Linux Audio

Check our new training course

Loading...
v4.17
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 *  arch/sparc64/mm/init.c
   4 *
   5 *  Copyright (C) 1996-1999 David S. Miller (davem@caip.rutgers.edu)
   6 *  Copyright (C) 1997-1999 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
   7 */
   8 
   9#include <linux/extable.h>
  10#include <linux/kernel.h>
  11#include <linux/sched.h>
  12#include <linux/string.h>
  13#include <linux/init.h>
  14#include <linux/bootmem.h>
  15#include <linux/mm.h>
  16#include <linux/hugetlb.h>
  17#include <linux/initrd.h>
  18#include <linux/swap.h>
  19#include <linux/pagemap.h>
  20#include <linux/poison.h>
  21#include <linux/fs.h>
  22#include <linux/seq_file.h>
  23#include <linux/kprobes.h>
  24#include <linux/cache.h>
  25#include <linux/sort.h>
  26#include <linux/ioport.h>
  27#include <linux/percpu.h>
  28#include <linux/memblock.h>
  29#include <linux/mmzone.h>
  30#include <linux/gfp.h>
  31
  32#include <asm/head.h>
  33#include <asm/page.h>
  34#include <asm/pgalloc.h>
  35#include <asm/pgtable.h>
  36#include <asm/oplib.h>
  37#include <asm/iommu.h>
  38#include <asm/io.h>
  39#include <linux/uaccess.h>
  40#include <asm/mmu_context.h>
  41#include <asm/tlbflush.h>
  42#include <asm/dma.h>
  43#include <asm/starfire.h>
  44#include <asm/tlb.h>
  45#include <asm/spitfire.h>
  46#include <asm/sections.h>
  47#include <asm/tsb.h>
  48#include <asm/hypervisor.h>
  49#include <asm/prom.h>
  50#include <asm/mdesc.h>
  51#include <asm/cpudata.h>
  52#include <asm/setup.h>
  53#include <asm/irq.h>
  54
  55#include "init_64.h"
  56
  57unsigned long kern_linear_pte_xor[4] __read_mostly;
  58static unsigned long page_cache4v_flag;
  59
  60/* A bitmap, two bits for every 256MB of physical memory.  These two
  61 * bits determine what page size we use for kernel linear
  62 * translations.  They form an index into kern_linear_pte_xor[].  The
  63 * value in the indexed slot is XOR'd with the TLB miss virtual
  64 * address to form the resulting TTE.  The mapping is:
  65 *
  66 *	0	==>	4MB
  67 *	1	==>	256MB
  68 *	2	==>	2GB
  69 *	3	==>	16GB
  70 *
  71 * All sun4v chips support 256MB pages.  Only SPARC-T4 and later
  72 * support 2GB pages, and hopefully future cpus will support the 16GB
  73 * pages as well.  For slots 2 and 3, we encode a 256MB TTE xor there
  74 * if these larger page sizes are not supported by the cpu.
  75 *
  76 * It would be nice to determine this from the machine description
  77 * 'cpu' properties, but we need to have this table setup before the
  78 * MDESC is initialized.
  79 */
  80
  81#ifndef CONFIG_DEBUG_PAGEALLOC
  82/* A special kernel TSB for 4MB, 256MB, 2GB and 16GB linear mappings.
  83 * Space is allocated for this right after the trap table in
  84 * arch/sparc64/kernel/head.S
  85 */
  86extern struct tsb swapper_4m_tsb[KERNEL_TSB4M_NENTRIES];
  87#endif
  88extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
  89
  90static unsigned long cpu_pgsz_mask;
  91
  92#define MAX_BANKS	1024
  93
  94static struct linux_prom64_registers pavail[MAX_BANKS];
  95static int pavail_ents;
  96
  97u64 numa_latency[MAX_NUMNODES][MAX_NUMNODES];
  98
  99static int cmp_p64(const void *a, const void *b)
 100{
 101	const struct linux_prom64_registers *x = a, *y = b;
 102
 103	if (x->phys_addr > y->phys_addr)
 104		return 1;
 105	if (x->phys_addr < y->phys_addr)
 106		return -1;
 107	return 0;
 108}
 109
 110static void __init read_obp_memory(const char *property,
 111				   struct linux_prom64_registers *regs,
 112				   int *num_ents)
 113{
 114	phandle node = prom_finddevice("/memory");
 115	int prop_size = prom_getproplen(node, property);
 116	int ents, ret, i;
 117
 118	ents = prop_size / sizeof(struct linux_prom64_registers);
 119	if (ents > MAX_BANKS) {
 120		prom_printf("The machine has more %s property entries than "
 121			    "this kernel can support (%d).\n",
 122			    property, MAX_BANKS);
 123		prom_halt();
 124	}
 125
 126	ret = prom_getproperty(node, property, (char *) regs, prop_size);
 127	if (ret == -1) {
 128		prom_printf("Couldn't get %s property from /memory.\n",
 129				property);
 130		prom_halt();
 131	}
 132
 133	/* Sanitize what we got from the firmware, by page aligning
 134	 * everything.
 135	 */
 136	for (i = 0; i < ents; i++) {
 137		unsigned long base, size;
 138
 139		base = regs[i].phys_addr;
 140		size = regs[i].reg_size;
 141
 142		size &= PAGE_MASK;
 143		if (base & ~PAGE_MASK) {
 144			unsigned long new_base = PAGE_ALIGN(base);
 145
 146			size -= new_base - base;
 147			if ((long) size < 0L)
 148				size = 0UL;
 149			base = new_base;
 150		}
 151		if (size == 0UL) {
 152			/* If it is empty, simply get rid of it.
 153			 * This simplifies the logic of the other
 154			 * functions that process these arrays.
 155			 */
 156			memmove(&regs[i], &regs[i + 1],
 157				(ents - i - 1) * sizeof(regs[0]));
 158			i--;
 159			ents--;
 160			continue;
 161		}
 162		regs[i].phys_addr = base;
 163		regs[i].reg_size = size;
 164	}
 165
 166	*num_ents = ents;
 167
 168	sort(regs, ents, sizeof(struct linux_prom64_registers),
 169	     cmp_p64, NULL);
 170}
 171
 172/* Kernel physical address base and size in bytes.  */
 173unsigned long kern_base __read_mostly;
 174unsigned long kern_size __read_mostly;
 175
 176/* Initial ramdisk setup */
 177extern unsigned long sparc_ramdisk_image64;
 178extern unsigned int sparc_ramdisk_image;
 179extern unsigned int sparc_ramdisk_size;
 180
 181struct page *mem_map_zero __read_mostly;
 182EXPORT_SYMBOL(mem_map_zero);
 183
 184unsigned int sparc64_highest_unlocked_tlb_ent __read_mostly;
 185
 186unsigned long sparc64_kern_pri_context __read_mostly;
 187unsigned long sparc64_kern_pri_nuc_bits __read_mostly;
 188unsigned long sparc64_kern_sec_context __read_mostly;
 189
 190int num_kernel_image_mappings;
 191
 192#ifdef CONFIG_DEBUG_DCFLUSH
 193atomic_t dcpage_flushes = ATOMIC_INIT(0);
 194#ifdef CONFIG_SMP
 195atomic_t dcpage_flushes_xcall = ATOMIC_INIT(0);
 196#endif
 197#endif
 198
 199inline void flush_dcache_page_impl(struct page *page)
 200{
 201	BUG_ON(tlb_type == hypervisor);
 202#ifdef CONFIG_DEBUG_DCFLUSH
 203	atomic_inc(&dcpage_flushes);
 204#endif
 205
 206#ifdef DCACHE_ALIASING_POSSIBLE
 207	__flush_dcache_page(page_address(page),
 208			    ((tlb_type == spitfire) &&
 209			     page_mapping_file(page) != NULL));
 210#else
 211	if (page_mapping_file(page) != NULL &&
 212	    tlb_type == spitfire)
 213		__flush_icache_page(__pa(page_address(page)));
 214#endif
 215}
 216
 217#define PG_dcache_dirty		PG_arch_1
 218#define PG_dcache_cpu_shift	32UL
 219#define PG_dcache_cpu_mask	\
 220	((1UL<<ilog2(roundup_pow_of_two(NR_CPUS)))-1UL)
 221
 222#define dcache_dirty_cpu(page) \
 223	(((page)->flags >> PG_dcache_cpu_shift) & PG_dcache_cpu_mask)
 224
 225static inline void set_dcache_dirty(struct page *page, int this_cpu)
 226{
 227	unsigned long mask = this_cpu;
 228	unsigned long non_cpu_bits;
 229
 230	non_cpu_bits = ~(PG_dcache_cpu_mask << PG_dcache_cpu_shift);
 231	mask = (mask << PG_dcache_cpu_shift) | (1UL << PG_dcache_dirty);
 232
 233	__asm__ __volatile__("1:\n\t"
 234			     "ldx	[%2], %%g7\n\t"
 235			     "and	%%g7, %1, %%g1\n\t"
 236			     "or	%%g1, %0, %%g1\n\t"
 237			     "casx	[%2], %%g7, %%g1\n\t"
 238			     "cmp	%%g7, %%g1\n\t"
 239			     "bne,pn	%%xcc, 1b\n\t"
 240			     " nop"
 241			     : /* no outputs */
 242			     : "r" (mask), "r" (non_cpu_bits), "r" (&page->flags)
 243			     : "g1", "g7");
 244}
 245
 246static inline void clear_dcache_dirty_cpu(struct page *page, unsigned long cpu)
 247{
 248	unsigned long mask = (1UL << PG_dcache_dirty);
 249
 250	__asm__ __volatile__("! test_and_clear_dcache_dirty\n"
 251			     "1:\n\t"
 252			     "ldx	[%2], %%g7\n\t"
 253			     "srlx	%%g7, %4, %%g1\n\t"
 254			     "and	%%g1, %3, %%g1\n\t"
 255			     "cmp	%%g1, %0\n\t"
 256			     "bne,pn	%%icc, 2f\n\t"
 257			     " andn	%%g7, %1, %%g1\n\t"
 258			     "casx	[%2], %%g7, %%g1\n\t"
 259			     "cmp	%%g7, %%g1\n\t"
 260			     "bne,pn	%%xcc, 1b\n\t"
 261			     " nop\n"
 262			     "2:"
 263			     : /* no outputs */
 264			     : "r" (cpu), "r" (mask), "r" (&page->flags),
 265			       "i" (PG_dcache_cpu_mask),
 266			       "i" (PG_dcache_cpu_shift)
 267			     : "g1", "g7");
 268}
 269
 270static inline void tsb_insert(struct tsb *ent, unsigned long tag, unsigned long pte)
 271{
 272	unsigned long tsb_addr = (unsigned long) ent;
 273
 274	if (tlb_type == cheetah_plus || tlb_type == hypervisor)
 275		tsb_addr = __pa(tsb_addr);
 276
 277	__tsb_insert(tsb_addr, tag, pte);
 278}
 279
 280unsigned long _PAGE_ALL_SZ_BITS __read_mostly;
 281
 282static void flush_dcache(unsigned long pfn)
 283{
 284	struct page *page;
 285
 286	page = pfn_to_page(pfn);
 287	if (page) {
 288		unsigned long pg_flags;
 289
 290		pg_flags = page->flags;
 291		if (pg_flags & (1UL << PG_dcache_dirty)) {
 292			int cpu = ((pg_flags >> PG_dcache_cpu_shift) &
 293				   PG_dcache_cpu_mask);
 294			int this_cpu = get_cpu();
 295
 296			/* This is just to optimize away some function calls
 297			 * in the SMP case.
 298			 */
 299			if (cpu == this_cpu)
 300				flush_dcache_page_impl(page);
 301			else
 302				smp_flush_dcache_page_impl(page, cpu);
 303
 304			clear_dcache_dirty_cpu(page, cpu);
 305
 306			put_cpu();
 307		}
 308	}
 309}
 310
 311/* mm->context.lock must be held */
 312static void __update_mmu_tsb_insert(struct mm_struct *mm, unsigned long tsb_index,
 313				    unsigned long tsb_hash_shift, unsigned long address,
 314				    unsigned long tte)
 315{
 316	struct tsb *tsb = mm->context.tsb_block[tsb_index].tsb;
 317	unsigned long tag;
 318
 319	if (unlikely(!tsb))
 320		return;
 321
 322	tsb += ((address >> tsb_hash_shift) &
 323		(mm->context.tsb_block[tsb_index].tsb_nentries - 1UL));
 324	tag = (address >> 22UL);
 325	tsb_insert(tsb, tag, tte);
 326}
 327
 328#ifdef CONFIG_HUGETLB_PAGE
 329static void __init add_huge_page_size(unsigned long size)
 330{
 331	unsigned int order;
 332
 333	if (size_to_hstate(size))
 334		return;
 335
 336	order = ilog2(size) - PAGE_SHIFT;
 337	hugetlb_add_hstate(order);
 338}
 339
 340static int __init hugetlbpage_init(void)
 341{
 342	add_huge_page_size(1UL << HPAGE_64K_SHIFT);
 343	add_huge_page_size(1UL << HPAGE_SHIFT);
 344	add_huge_page_size(1UL << HPAGE_256MB_SHIFT);
 345	add_huge_page_size(1UL << HPAGE_2GB_SHIFT);
 346
 347	return 0;
 348}
 349
 350arch_initcall(hugetlbpage_init);
 351
 352static void __init pud_huge_patch(void)
 353{
 354	struct pud_huge_patch_entry *p;
 355	unsigned long addr;
 356
 357	p = &__pud_huge_patch;
 358	addr = p->addr;
 359	*(unsigned int *)addr = p->insn;
 360
 361	__asm__ __volatile__("flush %0" : : "r" (addr));
 362}
 363
 364static int __init setup_hugepagesz(char *string)
 365{
 366	unsigned long long hugepage_size;
 367	unsigned int hugepage_shift;
 368	unsigned short hv_pgsz_idx;
 369	unsigned int hv_pgsz_mask;
 370	int rc = 0;
 371
 372	hugepage_size = memparse(string, &string);
 373	hugepage_shift = ilog2(hugepage_size);
 374
 375	switch (hugepage_shift) {
 376	case HPAGE_16GB_SHIFT:
 377		hv_pgsz_mask = HV_PGSZ_MASK_16GB;
 378		hv_pgsz_idx = HV_PGSZ_IDX_16GB;
 379		pud_huge_patch();
 380		break;
 381	case HPAGE_2GB_SHIFT:
 382		hv_pgsz_mask = HV_PGSZ_MASK_2GB;
 383		hv_pgsz_idx = HV_PGSZ_IDX_2GB;
 384		break;
 385	case HPAGE_256MB_SHIFT:
 386		hv_pgsz_mask = HV_PGSZ_MASK_256MB;
 387		hv_pgsz_idx = HV_PGSZ_IDX_256MB;
 388		break;
 389	case HPAGE_SHIFT:
 390		hv_pgsz_mask = HV_PGSZ_MASK_4MB;
 391		hv_pgsz_idx = HV_PGSZ_IDX_4MB;
 392		break;
 393	case HPAGE_64K_SHIFT:
 394		hv_pgsz_mask = HV_PGSZ_MASK_64K;
 395		hv_pgsz_idx = HV_PGSZ_IDX_64K;
 396		break;
 397	default:
 398		hv_pgsz_mask = 0;
 399	}
 400
 401	if ((hv_pgsz_mask & cpu_pgsz_mask) == 0U) {
 402		hugetlb_bad_size();
 403		pr_err("hugepagesz=%llu not supported by MMU.\n",
 404			hugepage_size);
 405		goto out;
 406	}
 407
 408	add_huge_page_size(hugepage_size);
 409	rc = 1;
 410
 411out:
 412	return rc;
 413}
 414__setup("hugepagesz=", setup_hugepagesz);
 415#endif	/* CONFIG_HUGETLB_PAGE */
 416
 417void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
 418{
 419	struct mm_struct *mm;
 420	unsigned long flags;
 421	bool is_huge_tsb;
 422	pte_t pte = *ptep;
 423
 424	if (tlb_type != hypervisor) {
 425		unsigned long pfn = pte_pfn(pte);
 426
 427		if (pfn_valid(pfn))
 428			flush_dcache(pfn);
 429	}
 430
 431	mm = vma->vm_mm;
 432
 433	/* Don't insert a non-valid PTE into the TSB, we'll deadlock.  */
 434	if (!pte_accessible(mm, pte))
 435		return;
 436
 437	spin_lock_irqsave(&mm->context.lock, flags);
 438
 439	is_huge_tsb = false;
 440#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 441	if (mm->context.hugetlb_pte_count || mm->context.thp_pte_count) {
 442		unsigned long hugepage_size = PAGE_SIZE;
 443
 444		if (is_vm_hugetlb_page(vma))
 445			hugepage_size = huge_page_size(hstate_vma(vma));
 446
 447		if (hugepage_size >= PUD_SIZE) {
 448			unsigned long mask = 0x1ffc00000UL;
 449
 450			/* Transfer bits [32:22] from address to resolve
 451			 * at 4M granularity.
 452			 */
 453			pte_val(pte) &= ~mask;
 454			pte_val(pte) |= (address & mask);
 455		} else if (hugepage_size >= PMD_SIZE) {
 456			/* We are fabricating 8MB pages using 4MB
 457			 * real hw pages.
 458			 */
 459			pte_val(pte) |= (address & (1UL << REAL_HPAGE_SHIFT));
 460		}
 461
 462		if (hugepage_size >= PMD_SIZE) {
 463			__update_mmu_tsb_insert(mm, MM_TSB_HUGE,
 464				REAL_HPAGE_SHIFT, address, pte_val(pte));
 465			is_huge_tsb = true;
 466		}
 467	}
 468#endif
 469	if (!is_huge_tsb)
 470		__update_mmu_tsb_insert(mm, MM_TSB_BASE, PAGE_SHIFT,
 471					address, pte_val(pte));
 472
 473	spin_unlock_irqrestore(&mm->context.lock, flags);
 474}
 475
 476void flush_dcache_page(struct page *page)
 477{
 478	struct address_space *mapping;
 479	int this_cpu;
 480
 481	if (tlb_type == hypervisor)
 482		return;
 483
 484	/* Do not bother with the expensive D-cache flush if it
 485	 * is merely the zero page.  The 'bigcore' testcase in GDB
 486	 * causes this case to run millions of times.
 487	 */
 488	if (page == ZERO_PAGE(0))
 489		return;
 490
 491	this_cpu = get_cpu();
 492
 493	mapping = page_mapping_file(page);
 494	if (mapping && !mapping_mapped(mapping)) {
 495		int dirty = test_bit(PG_dcache_dirty, &page->flags);
 496		if (dirty) {
 497			int dirty_cpu = dcache_dirty_cpu(page);
 498
 499			if (dirty_cpu == this_cpu)
 500				goto out;
 501			smp_flush_dcache_page_impl(page, dirty_cpu);
 502		}
 503		set_dcache_dirty(page, this_cpu);
 504	} else {
 505		/* We could delay the flush for the !page_mapping
 506		 * case too.  But that case is for exec env/arg
 507		 * pages and those are %99 certainly going to get
 508		 * faulted into the tlb (and thus flushed) anyways.
 509		 */
 510		flush_dcache_page_impl(page);
 511	}
 512
 513out:
 514	put_cpu();
 515}
 516EXPORT_SYMBOL(flush_dcache_page);
 517
 518void __kprobes flush_icache_range(unsigned long start, unsigned long end)
 519{
 520	/* Cheetah and Hypervisor platform cpus have coherent I-cache. */
 521	if (tlb_type == spitfire) {
 522		unsigned long kaddr;
 523
 524		/* This code only runs on Spitfire cpus so this is
 525		 * why we can assume _PAGE_PADDR_4U.
 526		 */
 527		for (kaddr = start; kaddr < end; kaddr += PAGE_SIZE) {
 528			unsigned long paddr, mask = _PAGE_PADDR_4U;
 529
 530			if (kaddr >= PAGE_OFFSET)
 531				paddr = kaddr & mask;
 532			else {
 533				pgd_t *pgdp = pgd_offset_k(kaddr);
 534				pud_t *pudp = pud_offset(pgdp, kaddr);
 535				pmd_t *pmdp = pmd_offset(pudp, kaddr);
 536				pte_t *ptep = pte_offset_kernel(pmdp, kaddr);
 537
 538				paddr = pte_val(*ptep) & mask;
 539			}
 540			__flush_icache_page(paddr);
 541		}
 542	}
 543}
 544EXPORT_SYMBOL(flush_icache_range);
 545
 546void mmu_info(struct seq_file *m)
 547{
 548	static const char *pgsz_strings[] = {
 549		"8K", "64K", "512K", "4MB", "32MB",
 550		"256MB", "2GB", "16GB",
 551	};
 552	int i, printed;
 553
