r" (cpu), "r" (mask), "r" (&page->flags),
262			       "i" (PG_dcache_cpu_mask),
263			       "i" (PG_dcache_cpu_shift)
264			     : "g1", "g7");
265}
266
267static inline void tsb_insert(struct tsb *ent, unsigned long tag, unsigned long pte)
268{
269	unsigned long tsb_addr = (unsigned long) ent;
270
271	if (tlb_type == cheetah_plus || tlb_type == hypervisor)
272		tsb_addr = __pa(tsb_addr);
273
274	__tsb_insert(tsb_addr, tag, pte);
275}
276
277unsigned long _PAGE_ALL_SZ_BITS __read_mostly;
278
279static void flush_dcache(unsigned long pfn)
280{
281	struct page *page;
282
283	page = pfn_to_page(pfn);
284	if (page) {
285		unsigned long pg_flags;
286
287		pg_flags = page->flags;
288		if (pg_flags & (1UL << PG_dcache_dirty)) {
289			int cpu = ((pg_flags >> PG_dcache_cpu_shift) &
290				   PG_dcache_cpu_mask);
291			int this_cpu = get_cpu();
292
293			/* This is just to optimize away some function calls
294			 * in the SMP case.
295			 */
296			if (cpu == this_cpu)
297				flush_dcache_page_impl(page);
298			else
299				smp_flush_dcache_page_impl(page, cpu);
300
301			clear_dcache_dirty_cpu(page, cpu);
302
303			put_cpu();
304		}
305	}
306}
307
308/* mm->context.lock must be held */
309static void __update_mmu_tsb_insert(struct mm_struct *mm, unsigned long tsb_index,
310				    unsigned long tsb_hash_shift, unsigned long address,
311				    unsigned long tte)
312{
313	struct tsb *tsb = mm->context.tsb_block[tsb_index].tsb;
314	unsigned long tag;
315
316	if (unlikely(!tsb))
317		return;
318
319	tsb += ((address >> tsb_hash_shift) &
320		(mm->context.tsb_block[tsb_index].tsb_nentries - 1UL));
321	tag = (address >> 22UL);
322	tsb_insert(tsb, tag, tte);
323}
324
325#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
326static inline bool is_hugetlb_pte(pte_t pte)
327{
328	if ((tlb_type == hypervisor &&
329	     (pte_val(pte) & _PAGE_SZALL_4V) == _PAGE_SZHUGE_4V) ||
330	    (tlb_type != hypervisor &&
331	     (pte_val(pte) & _PAGE_SZALL_4U) == _PAGE_SZHUGE_4U))
332		return true;
333	return false;
334}
335#endif
336
337void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
338{
339	struct mm_struct *mm;
340	unsigned long flags;
341	pte_t pte = *ptep;
342
343	if (tlb_type != hypervisor) {
344		unsigned long pfn = pte_pfn(pte);
345
346		if (pfn_valid(pfn))
347			flush_dcache(pfn);
348	}
349
350	mm = vma->vm_mm;
351
352	/* Don't insert a non-valid PTE into the TSB, we'll deadlock.  */
353	if (!pte_accessible(mm, pte))
354		return;
355
356	spin_lock_irqsave(&mm->context.lock, flags);
357
358#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
359	if (mm->context.huge_pte_count && is_hugetlb_pte(pte))
360		__update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
361					address, pte_val(pte));
362	else
363#endif
364		__update_mmu_tsb_insert(mm, MM_TSB_BASE, PAGE_SHIFT,
365					address, pte_val(pte));
366
367	spin_unlock_irqrestore(&mm->context.lock, flags);
368}
369
370void flush_dcache_page(struct page *page)
371{
372	struct address_space *mapping;
373	int this_cpu;
374
375	if (tlb_type == hypervisor)
376		return;
377
378	/* Do not bother with the expensive D-cache flush if it
379	 * is merely the zero page.  The 'bigcore' testcase in GDB
380	 * causes this case to run millions of times.
381	 */
382	if (page == ZERO_PAGE(0))
383		return;
384
385	this_cpu = get_cpu();
386
387	mapping = page_mapping(page);
388	if (mapping && !mapping_mapped(mapping)) {
389		int dirty = test_bit(PG_dcache_dirty, &page->flags);
390		if (dirty) {
391			int dirty_cpu = dcache_dirty_cpu(page);
392
393			if (dirty_cpu == this_cpu)
394				goto out;
395			smp_flush_dcache_page_impl(page, dirty_cpu);
396		}
397		set_dcache_dirty(page, this_cpu);
398	} else {
399		/* We could delay the flush for the !page_mapping
400		 * case too.  But that case is for exec env/arg
401		 * pages and those are %99 certainly going to get
402		 * faulted into the tlb (and thus flushed) anyways.
403		 */
404		flush_dcache_page_impl(page);
405	}
406
407out:
408	put_cpu();
409}
410EXPORT_SYMBOL(flush_dcache_page);
411
412void __kprobes flush_icache_range(unsigned long start, unsigned long end)
413{
414	/* Cheetah and Hypervisor platform cpus have coherent I-cache. */
415	if (tlb_type == spitfire) {
416		unsigned long kaddr;
417
418		/* This code only runs on Spitfire cpus so this is
419		 * why we can assume _PAGE_PADDR_4U.
420		 */
421		for (kaddr = start; kaddr < end; kaddr += PAGE_SIZE) {
422			unsigned long paddr, mask = _PAGE_PADDR_4U;
423
424			if (kaddr >= PAGE_OFFSET)
425				paddr = kaddr & mask;
426			else {
427				pgd_t *pgdp = pgd_offset_k(kaddr);
428				pud_t *pudp = pud_offset(pgdp, kaddr);
429				pmd_t *pmdp = pmd_offset(pudp, kaddr);
430				pte_t *ptep = pte_offset_kernel(pmdp, kaddr);
431
432				paddr = pte_val(*ptep) & mask;
433			}
434			__flush_icache_page(paddr);
435		}
436	}
437}
438EXPORT_SYMBOL(flush_icache_range);
439
440void mmu_info(struct seq_file *m)
441{
442	static const char *pgsz_strings[] = {
443		"8K", "64K", "512K", "4MB", "32MB",
444		"256MB", "2GB", "16GB",
445	};
446	int i, printed;
447
448	if (tlb_type == cheetah)
449		seq_printf(m, "MMU Type\t: Cheetah\n");
450	else if (tlb_type == cheetah_plus)
451		seq_printf(m, "MMU Type\t: Cheetah+\n");
452	else if (tlb_type == spitfire)
453		seq_printf(m, "MMU Type\t: Spitfire\n");
454	else if (tlb_type == hypervisor)
455		seq_printf(m, "MMU Type\t: Hypervisor (sun4v)\n");
456	else
457		seq_printf(m, "MMU Type\t: ???\n");
458
459	seq_printf(m, "MMU PGSZs\t: ");
460	printed = 0;
461	for (i = 0; i < ARRAY_SIZE(pgsz_strings); i++) {
462		if (cpu_pgsz_mask & (1UL << i)) {
463			seq_printf(m, "%s%s",
464				   printed ? "," : "", pgsz_strings[i]);
465			printed++;
466		}
467	}
468	seq_putc(m, '\n');
469
470#ifdef CONFIG_DEBUG_DCFLUSH
471	seq_printf(m, "DCPageFlushes\t: %d\n",
472		   atomic_read(&dcpage_flushes));
473#ifdef CONFIG_SMP
474	seq_printf(m, "DCPageFlushesXC\t: %d\n",
475		   atomic_read(&dcpage_flushes_xcall));
476#endif /* CONFIG_SMP */
477#endif /* CONFIG_DEBUG_DCFLUSH */
478}
479
480struct linux_prom_translation prom_trans[512] __read_mostly;
481unsigned int prom_trans_ents __read_mostly;
482
483unsigned long kern_locked_tte_data;
484
485/* The obp translations are saved based on 8k pagesize, since obp can
486 * use a mixture of pagesizes. Misses to the LOW_OBP_ADDRESS ->
487 * HI_OBP_ADDRESS range are handled in ktlb.S.
488 */
489static inline int in_obp_range(unsigned long vaddr)
490{
491	return (vaddr >= LOW_OBP_ADDRESS &&
492		vaddr < HI_OBP_ADDRESS);
493}
494
495static int cmp_ptrans(const void *a, const void *b)
496{
497	const struct linux_prom_translation *x = a, *y = b;
498
499	if (x->virt > y->virt)
500		return 1;
501	if (x->virt < y->virt)
502		return -1;
503	return 0;
504}
505
506/* Read OBP translations property into 'prom_trans[]'.  */
507static void __init read_obp_translations(void)
508{
509	int n, node, ents, first, last, i;
510
511	node = prom_finddevice("/virtual-memory");
512	n = prom_getproplen(node, "translations");
513	if (unlikely(n == 0 || n == -1)) {
514		prom_printf("prom_mappings: Couldn't get size.\n");
515		prom_halt();
516	}
517	if (unlikely(n > sizeof(prom_trans))) {
518		prom_printf("prom_mappings: Size %d is too big.\n", n);
519		prom_halt();
520	}
521
522	if ((n = prom_getproperty(node, "translations",
523				  (char *)&prom_trans[0],
524				  sizeof(prom_trans))) == -1) {
525		prom_printf("prom_mappings: Couldn't get property.\n");
526		prom_halt();
527	}
528
529	n = n / sizeof(struct linux_prom_translation);
530
531	ents = n;
532
533	sort(prom_trans, ents, sizeof(struct linux_prom_translation),
534	     cmp_ptrans, NULL);
535
536	/* Now kick out all the non-OBP entries.  */
537	for (i = 0; i < ents; i++) {
538		if (in_obp_range(prom_trans[i].virt))
539			break;
540	}
541	first = i;
542	for (; i < ents; i++) {
543		if (!in_obp_range(prom_trans[i].virt))
544			break;
545	}
546	last = i;
547
548	for (i = 0; i < (last - first); i++) {
549		struct linux_prom_translation *src = &prom_trans[i + first];
550		struct linux_prom_translation *dest = &prom_trans[i];
551
552		*dest = *src;
553	}
554	for (; i < ents; i++) {
555		struct linux_prom_translation *dest = &prom_trans[i];
556		dest->virt = dest->size = dest->data = 0x0UL;
557	}
558
559	prom_trans_ents = last - first;
560
561	if (tlb_type == spitfire) {
562		/* Clear diag TTE bits. */
563		for (i = 0; i < prom_trans_ents; i++)
564			prom_trans[i].data &= ~0x0003fe0000000000UL;
565	}
566
567	/* Force execute bit on.  */
568	for (i = 0; i < prom_trans_ents; i++)
569		prom_trans[i].data |= (tlb_type == hypervisor ?
570				       _PAGE_EXEC_4V : _PAGE_EXEC_4U);
571}
572
573static void __init hypervisor_tlb_lock(unsigned long vaddr,
574				       unsigned long pte,
575				       unsigned long mmu)
576{
577	unsigned long ret = sun4v_mmu_map_perm_addr(vaddr, 0, pte, mmu);
578
579	if (ret != 0) {
580		prom_printf("hypervisor_tlb_lock[%lx:%x:%lx:%lx]: "
581			    "errors with %lx\n", vaddr, 0, pte, mmu, ret);
582		prom_halt();
583	}
584}
585
586static unsigned long kern_large_tte(unsigned long paddr);
587
588static void __init remap_kernel(void)
589{
590	unsigned long phys_page, tte_vaddr, tte_data;
591	int i, tlb_ent = sparc64_highest_locked_tlbent();
592
593	tte_vaddr = (unsigned long) KERNBASE;
594	phys_page = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
595	tte_data = kern_large_tte(phys_page);
596
597	kern_locked_tte_data = tte_data;
598
599	/* Now lock us into the TLBs via Hypervisor or OBP. */
600	if (tlb_type == hypervisor) {
601		for (i = 0; i < num_kernel_image_mappings; i++) {
602			hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_DMMU);
603			hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_IMMU);
604			tte_vaddr += 0x400000;
605			tte_data += 0x400000;
606		}
607	} else {
608		for (i = 0; i < num_kernel_image_mappings; i++) {
609			prom_dtlb_load(tlb_ent - i, tte_data, tte_vaddr);
610			prom_itlb_load(tlb_ent - i, tte_data, tte_vaddr);
611			tte_vaddr += 0x400000;
612			tte_data += 0x400000;
613		}
614		sparc64_highest_unlocked_tlb_ent = tlb_ent - i;
615	}
616	if (tlb_type == cheetah_plus) {
617		sparc64_kern_pri_context = (CTX_CHEETAH_PLUS_CTX0 |
618					    CTX_CHEETAH_PLUS_NUC);
619		sparc64_kern_pri_nuc_bits = CTX_CHEETAH_PLUS_NUC;
620		sparc64_kern_sec_context = CTX_CHEETAH_PLUS_CTX0;
621	}
622}
623
624
625static void __init inherit_prom_mappings(void)
626{
627	/* Now fixup OBP's idea about where we really are mapped. */
628	printk("Remapping the kernel... ");
629	remap_kernel();
630	printk("done.\n");
631}
632
633void prom_world(int enter)
634{
635	if (!enter)
636		set_fs(get_fs());
637
638	__asm__ __volatile__("flushw");
639}
640
641void __flush_dcache_range(unsigned long start, unsigned long end)
642{
643	unsigned long va;
644
645	if (tlb_type == spitfire) {
646		int n = 0;
647
648		for (va = start; va < end; va += 32) {
649			spitfire_put_dcache_tag(va & 0x3fe0, 0x0);
650			if (++n >= 512)
651				break;
652		}
653	} else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
654		start = __pa(start);
655		end = __pa(end);
656		for (va = start; va < end; va += 32)
657			__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
658					     "membar #Sync"
659					     : /* no outputs */
660					     : "r" (va),
661					       "i" (ASI_DCACHE_INVALIDATE));
662	}
663}
664EXPORT_SYMBOL(__flush_dcache_range);
665
666/* get_new_mmu_context() uses "cache + 1".  */
667DEFINE_SPINLOCK(ctx_alloc_lock);
668unsigned long tlb_context_cache = CTX_FIRST_VERSION - 1;
669#define MAX_CTX_NR	(1UL << CTX_NR_BITS)
670#define CTX_BMAP_SLOTS	BITS_TO_LONGS(MAX_CTX_NR)
671DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR);
672
673/* Caller does TLB context flushing on local CPU if necessary.
674 * The caller also ensures that CTX_VALID(mm->context) is false.
675 *
676 * We must be careful about boundary cases so that we never
677 * let the user have CTX 0 (nucleus) or we ever use a CTX
678 * version of zero (and thus NO_CONTEXT would not be caught
679 * by version mis-match tests in mmu_context.h).
680 *
681 * Always invoked with interrupts disabled.
682 */
683void get_new_mmu_context(struct mm_struct *mm)
684{
685	unsigned long ctx, new_ctx;
686	unsigned long orig_pgsz_bits;
687	int new_version;
688
689	spin_lock(&ctx_alloc_lock);
690	orig_pgsz_bits = (mm->context.sparc64_ctx_val & CTX_PGSZ_MASK);
691	ctx = (tlb_context_cache + 1) & CTX_NR_MASK;
692	new_ctx = find_next_zero_bit(mmu_context_bmap, 1 << CTX_NR_BITS, ctx);
693	new_version = 0;
694	if (new_ctx >= (1 << CTX_NR_BITS)) {
695		new_ctx = find_next_zero_bit(mmu_context_bmap, ctx, 1);
696		if (new_ctx >= ctx) {
697			int i;
698			new_ctx = (tlb_context_cache & CTX_VERSION_MASK) +
699				CTX_FIRST_VERSION;
700			if (new_ctx == 1)
701				new_ctx = CTX_FIRST_VERSION;
702
703			/* Don't call memset, for 16 entries that's just
704			 * plain silly...
705			 */
706			mmu_context_bmap[0] = 3;
707			mmu_context_bmap[1] = 0;
708			mmu_context_bmap[2] = 0;
709			mmu_context_bmap[3] = 0;
710			for (i = 4; i < CTX_BMAP_SLOTS; i += 4) {
711				mmu_context_bmap[i + 0] = 0;
712				mmu_context_bmap[i + 1] = 0;
713				mmu_context_bmap[i + 2] = 0;
714				mmu_context_bmap[i + 3] = 0;
715			}
716			new_version = 1;
717			goto out;
718		}
719	}
720	mmu_context_bmap[new_ctx>>6] |= (1UL << (new_ctx & 63));
721	new_ctx |= (tlb_context_cache & CTX_VERSION_MASK);
722out:
723	tlb_context_cache = new_ctx;
724	mm->context.sparc64_ctx_val = new_ctx | orig_pgsz_bits;
725	spin_unlock(&ctx_alloc_lock);
726
727	if (unlikely(new_version))
728		smp_new_mmu_context_version();
729}
730
731static int numa_enabled = 1;
732static int numa_debug;
733
734static int __init early_numa(char *p)
735{
736	if (!p)
737		return 0;
738
739	if (strstr(p, "off"))
740		numa_enabled = 0;
741
742	if (strstr(p, "debug"))
743		numa_debug = 1;
744
745	return 0;
746}
747early_param("numa", early_numa);
748
749#define numadbg(f, a...) \
750do {	if (numa_debug) \
751		printk(KERN_INFO f, ## a); \
752} while (0)
753
754static void __init find_ramdisk(unsigned long phys_base)
755{
756#ifdef CONFIG_BLK_DEV_INITRD
757	if (sparc_ramdisk_image || sparc_ramdisk_image64) {
758		unsigned long ramdisk_image;
759
760		/* Older versions of the bootloader only supported a
761		 * 32-bit physical address for the ramdisk image
762		 * location, stored at sparc_ramdisk_image.  Newer
763		 * SILO versions set sparc_ramdisk_image to zero and
764		 * provide a full 64-bit physical address at
765		 * sparc_ramdisk_image64.
766		 */
767		ramdisk_image = sparc_ramdisk_image;
768		if (!ramdisk_image)
769			ramdisk_image = sparc_ramdisk_image64;
770
771		/* Another bootloader quirk.  The bootloader normalizes
772		 * the physical address to KERNBASE, so we have to
773		 * factor that back out and add in the lowest valid
774		 * physical page address to get the true physical address.
775		 */
776		ramdisk_image -= KERNBASE;
777		ramdisk_image += phys_base;
778
779		numadbg("Found ramdisk at physical address 0x%lx, size %u\n",
780			ramdisk_image, sparc_ramdisk_size);
781
782		initrd_start = ramdisk_image;
783		initrd_end = ramdisk_image + sparc_ramdisk_size;
784
785		memblock_reserve(initrd_start, sparc_ramdisk_size);
786
787		initrd_start += PAGE_OFFSET;
788		initrd_end += PAGE_OFFSET;
789	}
790#endif
791}
792
793struct node_mem_mask {
794	unsigned long mask;
795	unsigned long val;
796};
797static struct node_mem_mask node_masks[MAX_NUMNODES];
798static int num_node_masks;
799
800#ifdef CONFIG_NEED_MULTIPLE_NODES
801
802int numa_cpu_lookup_table[NR_CPUS];
803cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
804
805struct mdesc_mblock {
806	u64	base;
807	u64	size;
808	u64	offset; /* RA-to-PA */
809};
810static struct mdesc_mblock *mblocks;
811static int num_mblocks;
812
813static unsigned long ra_to_pa(unsigned long addr)
814{
815	int i;
816
817	for (i = 0; i < num_mblocks; i++) {
818		struct mdesc_mblock *m = &mblocks[i];
819
820		if (addr >= m->base &&
821		    addr < (m->base + m->size)) {
822			addr += m->offset;
823			break;
824		}
825	}
826	return addr;
827}
828
829static int find_node(unsigned long addr)
830{
831	int i;
832
833	addr = ra_to_pa(addr);
834	for (i = 0; i < num_node_masks; i++) {
835		struct node_mem_mask *p = &node_masks[i];
836
837		if ((addr & p->mask) == p->val)
838			return i;
839	}
840	/* The following condition has been observed on LDOM guests.*/
841	WARN_ONCE(1, "find_node: A physical address doesn't match a NUMA node"
842		" rule. Some physical memory will be owned by node 0.");
843	return 0;
844}
845
846static u64 memblock_nid_range(u64 start, u64 end, int *nid)
847{
848	*nid = find_node(start);
849	start += PAGE_SIZE;
850	while (start < end) {
851		int n = find_node(start);
852
853		if (n != *nid)
854			break;
855		start += PAGE_SIZE;
856	}
857
858	if (start > end)
859		start = end;
860
861	return start;
862}
863#endif
864
865/* This must be invoked after performing all of the necessary
866 * memblock_set_node() calls for 'nid'.  We need to be able to get
867 * correct data from get_pfn_range_for_nid().
868 */
869static void __init allocate_node_data(int nid)
870{
871	struct pglist_data *p;
872	unsigned long start_pfn, end_pfn;
873#ifdef CONFIG_NEED_MULTIPLE_NODES
874	unsigned long paddr;
875
876	paddr = memblock_alloc_try_nid(sizeof(struct pglist_data), SMP_CACHE_BYTES, nid);
877	if (!paddr) {
878		prom_printf("Cannot allocate pglist_data for nid[%d]\n", nid);
879		prom_halt();
880	}
881	NODE_DATA(nid) = __va(paddr);
882	memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
883
884	NODE_DATA(nid)->node_id = nid;
885#endif
886
887	p = NODE_DATA(nid);
888
889	get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
890	p->node_start_pfn = start_pfn;
891	p->node_spanned_pages = end_pfn - start_pfn;
892}
893
894static void init_node_masks_nonnuma(void)
895{
896#ifdef CONFIG_NEED_MULTIPLE_NODES
897	int i;
898#endif
899
900	numadbg("Initializing tables for non-numa.\n");
901
902	node_masks[0].mask = node_masks[0].val = 0;
903	num_node_masks = 1;
904
905#ifdef CONFIG_NEED_MULTIPLE_NODES
906	for (i = 0; i < NR_CPUS; i++)
907		numa_cpu_lookup_table[i] = 0;
908
909	cpumask_setall(&numa_cpumask_lookup_table[0]);
910#endif
911}
912
913#ifdef CONFIG_NEED_MULTIPLE_NODES
914struct pglist_data *node_data[MAX_NUMNODES];
915
916EXPORT_SYMBOL(numa_cpu_lookup_table);
917EXPORT_SYMBOL(numa_cpumask_lookup_table);
918EXPORT_SYMBOL(node_data);
919
920struct mdesc_mlgroup {
921	u64	node;
922	u64	latency;
923	u64	match;
924	u64	mask;
925};
926static struct mdesc_mlgroup *mlgroups;
927static int num_mlgroups;
928
929static int scan_pio_for_cfg_handle(struct mdesc_handle *md, u64 pio,
930				   u32 cfg_handle)
931{
932	u64 arc;
933
934	mdesc_for_each_arc(arc, md, pio, MDESC_ARC_TYPE_FWD) {
935		u64 target = mdesc_arc_target(md, arc);
936		const u64 *val;
937
938		val = mdesc_get_property(md, target,
939					 "cfg-handle", NULL);
940		if (val && *val == cfg_handle)
941			return 0;
942	}
943	return -ENODEV;
944}
945
946static int scan_arcs_for_cfg_handle(struct mdesc_handle *md, u64 grp,
947				    u32 cfg_handle)
948{
949	u64 arc, candidate, best_latency = ~(u64)0;
950
951	candidate = MDESC_NODE_NULL;
952	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
953		u64 target = mdesc_arc_target(md, arc);
954		const char *name = mdesc_node_name(md, target);
955		const u64 *val;
956
957		if (strcmp(name, "pio-latency-group"))
958			continue;
959
960		val = mdesc_get_property(md, target, "latency", NULL);
961		if (!val)
962			continue;
963
964		if (*val < best_latency) {
965			candidate = target;
966			best_latency = *val;
967		}
968	}
969
970	if (candidate == MDESC_NODE_NULL)
971		return -ENODEV;
972
973	return scan_pio_for_cfg_handle(md, candidate, cfg_handle);
974}
975
976int of_node_to_nid(struct device_node *dp)
977{
978	const struct linux_prom64_registers *regs;
979	struct mdesc_handle *md;
980	u32 cfg_handle;
981	int count, nid;
982	u64 grp;
983
984	/* This is the right thing to do on currently supported
985	 * SUN4U NUMA platforms as well, as the PCI controller does
986	 * not sit behind any particular memory controller.