 554	if (tlb_type == cheetah)
 555		seq_printf(m, "MMU Type\t: Cheetah\n");
 556	else if (tlb_type == cheetah_plus)
 557		seq_printf(m, "MMU Type\t: Cheetah+\n");
 558	else if (tlb_type == spitfire)
 559		seq_printf(m, "MMU Type\t: Spitfire\n");
 560	else if (tlb_type == hypervisor)
 561		seq_printf(m, "MMU Type\t: Hypervisor (sun4v)\n");
 562	else
 563		seq_printf(m, "MMU Type\t: ???\n");
 564
 565	seq_printf(m, "MMU PGSZs\t: ");
 566	printed = 0;
 567	for (i = 0; i < ARRAY_SIZE(pgsz_strings); i++) {
 568		if (cpu_pgsz_mask & (1UL << i)) {
 569			seq_printf(m, "%s%s",
 570				   printed ? "," : "", pgsz_strings[i]);
 571			printed++;
 572		}
 573	}
 574	seq_putc(m, '\n');
 575
 576#ifdef CONFIG_DEBUG_DCFLUSH
 577	seq_printf(m, "DCPageFlushes\t: %d\n",
 578		   atomic_read(&dcpage_flushes));
 579#ifdef CONFIG_SMP
 580	seq_printf(m, "DCPageFlushesXC\t: %d\n",
 581		   atomic_read(&dcpage_flushes_xcall));
 582#endif /* CONFIG_SMP */
 583#endif /* CONFIG_DEBUG_DCFLUSH */
 584}
 585
 586struct linux_prom_translation prom_trans[512] __read_mostly;
 587unsigned int prom_trans_ents __read_mostly;
 588
 589unsigned long kern_locked_tte_data;
 590
 591/* The obp translations are saved based on 8k pagesize, since obp can
 592 * use a mixture of pagesizes. Misses to the LOW_OBP_ADDRESS ->
 593 * HI_OBP_ADDRESS range are handled in ktlb.S.
 594 */
 595static inline int in_obp_range(unsigned long vaddr)
 596{
 597	return (vaddr >= LOW_OBP_ADDRESS &&
 598		vaddr < HI_OBP_ADDRESS);
 599}
 600
 601static int cmp_ptrans(const void *a, const void *b)
 602{
 603	const struct linux_prom_translation *x = a, *y = b;
 604
 605	if (x->virt > y->virt)
 606		return 1;
 607	if (x->virt < y->virt)
 608		return -1;
 609	return 0;
 610}
 611
 612/* Read OBP translations property into 'prom_trans[]'.  */
 613static void __init read_obp_translations(void)
 614{
 615	int n, node, ents, first, last, i;
 616
 617	node = prom_finddevice("/virtual-memory");
 618	n = prom_getproplen(node, "translations");
 619	if (unlikely(n == 0 || n == -1)) {
 620		prom_printf("prom_mappings: Couldn't get size.\n");
 621		prom_halt();
 622	}
 623	if (unlikely(n > sizeof(prom_trans))) {
 624		prom_printf("prom_mappings: Size %d is too big.\n", n);
 625		prom_halt();
 626	}
 627
 628	if ((n = prom_getproperty(node, "translations",
 629				  (char *)&prom_trans[0],
 630				  sizeof(prom_trans))) == -1) {
 631		prom_printf("prom_mappings: Couldn't get property.\n");
 632		prom_halt();
 633	}
 634
 635	n = n / sizeof(struct linux_prom_translation);
 636
 637	ents = n;
 638
 639	sort(prom_trans, ents, sizeof(struct linux_prom_translation),
 640	     cmp_ptrans, NULL);
 641
 642	/* Now kick out all the non-OBP entries.  */
 643	for (i = 0; i < ents; i++) {
 644		if (in_obp_range(prom_trans[i].virt))
 645			break;
 646	}
 647	first = i;
 648	for (; i < ents; i++) {
 649		if (!in_obp_range(prom_trans[i].virt))
 650			break;
 651	}
 652	last = i;
 653
 654	for (i = 0; i < (last - first); i++) {
 655		struct linux_prom_translation *src = &prom_trans[i + first];
 656		struct linux_prom_translation *dest = &prom_trans[i];
 657
 658		*dest = *src;
 659	}
 660	for (; i < ents; i++) {
 661		struct linux_prom_translation *dest = &prom_trans[i];
 662		dest->virt = dest->size = dest->data = 0x0UL;
 663	}
 664
 665	prom_trans_ents = last - first;
 666
 667	if (tlb_type == spitfire) {
 668		/* Clear diag TTE bits. */
 669		for (i = 0; i < prom_trans_ents; i++)
 670			prom_trans[i].data &= ~0x0003fe0000000000UL;
 671	}
 672
 673	/* Force execute bit on.  */
 674	for (i = 0; i < prom_trans_ents; i++)
 675		prom_trans[i].data |= (tlb_type == hypervisor ?
 676				       _PAGE_EXEC_4V : _PAGE_EXEC_4U);
 677}
 678
 679static void __init hypervisor_tlb_lock(unsigned long vaddr,
 680				       unsigned long pte,
 681				       unsigned long mmu)
 682{
 683	unsigned long ret = sun4v_mmu_map_perm_addr(vaddr, 0, pte, mmu);
 684
 685	if (ret != 0) {
 686		prom_printf("hypervisor_tlb_lock[%lx:%x:%lx:%lx]: "
 687			    "errors with %lx\n", vaddr, 0, pte, mmu, ret);
 688		prom_halt();
 689	}
 690}
 691
 692static unsigned long kern_large_tte(unsigned long paddr);
 693
 694static void __init remap_kernel(void)
 695{
 696	unsigned long phys_page, tte_vaddr, tte_data;
 697	int i, tlb_ent = sparc64_highest_locked_tlbent();
 698
 699	tte_vaddr = (unsigned long) KERNBASE;
 700	phys_page = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
 701	tte_data = kern_large_tte(phys_page);
 702
 703	kern_locked_tte_data = tte_data;
 704
 705	/* Now lock us into the TLBs via Hypervisor or OBP. */
 706	if (tlb_type == hypervisor) {
 707		for (i = 0; i < num_kernel_image_mappings; i++) {
 708			hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_DMMU);
 709			hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_IMMU);
 710			tte_vaddr += 0x400000;
 711			tte_data += 0x400000;
 712		}
 713	} else {
 714		for (i = 0; i < num_kernel_image_mappings; i++) {
 715			prom_dtlb_load(tlb_ent - i, tte_data, tte_vaddr);
 716			prom_itlb_load(tlb_ent - i, tte_data, tte_vaddr);
 717			tte_vaddr += 0x400000;
 718			tte_data += 0x400000;
 719		}
 720		sparc64_highest_unlocked_tlb_ent = tlb_ent - i;
 721	}
 722	if (tlb_type == cheetah_plus) {
 723		sparc64_kern_pri_context = (CTX_CHEETAH_PLUS_CTX0 |
 724					    CTX_CHEETAH_PLUS_NUC);
 725		sparc64_kern_pri_nuc_bits = CTX_CHEETAH_PLUS_NUC;
 726		sparc64_kern_sec_context = CTX_CHEETAH_PLUS_CTX0;
 727	}
 728}
 729
 730
 731static void __init inherit_prom_mappings(void)
 732{
 733	/* Now fixup OBP's idea about where we really are mapped. */
 734	printk("Remapping the kernel... ");
 735	remap_kernel();
 736	printk("done.\n");
 737}
 738
 739void prom_world(int enter)
 740{
 741	if (!enter)
 742		set_fs(get_fs());
 743
 744	__asm__ __volatile__("flushw");
 745}
 746
 747void __flush_dcache_range(unsigned long start, unsigned long end)
 748{
 749	unsigned long va;
 750
 751	if (tlb_type == spitfire) {
 752		int n = 0;
 753
 754		for (va = start; va < end; va += 32) {
 755			spitfire_put_dcache_tag(va & 0x3fe0, 0x0);
 756			if (++n >= 512)
 757				break;
 758		}
 759	} else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
 760		start = __pa(start);
 761		end = __pa(end);
 762		for (va = start; va < end; va += 32)
 763			__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
 764					     "membar #Sync"
 765					     : /* no outputs */
 766					     : "r" (va),
 767					       "i" (ASI_DCACHE_INVALIDATE));
 768	}
 769}
 770EXPORT_SYMBOL(__flush_dcache_range);
 771
 772/* get_new_mmu_context() uses "cache + 1".  */
 773DEFINE_SPINLOCK(ctx_alloc_lock);
 774unsigned long tlb_context_cache = CTX_FIRST_VERSION;
 775#define MAX_CTX_NR	(1UL << CTX_NR_BITS)
 776#define CTX_BMAP_SLOTS	BITS_TO_LONGS(MAX_CTX_NR)
 777DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR);
 778DEFINE_PER_CPU(struct mm_struct *, per_cpu_secondary_mm) = {0};
 779
 780static void mmu_context_wrap(void)
 781{
 782	unsigned long old_ver = tlb_context_cache & CTX_VERSION_MASK;
 783	unsigned long new_ver, new_ctx, old_ctx;
 784	struct mm_struct *mm;
 785	int cpu;
 786
 787	bitmap_zero(mmu_context_bmap, 1 << CTX_NR_BITS);
 788
 789	/* Reserve kernel context */
 790	set_bit(0, mmu_context_bmap);
 791
 792	new_ver = (tlb_context_cache & CTX_VERSION_MASK) + CTX_FIRST_VERSION;
 793	if (unlikely(new_ver == 0))
 794		new_ver = CTX_FIRST_VERSION;
 795	tlb_context_cache = new_ver;
 796
 797	/*
 798	 * Make sure that any new mm that are added into per_cpu_secondary_mm,
 799	 * are going to go through get_new_mmu_context() path.
 800	 */
 801	mb();
 802
 803	/*
 804	 * Updated versions to current on those CPUs that had valid secondary
 805	 * contexts
 806	 */
 807	for_each_online_cpu(cpu) {
 808		/*
 809		 * If a new mm is stored after we took this mm from the array,
 810		 * it will go into get_new_mmu_context() path, because we
 811		 * already bumped the version in tlb_context_cache.
 812		 */
 813		mm = per_cpu(per_cpu_secondary_mm, cpu);
 814
 815		if (unlikely(!mm || mm == &init_mm))
 816			continue;
 817
 818		old_ctx = mm->context.sparc64_ctx_val;
 819		if (likely((old_ctx & CTX_VERSION_MASK) == old_ver)) {
 820			new_ctx = (old_ctx & ~CTX_VERSION_MASK) | new_ver;
 821			set_bit(new_ctx & CTX_NR_MASK, mmu_context_bmap);
 822			mm->context.sparc64_ctx_val = new_ctx;
 823		}
 824	}
 825}
 826
 827/* Caller does TLB context flushing on local CPU if necessary.
 828 * The caller also ensures that CTX_VALID(mm->context) is false.
 829 *
 830 * We must be careful about boundary cases so that we never
 831 * let the user have CTX 0 (nucleus) or we ever use a CTX
 832 * version of zero (and thus NO_CONTEXT would not be caught
 833 * by version mis-match tests in mmu_context.h).
 834 *
 835 * Always invoked with interrupts disabled.
 836 */
 837void get_new_mmu_context(struct mm_struct *mm)
 838{
 839	unsigned long ctx, new_ctx;
 840	unsigned long orig_pgsz_bits;
 841
 842	spin_lock(&ctx_alloc_lock);
 843retry:
 844	/* wrap might have happened, test again if our context became valid */
 845	if (unlikely(CTX_VALID(mm->context)))
 846		goto out;
 847	orig_pgsz_bits = (mm->context.sparc64_ctx_val & CTX_PGSZ_MASK);
 848	ctx = (tlb_context_cache + 1) & CTX_NR_MASK;
 849	new_ctx = find_next_zero_bit(mmu_context_bmap, 1 << CTX_NR_BITS, ctx);
 850	if (new_ctx >= (1 << CTX_NR_BITS)) {
 851		new_ctx = find_next_zero_bit(mmu_context_bmap, ctx, 1);
 852		if (new_ctx >= ctx) {
 853			mmu_context_wrap();
 854			goto retry;
 855		}
 856	}
 857	if (mm->context.sparc64_ctx_val)
 858		cpumask_clear(mm_cpumask(mm));
 859	mmu_context_bmap[new_ctx>>6] |= (1UL << (new_ctx & 63));
 860	new_ctx |= (tlb_context_cache & CTX_VERSION_MASK);
 861	tlb_context_cache = new_ctx;
 862	mm->context.sparc64_ctx_val = new_ctx | orig_pgsz_bits;
 863out:
 864	spin_unlock(&ctx_alloc_lock);
 865}
 866
 867static int numa_enabled = 1;
 868static int numa_debug;
 869
 870static int __init early_numa(char *p)
 871{
 872	if (!p)
 873		return 0;
 874
 875	if (strstr(p, "off"))
 876		numa_enabled = 0;
 877
 878	if (strstr(p, "debug"))
 879		numa_debug = 1;
 880
 881	return 0;
 882}
 883early_param("numa", early_numa);
 884
 885#define numadbg(f, a...) \
 886do {	if (numa_debug) \
 887		printk(KERN_INFO f, ## a); \
 888} while (0)
 889
 890static void __init find_ramdisk(unsigned long phys_base)
 891{
 892#ifdef CONFIG_BLK_DEV_INITRD
 893	if (sparc_ramdisk_image || sparc_ramdisk_image64) {
 894		unsigned long ramdisk_image;
 895
 896		/* Older versions of the bootloader only supported a
 897		 * 32-bit physical address for the ramdisk image
 898		 * location, stored at sparc_ramdisk_image.  Newer
 899		 * SILO versions set sparc_ramdisk_image to zero and
 900		 * provide a full 64-bit physical address at
 901		 * sparc_ramdisk_image64.
 902		 */
 903		ramdisk_image = sparc_ramdisk_image;
 904		if (!ramdisk_image)
 905			ramdisk_image = sparc_ramdisk_image64;
 906
 907		/* Another bootloader quirk.  The bootloader normalizes
 908		 * the physical address to KERNBASE, so we have to
 909		 * factor that back out and add in the lowest valid
 910		 * physical page address to get the true physical address.
 911		 */
 912		ramdisk_image -= KERNBASE;
 913		ramdisk_image += phys_base;
 914
 915		numadbg("Found ramdisk at physical address 0x%lx, size %u\n",
 916			ramdisk_image, sparc_ramdisk_size);
 917
 918		initrd_start = ramdisk_image;
 919		initrd_end = ramdisk_image + sparc_ramdisk_size;
 920
 921		memblock_reserve(initrd_start, sparc_ramdisk_size);
 922
 923		initrd_start += PAGE_OFFSET;
 924		initrd_end += PAGE_OFFSET;
 925	}
 926#endif
 927}
 928
 929struct node_mem_mask {
 930	unsigned long mask;
 931	unsigned long match;
 932};
 933static struct node_mem_mask node_masks[MAX_NUMNODES];
 934static int num_node_masks;
 935
 936#ifdef CONFIG_NEED_MULTIPLE_NODES
 937
 938struct mdesc_mlgroup {
 939	u64	node;
 940	u64	latency;
 941	u64	match;
 942	u64	mask;
 943};
 944
 945static struct mdesc_mlgroup *mlgroups;
 946static int num_mlgroups;
 947
 948int numa_cpu_lookup_table[NR_CPUS];
 949cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
 950
 951struct mdesc_mblock {
 952	u64	base;
 953	u64	size;
 954	u64	offset; /* RA-to-PA */
 955};
 956static struct mdesc_mblock *mblocks;
 957static int num_mblocks;
 958
 959static struct mdesc_mblock * __init addr_to_mblock(unsigned long addr)
 960{
 961	struct mdesc_mblock *m = NULL;
 962	int i;
 963
 964	for (i = 0; i < num_mblocks; i++) {
 965		m = &mblocks[i];
 966
 967		if (addr >= m->base &&
 968		    addr < (m->base + m->size)) {
 969			break;
 970		}
 971	}
 972
 973	return m;
 974}
 975
 976static u64 __init memblock_nid_range_sun4u(u64 start, u64 end, int *nid)
 977{
 978	int prev_nid, new_nid;
 979
 980	prev_nid = -1;
 981	for ( ; start < end; start += PAGE_SIZE) {
 982		for (new_nid = 0; new_nid < num_node_masks; new_nid++) {
 983			struct node_mem_mask *p = &node_masks[new_nid];
 984
 985			if ((start & p->mask) == p->match) {
 986				if (prev_nid == -1)
 987					prev_nid = new_nid;
 988				break;
 989			}
 990		}
 991
 992		if (new_nid == num_node_masks) {
 993			prev_nid = 0;
 994			WARN_ONCE(1, "addr[%Lx] doesn't match a NUMA node rule. Some memory will be owned by node 0.",
 995				  start);
 996			break;
 997		}
 998
 999		if (prev_nid != new_nid)
1000			break;
1001	}
1002	*nid = prev_nid;
1003
1004	return start > end ? end : start;
1005}
1006
1007static u64 __init memblock_nid_range(u64 start, u64 end, int *nid)
1008{
1009	u64 ret_end, pa_start, m_mask, m_match, m_end;
1010	struct mdesc_mblock *mblock;
1011	int _nid, i;
1012
1013	if (tlb_type != hypervisor)
1014		return memblock_nid_range_sun4u(start, end, nid);
1015
1016	mblock = addr_to_mblock(start);
1017	if (!mblock) {
1018		WARN_ONCE(1, "memblock_nid_range: Can't find mblock addr[%Lx]",
1019			  start);
1020
1021		_nid = 0;
1022		ret_end = end;
1023		goto done;
1024	}
1025
1026	pa_start = start + mblock->offset;
1027	m_match = 0;
1028	m_mask = 0;
1029
1030	for (_nid = 0; _nid < num_node_masks; _nid++) {
1031		struct node_mem_mask *const m = &node_masks[_nid];
1032
1033		if ((pa_start & m->mask) == m->match) {
1034			m_match = m->match;
1035			m_mask = m->mask;
1036			break;
1037		}
1038	}
1039
1040	if (num_node_masks == _nid) {
1041		/* We could not find NUMA group, so default to 0, but lets
1042		 * search for latency group, so we could calculate the correct
1043		 * end address that we return
1044		 */
1045		_nid = 0;
1046
1047		for (i = 0; i < num_mlgroups; i++) {
1048			struct mdesc_mlgroup *const m = &mlgroups[i];
1049
1050			if ((pa_start & m->mask) == m->match) {
1051				m_match = m->match;
1052				m_mask = m->mask;
1053				break;
1054			}
1055		}
1056
1057		if (i == num_mlgroups) {
1058			WARN_ONCE(1, "memblock_nid_range: Can't find latency group addr[%Lx]",
1059				  start);
1060
1061			ret_end = end;
1062			goto done;
1063		}
1064	}
1065
1066	/*
1067	 * Each latency group has match and mask, and each memory block has an
1068	 * offset.  An address belongs to a latency group if its address matches
1069	 * the following formula: ((addr + offset) & mask) == match
1070	 * It is, however, slow to check every single page if it matches a
1071	 * particular latency group. As optimization we calculate end value by
1072	 * using bit arithmetics.
1073	 */
1074	m_end = m_match + (1ul << __ffs(m_mask)) - mblock->offset;
1075	m_end += pa_start & ~((1ul << fls64(m_mask)) - 1);
1076	ret_end = m_end > end ? end : m_end;
1077
1078done:
1079	*nid = _nid;
1080	return ret_end;
1081}
1082#endif
1083
1084/* This must be invoked after performing all of the necessary
1085 * memblock_set_node() calls for 'nid'.  We need to be able to get
1086 * correct data from get_pfn_range_for_nid().
1087 */
1088static void __init allocate_node_data(int nid)
1089{
1090	struct pglist_data *p;
1091	unsigned long start_pfn, end_pfn;
1092#ifdef CONFIG_NEED_MULTIPLE_NODES
1093	unsigned long paddr;
1094
1095	paddr = memblock_alloc_try_nid(sizeof(struct pglist_data), SMP_CACHE_BYTES, nid);
1096	if (!paddr) {
 