987	 */
988	if (!mlgroups)
989		return -1;
990
991	regs = of_get_property(dp, "reg", NULL);
992	if (!regs)
993		return -1;
994
995	cfg_handle = (regs->phys_addr >> 32UL) & 0x0fffffff;
996
997	md = mdesc_grab();
998
999	count = 0;
1000	nid = -1;
1001	mdesc_for_each_node_by_name(md, grp, "group") {
1002		if (!scan_arcs_for_cfg_handle(md, grp, cfg_handle)) {
1003			nid = count;
1004			break;
1005		}
1006		count++;
1007	}
1008
1009	mdesc_release(md);
1010
1011	return nid;
1012}
1013
1014static void __init add_node_ranges(void)
1015{
1016	struct memblock_region *reg;
1017
1018	for_each_memblock(memory, reg) {
1019		unsigned long size = reg->size;
1020		unsigned long start, end;
1021
1022		start = reg->base;
1023		end = start + size;
1024		while (start < end) {
1025			unsigned long this_end;
1026			int nid;
1027
1028			this_end = memblock_nid_range(start, end, &nid);
1029
1030			numadbg("Setting memblock NUMA node nid[%d] "
1031				"start[%lx] end[%lx]\n",
1032				nid, start, this_end);
1033
1034			memblock_set_node(start, this_end - start,
1035					  &memblock.memory, nid);
1036			start = this_end;
1037		}
1038	}
1039}
1040
1041static int __init grab_mlgroups(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, "memory-latency-group")
1048		count++;
1049	if (!count)
1050		return -ENOENT;
1051
1052	paddr = memblock_alloc(count * sizeof(struct mdesc_mlgroup),
1053			  SMP_CACHE_BYTES);
1054	if (!paddr)
1055		return -ENOMEM;
1056
1057	mlgroups = __va(paddr);
1058	num_mlgroups = count;
1059
1060	count = 0;
1061	mdesc_for_each_node_by_name(md, node, "memory-latency-group") {
1062		struct mdesc_mlgroup *m = &mlgroups[count++];
1063		const u64 *val;
1064
1065		m->node = node;
1066
1067		val = mdesc_get_property(md, node, "latency", NULL);
1068		m->latency = *val;
1069		val = mdesc_get_property(md, node, "address-match", NULL);
1070		m->match = *val;
1071		val = mdesc_get_property(md, node, "address-mask", NULL);
1072		m->mask = *val;
1073
1074		numadbg("MLGROUP[%d]: node[%llx] latency[%llx] "
1075			"match[%llx] mask[%llx]\n",
1076			count - 1, m->node, m->latency, m->match, m->mask);
1077	}
1078
1079	return 0;
1080}
1081
1082static int __init grab_mblocks(struct mdesc_handle *md)
1083{
1084	unsigned long paddr;
1085	int count = 0;
1086	u64 node;
1087
1088	mdesc_for_each_node_by_name(md, node, "mblock")
1089		count++;
1090	if (!count)
1091		return -ENOENT;
1092
1093	paddr = memblock_alloc(count * sizeof(struct mdesc_mblock),
1094			  SMP_CACHE_BYTES);
1095	if (!paddr)
1096		return -ENOMEM;
1097
1098	mblocks = __va(paddr);
1099	num_mblocks = count;
1100
1101	count = 0;
1102	mdesc_for_each_node_by_name(md, node, "mblock") {
1103		struct mdesc_mblock *m = &mblocks[count++];
1104		const u64 *val;
1105
1106		val = mdesc_get_property(md, node, "base", NULL);
1107		m->base = *val;
1108		val = mdesc_get_property(md, node, "size", NULL);
1109		m->size = *val;
1110		val = mdesc_get_property(md, node,
1111					 "address-congruence-offset", NULL);
1112
1113		/* The address-congruence-offset property is optional.
1114		 * Explicity zero it be identifty this.
1115		 */
1116		if (val)
1117			m->offset = *val;
1118		else
1119			m->offset = 0UL;
1120
1121		numadbg("MBLOCK[%d]: base[%llx] size[%llx] offset[%llx]\n",
1122			count - 1, m->base, m->size, m->offset);
1123	}
1124
1125	return 0;
1126}
1127
1128static void __init numa_parse_mdesc_group_cpus(struct mdesc_handle *md,
1129					       u64 grp, cpumask_t *mask)
1130{
1131	u64 arc;
1132
1133	cpumask_clear(mask);
1134
1135	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_BACK) {
1136		u64 target = mdesc_arc_target(md, arc);
1137		const char *name = mdesc_node_name(md, target);
1138		const u64 *id;
1139
1140		if (strcmp(name, "cpu"))
1141			continue;
1142		id = mdesc_get_property(md, target, "id", NULL);
1143		if (*id < nr_cpu_ids)
1144			cpumask_set_cpu(*id, mask);
1145	}
1146}
1147
1148static struct mdesc_mlgroup * __init find_mlgroup(u64 node)
1149{
1150	int i;
1151
1152	for (i = 0; i < num_mlgroups; i++) {
1153		struct mdesc_mlgroup *m = &mlgroups[i];
1154		if (m->node == node)
1155			return m;
1156	}
1157	return NULL;
1158}
1159
1160static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp,
1161				      int index)
1162{
1163	struct mdesc_mlgroup *candidate = NULL;
1164	u64 arc, best_latency = ~(u64)0;
1165	struct node_mem_mask *n;
1166
1167	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1168		u64 target = mdesc_arc_target(md, arc);
1169		struct mdesc_mlgroup *m = find_mlgroup(target);
1170		if (!m)
1171			continue;
1172		if (m->latency < best_latency) {
1173			candidate = m;
1174			best_latency = m->latency;
1175		}
1176	}
1177	if (!candidate)
1178		return -ENOENT;
1179
1180	if (num_node_masks != index) {
1181		printk(KERN_ERR "Inconsistent NUMA state, "
1182		       "index[%d] != num_node_masks[%d]\n",
1183		       index, num_node_masks);
1184		return -EINVAL;
1185	}
1186
1187	n = &node_masks[num_node_masks++];
1188
1189	n->mask = candidate->mask;
1190	n->val = candidate->match;
1191
1192	numadbg("NUMA NODE[%d]: mask[%lx] val[%lx] (latency[%llx])\n",
1193		index, n->mask, n->val, candidate->latency);
1194
1195	return 0;
1196}
1197
1198static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp,
1199					 int index)
1200{
1201	cpumask_t mask;
1202	int cpu;
1203
1204	numa_parse_mdesc_group_cpus(md, grp, &mask);
1205
1206	for_each_cpu(cpu, &mask)
1207		numa_cpu_lookup_table[cpu] = index;
1208	cpumask_copy(&numa_cpumask_lookup_table[index], &mask);
1209
1210	if (numa_debug) {
1211		printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index);
1212		for_each_cpu(cpu, &mask)
1213			printk("%d ", cpu);
1214		printk("]\n");
1215	}
1216
1217	return numa_attach_mlgroup(md, grp, index);
1218}
1219
1220static int __init numa_parse_mdesc(void)
1221{
1222	struct mdesc_handle *md = mdesc_grab();
1223	int i, err, count;
1224	u64 node;
1225
1226	node = mdesc_node_by_name(md, MDESC_NODE_NULL, "latency-groups");
1227	if (node == MDESC_NODE_NULL) {
1228		mdesc_release(md);
1229		return -ENOENT;
1230	}
1231
1232	err = grab_mblocks(md);
1233	if (err < 0)
1234		goto out;
1235
1236	err = grab_mlgroups(md);
1237	if (err < 0)
1238		goto out;
1239
1240	count = 0;
1241	mdesc_for_each_node_by_name(md, node, "group") {
1242		err = numa_parse_mdesc_group(md, node, count);
1243		if (err < 0)
1244			break;
1245		count++;
1246	}
1247
1248	add_node_ranges();
1249
1250	for (i = 0; i < num_node_masks; i++) {
1251		allocate_node_data(i);
1252		node_set_online(i);
1253	}
1254
1255	err = 0;
1256out:
1257	mdesc_release(md);
1258	return err;
1259}
1260
1261static int __init numa_parse_jbus(void)
1262{
1263	unsigned long cpu, index;
1264
1265	/* NUMA node id is encoded in bits 36 and higher, and there is
1266	 * a 1-to-1 mapping from CPU ID to NUMA node ID.
1267	 */
1268	index = 0;
1269	for_each_present_cpu(cpu) {
1270		numa_cpu_lookup_table[cpu] = index;
1271		cpumask_copy(&numa_cpumask_lookup_table[index], cpumask_of(cpu));
1272		node_masks[index].mask = ~((1UL << 36UL) - 1UL);
1273		node_masks[index].val = cpu << 36UL;
1274
1275		index++;
1276	}
1277	num_node_masks = index;
1278
1279	add_node_ranges();
1280
1281	for (index = 0; index < num_node_masks; index++) {
1282		allocate_node_data(index);
1283		node_set_online(index);
1284	}
1285
1286	return 0;
1287}
1288
1289static int __init numa_parse_sun4u(void)
1290{
1291	if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1292		unsigned long ver;
1293
1294		__asm__ ("rdpr %%ver, %0" : "=r" (ver));
1295		if ((ver >> 32UL) == __JALAPENO_ID ||
1296		    (ver >> 32UL) == __SERRANO_ID)
1297			return numa_parse_jbus();
1298	}
1299	return -1;
1300}
1301
1302static int __init bootmem_init_numa(void)
1303{
1304	int err = -1;
1305
1306	numadbg("bootmem_init_numa()\n");
1307
1308	if (numa_enabled) {
1309		if (tlb_type == hypervisor)
1310			err = numa_parse_mdesc();
1311		else
1312			err = numa_parse_sun4u();
1313	}
1314	return err;
1315}
1316
1317#else
1318
1319static int bootmem_init_numa(void)
1320{
1321	return -1;
1322}
1323
1324#endif
1325
1326static void __init bootmem_init_nonnuma(void)
1327{
1328	unsigned long top_of_ram = memblock_end_of_DRAM();
1329	unsigned long total_ram = memblock_phys_mem_size();
1330
1331	numadbg("bootmem_init_nonnuma()\n");
1332
1333	printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
1334	       top_of_ram, total_ram);
1335	printk(KERN_INFO "Memory hole size: %ldMB\n",
1336	       (top_of_ram - total_ram) >> 20);
1337
1338	init_node_masks_nonnuma();
1339	memblock_set_node(0, (phys_addr_t)ULLONG_MAX, &memblock.memory, 0);
1340	allocate_node_data(0);
1341	node_set_online(0);
1342}
1343
1344static unsigned long __init bootmem_init(unsigned long phys_base)
1345{
1346	unsigned long end_pfn;
1347
1348	end_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1349	max_pfn = max_low_pfn = end_pfn;
1350	min_low_pfn = (phys_base >> PAGE_SHIFT);
1351
1352	if (bootmem_init_numa() < 0)
1353		bootmem_init_nonnuma();
1354
1355	/* Dump memblock with node info. */
1356	memblock_dump_all();
1357
1358	/* XXX cpu notifier XXX */
1359
1360	sparse_memory_present_with_active_regions(MAX_NUMNODES);
1361	sparse_init();
1362
1363	return end_pfn;
1364}
1365
1366static struct linux_prom64_registers pall[MAX_BANKS] __initdata;
1367static int pall_ents __initdata;
1368
1369static unsigned long max_phys_bits = 40;
1370
1371bool kern_addr_valid(unsigned long addr)
1372{
1373	pgd_t *pgd;
1374	pud_t *pud;
1375	pmd_t *pmd;
1376	pte_t *pte;
1377
1378	if ((long)addr < 0L) {
1379		unsigned long pa = __pa(addr);
1380
1381		if ((addr >> max_phys_bits) != 0UL)
1382			return false;
1383
1384		return pfn_valid(pa >> PAGE_SHIFT);
1385	}
1386
1387	if (addr >= (unsigned long) KERNBASE &&
1388	    addr < (unsigned long)&_end)
1389		return true;
1390
1391	pgd = pgd_offset_k(addr);
1392	if (pgd_none(*pgd))
1393		return 0;
1394
1395	pud = pud_offset(pgd, addr);
1396	if (pud_none(*pud))
1397		return 0;
1398
1399	if (pud_large(*pud))
1400		return pfn_valid(pud_pfn(*pud));
1401
1402	pmd = pmd_offset(pud, addr);
1403	if (pmd_none(*pmd))
1404		return 0;
1405
1406	if (pmd_large(*pmd))
1407		return pfn_valid(pmd_pfn(*pmd));
1408
1409	pte = pte_offset_kernel(pmd, addr);
1410	if (pte_none(*pte))
1411		return 0;
1412
1413	return pfn_valid(pte_pfn(*pte));
1414}
1415EXPORT_SYMBOL(kern_addr_valid);
1416
1417static unsigned long __ref kernel_map_hugepud(unsigned long vstart,
1418					      unsigned long vend,
1419					      pud_t *pud)
1420{
1421	const unsigned long mask16gb = (1UL << 34) - 1UL;
1422	u64 pte_val = vstart;
1423
1424	/* Each PUD is 8GB */
1425	if ((vstart & mask16gb) ||
1426	    (vend - vstart <= mask16gb)) {
1427		pte_val ^= kern_linear_pte_xor[2];
1428		pud_val(*pud) = pte_val | _PAGE_PUD_HUGE;
1429
1430		return vstart + PUD_SIZE;
1431	}
1432
1433	pte_val ^= kern_linear_pte_xor[3];
1434	pte_val |= _PAGE_PUD_HUGE;
1435
1436	vend = vstart + mask16gb + 1UL;
1437	while (vstart < vend) {
1438		pud_val(*pud) = pte_val;
1439
1440		pte_val += PUD_SIZE;
1441		vstart += PUD_SIZE;
1442		pud++;
1443	}
1444	return vstart;
1445}
1446
1447static bool kernel_can_map_hugepud(unsigned long vstart, unsigned long vend,
1448				   bool guard)
1449{
1450	if (guard && !(vstart & ~PUD_MASK) && (vend - vstart) >= PUD_SIZE)
1451		return true;
1452
1453	return false;
1454}
1455
1456static unsigned long __ref kernel_map_hugepmd(unsigned long vstart,
1457					      unsigned long vend,
1458					      pmd_t *pmd)
1459{
1460	const unsigned long mask256mb = (1UL << 28) - 1UL;
1461	const unsigned long mask2gb = (1UL << 31) - 1UL;
1462	u64 pte_val = vstart;
1463
1464	/* Each PMD is 8MB */
1465	if ((vstart & mask256mb) ||
1466	    (vend - vstart <= mask256mb)) {
1467		pte_val ^= kern_linear_pte_xor[0];
1468		pmd_val(*pmd) = pte_val | _PAGE_PMD_HUGE;
1469
1470		return vstart + PMD_SIZE;
1471	}
1472
1473	if ((vstart & mask2gb) ||
1474	    (vend - vstart <= mask2gb)) {
1475		pte_val ^= kern_linear_pte_xor[1];
1476		pte_val |= _PAGE_PMD_HUGE;
1477		vend = vstart + mask256mb + 1UL;
1478	} else {
1479		pte_val ^= kern_linear_pte_xor[2];
1480		pte_val |= _PAGE_PMD_HUGE;
1481		vend = vstart + mask2gb + 1UL;
1482	}
1483
1484	while (vstart < vend) {
1485		pmd_val(*pmd) = pte_val;
1486
1487		pte_val += PMD_SIZE;
1488		vstart += PMD_SIZE;
1489		pmd++;
1490	}
1491
1492	return vstart;
1493}
1494
1495static bool kernel_can_map_hugepmd(unsigned long vstart, unsigned long vend,
1496				   bool guard)
1497{
1498	if (guard && !(vstart & ~PMD_MASK) && (vend - vstart) >= PMD_SIZE)
1499		return true;
1500
1501	return false;
1502}
1503
1504static unsigned long __ref kernel_map_range(unsigned long pstart,
1505					    unsigned long pend, pgprot_t prot,
1506					    bool use_huge)
1507{
1508	unsigned long vstart = PAGE_OFFSET + pstart;
1509	unsigned long vend = PAGE_OFFSET + pend;
1510	unsigned long alloc_bytes = 0UL;
1511
1512	if ((vstart & ~PAGE_MASK) || (vend & ~PAGE_MASK)) {
1513		prom_printf("kernel_map: Unaligned physmem[%lx:%lx]\n",
1514			    vstart, vend);
1515		prom_halt();
1516	}
1517
1518	while (vstart < vend) {
1519		unsigned long this_end, paddr = __pa(vstart);
1520		pgd_t *pgd = pgd_offset_k(vstart);
1521		pud_t *pud;
1522		pmd_t *pmd;
1523		pte_t *pte;
1524
1525		if (pgd_none(*pgd)) {
1526			pud_t *new;
1527
1528			new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1529			alloc_bytes += PAGE_SIZE;
1530			pgd_populate(&init_mm, pgd, new);
1531		}
1532		pud = pud_offset(pgd, vstart);
1533		if (pud_none(*pud)) {
1534			pmd_t *new;
1535
1536			if (kernel_can_map_hugepud(vstart, vend, use_huge)) {
1537				vstart = kernel_map_hugepud(vstart, vend, pud);
1538				continue;
1539			}
1540			new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1541			alloc_bytes += PAGE_SIZE;
1542			pud_populate(&init_mm, pud, new);
1543		}
1544
1545		pmd = pmd_offset(pud, vstart);
1546		if (pmd_none(*pmd)) {
1547			pte_t *new;
1548
1549			if (kernel_can_map_hugepmd(vstart, vend, use_huge)) {
1550				vstart = kernel_map_hugepmd(vstart, vend, pmd);
1551				continue;
1552			}
1553			new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1554			alloc_bytes += PAGE_SIZE;
1555			pmd_populate_kernel(&init_mm, pmd, new);
1556		}
1557
1558		pte = pte_offset_kernel(pmd, vstart);
1559		this_end = (vstart + PMD_SIZE) & PMD_MASK;
1560		if (this_end > vend)
1561			this_end = vend;
1562
1563		while (vstart < this_end) {
1564			pte_val(*pte) = (paddr | pgprot_val(prot));
1565
1566			vstart += PAGE_SIZE;
1567			paddr += PAGE_SIZE;
1568			pte++;
1569		}
1570	}
1571
1572	return alloc_bytes;
1573}
1574
1575static void __init flush_all_kernel_tsbs(void)
1576{
1577	int i;
1578
1579	for (i = 0; i < KERNEL_TSB_NENTRIES; i++) {
1580		struct tsb *ent = &swapper_tsb[i];
1581
1582		ent->tag = (1UL << TSB_TAG_INVALID_BIT);
1583	}
1584#ifndef CONFIG_DEBUG_PAGEALLOC
1585	for (i = 0; i < KERNEL_TSB4M_NENTRIES; i++) {
1586		struct tsb *ent = &swapper_4m_tsb[i];
1587
1588		ent->tag = (1UL << TSB_TAG_INVALID_BIT);
1589	}
1590#endif
1591}
1592
1593extern unsigned int kvmap_linear_patch[1];
1594
1595static void __init kernel_physical_mapping_init(void)
1596{
1597	unsigned long i, mem_alloced = 0UL;
1598	bool use_huge = true;
1599
1600#ifdef CONFIG_DEBUG_PAGEALLOC
1601	use_huge = false;
1602#endif
1603	for (i = 0; i < pall_ents; i++) {
1604		unsigned long phys_start, phys_end;
1605
1606		phys_start = pall[i].phys_addr;
1607		phys_end = phys_start + pall[i].reg_size;
1608
1609		mem_alloced += kernel_map_range(phys_start, phys_end,
1610						PAGE_KERNEL, use_huge);
1611	}
1612
1613	printk("Allocated %ld bytes for kernel page tables.\n",
1614	       mem_alloced);
1615
1616	kvmap_linear_patch[0] = 0x01000000; /* nop */
1617	flushi(&kvmap_linear_patch[0]);
1618
1619	flush_all_kernel_tsbs();
1620
1621	__flush_tlb_all();
1622}
1623
1624#ifdef CONFIG_DEBUG_PAGEALLOC
1625void __kernel_map_pages(struct page *page, int numpages, int enable)
1626{
1627	unsigned long phys_start = page_to_pfn(page) << PAGE_SHIFT;
1628	unsigned long phys_end = phys_start + (numpages * PAGE_SIZE);
1629
1630	kernel_map_range(phys_start, phys_end,
1631			 (enable ? PAGE_KERNEL : __pgprot(0)), false);
1632
1633	flush_tsb_kernel_range(PAGE_OFFSET + phys_start,
1634			       PAGE_OFFSET + phys_end);
1635
1636	/* we should perform an IPI and flush all tlbs,
1637	 * but that can deadlock->flush only current cpu.
1638	 */
1639	__flush_tlb_kernel_range(PAGE_OFFSET + phys_start,
1640				 PAGE_OFFSET + phys_end);
1641}
1642#endif
1643
1644unsigned long __init find_ecache_flush_span(unsigned long size)
1645{
1646	int i;
1647
1648	for (i = 0; i < pavail_ents; i++) {
1649		if (pavail[i].reg_size >= size)
1650			return pavail[i].phys_addr;
1651	}
1652
1653	return ~0UL;
1654}
1655
1656unsigned long PAGE_OFFSET;
1657EXPORT_SYMBOL(PAGE_OFFSET);
1658
1659unsigned long VMALLOC_END   = 0x0000010000000000UL;
1660EXPORT_SYMBOL(VMALLOC_END);
1661
1662unsigned long sparc64_va_hole_top =    0xfffff80000000000UL;
1663unsigned long sparc64_va_hole_bottom = 0x0000080000000000UL;
1664
1665static void __init setup_page_offset(void)
1666{
1667	if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1668		/* Cheetah/Panther support a full 64-bit virtual
1669		 * address, so we can use all that our page tables
1670		 * support.