1097		prom_printf("Cannot allocate pglist_data for nid[%d]\n", nid);
1098		prom_halt();
1099	}
1100	NODE_DATA(nid) = __va(paddr);
1101	memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
1102
1103	NODE_DATA(nid)->node_id = nid;
1104#endif
1105
1106	p = NODE_DATA(nid);
1107
1108	get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1109	p->node_start_pfn = start_pfn;
1110	p->node_spanned_pages = end_pfn - start_pfn;
1111}
1112
1113static void init_node_masks_nonnuma(void)
1114{
1115#ifdef CONFIG_NEED_MULTIPLE_NODES
1116	int i;
1117#endif
1118
1119	numadbg("Initializing tables for non-numa.\n");
1120
1121	node_masks[0].mask = 0;
1122	node_masks[0].match = 0;
1123	num_node_masks = 1;
1124
1125#ifdef CONFIG_NEED_MULTIPLE_NODES
1126	for (i = 0; i < NR_CPUS; i++)
1127		numa_cpu_lookup_table[i] = 0;
1128
1129	cpumask_setall(&numa_cpumask_lookup_table[0]);
1130#endif
1131}
1132
1133#ifdef CONFIG_NEED_MULTIPLE_NODES
1134struct pglist_data *node_data[MAX_NUMNODES];
1135
1136EXPORT_SYMBOL(numa_cpu_lookup_table);
1137EXPORT_SYMBOL(numa_cpumask_lookup_table);
1138EXPORT_SYMBOL(node_data);
1139
1140static int scan_pio_for_cfg_handle(struct mdesc_handle *md, u64 pio,
1141				   u32 cfg_handle)
1142{
1143	u64 arc;
1144
1145	mdesc_for_each_arc(arc, md, pio, MDESC_ARC_TYPE_FWD) {
1146		u64 target = mdesc_arc_target(md, arc);
1147		const u64 *val;
1148
1149		val = mdesc_get_property(md, target,
1150					 "cfg-handle", NULL);
1151		if (val && *val == cfg_handle)
1152			return 0;
1153	}
1154	return -ENODEV;
1155}
1156
1157static int scan_arcs_for_cfg_handle(struct mdesc_handle *md, u64 grp,
1158				    u32 cfg_handle)
1159{
1160	u64 arc, candidate, best_latency = ~(u64)0;
1161
1162	candidate = MDESC_NODE_NULL;
1163	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1164		u64 target = mdesc_arc_target(md, arc);
1165		const char *name = mdesc_node_name(md, target);
1166		const u64 *val;
1167
1168		if (strcmp(name, "pio-latency-group"))
1169			continue;
1170
1171		val = mdesc_get_property(md, target, "latency", NULL);
1172		if (!val)
1173			continue;
1174
1175		if (*val < best_latency) {
1176			candidate = target;
1177			best_latency = *val;
1178		}
1179	}
1180
1181	if (candidate == MDESC_NODE_NULL)
1182		return -ENODEV;
1183
1184	return scan_pio_for_cfg_handle(md, candidate, cfg_handle);
1185}
1186
1187int of_node_to_nid(struct device_node *dp)
1188{
1189	const struct linux_prom64_registers *regs;
1190	struct mdesc_handle *md;
1191	u32 cfg_handle;
1192	int count, nid;
1193	u64 grp;
1194
1195	/* This is the right thing to do on currently supported
1196	 * SUN4U NUMA platforms as well, as the PCI controller does
1197	 * not sit behind any particular memory controller.
1198	 */
1199	if (!mlgroups)
1200		return -1;
1201
1202	regs = of_get_property(dp, "reg", NULL);
1203	if (!regs)
1204		return -1;
1205
1206	cfg_handle = (regs->phys_addr >> 32UL) & 0x0fffffff;
1207
1208	md = mdesc_grab();
1209
1210	count = 0;
1211	nid = -1;
1212	mdesc_for_each_node_by_name(md, grp, "group") {
1213		if (!scan_arcs_for_cfg_handle(md, grp, cfg_handle)) {
1214			nid = count;
1215			break;
1216		}
1217		count++;
1218	}
1219
1220	mdesc_release(md);
1221
1222	return nid;
1223}
1224
1225static void __init add_node_ranges(void)
1226{
1227	struct memblock_region *reg;
1228	unsigned long prev_max;
1229
1230memblock_resized:
1231	prev_max = memblock.memory.max;
1232
1233	for_each_memblock(memory, reg) {
1234		unsigned long size = reg->size;
1235		unsigned long start, end;
1236
1237		start = reg->base;
1238		end = start + size;
1239		while (start < end) {
1240			unsigned long this_end;
1241			int nid;
1242
1243			this_end = memblock_nid_range(start, end, &nid);
1244
1245			numadbg("Setting memblock NUMA node nid[%d] "
1246				"start[%lx] end[%lx]\n",
1247				nid, start, this_end);
1248
1249			memblock_set_node(start, this_end - start,
1250					  &memblock.memory, nid);
1251			if (memblock.memory.max != prev_max)
1252				goto memblock_resized;
1253			start = this_end;
1254		}
1255	}
1256}
1257
1258static int __init grab_mlgroups(struct mdesc_handle *md)
1259{
1260	unsigned long paddr;
1261	int count = 0;
1262	u64 node;
1263
1264	mdesc_for_each_node_by_name(md, node, "memory-latency-group")
1265		count++;
1266	if (!count)
1267		return -ENOENT;
1268
1269	paddr = memblock_alloc(count * sizeof(struct mdesc_mlgroup),
1270			  SMP_CACHE_BYTES);
1271	if (!paddr)
1272		return -ENOMEM;
1273
1274	mlgroups = __va(paddr);
1275	num_mlgroups = count;
1276
1277	count = 0;
1278	mdesc_for_each_node_by_name(md, node, "memory-latency-group") {
1279		struct mdesc_mlgroup *m = &mlgroups[count++];
1280		const u64 *val;
1281
1282		m->node = node;
1283
1284		val = mdesc_get_property(md, node, "latency", NULL);
1285		m->latency = *val;
1286		val = mdesc_get_property(md, node, "address-match", NULL);
1287		m->match = *val;
1288		val = mdesc_get_property(md, node, "address-mask", NULL);
1289		m->mask = *val;
1290
1291		numadbg("MLGROUP[%d]: node[%llx] latency[%llx] "
1292			"match[%llx] mask[%llx]\n",
1293			count - 1, m->node, m->latency, m->match, m->mask);
1294	}
1295
1296	return 0;
1297}
1298
1299static int __init grab_mblocks(struct mdesc_handle *md)
1300{
1301	unsigned long paddr;
1302	int count = 0;
1303	u64 node;
1304
1305	mdesc_for_each_node_by_name(md, node, "mblock")
1306		count++;
1307	if (!count)
1308		return -ENOENT;
1309
1310	paddr = memblock_alloc(count * sizeof(struct mdesc_mblock),
1311			  SMP_CACHE_BYTES);
1312	if (!paddr)
1313		return -ENOMEM;
1314
1315	mblocks = __va(paddr);
1316	num_mblocks = count;
1317
1318	count = 0;
1319	mdesc_for_each_node_by_name(md, node, "mblock") {
1320		struct mdesc_mblock *m = &mblocks[count++];
1321		const u64 *val;
1322
1323		val = mdesc_get_property(md, node, "base", NULL);
1324		m->base = *val;
1325		val = mdesc_get_property(md, node, "size", NULL);
1326		m->size = *val;
1327		val = mdesc_get_property(md, node,
1328					 "address-congruence-offset", NULL);
1329
1330		/* The address-congruence-offset property is optional.
1331		 * Explicity zero it be identifty this.
1332		 */
1333		if (val)
1334			m->offset = *val;
1335		else
1336			m->offset = 0UL;
1337
1338		numadbg("MBLOCK[%d]: base[%llx] size[%llx] offset[%llx]\n",
1339			count - 1, m->base, m->size, m->offset);
1340	}
1341
1342	return 0;
1343}
1344
1345static void __init numa_parse_mdesc_group_cpus(struct mdesc_handle *md,
1346					       u64 grp, cpumask_t *mask)
1347{
1348	u64 arc;
1349
1350	cpumask_clear(mask);
1351
1352	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_BACK) {
1353		u64 target = mdesc_arc_target(md, arc);
1354		const char *name = mdesc_node_name(md, target);
1355		const u64 *id;
1356
1357		if (strcmp(name, "cpu"))
1358			continue;
1359		id = mdesc_get_property(md, target, "id", NULL);
1360		if (*id < nr_cpu_ids)
1361			cpumask_set_cpu(*id, mask);
1362	}
1363}
1364
1365static struct mdesc_mlgroup * __init find_mlgroup(u64 node)
1366{
1367	int i;
1368
1369	for (i = 0; i < num_mlgroups; i++) {
1370		struct mdesc_mlgroup *m = &mlgroups[i];
1371		if (m->node == node)
1372			return m;
1373	}
1374	return NULL;
1375}
1376
1377int __node_distance(int from, int to)
1378{
1379	if ((from >= MAX_NUMNODES) || (to >= MAX_NUMNODES)) {
1380		pr_warn("Returning default NUMA distance value for %d->%d\n",
1381			from, to);
1382		return (from == to) ? LOCAL_DISTANCE : REMOTE_DISTANCE;
1383	}
1384	return numa_latency[from][to];
1385}
 