1671		 */
1672		sparc64_va_hole_top =    0xfff0000000000000UL;
1673		sparc64_va_hole_bottom = 0x0010000000000000UL;
1674
1675		max_phys_bits = 42;
1676	} else if (tlb_type == hypervisor) {
1677		switch (sun4v_chip_type) {
1678		case SUN4V_CHIP_NIAGARA1:
1679		case SUN4V_CHIP_NIAGARA2:
1680			/* T1 and T2 support 48-bit virtual addresses.  */
1681			sparc64_va_hole_top =    0xffff800000000000UL;
1682			sparc64_va_hole_bottom = 0x0000800000000000UL;
1683
1684			max_phys_bits = 39;
1685			break;
1686		case SUN4V_CHIP_NIAGARA3:
1687			/* T3 supports 48-bit virtual addresses.  */
1688			sparc64_va_hole_top =    0xffff800000000000UL;
1689			sparc64_va_hole_bottom = 0x0000800000000000UL;
1690
1691			max_phys_bits = 43;
1692			break;
1693		case SUN4V_CHIP_NIAGARA4:
1694		case SUN4V_CHIP_NIAGARA5:
1695		case SUN4V_CHIP_SPARC64X:
1696		case SUN4V_CHIP_SPARC_M6:
1697			/* T4 and later support 52-bit virtual addresses.  */
1698			sparc64_va_hole_top =    0xfff8000000000000UL;
1699			sparc64_va_hole_bottom = 0x0008000000000000UL;
1700			max_phys_bits = 47;
1701			break;
1702		case SUN4V_CHIP_SPARC_M7:
1703		default:
1704			/* M7 and later support 52-bit virtual addresses.  */
1705			sparc64_va_hole_top =    0xfff8000000000000UL;
1706			sparc64_va_hole_bottom = 0x0008000000000000UL;
1707			max_phys_bits = 49;
1708			break;
1709		}
1710	}
1711
1712	if (max_phys_bits > MAX_PHYS_ADDRESS_BITS) {
1713		prom_printf("MAX_PHYS_ADDRESS_BITS is too small, need %lu\n",
1714			    max_phys_bits);
1715		prom_halt();
1716	}
1717
1718	PAGE_OFFSET = sparc64_va_hole_top;
1719	VMALLOC_END = ((sparc64_va_hole_bottom >> 1) +
1720		       (sparc64_va_hole_bottom >> 2));
1721
1722	pr_info("MM: PAGE_OFFSET is 0x%016lx (max_phys_bits == %lu)\n",
1723		PAGE_OFFSET, max_phys_bits);
1724	pr_info("MM: VMALLOC [0x%016lx --> 0x%016lx]\n",
1725		VMALLOC_START, VMALLOC_END);
1726	pr_info("MM: VMEMMAP [0x%016lx --> 0x%016lx]\n",
1727		VMEMMAP_BASE, VMEMMAP_BASE << 1);
1728}
1729
1730static void __init tsb_phys_patch(void)
1731{
1732	struct tsb_ldquad_phys_patch_entry *pquad;
1733	struct tsb_phys_patch_entry *p;
1734
1735	pquad = &__tsb_ldquad_phys_patch;
1736	while (pquad < &__tsb_ldquad_phys_patch_end) {
1737		unsigned long addr = pquad->addr;
1738
1739		if (tlb_type == hypervisor)
1740			*(unsigned int *) addr = pquad->sun4v_insn;
1741		else
1742			*(unsigned int *) addr = pquad->sun4u_insn;
1743		wmb();
1744		__asm__ __volatile__("flush	%0"
1745				     : /* no outputs */
1746				     : "r" (addr));
1747
1748		pquad++;
1749	}
1750
1751	p = &__tsb_phys_patch;
1752	while (p < &__tsb_phys_patch_end) {
1753		unsigned long addr = p->addr;
1754
1755		*(unsigned int *) addr = p->insn;
1756		wmb();
1757		__asm__ __volatile__("flush	%0"
1758				     : /* no outputs */
1759				     : "r" (addr));
1760
1761		p++;
1762	}
1763}
1764
1765/* Don't mark as init, we give this to the Hypervisor.  */
1766#ifndef CONFIG_DEBUG_PAGEALLOC
1767#define NUM_KTSB_DESCR	2
1768#else
1769#define NUM_KTSB_DESCR	1
1770#endif
1771static struct hv_tsb_descr ktsb_descr[NUM_KTSB_DESCR];
1772
1773/* The swapper TSBs are loaded with a base sequence of:
1774 *
1775 *	sethi	%uhi(SYMBOL), REG1
1776 *	sethi	%hi(SYMBOL), REG2
1777 *	or	REG1, %ulo(SYMBOL), REG1
1778 *	or	REG2, %lo(SYMBOL), REG2
1779 *	sllx	REG1, 32, REG1
1780 *	or	REG1, REG2, REG1
1781 *
1782 * When we use physical addressing for the TSB accesses, we patch the
1783 * first four instructions in the above sequence.
1784 */
1785
1786static void patch_one_ktsb_phys(unsigned int *start, unsigned int *end, unsigned long pa)
1787{
1788	unsigned long high_bits, low_bits;
1789
1790	high_bits = (pa >> 32) & 0xffffffff;
1791	low_bits = (pa >> 0) & 0xffffffff;
1792
1793	while (start < end) {
1794		unsigned int *ia = (unsigned int *)(unsigned long)*start;
1795
1796		ia[0] = (ia[0] & ~0x3fffff) | (high_bits >> 10);
1797		__asm__ __volatile__("flush	%0" : : "r" (ia));
1798
1799		ia[1] = (ia[1] & ~0x3fffff) | (low_bits >> 10);
1800		__asm__ __volatile__("flush	%0" : : "r" (ia + 1));
1801
1802		ia[2] = (ia[2] & ~0x1fff) | (high_bits & 0x3ff);
1803		__asm__ __volatile__("flush	%0" : : "r" (ia + 2));
1804
1805		ia[3] = (ia[3] & ~0x1fff) | (low_bits & 0x3ff);
1806		__asm__ __volatile__("flush	%0" : : "r" (ia + 3));
1807
1808		start++;
1809	}
1810}
1811
1812static void ktsb_phys_patch(void)
1813{
1814	extern unsigned int __swapper_tsb_phys_patch;
1815	extern unsigned int __swapper_tsb_phys_patch_end;
1816	unsigned long ktsb_pa;
1817
1818	ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
1819	patch_one_ktsb_phys(&__swapper_tsb_phys_patch,
1820			    &__swapper_tsb_phys_patch_end, ktsb_pa);
1821#ifndef CONFIG_DEBUG_PAGEALLOC
1822	{
1823	extern unsigned int __swapper_4m_tsb_phys_patch;
1824	extern unsigned int __swapper_4m_tsb_phys_patch_end;
1825	ktsb_pa = (kern_base +
1826		   ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
1827	patch_one_ktsb_phys(&__swapper_4m_tsb_phys_patch,
1828			    &__swapper_4m_tsb_phys_patch_end, ktsb_pa);
1829	}
1830#endif
1831}
1832
1833static void __init sun4v_ktsb_init(void)
1834{
1835	unsigned long ktsb_pa;
1836
1837	/* First KTSB for PAGE_SIZE mappings.  */
1838	ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
1839
1840	switch (PAGE_SIZE) {
1841	case 8 * 1024:
1842	default:
1843		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_8K;
1844		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_8K;
1845		break;
1846
1847	case 64 * 1024:
1848		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_64K;
1849		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_64K;
1850		break;
1851
1852	case 512 * 1024:
1853		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_512K;
1854		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_512K;
1855		break;
1856
1857	case 4 * 1024 * 1024:
1858		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_4MB;
1859		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_4MB;
1860		break;
1861	}
1862
1863	ktsb_descr[0].assoc = 1;
1864	ktsb_descr[0].num_ttes = KERNEL_TSB_NENTRIES;
1865	ktsb_descr[0].ctx_idx = 0;
1866	ktsb_descr[0].tsb_base = ktsb_pa;
1867	ktsb_descr[0].resv = 0;
1868
1869#ifndef CONFIG_DEBUG_PAGEALLOC
1870	/* Second KTSB for 4MB/256MB/2GB/16GB mappings.  */
1871	ktsb_pa = (kern_base +
1872		   ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
1873
1874	ktsb_descr[1].pgsz_idx = HV_PGSZ_IDX_4MB;
1875	ktsb_descr[1].pgsz_mask = ((HV_PGSZ_MASK_4MB |
1876				    HV_PGSZ_MASK_256MB |
1877				    HV_PGSZ_MASK_2GB |
1878				    HV_PGSZ_MASK_16GB) &
1879				   cpu_pgsz_mask);
1880	ktsb_descr[1].assoc = 1;
1881	ktsb_descr[1].num_ttes = KERNEL_TSB4M_NENTRIES;
1882	ktsb_descr[1].ctx_idx = 0;
1883	ktsb_descr[1].tsb_base = ktsb_pa;
1884	ktsb_descr[1].resv = 0;
1885#endif
1886}
1887
1888void sun4v_ktsb_register(void)
1889{
1890	unsigned long pa, ret;
1891
1892	pa = kern_base + ((unsigned long)&ktsb_descr[0] - KERNBASE);
1893
1894	ret = sun4v_mmu_tsb_ctx0(NUM_KTSB_DESCR, pa);
1895	if (ret != 0) {
1896		prom_printf("hypervisor_mmu_tsb_ctx0[%lx]: "
1897			    "errors with %lx\n", pa, ret);
1898		prom_halt();
1899	}
1900}
1901
1902static void __init sun4u_linear_pte_xor_finalize(void)
1903{
1904#ifndef CONFIG_DEBUG_PAGEALLOC
1905	/* This is where we would add Panther support for
1906	 * 32MB and 256MB pages.
1907	 */
1908#endif
1909}
1910
1911static void __init sun4v_linear_pte_xor_finalize(void)
1912{
1913	unsigned long pagecv_flag;
1914
1915	/* Bit 9 of TTE is no longer CV bit on M7 processor and it instead
1916	 * enables MCD error. Do not set bit 9 on M7 processor.