1386
1387static int __init find_best_numa_node_for_mlgroup(struct mdesc_mlgroup *grp)
1388{
1389	int i;
1390
1391	for (i = 0; i < MAX_NUMNODES; i++) {
1392		struct node_mem_mask *n = &node_masks[i];
1393
1394		if ((grp->mask == n->mask) && (grp->match == n->match))
1395			break;
1396	}
1397	return i;
1398}
1399
1400static void __init find_numa_latencies_for_group(struct mdesc_handle *md,
1401						 u64 grp, int index)
1402{
1403	u64 arc;
1404
1405	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1406		int tnode;
1407		u64 target = mdesc_arc_target(md, arc);
1408		struct mdesc_mlgroup *m = find_mlgroup(target);
1409
1410		if (!m)
1411			continue;
1412		tnode = find_best_numa_node_for_mlgroup(m);
1413		if (tnode == MAX_NUMNODES)
1414			continue;
1415		numa_latency[index][tnode] = m->latency;
1416	}
1417}
1418
1419static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp,
1420				      int index)
1421{
1422	struct mdesc_mlgroup *candidate = NULL;
1423	u64 arc, best_latency = ~(u64)0;
1424	struct node_mem_mask *n;
1425
1426	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1427		u64 target = mdesc_arc_target(md, arc);
1428		struct mdesc_mlgroup *m = find_mlgroup(target);
1429		if (!m)
1430			continue;
1431		if (m->latency < best_latency) {
1432			candidate = m;
1433			best_latency = m->latency;
1434		}
1435	}
1436	if (!candidate)
1437		return -ENOENT;
1438
1439	if (num_node_masks != index) {
1440		printk(KERN_ERR "Inconsistent NUMA state, "
1441		       "index[%d] != num_node_masks[%d]\n",
1442		       index, num_node_masks);
1443		return -EINVAL;
1444	}
1445
1446	n = &node_masks[num_node_masks++];
1447
1448	n->mask = candidate->mask;
1449	n->match = candidate->match;
1450
1451	numadbg("NUMA NODE[%d]: mask[%lx] match[%lx] (latency[%llx])\n",
1452		index, n->mask, n->match, candidate->latency);
1453
1454	return 0;
1455}
1456
1457static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp,
1458					 int index)
1459{
1460	cpumask_t mask;
1461	int cpu;
1462
1463	numa_parse_mdesc_group_cpus(md, grp, &mask);
1464
1465	for_each_cpu(cpu, &mask)
1466		numa_cpu_lookup_table[cpu] = index;
1467	cpumask_copy(&numa_cpumask_lookup_table[index], &mask);
1468
1469	if (numa_debug) {
1470		printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index);
1471		for_each_cpu(cpu, &mask)
1472			printk("%d ", cpu);
1473		printk("]\n");
1474	}
1475
1476	return numa_attach_mlgroup(md, grp, index);
1477}
1478
1479static int __init numa_parse_mdesc(void)
1480{
1481	struct mdesc_handle *md = mdesc_grab();
1482	int i, j, err, count;
1483	u64 node;
1484
1485	node = mdesc_node_by_name(md, MDESC_NODE_NULL, "latency-groups");
1486	if (node == MDESC_NODE_NULL) {
1487		mdesc_release(md);
1488		return -ENOENT;
1489	}
1490
1491	err = grab_mblocks(md);
1492	if (err < 0)
1493		goto out;
1494
1495	err = grab_mlgroups(md);
1496	if (err < 0)
1497		goto out;
1498
1499	count = 0;
1500	mdesc_for_each_node_by_name(md, node, "group") {
1501		err = numa_parse_mdesc_group(md, node, count);
1502		if (err < 0)
1503			break;
1504		count++;
1505	}
1506
1507	count = 0;
1508	mdesc_for_each_node_by_name(md, node, "group") {
1509		find_numa_latencies_for_group(md, node, count);
1510		count++;
1511	}
1512
1513	/* Normalize numa latency matrix according to ACPI SLIT spec. */
1514	for (i = 0; i < MAX_NUMNODES; i++) {
1515		u64 self_latency = numa_latency[i][i];
1516
1517		for (j = 0; j < MAX_NUMNODES; j++) {
1518			numa_latency[i][j] =
1519				(numa_latency[i][j] * LOCAL_DISTANCE) /
1520				self_latency;
1521		}
1522	}
1523
1524	add_node_ranges();
1525
1526	for (i = 0; i < num_node_masks; i++) {
1527		allocate_node_data(i);
1528		node_set_online(i);
1529	}
1530
1531	err = 0;
1532out:
1533	mdesc_release(md);
1534	return err;
1535}
1536
1537static int __init numa_parse_jbus(void)
1538{
1539	unsigned long cpu, index;
1540
1541	/* NUMA node id is encoded in bits 36 and higher, and there is
1542	 * a 1-to-1 mapping from CPU ID to NUMA node ID.
1543	 */
1544	index = 0;
1545	for_each_present_cpu(cpu) {
1546		numa_cpu_lookup_table[cpu] = index;
1547		cpumask_copy(&numa_cpumask_lookup_table[index], cpumask_of(cpu));
1548		node_masks[index].mask = ~((1UL << 36UL) - 1UL);
1549		node_masks[index].match = cpu << 36UL;
1550
1551		index++;
1552	}
1553	num_node_masks = index;
1554
1555	add_node_ranges();
1556
1557	for (index = 0; index < num_node_masks; index++) {
1558		allocate_node_data(index);
1559		node_set_online(index);
1560	}
1561
1562	return 0;
1563}
1564
1565static int __init numa_parse_sun4u(void)
1566{
1567	if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1568		unsigned long ver;
1569
1570		__asm__ ("rdpr %%ver, %0" : "=r" (ver));
1571		if ((ver >> 32UL) == __JALAPENO_ID ||
1572		    (ver >> 32UL) == __SERRANO_ID)
1573			return numa_parse_jbus();
1574	}
1575	return -1;
1576}
1577
1578static int __init bootmem_init_numa(void)
1579{
1580	int i, j;
1581	int err = -1;
1582
1583	numadbg("bootmem_init_numa()\n");
1584
1585	/* Some sane defaults for numa latency values */
1586	for (i = 0; i < MAX_NUMNODES; i++) {
1587		for (j = 0; j < MAX_NUMNODES; j++)
1588			numa_latency[i][j] = (i == j) ?
1589				LOCAL_DISTANCE : REMOTE_DISTANCE;
1590	}
1591
1592	if (numa_enabled) {
1593		if (tlb_type == hypervisor)
1594			err = numa_parse_mdesc();
1595		else
1596			err = numa_parse_sun4u();
1597	}
1598	return err;
1599}
1600
1601#else
1602
1603static int bootmem_init_numa(void)
1604{
1605	return -1;
1606}
1607
1608#endif
1609
1610static void __init bootmem_init_nonnuma(void)
1611{
1612	unsigned long top_of_ram = memblock_end_of_DRAM();
1613	unsigned long total_ram = memblock_phys_mem_size();
1614
1615	numadbg("bootmem_init_nonnuma()\n");
1616
1617	printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
1618	       top_of_ram, total_ram);
1619	printk(KERN_INFO "Memory hole size: %ldMB\n",
1620	       (top_of_ram - total_ram) >> 20);
1621
1622	init_node_masks_nonnuma();
1623	memblock_set_node(0, (phys_addr_t)ULLONG_MAX, &memblock.memory, 0);
1624	allocate_node_data(0);
1625	node_set_online(0);
1626}
1627
1628static unsigned long __init bootmem_init(unsigned long phys_base)
1629{
1630	unsigned long end_pfn;
1631
1632	end_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1633	max_pfn = max_low_pfn = end_pfn;
1634	min_low_pfn = (phys_base >> PAGE_SHIFT);
1635
1636	if (bootmem_init_numa() < 0)
1637		bootmem_init_nonnuma();
1638
1639	/* Dump memblock with node info. */
1640	memblock_dump_all();
1641
1642	/* XXX cpu notifier XXX */
1643
1644	sparse_memory_present_with_active_regions(MAX_NUMNODES);
1645	sparse_init();
1646
1647	return end_pfn;
1648}
1649
1650static struct linux_prom64_registers pall[MAX_BANKS] __initdata;
1651static int pall_ents __initdata;
1652
1653static unsigned long max_phys_bits = 40;
1654
1655bool kern_addr_valid(unsigned long addr)
1656{
1657	pgd_t *pgd;
 