1917	 */
1918	switch (sun4v_chip_type) {
1919	case SUN4V_CHIP_SPARC_M7:
1920		pagecv_flag = 0x00;
1921		break;
1922	default:
1923		pagecv_flag = _PAGE_CV_4V;
1924		break;
1925	}
1926#ifndef CONFIG_DEBUG_PAGEALLOC
1927	if (cpu_pgsz_mask & HV_PGSZ_MASK_256MB) {
1928		kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZ256MB_4V) ^
1929			PAGE_OFFSET;
1930		kern_linear_pte_xor[1] |= (_PAGE_CP_4V | pagecv_flag |
1931					   _PAGE_P_4V | _PAGE_W_4V);
1932	} else {
1933		kern_linear_pte_xor[1] = kern_linear_pte_xor[0];
1934	}
1935
1936	if (cpu_pgsz_mask & HV_PGSZ_MASK_2GB) {
1937		kern_linear_pte_xor[2] = (_PAGE_VALID | _PAGE_SZ2GB_4V) ^
1938			PAGE_OFFSET;
1939		kern_linear_pte_xor[2] |= (_PAGE_CP_4V | pagecv_flag |
1940					   _PAGE_P_4V | _PAGE_W_4V);
1941	} else {
1942		kern_linear_pte_xor[2] = kern_linear_pte_xor[1];
1943	}
1944
1945	if (cpu_pgsz_mask & HV_PGSZ_MASK_16GB) {
1946		kern_linear_pte_xor[3] = (_PAGE_VALID | _PAGE_SZ16GB_4V) ^
1947			PAGE_OFFSET;
1948		kern_linear_pte_xor[3] |= (_PAGE_CP_4V | pagecv_flag |
1949					   _PAGE_P_4V | _PAGE_W_4V);
1950	} else {
1951		kern_linear_pte_xor[3] = kern_linear_pte_xor[2];
1952	}
1953#endif
1954}
1955
1956/* paging_init() sets up the page tables */
1957
1958static unsigned long last_valid_pfn;
1959
1960static void sun4u_pgprot_init(void);
1961static void sun4v_pgprot_init(void);
1962
1963static phys_addr_t __init available_memory(void)
1964{
1965	phys_addr_t available = 0ULL;
1966	phys_addr_t pa_start, pa_end;
1967	u64 i;
1968
1969	for_each_free_mem_range(i, NUMA_NO_NODE, &pa_start, &pa_end, NULL)
1970		available = available + (pa_end  - pa_start);
1971
1972	return available;
1973}
1974
1975#define _PAGE_CACHE_4U	(_PAGE_CP_4U | _PAGE_CV_4U)
1976#define _PAGE_CACHE_4V	(_PAGE_CP_4V | _PAGE_CV_4V)
1977#define __DIRTY_BITS_4U	 (_PAGE_MODIFIED_4U | _PAGE_WRITE_4U | _PAGE_W_4U)
1978#define __DIRTY_BITS_4V	 (_PAGE_MODIFIED_4V | _PAGE_WRITE_4V | _PAGE_W_4V)
1979#define __ACCESS_BITS_4U (_PAGE_ACCESSED_4U | _PAGE_READ_4U | _PAGE_R)
1980#define __ACCESS_BITS_4V (_PAGE_ACCESSED_4V | _PAGE_READ_4V | _PAGE_R)
1981
1982/* We need to exclude reserved regions. This exclusion will include
1983 * vmlinux and initrd. To be more precise the initrd size could be used to
1984 * compute a new lower limit because it is freed later during initialization.
1985 */
1986static void __init reduce_memory(phys_addr_t limit_ram)
1987{
1988	phys_addr_t avail_ram = available_memory();
1989	phys_addr_t pa_start, pa_end;
1990	u64 i;
1991
1992	if (limit_ram >= avail_ram)
1993		return;
1994
1995	for_each_free_mem_range(i, NUMA_NO_NODE, &pa_start, &pa_end, NULL) {
1996		phys_addr_t region_size = pa_end - pa_start;
1997		phys_addr_t clip_start = pa_start;
1998
1999		avail_ram = avail_ram - region_size;
2000		/* Are we consuming too much? */
2001		if (avail_ram < limit_ram) {
2002			phys_addr_t give_back = limit_ram - avail_ram;
2003
2004			region_size = region_size - give_back;
2005			clip_start = clip_start + give_back;
2006		}
2007
2008		memblock_remove(clip_start, region_size);
2009
2010		if (avail_ram <= limit_ram)
2011			break;
2012		i = 0UL;
2013	}
2014}
2015
2016void __init paging_init(void)
2017{
2018	unsigned long end_pfn, shift, phys_base;
2019	unsigned long real_end, i;
2020	int node;
2021
2022	setup_page_offset();
2023
2024	/* These build time checkes make sure that the dcache_dirty_cpu()
2025	 * page->flags usage will work.
2026	 *
2027	 * When a page gets marked as dcache-dirty, we store the
2028	 * cpu number starting at bit 32 in the page->flags.  Also,
2029	 * functions like clear_dcache_dirty_cpu use the cpu mask
2030	 * in 13-bit signed-immediate instruction fields.
2031	 */
2032
2033	/*
2034	 * Page flags must not reach into upper 32 bits that are used
2035	 * for the cpu number
2036	 */
2037	BUILD_BUG_ON(NR_PAGEFLAGS > 32);
2038
2039	/*
2040	 * The bit fields placed in the high range must not reach below
2041	 * the 32 bit boundary. Otherwise we cannot place the cpu field
2042	 * at the 32 bit boundary.
2043	 */
2044	BUILD_BUG_ON(SECTIONS_WIDTH + NODES_WIDTH + ZONES_WIDTH +
2045		ilog2(roundup_pow_of_two(NR_CPUS)) > 32);
2046
2047	BUILD_BUG_ON(NR_CPUS > 4096);
2048
2049	kern_base = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
2050	kern_size = (unsigned long)&_end - (unsigned long)KERNBASE;
2051
2052	/* Invalidate both kernel TSBs.  */
2053	memset(swapper_tsb, 0x40, sizeof(swapper_tsb));
2054#ifndef CONFIG_DEBUG_PAGEALLOC
2055	memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2056#endif
2057
2058	/* TTE.cv bit on sparc v9 occupies the same position as TTE.mcde
2059	 * bit on M7 processor. This is a conflicting usage of the same
2060	 * bit. Enabling TTE.cv on M7 would turn on Memory Corruption
2061	 * Detection error on all pages and this will lead to problems
2062	 * later. Kernel does not run with MCD enabled and hence rest
2063	 * of the required steps to fully configure memory corruption
2064	 * detection are not taken. We need to ensure TTE.mcde is not
2065	 * set on M7 processor. Compute the value of cacheability
2066	 * flag for use later taking this into consideration.
2067	 */
2068	switch (sun4v_chip_type) {
2069	case SUN4V_CHIP_SPARC_M7:
2070		page_cache4v_flag = _PAGE_CP_4V;
2071		break;
2072	default:
2073		page_cache4v_flag = _PAGE_CACHE_4V;
2074		break;
2075	}
2076
2077	if (tlb_type == hypervisor)
2078		sun4v_pgprot_init();
2079	else
2080		sun4u_pgprot_init();
2081
2082	if (tlb_type == cheetah_plus ||
2083	    tlb_type == hypervisor) {
2084		tsb_phys_patch();
2085		ktsb_phys_patch();
2086	}
2087
2088	if (tlb_type == hypervisor)
2089		sun4v_patch_tlb_handlers();
2090
2091	/* Find available physical memory...
2092	 *
2093	 * Read it twice in order to work around a bug in openfirmware.
2094	 * The call to grab this table itself can cause openfirmware to
2095	 * allocate memory, which in turn can take away some space from
2096	 * the list of available memory.  Reading it twice makes sure
2097	 * we really do get the final value.
2098	 */
2099	read_obp_translations();
2100	read_obp_memory("reg", &pall[0], &pall_ents);
2101	read_obp_memory("available", &pavail[0], &pavail_ents);
2102	read_obp_memory("available", &pavail[0], &pavail_ents);
2103
2104	phys_base = 0xffffffffffffffffUL;
2105	for (i = 0; i < pavail_ents; i++) {
2106		phys_base = min(phys_base, pavail[i].phys_addr);
2107		memblock_add(pavail[i].phys_addr, pavail[i].reg_size);
2108	}
2109
2110	memblock_reserve(kern_base, kern_size);
2111
2112	find_ramdisk(phys_base);
2113
2114	if (cmdline_memory_size)
2115		reduce_memory(cmdline_memory_size);
2116
2117	memblock_allow_resize();
2118	memblock_dump_all();
2119
2120	set_bit(0, mmu_context_bmap);
2121
2122	shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE);
2123
2124	real_end = (unsigned long)_end;
2125	num_kernel_image_mappings = DIV_ROUND_UP(real_end - KERNBASE, 1 << ILOG2_4MB);
2126	printk("Kernel: Using %d locked TLB entries for main kernel image.\n",
2127	       num_kernel_image_mappings);
2128
2129	/* Set kernel pgd to upper alias so physical page computations
2130	 * work.
2131	 */
2132	init_mm.pgd += ((shift) / (sizeof(pgd_t)));
2133
2134	memset(swapper_pg_dir, 0, sizeof(swapper_pg_dir));
2135
2136	inherit_prom_mappings();
2137
2138	/* Ok, we can use our TLB miss and window trap handlers safely.  */
2139	setup_tba();
2140
2141	__flush_tlb_all();
2142
2143	prom_build_devicetree();
2144	of_populate_present_mask();
2145#ifndef CONFIG_SMP
2146	of_fill_in_cpu_data();
2147#endif
2148
2149	if (tlb_type == hypervisor) {
2150		sun4v_mdesc_init();
2151		mdesc_populate_present_mask(cpu_all_mask);
2152#ifndef CONFIG_SMP
2153		mdesc_fill_in_cpu_data(cpu_all_mask);
2154#endif
2155		mdesc_get_page_sizes(cpu_all_mask, &cpu_pgsz_mask);
2156
2157		sun4v_linear_pte_xor_finalize();
2158
2159		sun4v_ktsb_init();
2160		sun4v_ktsb_register();
2161	} else {
2162		unsigned long impl, ver;
2163
2164		cpu_pgsz_mask = (HV_PGSZ_MASK_8K | HV_PGSZ_MASK_64K |
2165				 HV_PGSZ_MASK_512K | HV_PGSZ_MASK_4MB);
2166
2167		__asm__ __volatile__("rdpr %%ver, %0" : "=r" (ver));
2168		impl = ((ver >> 32) & 0xffff);
2169		if (impl == PANTHER_IMPL)
2170			cpu_pgsz_mask |= (HV_PGSZ_MASK_32MB |
2171					  HV_PGSZ_MASK_256MB);
2172
2173		sun4u_linear_pte_xor_finalize();
2174	}
2175
2176	/* Flush the TLBs and the 4M TSB so that the updated linear
2177	 * pte XOR settings are realized for all mappings.
2178	 */
2179	__flush_tlb_all();
2180#ifndef CONFIG_DEBUG_PAGEALLOC
2181	memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2182#endif
2183	__flush_tlb_all();
2184
2185	/* Setup bootmem... */
2186	last_valid_pfn = end_pfn = bootmem_init(phys_base);
2187
2188	/* Once the OF device tree and MDESC have been setup, we know
2189	 * the list of possible cpus.  Therefore we can allocate the
2190	 * IRQ stacks.
2191	 */
2192	for_each_possible_cpu(i) {
2193		node = cpu_to_node(i);
2194
2195		softirq_stack[i] = __alloc_bootmem_node(NODE_DATA(node),
2196							THREAD_SIZE,
2197							THREAD_SIZE, 0);
2198		hardirq_stack[i] = __alloc_bootmem_node(NODE_DATA(node),
2199							THREAD_SIZE,
2200							THREAD_SIZE, 0);
2201	}
2202
2203	kernel_physical_mapping_init();
2204
2205	{
2206		unsigned long max_zone_pfns[MAX_NR_ZONES];
2207
2208		memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
2209
2210		max_zone_pfns[ZONE_NORMAL] = end_pfn;
2211
2212		free_area_init_nodes(max_zone_pfns);
2213	}
2214
2215	printk("Booting Linux...\n");
2216}
2217
2218int page_in_phys_avail(unsigned long paddr)
2219{
2220	int i;
2221
2222	paddr &= PAGE_MASK;
2223
2224	for (i = 0; i < pavail_ents; i++) {
2225		unsigned long start, end;
2226
2227		start = pavail[i].phys_addr;
2228		end = start + pavail[i].reg_size;
2229
2230		if (paddr >= start && paddr < end)
2231			return 1;
2232	}
2233	if (paddr >= kern_base && paddr < (kern_base + kern_size))
2234		return 1;
2235#ifdef CONFIG_BLK_DEV_INITRD
2236	if (paddr >= __pa(initrd_start) &&
2237	    paddr < __pa(PAGE_ALIGN(initrd_end)))
2238		return 1;
2239#endif
2240
2241	return 0;
2242}
2243
2244static void __init register_page_bootmem_info(void)
2245{
2246#ifdef CONFIG_NEED_MULTIPLE_NODES
2247	int i;
2248
2249	for_each_online_node(i)
2250		if (NODE_DATA(i)->node_spanned_pages)
2251			register_page_bootmem_info_node(NODE_DATA(i));
2252#endif
2253}
2254void __init mem_init(void)
2255{
2256	high_memory = __va(last_valid_pfn << PAGE_SHIFT);
2257
2258	register_page_bootmem_info();
2259	free_all_bootmem();
2260
2261	/*
2262	 * Set up the zero page, mark it reserved, so that page count
2263	 * is not manipulated when freeing the page from user ptes.
2264	 */
2265	mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0);
2266	if (mem_map_zero == NULL) {
2267		prom_printf("paging_init: Cannot alloc zero page.\n");
2268		prom_halt();
2269	}
2270	mark_page_reserved(mem_map_zero);
2271
2272	mem_init_print_info(NULL);
2273
2274	if (tlb_type == cheetah || tlb_type == cheetah_plus)
2275		cheetah_ecache_flush_init();
2276}
2277
2278void free_initmem(void)
2279{
2280	unsigned long addr, initend;
2281	int do_free = 1;
2282
2283	/* If the physical memory maps were trimmed by kernel command
2284	 * line options, don't even try freeing this initmem stuff up.
2285	 * The kernel image could have been in the trimmed out region
2286	 * and if so the freeing below will free invalid page structs.
2287	 */
2288	if (cmdline_memory_size)
2289		do_free = 0;
2290
2291	/*
2292	 * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes.
2293	 */
2294	addr = PAGE_ALIGN((unsigned long)(__init_begin));
2295	initend = (unsigned long)(__init_end) & PAGE_MASK;
2296	for (; addr < initend; addr += PAGE_SIZE) {
2297		unsigned long page;
2298
2299		page = (addr +
2300			((unsigned long) __va(kern_base)) -
2301			((unsigned long) KERNBASE));
2302		memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
2303
2304		if (do_free)
2305			free_reserved_page(virt_to_page(page));
2306	}
2307}
2308
2309#ifdef CONFIG_BLK_DEV_INITRD
2310void free_initrd_mem(unsigned long start, unsigned long end)
2311{
2312	free_reserved_area((void *)start, (void *)end, POISON_FREE_INITMEM,
2313			   "initrd");
2314}
2315#endif
2316
2317pgprot_t PAGE_KERNEL __read_mostly;
2318EXPORT_SYMBOL(PAGE_KERNEL);
2319
2320pgprot_t PAGE_KERNEL_LOCKED __read_mostly;
2321pgprot_t PAGE_COPY __read_mostly;
2322
2323pgprot_t PAGE_SHARED __read_mostly;
2324EXPORT_SYMBOL(PAGE_SHARED);
2325
2326unsigned long pg_iobits __read_mostly;
2327
2328unsigned long _PAGE_IE __read_mostly;
2329EXPORT_SYMBOL(_PAGE_IE);
2330
2331unsigned long _PAGE_E __read_mostly;
2332EXPORT_SYMBOL(_PAGE_E);
2333
2334unsigned long _PAGE_CACHE __read_mostly;
2335EXPORT_SYMBOL(_PAGE_CACHE);
2336
2337#ifdef CONFIG_SPARSEMEM_VMEMMAP
2338int __meminit vmemmap_populate(unsigned long vstart, unsigned long vend,
2339			       int node)
2340{
2341	unsigned long pte_base;
2342
2343	pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2344		    _PAGE_CP_4U | _PAGE_CV_4U |
2345		    _PAGE_P_4U | _PAGE_W_4U);
2346	if (tlb_type == hypervisor)
2347		pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2348			    page_cache4v_flag | _PAGE_P_4V | _PAGE_W_4V);
2349
2350	pte_base |= _PAGE_PMD_HUGE;
2351
2352	vstart = vstart & PMD_MASK;
2353	vend = ALIGN(vend, PMD_SIZE);
2354	for (; vstart < vend; vstart += PMD_SIZE) {
2355		pgd_t *pgd = pgd_offset_k(vstart);
2356		unsigned long pte;
2357		pud_t *pud;
2358		pmd_t *pmd;
2359
2360		if (pgd_none(*pgd)) {
2361			pud_t *new = vmemmap_alloc_block(PAGE_SIZE, node);
2362
2363			if (!new)
2364				return -ENOMEM;
2365			pgd_populate(&init_mm, pgd, new);
2366		}
2367
2368		pud = pud_offset(pgd, vstart);
2369		if (pud_none(*pud)) {
2370			pmd_t *new = vmemmap_alloc_block(PAGE_SIZE, node);
2371
2372			if (!new)
2373				return -ENOMEM;
2374			pud_populate(&init_mm, pud, new);
2375		}
2376
2377		pmd = pmd_offset(pud, vstart);
2378
2379		pte = pmd_val(*pmd);
2380		if (!(pte & _PAGE_VALID)) {
2381			void *block = vmemmap_alloc_block(PMD_SIZE, node);
2382
2383			if (!block)
2384				return -ENOMEM;
2385
2386			pmd_val(*pmd) = pte_base | __pa(block);
2387		}
2388	}
2389
2390	return 0;
2391}
2392
2393void vmemmap_free(unsigned long start, unsigned long end)
2394{
2395}
2396#endif /* CONFIG_SPARSEMEM_VMEMMAP */
2397
2398static void prot_init_common(unsigned long page_none,
2399			     unsigned long page_shared,
2400			     unsigned long page_copy,
2401			     unsigned long page_readonly,
2402			     unsigned long page_exec_bit)
2403{
2404	PAGE_COPY = __pgprot(page_copy);
2405	PAGE_SHARED = __pgprot(page_shared);
2406
2407	protection_map[0x0] = __pgprot(page_none);
2408	protection_map[0x1] = __pgprot(page_readonly & ~page_exec_bit);
2409	protection_map[0x2] = __pgprot(page_copy & ~page_exec_bit);
2410	protection_map[0x3] = __pgprot(page_copy & ~page_exec_bit);
2411	protection_map[0x4] = __pgprot(page_readonly);
2412	protection_map[0x5] = __pgprot(page_readonly);
2413	protection_map[0x6] = __pgprot(page_copy);
2414	protection_map[0x7] = __pgprot(page_copy);
2415	protection_map[0x8] = __pgprot(page_none);
2416	protection_map[0x9] = __pgprot(page_readonly & ~page_exec_bit);
2417	protection_map[0xa] = __pgprot(page_shared & ~page_exec_bit);
2418	protection_map[0xb] = __pgprot(page_shared & ~page_exec_bit);
2419	protection_map[0xc] = __pgprot(page_readonly);
2420	protection_map[0xd] = __pgprot(page_readonly);
2421	protection_map[0xe] = __pgprot(page_shared);
2422	protection_map[0xf] = __pgprot(page_shared);
2423}
2424
2425static void __init sun4u_pgprot_init(void)
2426{
2427	unsigned long page_none, page_shared, page_copy, page_readonly;
2428	unsigned long page_exec_bit;
2429	int i;
2430
2431	PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2432				_PAGE_CACHE_4U | _PAGE_P_4U |
2433				__ACCESS_BITS_4U | __DIRTY_BITS_4U |
2434				_PAGE_EXEC_4U);
2435	PAGE_KERNEL_LOCKED = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2436				       _PAGE_CACHE_4U | _PAGE_P_4U |
2437				       __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2438				       _PAGE_EXEC_4U | _PAGE_L_4U);
2439
2440	_PAGE_IE = _PAGE_IE_4U;
2441	_PAGE_E = _PAGE_E_4U;
2442	_PAGE_CACHE = _PAGE_CACHE_4U;
2443
2444	pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4U | __DIRTY_BITS_4U |
2445		     __ACCESS_BITS_4U | _PAGE_E_4U);
2446
2447#ifdef CONFIG_DEBUG_PAGEALLOC
2448	kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2449#else
2450	kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4U) ^
2451		PAGE_OFFSET;
2452#endif
2453	kern_linear_pte_xor[0] |= (_PAGE_CP_4U | _PAGE_CV_4U |
2454				   _PAGE_P_4U | _PAGE_W_4U);
2455
2456	for (i = 1; i < 4; i++)
2457		kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2458
2459	_PAGE_ALL_SZ_BITS =  (_PAGE_SZ4MB_4U | _PAGE_SZ512K_4U |
2460			      _PAGE_SZ64K_4U | _PAGE_SZ8K_4U |
2461			      _PAGE_SZ32MB_4U | _PAGE_SZ256MB_4U);
2462
2463
2464	page_none = _PAGE_PRESENT_4U | _PAGE_ACCESSED_4U | _PAGE_CACHE_4U;
2465	page_shared = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2466		       __ACCESS_BITS_4U | _PAGE_WRITE_4U | _PAGE_EXEC_4U);
2467	page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2468		       __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2469	page_readonly   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2470			   __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2471
2472	page_exec_bit = _PAGE_EXEC_4U;
2473
2474	prot_init_common(page_none, page_shared, page_copy, page_readonly,
2475			 page_exec_bit);
2476}
2477
2478static void __init sun4v_pgprot_init(void)
2479{
2480	unsigned long page_none, page_shared, page_copy, page_readonly;
2481	unsigned long page_exec_bit;
2482	int i;
2483
2484	PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4V | _PAGE_VALID |
2485				page_cache4v_flag | _PAGE_P_4V |
2486				__ACCESS_BITS_4V | __DIRTY_BITS_4V |
2487				_PAGE_EXEC_4V);
2488	PAGE_KERNEL_LOCKED = PAGE_KERNEL;
2489
2490	_PAGE_IE = _PAGE_IE_4V;
2491	_PAGE_E = _PAGE_E_4V;
2492	_PAGE_CACHE = page_cache4v_flag;
2493
2494#ifdef CONFIG_DEBUG_PAGEALLOC
2495	kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2496#else
2497	kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4V) ^
2498		PAGE_OFFSET;
2499#endif
2500	kern_linear_pte_xor[0] |= (page_cache4v_flag | _PAGE_P_4V |
2501				   _PAGE_W_4V);
2502
2503	for (i = 1; i < 4; i++)
2504		kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2505
2506	pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4V | __DIRTY_BITS_4V |
2507		     __ACCESS_BITS_4V | _PAGE_E_4V);
2508
2509	_PAGE_ALL_SZ_BITS = (_PAGE_SZ16GB_4V | _PAGE_SZ2GB_4V |
2510			     _PAGE_SZ256MB_4V | _PAGE_SZ32MB_4V |
2511			     _PAGE_SZ4MB_4V | _PAGE_SZ512K_4V |
2512			     _PAGE_SZ64K_4V | _PAGE_SZ8K_4V);
2513
2514	page_none = _PAGE_PRESENT_4V | _PAGE_ACCESSED_4V | page_cache4v_flag;
2515	page_shared = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2516		       __ACCESS_BITS_4V | _PAGE_WRITE_4V | _PAGE_EXEC_4V);
2517	page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2518		       __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2519	page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2520			 __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2521
2522	page_exec_bit = _PAGE_EXEC_4V;
2523
2524	prot_init_common(page_none, page_shared, page_copy, page_readonly,
2525			 page_exec_bit);
2526}
2527
2528unsigned long pte_sz_bits(unsigned long sz)
2529{
2530	if (tlb_type == hypervisor) {
2531		switch (sz) {
2532		case 8 * 1024:
2533		default:
2534			return _PAGE_SZ8K_4V;
2535		case 64 * 1024:
2536			return _PAGE_SZ64K_4V;
2537		case 512 * 1024:
2538			return _PAGE_SZ512K_4V;
2539		case 4 * 1024 * 1024:
2540			return _PAGE_SZ4MB_4V;
2541		}
2542	} else {
2543		switch (sz) {
2544		case 8 * 1024:
2545		default:
2546			return _PAGE_SZ8K_4U;
2547		case 64 * 1024:
2548			return _PAGE_SZ64K_4U;
2549		case 512 * 1024:
2550			return _PAGE_SZ512K_4U;
2551		case 4 * 1024 * 1024:
2552			return _PAGE_SZ4MB_4U;
2553		}
2554	}
2555}
2556
2557pte_t mk_pte_io(unsigned long page, pgprot_t prot, int space, unsigned long page_size)
2558{
2559	pte_t pte;
2560
2561	pte_val(pte)  = page | pgprot_val(pgprot_noncached(prot));
2562	pte_val(pte) |= (((unsigned long)space) << 32);
2563	pte_val(pte) |= pte_sz_bits(page_size);
2564
2565	return pte;
2566}
2567
2568static unsigned long kern_large_tte(unsigned long paddr)
2569{
2570	unsigned long val;
2571
2572	val = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2573	       _PAGE_CP_4U | _PAGE_CV_4U | _PAGE_P_4U |
2574	       _PAGE_EXEC_4U | _PAGE_L_4U | _PAGE_W_4U);
2575	if (tlb_type == hypervisor)
2576		val = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2577		       page_cache4v_flag | _PAGE_P_4V |
2578		       _PAGE_EXEC_4V | _PAGE_W_4V);
2579
2580	return val | paddr;
2581}
2582
2583/* If not locked, zap it. */
2584void __flush_tlb_all(void)
2585{
2586	unsigned long pstate;
2587	int i;
2588
2589	__asm__ __volatile__("flushw\n\t"
2590			     "rdpr	%%pstate, %0\n\t"
2591			     "wrpr	%0, %1, %%pstate"
2592			     : "=r" (pstate)
2593			     : "i" (PSTATE_IE));
2594	if (tlb_type == hypervisor) {
2595		sun4v_mmu_demap_all();
2596	} else if (tlb_type == spitfire) {
2597		for (i = 0; i < 64; i++) {
2598			/* Spitfire Errata #32 workaround */
2599			/* NOTE: Always runs on spitfire, so no
2600			 *       cheetah+ page size encodings.