1658	pud_t *pud;
1659	pmd_t *pmd;
1660	pte_t *pte;
1661
1662	if ((long)addr < 0L) {
1663		unsigned long pa = __pa(addr);
1664
1665		if ((pa >> max_phys_bits) != 0UL)
1666			return false;
1667
1668		return pfn_valid(pa >> PAGE_SHIFT);
1669	}
1670
1671	if (addr >= (unsigned long) KERNBASE &&
1672	    addr < (unsigned long)&_end)
1673		return true;
1674
1675	pgd = pgd_offset_k(addr);
1676	if (pgd_none(*pgd))
1677		return 0;
1678
1679	pud = pud_offset(pgd, addr);
 
 
 
 
1680	if (pud_none(*pud))
1681		return 0;
1682
1683	if (pud_large(*pud))
1684		return pfn_valid(pud_pfn(*pud));
1685
1686	pmd = pmd_offset(pud, addr);
1687	if (pmd_none(*pmd))
1688		return 0;
1689
1690	if (pmd_large(*pmd))
1691		return pfn_valid(pmd_pfn(*pmd));
1692
1693	pte = pte_offset_kernel(pmd, addr);
1694	if (pte_none(*pte))
1695		return 0;
1696
1697	return pfn_valid(pte_pfn(*pte));
1698}
1699EXPORT_SYMBOL(kern_addr_valid);
1700
1701static unsigned long __ref kernel_map_hugepud(unsigned long vstart,
1702					      unsigned long vend,
1703					      pud_t *pud)
1704{
1705	const unsigned long mask16gb = (1UL << 34) - 1UL;
1706	u64 pte_val = vstart;
1707
1708	/* Each PUD is 8GB */
1709	if ((vstart & mask16gb) ||
1710	    (vend - vstart <= mask16gb)) {
1711		pte_val ^= kern_linear_pte_xor[2];
1712		pud_val(*pud) = pte_val | _PAGE_PUD_HUGE;
1713
1714		return vstart + PUD_SIZE;
1715	}
1716
1717	pte_val ^= kern_linear_pte_xor[3];
1718	pte_val |= _PAGE_PUD_HUGE;
1719
1720	vend = vstart + mask16gb + 1UL;
1721	while (vstart < vend) {
1722		pud_val(*pud) = pte_val;
1723
1724		pte_val += PUD_SIZE;
1725		vstart += PUD_SIZE;
1726		pud++;
1727	}
1728	return vstart;
1729}
1730
1731static bool kernel_can_map_hugepud(unsigned long vstart, unsigned long vend,
1732				   bool guard)
1733{
1734	if (guard && !(vstart & ~PUD_MASK) && (vend - vstart) >= PUD_SIZE)
1735		return true;
1736
1737	return false;
1738}
1739
1740static unsigned long __ref kernel_map_hugepmd(unsigned long vstart,
1741					      unsigned long vend,
1742					      pmd_t *pmd)
1743{
1744	const unsigned long mask256mb = (1UL << 28) - 1UL;
1745	const unsigned long mask2gb = (1UL << 31) - 1UL;
1746	u64 pte_val = vstart;
1747
1748	/* Each PMD is 8MB */
1749	if ((vstart & mask256mb) ||
1750	    (vend - vstart <= mask256mb)) {
1751		pte_val ^= kern_linear_pte_xor[0];
1752		pmd_val(*pmd) = pte_val | _PAGE_PMD_HUGE;
1753
1754		return vstart + PMD_SIZE;
1755	}
1756
1757	if ((vstart & mask2gb) ||
1758	    (vend - vstart <= mask2gb)) {
1759		pte_val ^= kern_linear_pte_xor[1];
1760		pte_val |= _PAGE_PMD_HUGE;
1761		vend = vstart + mask256mb + 1UL;
1762	} else {
1763		pte_val ^= kern_linear_pte_xor[2];
1764		pte_val |= _PAGE_PMD_HUGE;
1765		vend = vstart + mask2gb + 1UL;
1766	}
1767
1768	while (vstart < vend) {
1769		pmd_val(*pmd) = pte_val;
1770
1771		pte_val += PMD_SIZE;
1772		vstart += PMD_SIZE;
1773		pmd++;
1774	}
1775
1776	return vstart;
1777}
1778
1779static bool kernel_can_map_hugepmd(unsigned long vstart, unsigned long vend,
1780				   bool guard)
1781{
1782	if (guard && !(vstart & ~PMD_MASK) && (vend - vstart) >= PMD_SIZE)
1783		return true;
1784
1785	return false;
1786}
1787
1788static unsigned long __ref kernel_map_range(unsigned long pstart,
1789					    unsigned long pend, pgprot_t prot,
1790					    bool use_huge)
1791{
1792	unsigned long vstart = PAGE_OFFSET + pstart;
1793	unsigned long vend = PAGE_OFFSET + pend;
1794	unsigned long alloc_bytes = 0UL;
1795
1796	if ((vstart & ~PAGE_MASK) || (vend & ~PAGE_MASK)) {
1797		prom_printf("kernel_map: Unaligned physmem[%lx:%lx]\n",
1798			    vstart, vend);
1799		prom_halt();
1800	}
1801
1802	while (vstart < vend) {
1803		unsigned long this_end, paddr = __pa(vstart);
1804		pgd_t *pgd = pgd_offset_k(vstart);
 
1805		pud_t *pud;
1806		pmd_t *pmd;
1807		pte_t *pte;
1808
1809		if (pgd_none(*pgd)) {
1810			pud_t *new;
1811
1812			new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
 
 
 
1813			alloc_bytes += PAGE_SIZE;
1814			pgd_populate(&init_mm, pgd, new);
1815		}
1816		pud = pud_offset(pgd, vstart);
 
 
 
 
 
 
 
 
 
 
 
 
 
1817		if (pud_none(*pud)) {
1818			pmd_t *new;
1819
1820			if (kernel_can_map_hugepud(vstart, vend, use_huge)) {
1821				vstart = kernel_map_hugepud(vstart, vend, pud);
1822				continue;
1823			}
1824			new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
 
 
 
1825			alloc_bytes += PAGE_SIZE;
1826			pud_populate(&init_mm, pud, new);
1827		}
1828
1829		pmd = pmd_offset(pud, vstart);
1830		if (pmd_none(*pmd)) {
1831			pte_t *new;
1832
1833			if (kernel_can_map_hugepmd(vstart, vend, use_huge)) {
1834				vstart = kernel_map_hugepmd(vstart, vend, pmd);
1835				continue;
1836			}
1837			new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
 
 
 
1838			alloc_bytes += PAGE_SIZE;
1839			pmd_populate_kernel(&init_mm, pmd, new);
1840		}
1841
1842		pte = pte_offset_kernel(pmd, vstart);
1843		this_end = (vstart + PMD_SIZE) & PMD_MASK;
1844		if (this_end > vend)
1845			this_end = vend;
1846
1847		while (vstart < this_end) {
1848			pte_val(*pte) = (paddr | pgprot_val(prot));
1849
1850			vstart += PAGE_SIZE;
1851			paddr += PAGE_SIZE;
1852			pte++;
1853		}
1854	}
1855
1856	return alloc_bytes;
 
 
 
 
 
1857}
1858
1859static void __init flush_all_kernel_tsbs(void)
1860{
1861	int i;
1862
1863	for (i = 0; i < KERNEL_TSB_NENTRIES; i++) {
1864		struct tsb *ent = &swapper_tsb[i];
1865
1866		ent->tag = (1UL << TSB_TAG_INVALID_BIT);
1867	}
1868#ifndef CONFIG_DEBUG_PAGEALLOC
1869	for (i = 0; i < KERNEL_TSB4M_NENTRIES; i++) {
1870		struct tsb *ent = &swapper_4m_tsb[i];
1871
1872		ent->tag = (1UL << TSB_TAG_INVALID_BIT);
1873	}
1874#endif
1875}
1876
1877extern unsigned int kvmap_linear_patch[1];
1878
1879static void __init kernel_physical_mapping_init(void)
1880{
1881	unsigned long i, mem_alloced = 0UL;
1882	bool use_huge = true;
1883
1884#ifdef CONFIG_DEBUG_PAGEALLOC
1885	use_huge = false;
1886#endif
1887	for (i = 0; i < pall_ents; i++) {
1888		unsigned long phys_start, phys_end;
1889
1890		phys_start = pall[i].phys_addr;
1891		phys_end = phys_start + pall[i].reg_size;
1892
1893		mem_alloced += kernel_map_range(phys_start, phys_end,
1894						PAGE_KERNEL, use_huge);
1895	}
1896
1897	printk("Allocated %ld bytes for kernel page tables.\n",
1898	       mem_alloced);
1899
1900	kvmap_linear_patch[0] = 0x01000000; /* nop */
1901	flushi(&kvmap_linear_patch[0]);
1902
1903	flush_all_kernel_tsbs();
1904
1905	__flush_tlb_all();
1906}
1907
1908#ifdef CONFIG_DEBUG_PAGEALLOC
1909void __kernel_map_pages(struct page *page, int numpages, int enable)
1910{
1911	unsigned long phys_start = page_to_pfn(page) << PAGE_SHIFT;
1912	unsigned long phys_end = phys_start + (numpages * PAGE_SIZE);
1913
1914	kernel_map_range(phys_start, phys_end,
1915			 (enable ? PAGE_KERNEL : __pgprot(0)), false);
1916
1917	flush_tsb_kernel_range(PAGE_OFFSET + phys_start,
1918			       PAGE_OFFSET + phys_end);
1919
1920	/* we should perform an IPI and flush all tlbs,
1921	 * but that can deadlock->flush only current cpu.
1922	 */
1923	__flush_tlb_kernel_range(PAGE_OFFSET + phys_start,
1924				 PAGE_OFFSET + phys_end);
1925}
1926#endif
1927
1928unsigned long __init find_ecache_flush_span(unsigned long size)
1929{
1930	int i;
1931
1932	for (i = 0; i < pavail_ents; i++) {
1933		if (pavail[i].reg_size >= size)
1934			return pavail[i].phys_addr;
1935	}
1936
1937	return ~0UL;
1938}
1939
1940unsigned long PAGE_OFFSET;
1941EXPORT_SYMBOL(PAGE_OFFSET);
1942
1943unsigned long VMALLOC_END   = 0x0000010000000000UL;
1944EXPORT_SYMBOL(VMALLOC_END);
1945
1946unsigned long sparc64_va_hole_top =    0xfffff80000000000UL;
1947unsigned long sparc64_va_hole_bottom = 0x0000080000000000UL;
1948
1949static void __init setup_page_offset(void)
1950{
1951	if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1952		/* Cheetah/Panther support a full 64-bit virtual
1953		 * address, so we can use all that our page tables
1954		 * support.
1955		 */
1956		sparc64_va_hole_top =    0xfff0000000000000UL;
1957		sparc64_va_hole_bottom = 0x0010000000000000UL;
1958
1959		max_phys_bits = 42;
1960	} else if (tlb_type == hypervisor) {
1961		switch (sun4v_chip_type) {
1962		case SUN4V_CHIP_NIAGARA1:
1963		case SUN4V_CHIP_NIAGARA2:
1964			/* T1 and T2 support 48-bit virtual addresses.  */
1965			sparc64_va_hole_top =    0xffff800000000000UL;
1966			sparc64_va_hole_bottom = 0x0000800000000000UL;
1967
1968			max_phys_bits = 39;
1969			break;
1970		case SUN4V_CHIP_NIAGARA3:
1971			/* T3 supports 48-bit virtual addresses.  */
1972			sparc64_va_hole_top =    0xffff800000000000UL;
1973			sparc64_va_hole_bottom = 0x0000800000000000UL;
1974
1975			max_phys_bits = 43;
1976			break;
1977		case SUN4V_CHIP_NIAGARA4:
1978		case SUN4V_CHIP_NIAGARA5:
1979		case SUN4V_CHIP_SPARC64X:
1980		case SUN4V_CHIP_SPARC_M6:
1981			/* T4 and later support 52-bit virtual addresses.  */
1982			sparc64_va_hole_top =    0xfff8000000000000UL;
1983			sparc64_va_hole_bottom = 0x0008000000000000UL;
1984			max_phys_bits = 47;
1985			break;
1986		case SUN4V_CHIP_SPARC_M7:
1987		case SUN4V_CHIP_SPARC_SN:
1988			/* M7 and later support 52-bit virtual addresses.  */
1989			sparc64_va_hole_top =    0xfff8000000000000UL;
1990			sparc64_va_hole_bottom = 0x0008000000000000UL;
1991			max_phys_bits = 49;
1992			break;
1993		case SUN4V_CHIP_SPARC_M8:
1994		default:
1995			/* M8 and later support 54-bit virtual addresses.
1996			 * However, restricting M8 and above VA bits to 53
1997			 * as 4-level page table cannot support more than
1998			 * 53 VA bits.
1999			 */
2000			sparc64_va_hole_top =    0xfff0000000000000UL;
2001			sparc64_va_hole_bottom = 0x0010000000000000UL;
2002			max_phys_bits = 51;
2003			break;
2004		}
2005	}
2006
2007	if (max_phys_bits > MAX_PHYS_ADDRESS_BITS) {
2008		prom_printf("MAX_PHYS_ADDRESS_BITS is too small, need %lu\n",
2009			    max_phys_bits);
2010		prom_halt();
2011	}
2012
2013	PAGE_OFFSET = sparc64_va_hole_top;
2014	VMALLOC_END = ((sparc64_va_hole_bottom >> 1) +
2015		       (sparc64_va_hole_bottom >> 2));
2016
2017	pr_info("MM: PAGE_OFFSET is 0x%016lx (max_phys_bits == %lu)\n",
2018		PAGE_OFFSET, max_phys_bits);
2019	pr_info("MM: VMALLOC [0x%016lx --> 0x%016lx]\n",
2020		VMALLOC_START, VMALLOC_END);
2021	pr_info("MM: VMEMMAP [0x%016lx --> 0x%016lx]\n",
2022		VMEMMAP_BASE, VMEMMAP_BASE << 1);
2023}
2024
2025static void __init tsb_phys_patch(void)
2026{
2027	struct tsb_ldquad_phys_patch_entry *pquad;
2028	struct tsb_phys_patch_entry *p;
2029
2030	pquad = &__tsb_ldquad_phys_patch;
2031	while (pquad < &__tsb_ldquad_phys_patch_end) {
2032		unsigned long addr = pquad->addr;
2033
2034		if (tlb_type == hypervisor)
2035			*(unsigned int *) addr = pquad->sun4v_insn;
2036		else
2037			*(unsigned int *) addr = pquad->sun4u_insn;
2038		wmb();
2039		__asm__ __volatile__("flush	%0"
2040				     : /* no outputs */
2041				     : "r" (addr));
2042
2043		pquad++;
2044	}
2045
2046	p = &__tsb_phys_patch;
2047	while (p < &__tsb_phys_patch_end) {
2048		unsigned long addr = p->addr;
2049
2050		*(unsigned int *) addr = p->insn;
2051		wmb();
2052		__asm__ __volatile__("flush	%0"
2053				     : /* no outputs */
2054				     : "r" (addr));
2055
2056		p++;
2057	}
2058}
2059
2060/* Don't mark as init, we give this to the Hypervisor.  */
2061#ifndef CONFIG_DEBUG_PAGEALLOC
2062#define NUM_KTSB_DESCR	2
2063#else
2064#define NUM_KTSB_DESCR	1
2065#endif
2066static struct hv_tsb_descr ktsb_descr[NUM_KTSB_DESCR];
2067
2068/* The swapper TSBs are loaded with a base sequence of:
2069 *
2070 *	sethi	%uhi(SYMBOL), REG1
2071 *	sethi	%hi(SYMBOL), REG2
2072 *	or	REG1, %ulo(SYMBOL), REG1
2073 *	or	REG2, %lo(SYMBOL), REG2
2074 *	sllx	REG1, 32, REG1
2075 *	or	REG1, REG2, REG1
2076 *
2077 * When we use physical addressing for the TSB accesses, we patch the
2078 * first four instructions in the above sequence.
2079 */
2080
2081static void patch_one_ktsb_phys(unsigned int *start, unsigned int *end, unsigned long pa)
2082{
2083	unsigned long high_bits, low_bits;
2084
2085	high_bits = (pa >> 32) & 0xffffffff;
2086	low_bits = (pa >> 0) & 0xffffffff;
2087
2088	while (start < end) {
2089		unsigned int *ia = (unsigned int *)(unsigned long)*start;
2090
2091		ia[0] = (ia[0] & ~0x3fffff) | (high_bits >> 10);
2092		__asm__ __volatile__("flush	%0" : : "r" (ia));
2093
2094		ia[1] = (ia[1] & ~0x3fffff) | (low_bits >> 10);
2095		__asm__ __volatile__("flush	%0" : : "r" (ia + 1));
2096
2097		ia[2] = (ia[2] & ~0x1fff) | (high_bits & 0x3ff);
2098		__asm__ __volatile__("flush	%0" : : "r" (ia + 2));
2099
2100		ia[3] = (ia[3] & ~0x1fff) | (low_bits & 0x3ff);
2101		__asm__ __volatile__("flush	%0" : : "r" (ia + 3));
2102
2103		start++;
2104	}
2105}
2106
2107static void ktsb_phys_patch(void)
2108{
2109	extern unsigned int __swapper_tsb_phys_patch;
2110	extern unsigned int __swapper_tsb_phys_patch_end;
2111	unsigned long ktsb_pa;
2112
2113	ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
2114	patch_one_ktsb_phys(&__swapper_tsb_phys_patch,
2115			    &__swapper_tsb_phys_patch_end, ktsb_pa);
2116#ifndef CONFIG_DEBUG_PAGEALLOC
2117	{
2118	extern unsigned int __swapper_4m_tsb_phys_patch;
2119	extern unsigned int __swapper_4m_tsb_phys_patch_end;
2120	ktsb_pa = (kern_base +
2121		   ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
2122	patch_one_ktsb_phys(&__swapper_4m_tsb_phys_patch,
2123			    &__swapper_4m_tsb_phys_patch_end, ktsb_pa);
2124	}
2125#endif
2126}
2127
2128static void __init sun4v_ktsb_init(void)
2129{
2130	unsigned long ktsb_pa;
2131
2132	/* First KTSB for PAGE_SIZE mappings.  */
2133	ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
2134
2135	switch (PAGE_SIZE) {
2136	case 8 * 1024:
2137	default:
2138		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_8K;
2139		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_8K;
2140		break;
2141
2142	case 64 * 1024:
2143		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_64K;
2144		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_64K;
2145		break;
2146
2147	case 512 * 1024:
2148		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_512K;
2149		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_512K;
2150		break;
2151
2152	case 4 * 1024 * 1024:
2153		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_4MB;
2154		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_4MB;
2155		break;
2156	}
2157
2158	ktsb_descr[0].assoc = 1;
2159	ktsb_descr[0].num_ttes = KERNEL_TSB_NENTRIES;
2160	ktsb_descr[0].ctx_idx = 0;
2161	ktsb_descr[0].tsb_base = ktsb_pa;
2162	ktsb_descr[0].resv = 0;
2163
2164#ifndef CONFIG_DEBUG_PAGEALLOC
2165	/* Second KTSB for 4MB/256MB/2GB/16GB mappings.  */
2166	ktsb_pa = (kern_base +
2167		   ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
2168
2169	ktsb_descr[1].pgsz_idx = HV_PGSZ_IDX_4MB;
2170	ktsb_descr[1].pgsz_mask = ((HV_PGSZ_MASK_4MB |
2171				    HV_PGSZ_MASK_256MB |
2172				    HV_PGSZ_MASK_2GB |
2173				    HV_PGSZ_MASK_16GB) &
2174				   cpu_pgsz_mask);
2175	ktsb_descr[1].assoc = 1;
2176	ktsb_descr[1].num_ttes = KERNEL_TSB4M_NENTRIES;
2177	ktsb_descr[1].ctx_idx = 0;
2178	ktsb_descr[1].tsb_base = ktsb_pa;
2179	ktsb_descr[1].resv = 0;
2180#endif
2181}
2182
2183void sun4v_ktsb_register(void)
2184{
2185	unsigned long pa, ret;
2186
2187	pa = kern_base + ((unsigned long)&ktsb_descr[0] - KERNBASE);
2188
2189	ret = sun4v_mmu_tsb_ctx0(NUM_KTSB_DESCR, pa);
2190	if (ret != 0) {
2191		prom_printf("hypervisor_mmu_tsb_ctx0[%lx]: "
2192			    "errors with %lx\n", pa, ret);
2193		prom_halt();
2194	}
2195}
2196
2197static void __init sun4u_linear_pte_xor_finalize(void)
2198{
2199#ifndef CONFIG_DEBUG_PAGEALLOC
2200	/* This is where we would add Panther support for
2201	 * 32MB and 256MB pages.
2202	 */
2203#endif
2204}
2205
2206static void __init sun4v_linear_pte_xor_finalize(void)
2207{
2208	unsigned long pagecv_flag;
2209
2210	/* Bit 9 of TTE is no longer CV bit on M7 processor and it instead
2211	 * enables MCD error. Do not set bit 9 on M7 processor.
2212	 */
2213	switch (sun4v_chip_type) {
2214	case SUN4V_CHIP_SPARC_M7:
2215	case SUN4V_CHIP_SPARC_M8:
2216	case SUN4V_CHIP_SPARC_SN:
2217		pagecv_flag = 0x00;
2218		break;
2219	default:
2220		pagecv_flag = _PAGE_CV_4V;
2221		break;
2222	}
2223#ifndef CONFIG_DEBUG_PAGEALLOC
2224	if (cpu_pgsz_mask & HV_PGSZ_MASK_256MB) {
2225		kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZ256MB_4V) ^
2226			PAGE_OFFSET;
2227		kern_linear_pte_xor[1] |= (_PAGE_CP_4V | pagecv_flag |
2228					   _PAGE_P_4V | _PAGE_W_4V);
2229	} else {
2230		kern_linear_pte_xor[1] = kern_linear_pte_xor[0];
2231	}
2232
2233	if (cpu_pgsz_mask & HV_PGSZ_MASK_2GB) {
2234		kern_linear_pte_xor[2] = (_PAGE_VALID | _PAGE_SZ2GB_4V) ^
2235			PAGE_OFFSET;
2236		kern_linear_pte_xor[2] |= (_PAGE_CP_4V | pagecv_flag |
2237					   _PAGE_P_4V | _PAGE_W_4V);
2238	} else {
2239		kern_linear_pte_xor[2] = kern_linear_pte_xor[1];
2240	}
2241
2242	if (cpu_pgsz_mask & HV_PGSZ_MASK_16GB) {
2243		kern_linear_pte_xor[3] = (_PAGE_VALID | _PAGE_SZ16GB_4V) ^
2244			PAGE_OFFSET;
2245		kern_linear_pte_xor[3] |= (_PAGE_CP_4V | pagecv_flag |
2246					   _PAGE_P_4V | _PAGE_W_4V);
2247	} else {
2248		kern_linear_pte_xor[3] = kern_linear_pte_xor[2];
2249	}
2250#endif
2251}
2252
2253/* paging_init() sets up the page tables */
2254
2255static unsigned long last_valid_pfn;
2256