2601			 */
2602			__asm__ __volatile__("stxa	%0, [%1] %2\n\t"
2603					     "flush	%%g6"
2604					     : /* No outputs */
2605					     : "r" (0),
2606					     "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2607
2608			if (!(spitfire_get_dtlb_data(i) & _PAGE_L_4U)) {
2609				__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2610						     "membar #Sync"
2611						     : /* no outputs */
2612						     : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
2613				spitfire_put_dtlb_data(i, 0x0UL);
2614			}
2615
2616			/* Spitfire Errata #32 workaround */
2617			/* NOTE: Always runs on spitfire, so no
2618			 *       cheetah+ page size encodings.
2619			 */
2620			__asm__ __volatile__("stxa	%0, [%1] %2\n\t"
2621					     "flush	%%g6"
2622					     : /* No outputs */
2623					     : "r" (0),
2624					     "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2625
2626			if (!(spitfire_get_itlb_data(i) & _PAGE_L_4U)) {
2627				__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2628						     "membar #Sync"
2629						     : /* no outputs */
2630						     : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
2631				spitfire_put_itlb_data(i, 0x0UL);
2632			}
2633		}
2634	} else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
2635		cheetah_flush_dtlb_all();
2636		cheetah_flush_itlb_all();
2637	}
2638	__asm__ __volatile__("wrpr	%0, 0, %%pstate"
2639			     : : "r" (pstate));
2640}
2641
2642pte_t *pte_alloc_one_kernel(struct mm_struct *mm,
2643			    unsigned long address)
2644{
2645	struct page *page = alloc_page(GFP_KERNEL | __GFP_NOTRACK |
2646				       __GFP_REPEAT | __GFP_ZERO);
2647	pte_t *pte = NULL;
2648
2649	if (page)
2650		pte = (pte_t *) page_address(page);
2651
2652	return pte;
2653}
2654
2655pgtable_t pte_alloc_one(struct mm_struct *mm,
2656			unsigned long address)
2657{
2658	struct page *page = alloc_page(GFP_KERNEL | __GFP_NOTRACK |
2659				       __GFP_REPEAT | __GFP_ZERO);
2660	if (!page)
2661		return NULL;
2662	if (!pgtable_page_ctor(page)) {
2663		free_hot_cold_page(page, 0);
2664		return NULL;
2665	}
2666	return (pte_t *) page_address(page);
2667}
2668
2669void pte_free_kernel(struct mm_struct *mm, pte_t *pte)
2670{
2671	free_page((unsigned long)pte);
2672}
2673
2674static void __pte_free(pgtable_t pte)
2675{
2676	struct page *page = virt_to_page(pte);
2677
2678	pgtable_page_dtor(page);
2679	__free_page(page);
2680}
2681
2682void pte_free(struct mm_struct *mm, pgtable_t pte)
2683{
2684	__pte_free(pte);
2685}
2686
2687void pgtable_free(void *table, bool is_page)
2688{
2689	if (is_page)
2690		__pte_free(table);
2691	else
2692		kmem_cache_free(pgtable_cache, table);
2693}
2694
2695#ifdef CONFIG_TRANSPARENT_HUGEPAGE
2696void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
2697			  pmd_t *pmd)
2698{
2699	unsigned long pte, flags;
2700	struct mm_struct *mm;
2701	pmd_t entry = *pmd;
2702
2703	if (!pmd_large(entry) || !pmd_young(entry))
2704		return;
2705
2706	pte = pmd_val(entry);
2707
2708	/* Don't insert a non-valid PMD into the TSB, we'll deadlock.  */
2709	if (!(pte & _PAGE_VALID))
2710		return;
2711
2712	/* We are fabricating 8MB pages using 4MB real hw pages.  */
2713	pte |= (addr & (1UL << REAL_HPAGE_SHIFT));
2714
2715	mm = vma->vm_mm;
2716
2717	spin_lock_irqsave(&mm->context.lock, flags);
2718
2719	if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL)
2720		__update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
2721					addr, pte);
2722
2723	spin_unlock_irqrestore(&mm->context.lock, flags);
2724}
2725#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2726
2727#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
2728static void context_reload(void *__data)
2729{
2730	struct mm_struct *mm = __data;
2731
2732	if (mm == current->mm)
2733		load_secondary_context(mm);
2734}
2735
2736void hugetlb_setup(struct pt_regs *regs)
2737{
2738	struct mm_struct *mm = current->mm;
2739	struct tsb_config *tp;
2740
2741	if (in_atomic() || !mm) {
2742		const struct exception_table_entry *entry;
2743
2744		entry = search_exception_tables(regs->tpc);
2745		if (entry) {
2746			regs->tpc = entry->fixup;
2747			regs->tnpc = regs->tpc + 4;
2748			return;
2749		}
2750		pr_alert("Unexpected HugeTLB setup in atomic context.\n");
2751		die_if_kernel("HugeTSB in atomic", regs);
2752	}
2753
2754	tp = &mm->context.tsb_block[MM_TSB_HUGE];
2755	if (likely(tp->tsb == NULL))
2756		tsb_grow(mm, MM_TSB_HUGE, 0);
2757
2758	tsb_context_switch(mm);
2759	smp_tsb_sync(mm);
2760
2761	/* On UltraSPARC-III+ and later, configure the second half of
2762	 * the Data-TLB for huge pages.
2763	 */
2764	if (tlb_type == cheetah_plus) {
2765		unsigned long ctx;
2766
2767		spin_lock(&ctx_alloc_lock);
2768		ctx = mm->context.sparc64_ctx_val;
2769		ctx &= ~CTX_PGSZ_MASK;
2770		ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT;
2771		ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT;
2772
2773		if (ctx != mm->context.sparc64_ctx_val) {
2774			/* When changing the page size fields, we
2775			 * must perform a context flush so that no
2776			 * stale entries match.  This flush must
2777			 * occur with the original context register
2778			 * settings.
2779			 */
2780			do_flush_tlb_mm(mm);
2781
2782			/* Reload the context register of all processors
2783			 * also executing in this address space.
2784			 */
2785			mm->context.sparc64_ctx_val = ctx;
2786			on_each_cpu(context_reload, mm, 0);
2787		}
2788		spin_unlock(&ctx_alloc_lock);
2789	}
2790}
2791#endif
2792
2793static struct resource code_resource = {
2794	.name	= "Kernel code",
2795	.flags	= IORESOURCE_BUSY | IORESOURCE_MEM
2796};
2797
2798static struct resource data_resource = {
2799	.name	= "Kernel data",
2800	.flags	= IORESOURCE_BUSY | IORESOURCE_MEM
2801};
2802
2803static struct resource bss_resource = {
2804	.name	= "Kernel bss",
2805	.flags	= IORESOURCE_BUSY | IORESOURCE_MEM
2806};
2807
2808static inline resource_size_t compute_kern_paddr(void *addr)
2809{
2810	return (resource_size_t) (addr - KERNBASE + kern_base);
2811}
2812
2813static void __init kernel_lds_init(void)
2814{
2815	code_resource.start = compute_kern_paddr(_text);
2816	code_resource.end   = compute_kern_paddr(_etext - 1);
2817	data_resource.start = compute_kern_paddr(_etext);
2818	data_resource.end   = compute_kern_paddr(_edata - 1);
2819	bss_resource.start  = compute_kern_paddr(__bss_start);
2820	bss_resource.end    = compute_kern_paddr(_end - 1);
2821}
2822
2823static int __init report_memory(void)
2824{
2825	int i;
2826	struct resource *res;
2827
2828	kernel_lds_init();
2829
2830	for (i = 0; i < pavail_ents; i++) {
2831		res = kzalloc(sizeof(struct resource), GFP_KERNEL);
2832
2833		if (!res) {
2834			pr_warn("Failed to allocate source.\n");
2835			break;
2836		}
2837
2838		res->name = "System RAM";
2839		res->start = pavail[i].phys_addr;
2840		res->end = pavail[i].phys_addr + pavail[i].reg_size - 1;
2841		res->flags = IORESOURCE_BUSY | IORESOURCE_MEM;
2842
2843		if (insert_resource(&iomem_resource, res) < 0) {
2844			pr_warn("Resource insertion failed.\n");
2845			break;
2846		}
2847
2848		insert_resource(res, &code_resource);
2849		insert_resource(res, &data_resource);
2850		insert_resource(res, &bss_resource);
2851	}
2852
2853	return 0;
2854}
2855arch_initcall(report_memory);
2856
2857#ifdef CONFIG_SMP
2858#define do_flush_tlb_kernel_range	smp_flush_tlb_kernel_range
2859#else
2860#define do_flush_tlb_kernel_range	__flush_tlb_kernel_range
2861#endif
2862
2863void flush_tlb_kernel_range(unsigned long start, unsigned long end)
2864{
2865	if (start < HI_OBP_ADDRESS && end > LOW_OBP_ADDRESS) {
2866		if (start < LOW_OBP_ADDRESS) {
2867			flush_tsb_kernel_range(start, LOW_OBP_ADDRESS);
2868			do_flush_tlb_kernel_range(start, LOW_OBP_ADDRESS);
2869		}
2870		if (end > HI_OBP_ADDRESS) {
2871			flush_tsb_kernel_range(HI_OBP_ADDRESS, end);
2872			do_flush_tlb_kernel_range(HI_OBP_ADDRESS, end);
2873		}
2874	} else {
2875		flush_tsb_kernel_range(start, end);
2876		do_flush_tlb_kernel_range(start, end);
2877	}
2878}
2879