2257static void sun4u_pgprot_init(void);
2258static void sun4v_pgprot_init(void);
2259
2260static phys_addr_t __init available_memory(void)
2261{
2262	phys_addr_t available = 0ULL;
2263	phys_addr_t pa_start, pa_end;
2264	u64 i;
2265
2266	for_each_free_mem_range(i, NUMA_NO_NODE, MEMBLOCK_NONE, &pa_start,
2267				&pa_end, NULL)
2268		available = available + (pa_end  - pa_start);
2269
2270	return available;
2271}
2272
2273#define _PAGE_CACHE_4U	(_PAGE_CP_4U | _PAGE_CV_4U)
2274#define _PAGE_CACHE_4V	(_PAGE_CP_4V | _PAGE_CV_4V)
2275#define __DIRTY_BITS_4U	 (_PAGE_MODIFIED_4U | _PAGE_WRITE_4U | _PAGE_W_4U)
2276#define __DIRTY_BITS_4V	 (_PAGE_MODIFIED_4V | _PAGE_WRITE_4V | _PAGE_W_4V)
2277#define __ACCESS_BITS_4U (_PAGE_ACCESSED_4U | _PAGE_READ_4U | _PAGE_R)
2278#define __ACCESS_BITS_4V (_PAGE_ACCESSED_4V | _PAGE_READ_4V | _PAGE_R)
2279
2280/* We need to exclude reserved regions. This exclusion will include
2281 * vmlinux and initrd. To be more precise the initrd size could be used to
2282 * compute a new lower limit because it is freed later during initialization.
2283 */
2284static void __init reduce_memory(phys_addr_t limit_ram)
2285{
2286	phys_addr_t avail_ram = available_memory();
2287	phys_addr_t pa_start, pa_end;
2288	u64 i;
2289
2290	if (limit_ram >= avail_ram)
2291		return;
2292
2293	for_each_free_mem_range(i, NUMA_NO_NODE, MEMBLOCK_NONE, &pa_start,
2294				&pa_end, NULL) {
2295		phys_addr_t region_size = pa_end - pa_start;
2296		phys_addr_t clip_start = pa_start;
2297
2298		avail_ram = avail_ram - region_size;
2299		/* Are we consuming too much? */
2300		if (avail_ram < limit_ram) {
2301			phys_addr_t give_back = limit_ram - avail_ram;
2302
2303			region_size = region_size - give_back;
2304			clip_start = clip_start + give_back;
2305		}
2306
2307		memblock_remove(clip_start, region_size);
2308
2309		if (avail_ram <= limit_ram)
2310			break;
2311		i = 0UL;
2312	}
2313}
2314
2315void __init paging_init(void)
2316{
2317	unsigned long end_pfn, shift, phys_base;
2318	unsigned long real_end, i;
2319
2320	setup_page_offset();
2321
2322	/* These build time checkes make sure that the dcache_dirty_cpu()
2323	 * page->flags usage will work.
2324	 *
2325	 * When a page gets marked as dcache-dirty, we store the
2326	 * cpu number starting at bit 32 in the page->flags.  Also,
2327	 * functions like clear_dcache_dirty_cpu use the cpu mask
2328	 * in 13-bit signed-immediate instruction fields.
2329	 */
2330
2331	/*
2332	 * Page flags must not reach into upper 32 bits that are used
2333	 * for the cpu number
2334	 */
2335	BUILD_BUG_ON(NR_PAGEFLAGS > 32);
2336
2337	/*
2338	 * The bit fields placed in the high range must not reach below
2339	 * the 32 bit boundary. Otherwise we cannot place the cpu field
2340	 * at the 32 bit boundary.
2341	 */
2342	BUILD_BUG_ON(SECTIONS_WIDTH + NODES_WIDTH + ZONES_WIDTH +
2343		ilog2(roundup_pow_of_two(NR_CPUS)) > 32);
2344
2345	BUILD_BUG_ON(NR_CPUS > 4096);
2346
2347	kern_base = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
2348	kern_size = (unsigned long)&_end - (unsigned long)KERNBASE;
2349
2350	/* Invalidate both kernel TSBs.  */
2351	memset(swapper_tsb, 0x40, sizeof(swapper_tsb));
2352#ifndef CONFIG_DEBUG_PAGEALLOC
2353	memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2354#endif
2355
2356	/* TTE.cv bit on sparc v9 occupies the same position as TTE.mcde
2357	 * bit on M7 processor. This is a conflicting usage of the same
2358	 * bit. Enabling TTE.cv on M7 would turn on Memory Corruption
2359	 * Detection error on all pages and this will lead to problems
2360	 * later. Kernel does not run with MCD enabled and hence rest
2361	 * of the required steps to fully configure memory corruption
2362	 * detection are not taken. We need to ensure TTE.mcde is not
2363	 * set on M7 processor. Compute the value of cacheability
2364	 * flag for use later taking this into consideration.
2365	 */
2366	switch (sun4v_chip_type) {
2367	case SUN4V_CHIP_SPARC_M7:
2368	case SUN4V_CHIP_SPARC_M8:
2369	case SUN4V_CHIP_SPARC_SN:
2370		page_cache4v_flag = _PAGE_CP_4V;
2371		break;
2372	default:
2373		page_cache4v_flag = _PAGE_CACHE_4V;
2374		break;
2375	}
2376
2377	if (tlb_type == hypervisor)
2378		sun4v_pgprot_init();
2379	else
2380		sun4u_pgprot_init();
2381
2382	if (tlb_type == cheetah_plus ||
2383	    tlb_type == hypervisor) {
2384		tsb_phys_patch();
2385		ktsb_phys_patch();
2386	}
2387
2388	if (tlb_type == hypervisor)
2389		sun4v_patch_tlb_handlers();
2390
2391	/* Find available physical memory...
2392	 *
2393	 * Read it twice in order to work around a bug in openfirmware.
2394	 * The call to grab this table itself can cause openfirmware to
2395	 * allocate memory, which in turn can take away some space from
2396	 * the list of available memory.  Reading it twice makes sure
2397	 * we really do get the final value.
2398	 */
2399	read_obp_translations();
2400	read_obp_memory("reg", &pall[0], &pall_ents);
2401	read_obp_memory("available", &pavail[0], &pavail_ents);
2402	read_obp_memory("available", &pavail[0], &pavail_ents);
2403
2404	phys_base = 0xffffffffffffffffUL;
2405	for (i = 0; i < pavail_ents; i++) {
2406		phys_base = min(phys_base, pavail[i].phys_addr);
2407		memblock_add(pavail[i].phys_addr, pavail[i].reg_size);
2408	}
2409
2410	memblock_reserve(kern_base, kern_size);
2411
2412	find_ramdisk(phys_base);
2413
2414	if (cmdline_memory_size)
2415		reduce_memory(cmdline_memory_size);
2416
2417	memblock_allow_resize();
2418	memblock_dump_all();
2419
2420	set_bit(0, mmu_context_bmap);
2421
2422	shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE);
2423
2424	real_end = (unsigned long)_end;
2425	num_kernel_image_mappings = DIV_ROUND_UP(real_end - KERNBASE, 1 << ILOG2_4MB);
2426	printk("Kernel: Using %d locked TLB entries for main kernel image.\n",
2427	       num_kernel_image_mappings);
2428
2429	/* Set kernel pgd to upper alias so physical page computations
2430	 * work.
2431	 */
2432	init_mm.pgd += ((shift) / (sizeof(pgd_t)));
2433	
2434	memset(swapper_pg_dir, 0, sizeof(swapper_pg_dir));
2435
2436	inherit_prom_mappings();
2437	
2438	/* Ok, we can use our TLB miss and window trap handlers safely.  */
2439	setup_tba();
2440
2441	__flush_tlb_all();
2442
2443	prom_build_devicetree();
2444	of_populate_present_mask();
2445#ifndef CONFIG_SMP
2446	of_fill_in_cpu_data();
2447#endif
2448
2449	if (tlb_type == hypervisor) {
2450		sun4v_mdesc_init();
2451		mdesc_populate_present_mask(cpu_all_mask);
2452#ifndef CONFIG_SMP
2453		mdesc_fill_in_cpu_data(cpu_all_mask);
2454#endif
2455		mdesc_get_page_sizes(cpu_all_mask, &cpu_pgsz_mask);
2456
2457		sun4v_linear_pte_xor_finalize();
2458
2459		sun4v_ktsb_init();
2460		sun4v_ktsb_register();
2461	} else {
2462		unsigned long impl, ver;
2463
2464		cpu_pgsz_mask = (HV_PGSZ_MASK_8K | HV_PGSZ_MASK_64K |
2465				 HV_PGSZ_MASK_512K | HV_PGSZ_MASK_4MB);
2466
2467		__asm__ __volatile__("rdpr %%ver, %0" : "=r" (ver));
2468		impl = ((ver >> 32) & 0xffff);
2469		if (impl == PANTHER_IMPL)
2470			cpu_pgsz_mask |= (HV_PGSZ_MASK_32MB |
2471					  HV_PGSZ_MASK_256MB);
2472
2473		sun4u_linear_pte_xor_finalize();
2474	}
2475
2476	/* Flush the TLBs and the 4M TSB so that the updated linear
2477	 * pte XOR settings are realized for all mappings.
2478	 */
2479	__flush_tlb_all();
2480#ifndef CONFIG_DEBUG_PAGEALLOC
2481	memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2482#endif
2483	__flush_tlb_all();
2484
2485	/* Setup bootmem... */
2486	last_valid_pfn = end_pfn = bootmem_init(phys_base);
2487
2488	kernel_physical_mapping_init();
2489
2490	{
2491		unsigned long max_zone_pfns[MAX_NR_ZONES];
2492
2493		memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
2494
2495		max_zone_pfns[ZONE_NORMAL] = end_pfn;
2496
2497		free_area_init_nodes(max_zone_pfns);
2498	}
2499
2500	printk("Booting Linux...\n");
2501}
2502
2503int page_in_phys_avail(unsigned long paddr)
2504{
2505	int i;
2506
2507	paddr &= PAGE_MASK;
2508
2509	for (i = 0; i < pavail_ents; i++) {
2510		unsigned long start, end;
2511
2512		start = pavail[i].phys_addr;
2513		end = start + pavail[i].reg_size;
2514
2515		if (paddr >= start && paddr < end)
2516			return 1;
2517	}
2518	if (paddr >= kern_base && paddr < (kern_base + kern_size))
2519		return 1;
2520#ifdef CONFIG_BLK_DEV_INITRD
2521	if (paddr >= __pa(initrd_start) &&
2522	    paddr < __pa(PAGE_ALIGN(initrd_end)))
2523		return 1;
2524#endif
2525
2526	return 0;
2527}
2528
2529static void __init register_page_bootmem_info(void)
2530{
2531#ifdef CONFIG_NEED_MULTIPLE_NODES
2532	int i;
2533
2534	for_each_online_node(i)
2535		if (NODE_DATA(i)->node_spanned_pages)
2536			register_page_bootmem_info_node(NODE_DATA(i));
2537#endif
2538}
2539void __init mem_init(void)
2540{
2541	high_memory = __va(last_valid_pfn << PAGE_SHIFT);
2542
2543	free_all_bootmem();
2544
2545	/*
2546	 * Must be done after boot memory is put on freelist, because here we
2547	 * might set fields in deferred struct pages that have not yet been
2548	 * initialized, and free_all_bootmem() initializes all the reserved
2549	 * deferred pages for us.
2550	 */
2551	register_page_bootmem_info();
2552
2553	/*
2554	 * Set up the zero page, mark it reserved, so that page count
2555	 * is not manipulated when freeing the page from user ptes.
2556	 */
2557	mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0);
2558	if (mem_map_zero == NULL) {
2559		prom_printf("paging_init: Cannot alloc zero page.\n");
2560		prom_halt();
2561	}
2562	mark_page_reserved(mem_map_zero);
2563
2564	mem_init_print_info(NULL);
2565
2566	if (tlb_type == cheetah || tlb_type == cheetah_plus)
2567		cheetah_ecache_flush_init();
2568}
2569
2570void free_initmem(void)
2571{
2572	unsigned long addr, initend;
2573	int do_free = 1;
2574
2575	/* If the physical memory maps were trimmed by kernel command
2576	 * line options, don't even try freeing this initmem stuff up.
2577	 * The kernel image could have been in the trimmed out region
2578	 * and if so the freeing below will free invalid page structs.
2579	 */
2580	if (cmdline_memory_size)
2581		do_free = 0;
2582
2583	/*
2584	 * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes.
2585	 */
2586	addr = PAGE_ALIGN((unsigned long)(__init_begin));
2587	initend = (unsigned long)(__init_end) & PAGE_MASK;
2588	for (; addr < initend; addr += PAGE_SIZE) {
2589		unsigned long page;
2590
2591		page = (addr +
2592			((unsigned long) __va(kern_base)) -
2593			((unsigned long) KERNBASE));
2594		memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
2595
2596		if (do_free)
2597			free_reserved_page(virt_to_page(page));
2598	}
2599}
2600
2601#ifdef CONFIG_BLK_DEV_INITRD
2602void free_initrd_mem(unsigned long start, unsigned long end)
2603{
2604	free_reserved_area((void *)start, (void *)end, POISON_FREE_INITMEM,
2605			   "initrd");
2606}
2607#endif
2608
2609pgprot_t PAGE_KERNEL __read_mostly;
2610EXPORT_SYMBOL(PAGE_KERNEL);
2611
2612pgprot_t PAGE_KERNEL_LOCKED __read_mostly;
2613pgprot_t PAGE_COPY __read_mostly;
2614
2615pgprot_t PAGE_SHARED __read_mostly;
2616EXPORT_SYMBOL(PAGE_SHARED);
2617
2618unsigned long pg_iobits __read_mostly;
2619
2620unsigned long _PAGE_IE __read_mostly;
2621EXPORT_SYMBOL(_PAGE_IE);
2622
2623unsigned long _PAGE_E __read_mostly;
2624EXPORT_SYMBOL(_PAGE_E);
2625
2626unsigned long _PAGE_CACHE __read_mostly;
2627EXPORT_SYMBOL(_PAGE_CACHE);
2628
2629#ifdef CONFIG_SPARSEMEM_VMEMMAP
2630int __meminit vmemmap_populate(unsigned long vstart, unsigned long vend,
2631			       int node, struct vmem_altmap *altmap)
2632{
2633	unsigned long pte_base;
2634
2635	pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2636		    _PAGE_CP_4U | _PAGE_CV_4U |
2637		    _PAGE_P_4U | _PAGE_W_4U);
2638	if (tlb_type == hypervisor)
2639		pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2640			    page_cache4v_flag | _PAGE_P_4V | _PAGE_W_4V);
2641
2642	pte_base |= _PAGE_PMD_HUGE;
2643
2644	vstart = vstart & PMD_MASK;
2645	vend = ALIGN(vend, PMD_SIZE);
2646	for (; vstart < vend; vstart += PMD_SIZE) {
2647		pgd_t *pgd = vmemmap_pgd_populate(vstart, node);
2648		unsigned long pte;
 
2649		pud_t *pud;
2650		pmd_t *pmd;
2651
2652		if (!pgd)
2653			return -ENOMEM;
2654
2655		pud = vmemmap_pud_populate(pgd, vstart, node);
 
 
 
 
2656		if (!pud)
2657			return -ENOMEM;
2658
2659		pmd = pmd_offset(pud, vstart);
2660		pte = pmd_val(*pmd);
2661		if (!(pte & _PAGE_VALID)) {
2662			void *block = vmemmap_alloc_block(PMD_SIZE, node);
2663
2664			if (!block)
2665				return -ENOMEM;
2666
2667			pmd_val(*pmd) = pte_base | __pa(block);
2668		}
2669	}
2670
2671	return 0;
2672}
2673
2674void vmemmap_free(unsigned long start, unsigned long end,
2675		struct vmem_altmap *altmap)
2676{
2677}
2678#endif /* CONFIG_SPARSEMEM_VMEMMAP */
2679
2680static void prot_init_common(unsigned long page_none,
2681			     unsigned long page_shared,
2682			     unsigned long page_copy,
2683			     unsigned long page_readonly,
2684			     unsigned long page_exec_bit)
2685{
2686	PAGE_COPY = __pgprot(page_copy);
2687	PAGE_SHARED = __pgprot(page_shared);
2688
2689	protection_map[0x0] = __pgprot(page_none);
2690	protection_map[0x1] = __pgprot(page_readonly & ~page_exec_bit);
2691	protection_map[0x2] = __pgprot(page_copy & ~page_exec_bit);
2692	protection_map[0x3] = __pgprot(page_copy & ~page_exec_bit);
2693	protection_map[0x4] = __pgprot(page_readonly);
2694	protection_map[0x5] = __pgprot(page_readonly);
2695	protection_map[0x6] = __pgprot(page_copy);
2696	protection_map[0x7] = __pgprot(page_copy);
2697	protection_map[0x8] = __pgprot(page_none);
2698	protection_map[0x9] = __pgprot(page_readonly & ~page_exec_bit);
2699	protection_map[0xa] = __pgprot(page_shared & ~page_exec_bit);
2700	protection_map[0xb] = __pgprot(page_shared & ~page_exec_bit);
2701	protection_map[0xc] = __pgprot(page_readonly);
2702	protection_map[0xd] = __pgprot(page_readonly);
2703	protection_map[0xe] = __pgprot(page_shared);
2704	protection_map[0xf] = __pgprot(page_shared);
2705}
2706
2707static void __init sun4u_pgprot_init(void)
2708{
2709	unsigned long page_none, page_shared, page_copy, page_readonly;
2710	unsigned long page_exec_bit;
2711	int i;
2712
2713	PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2714				_PAGE_CACHE_4U | _PAGE_P_4U |
2715				__ACCESS_BITS_4U | __DIRTY_BITS_4U |
2716				_PAGE_EXEC_4U);
2717	PAGE_KERNEL_LOCKED = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2718				       _PAGE_CACHE_4U | _PAGE_P_4U |
2719				       __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2720				       _PAGE_EXEC_4U | _PAGE_L_4U);
2721
2722	_PAGE_IE = _PAGE_IE_4U;
2723	_PAGE_E = _PAGE_E_4U;
2724	_PAGE_CACHE = _PAGE_CACHE_4U;
2725
2726	pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4U | __DIRTY_BITS_4U |
2727		     __ACCESS_BITS_4U | _PAGE_E_4U);
2728
2729#ifdef CONFIG_DEBUG_PAGEALLOC
2730	kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2731#else
2732	kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4U) ^
2733		PAGE_OFFSET;
2734#endif
2735	kern_linear_pte_xor[0] |= (_PAGE_CP_4U | _PAGE_CV_4U |
2736				   _PAGE_P_4U | _PAGE_W_4U);
2737
2738	for (i = 1; i < 4; i++)
2739		kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2740
2741	_PAGE_ALL_SZ_BITS =  (_PAGE_SZ4MB_4U | _PAGE_SZ512K_4U |
2742			      _PAGE_SZ64K_4U | _PAGE_SZ8K_4U |
2743			      _PAGE_SZ32MB_4U | _PAGE_SZ256MB_4U);
2744
2745
2746	page_none = _PAGE_PRESENT_4U | _PAGE_ACCESSED_4U | _PAGE_CACHE_4U;
2747	page_shared = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2748		       __ACCESS_BITS_4U | _PAGE_WRITE_4U | _PAGE_EXEC_4U);
2749	page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2750		       __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2751	page_readonly   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2752			   __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2753
2754	page_exec_bit = _PAGE_EXEC_4U;
2755
2756	prot_init_common(page_none, page_shared, page_copy, page_readonly,
2757			 page_exec_bit);
2758}
2759
2760static void __init sun4v_pgprot_init(void)
2761{
2762	unsigned long page_none, page_shared, page_copy, page_readonly;
2763	unsigned long page_exec_bit;
2764	int i;
2765
2766	PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4V | _PAGE_VALID |
2767				page_cache4v_flag | _PAGE_P_4V |
2768				__ACCESS_BITS_4V | __DIRTY_BITS_4V |
2769				_PAGE_EXEC_4V);
2770	PAGE_KERNEL_LOCKED = PAGE_KERNEL;
2771
2772	_PAGE_IE = _PAGE_IE_4V;
2773	_PAGE_E = _PAGE_E_4V;
2774	_PAGE_CACHE = page_cache4v_flag;
2775
2776#ifdef CONFIG_DEBUG_PAGEALLOC
2777	kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2778#else
2779	kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4V) ^
2780		PAGE_OFFSET;
2781#endif
2782	kern_linear_pte_xor[0] |= (page_cache4v_flag | _PAGE_P_4V |
2783				   _PAGE_W_4V);
2784
2785	for (i = 1; i < 4; i++)
2786		kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2787
2788	pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4V | __DIRTY_BITS_4V |
2789		     __ACCESS_BITS_4V | _PAGE_E_4V);
2790
2791	_PAGE_ALL_SZ_BITS = (_PAGE_SZ16GB_4V | _PAGE_SZ2GB_4V |
2792			     _PAGE_SZ256MB_4V | _PAGE_SZ32MB_4V |
2793			     _PAGE_SZ4MB_4V | _PAGE_SZ512K_4V |
2794			     _PAGE_SZ64K_4V | _PAGE_SZ8K_4V);
2795
2796	page_none = _PAGE_PRESENT_4V | _PAGE_ACCESSED_4V | page_cache4v_flag;
2797	page_shared = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2798		       __ACCESS_BITS_4V | _PAGE_WRITE_4V | _PAGE_EXEC_4V);
2799	page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2800		       __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2801	page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2802			 __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2803
2804	page_exec_bit = _PAGE_EXEC_4V;
2805
2806	prot_init_common(page_none, page_shared, page_copy, page_readonly,
2807			 page_exec_bit);
2808}
2809
2810unsigned long pte_sz_bits(unsigned long sz)
2811{
2812	if (tlb_type == hypervisor) {
2813		switch (sz) {
2814		case 8 * 1024:
2815		default:
2816			return _PAGE_SZ8K_4V;
2817		case 64 * 1024:
2818			return _PAGE_SZ64K_4V;
2819		case 512 * 1024:
2820			return _PAGE_SZ512K_4V;
2821		case 4 * 1024 * 1024:
2822			return _PAGE_SZ4MB_4V;
2823		}
2824	} else {
2825		switch (sz) {
2826		case 8 * 1024:
2827		default:
2828			return _PAGE_SZ8K_4U;
2829		case 64 * 1024:
2830			return _PAGE_SZ64K_4U;
2831		case 512 * 1024:
2832			return _PAGE_SZ512K_4U;
2833		case 4 * 1024 * 1024:
2834			return _PAGE_SZ4MB_4U;
2835		}
2836	}
2837}
2838
2839pte_t mk_pte_io(unsigned long page, pgprot_t prot, int space, unsigned long page_size)
2840{
2841	pte_t pte;
2842
2843	pte_val(pte)  = page | pgprot_val(pgprot_noncached(prot));
2844	pte_val(pte) |= (((unsigned long)space) << 32);
2845	pte_val(pte) |= pte_sz_bits(page_size);
2846
2847	return pte;
2848}
2849
2850static unsigned long kern_large_tte(unsigned long paddr)
2851{
2852	unsigned long val;
2853
2854	val = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2855	       _PAGE_CP_4U | _PAGE_CV_4U | _PAGE_P_4U |
2856	       _PAGE_EXEC_4U | _PAGE_L_4U | _PAGE_W_4U);
2857	if (tlb_type == hypervisor)
2858		val = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2859		       page_cache4v_flag | _PAGE_P_4V |
2860		       _PAGE_EXEC_4V | _PAGE_W_4V);
2861
2862	return val | paddr;
2863}
2864
2865/* If not locked, zap it. */
2866void __flush_tlb_all(void)
2867{
2868	unsigned long pstate;
2869	int i;
2870
2871	__asm__ __volatile__("flushw\n\t"
2872			     "rdpr	%%pstate, %0\n\t"
2873			     "wrpr	%0, %1, %%pstate"
2874			     : "=r" (pstate)
2875			     : "i" (PSTATE_IE));
2876	if (tlb_type == hypervisor) {
2877		sun4v_mmu_demap_all();
2878	} else if (tlb_type == spitfire) {
2879		for (i = 0; i < 64; i++) {
2880			/* Spitfire Errata #32 workaround */
2881			/* NOTE: Always runs on spitfire, so no
2882			 *       cheetah+ page size encodings.
2883			 */
2884			__asm__ __volatile__("stxa	%0, [%1] %2\n\t"
2885					     "flush	%%g6"
2886					     : /* No outputs */
2887					     : "r" (0),
2888					     "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2889
2890			if (!(spitfire_get_dtlb_data(i) & _PAGE_L_4U)) {
2891				__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2892						     "membar #Sync"
2893						     : /* no outputs */
2894						     : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
2895				spitfire_put_dtlb_data(i, 0x0UL);
2896			}
2897
2898			/* Spitfire Errata #32 workaround */
2899			/* NOTE: Always runs on spitfire, so no
2900			 *       cheetah+ page size encodings.
2901			 */
2902			__asm__ __volatile__("stxa	%0, [%1] %2\n\t"
2903					     "flush	%%g6"
2904					     : /* No outputs */
2905					     : "r" (0),
2906					     "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2907
2908			if (!(spitfire_get_itlb_data(i) & _PAGE_L_4U)) {
2909				__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2910						     "membar #Sync"
2911						     : /* no outputs */
2912						     : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
2913				spitfire_put_itlb_data(i, 0x0UL);
2914			}
2915		}
2916	} else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
2917		cheetah_flush_dtlb_all();
2918		cheetah_flush_itlb_all();
2919	}
2920	__asm__ __volatile__("wrpr	%0, 0, %%pstate"
2921			     : : "r" (pstate));
2922}
2923
2924pte_t *pte_alloc_one_kernel(struct mm_struct *mm,
2925			    unsigned long address)
2926{
2927	struct page *page = alloc_page(GFP_KERNEL | __GFP_ZERO);
2928	pte_t *pte = NULL;
2929
2930	if (page)
2931		pte = (pte_t *) page_address(page);
2932
2933	return pte;
2934}
2935
2936pgtable_t pte_alloc_one(struct mm_struct *mm,
2937			unsigned long address)
2938{
2939	struct page *page = alloc_page(GFP_KERNEL | __GFP_ZERO);
2940	if (!page)
2941		return NULL;
2942	if (!pgtable_page_ctor(page)) {
2943		free_unref_page(page);
2944		return NULL;
2945	}
2946	return (pte_t *) page_address(page);
2947}
2948
2949void pte_free_kernel(struct mm_struct *mm, pte_t *pte)
2950{
2951	free_page((unsigned long)pte);
2952}
2953
2954static void __pte_free(pgtable_t pte)
2955{
2956	struct page *page = virt_to_page(pte);
2957
2958	pgtable_page_dtor(page);
2959	__free_page(page);
2960}
2961
2962void pte_free(struct mm_struct *mm, pgtable_t pte)
2963{
2964	__pte_free(pte);
2965}
2966
2967void pgtable_free(void *table, bool is_page)
2968{
2969	if (is_page)
2970		__pte_free(table);
2971	else
2972		kmem_cache_free(pgtable_cache, table);
2973}
2974
2975#ifdef CONFIG_TRANSPARENT_HUGEPAGE
2976void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
2977			  pmd_t *pmd)
2978{
2979	unsigned long pte, flags;
2980	struct mm_struct *mm;
2981	pmd_t entry = *pmd;
2982
2983	if (!pmd_large(entry) || !pmd_young(entry))
2984		return;
2985
2986	pte = pmd_val(entry);
2987
2988	/* Don't insert a non-valid PMD into the TSB, we'll deadlock.  */
2989	if (!(pte & _PAGE_VALID))
2990		return;
2991
2992	/* We are fabricating 8MB pages using 4MB real hw pages.  */
2993	pte |= (addr & (1UL << REAL_HPAGE_SHIFT));
2994
2995	mm = vma->vm_mm;
2996
2997	spin_lock_irqsave(&mm->context.lock, flags);
2998
2999	if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL)
3000		__update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
3001					addr, pte);
3002
3003	spin_unlock_irqrestore(&mm->context.lock, flags);
3004}
3005#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3006
3007#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
3008static void context_reload(void *__data)
3009{
3010	struct mm_struct *mm = __data;
3011
3012	if (mm == current->mm)
3013		load_secondary_context(mm);
3014}
3015
3016void hugetlb_setup(struct pt_regs *regs)
3017{
3018	struct mm_struct *mm = current->mm;
3019	struct tsb_config *tp;
3020
3021	if (faulthandler_disabled() || !mm) {
3022		const struct exception_table_entry *entry;
3023
3024		entry = search_exception_tables(regs->tpc);
3025		if (entry) {
3026			regs->tpc = entry->fixup;
3027			regs->tnpc = regs->tpc + 4;
3028			return;
3029		}
3030		pr_alert("Unexpected HugeTLB setup in atomic context.\n");
3031		die_if_kernel("HugeTSB in atomic", regs);
3032	}
3033
3034	tp = &mm->context.tsb_block[MM_TSB_HUGE];
3035	if (likely(tp->tsb == NULL))
3036		tsb_grow(mm, MM_TSB_HUGE, 0);
3037
3038	tsb_context_switch(mm);
3039	smp_tsb_sync(mm);
3040
3041	/* On UltraSPARC-III+ and later, configure the second half of
3042	 * the Data-TLB for huge pages.
3043	 */
3044	if (tlb_type == cheetah_plus) {
3045		bool need_context_reload = false;
3046		unsigned long ctx;
3047
3048		spin_lock_irq(&ctx_alloc_lock);
3049		ctx = mm->context.sparc64_ctx_val;
3050		ctx &= ~CTX_PGSZ_MASK;
3051		ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT;
3052		ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT;
3053
3054		if (ctx != mm->context.sparc64_ctx_val) {
3055			/* When changing the page size fields, we
3056			 * must perform a context flush so that no
3057			 * stale entries match.  This flush must
3058			 * occur with the original context register
3059			 * settings.
3060			 */
3061			do_flush_tlb_mm(mm);
3062
3063			/* Reload the context register of all processors
3064			 * also executing in this address space.
3065			 */
3066			mm->context.sparc64_ctx_val = ctx;
3067			need_context_reload = true;
3068		}
3069		spin_unlock_irq(&ctx_alloc_lock);
3070
3071		if (need_context_reload)
3072			on_each_cpu(context_reload, mm, 0);
3073	}
3074}
3075#endif
3076
3077static struct resource code_resource = {
3078	.name	= "Kernel code",
3079	.flags	= IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
3080};
3081
3082static struct resource data_resource = {
3083	.name	= "Kernel data",
3084	.flags	= IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
3085};
3086
3087static struct resource bss_resource = {
3088	.name	= "Kernel bss",
3089	.flags	= IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
3090};
3091
3092static inline resource_size_t compute_kern_paddr(void *addr)
3093{
3094	return (resource_size_t) (addr - KERNBASE + kern_base);
3095}
3096
3097static void __init kernel_lds_init(void)
3098{
3099	code_resource.start = compute_kern_paddr(_text);
3100	code_resource.end   = compute_kern_paddr(_etext - 1);
3101	data_resource.start = compute_kern_paddr(_etext);
3102	data_resource.end   = compute_kern_paddr(_edata - 1);
3103	bss_resource.start  = compute_kern_paddr(__bss_start);
3104	bss_resource.end    = compute_kern_paddr(_end - 1);
3105}
3106
3107static int __init report_memory(void)
3108{
3109	int i;
3110	struct resource *res;
3111
3112	kernel_lds_init();
3113
3114	for (i = 0; i < pavail_ents; i++) {
3115		res = kzalloc(sizeof(struct resource), GFP_KERNEL);
3116
3117		if (!res) {
3118			pr_warn("Failed to allocate source.\n");
3119			break;
3120		}
3121
3122		res->name = "System RAM";
3123		res->start = pavail[i].phys_addr;
3124		res->end = pavail[i].phys_addr + pavail[i].reg_size - 1;
3125		res->flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM;
3126
3127		if (insert_resource(&iomem_resource, res) < 0) {
3128			pr_warn("Resource insertion failed.\n");
3129			break;
3130		}
3131
3132		insert_resource(res, &code_resource);
3133		insert_resource(res, &data_resource);
3134		insert_resource(res, &bss_resource);
3135	}
3136
3137	return 0;
3138}
3139arch_initcall(report_memory);
3140
3141#ifdef CONFIG_SMP
3142#define do_flush_tlb_kernel_range	smp_flush_tlb_kernel_range
3143#else
3144#define do_flush_tlb_kernel_range	__flush_tlb_kernel_range
3145#endif
3146
3147void flush_tlb_kernel_range(unsigned long start, unsigned long end)
3148{
3149	if (start < HI_OBP_ADDRESS && end > LOW_OBP_ADDRESS) {
3150		if (start < LOW_OBP_ADDRESS) {
3151			flush_tsb_kernel_range(start, LOW_OBP_ADDRESS);
3152			do_flush_tlb_kernel_range(start, LOW_OBP_ADDRESS);
3153		}
3154		if (end > HI_OBP_ADDRESS) {
3155			flush_tsb_kernel_range(HI_OBP_ADDRESS, end);
3156			do_flush_tlb_kernel_range(HI_OBP_ADDRESS, end);
3157		}
3158	} else {
3159		flush_tsb_kernel_range(start, end);
3160		do_flush_tlb_kernel_range(start, end);
3161	}
3162}
3163
3164void copy_user_highpage(struct page *to, struct page *from,
3165	unsigned long vaddr, struct vm_area_struct *vma)
3166{
3167	char *vfrom, *vto;
3168
3169	vfrom = kmap_atomic(from);
3170	vto = kmap_atomic(to);
3171	copy_user_page(vto, vfrom, vaddr, to);
3172	kunmap_atomic(vto);
3173	kunmap_atomic(vfrom);
3174
3175	/* If this page has ADI enabled, copy over any ADI tags
3176	 * as well
3177	 */
3178	if (vma->vm_flags & VM_SPARC_ADI) {
3179		unsigned long pfrom, pto, i, adi_tag;
3180
3181		pfrom = page_to_phys(from);
3182		pto = page_to_phys(to);
3183
3184		for (i = pfrom; i < (pfrom + PAGE_SIZE); i += adi_blksize()) {
3185			asm volatile("ldxa [%1] %2, %0\n\t"
3186					: "=r" (adi_tag)
3187					:  "r" (i), "i" (ASI_MCD_REAL));
3188			asm volatile("stxa %0, [%1] %2\n\t"
3189					:
3190					: "r" (adi_tag), "r" (pto),
3191					  "i" (ASI_MCD_REAL));
3192			pto += adi_blksize();
3193		}
3194		asm volatile("membar #Sync\n\t");
3195	}
3196}
3197EXPORT_SYMBOL(copy_user_highpage);
3198
3199void copy_highpage(struct page *to, struct page *from)
3200{
3201	char *vfrom, *vto;
3202
3203	vfrom = kmap_atomic(from);
3204	vto = kmap_atomic(to);
3205	copy_page(vto, vfrom);
3206	kunmap_atomic(vto);
3207	kunmap_atomic(vfrom);
3208
3209	/* If this platform is ADI enabled, copy any ADI tags
3210	 * as well
3211	 */
3212	if (adi_capable()) {
3213		unsigned long pfrom, pto, i, adi_tag;
3214
3215		pfrom = page_to_phys(from);
3216		pto = page_to_phys(to);
3217
3218		for (i = pfrom; i < (pfrom + PAGE_SIZE); i += adi_blksize()) {
3219			asm volatile("ldxa [%1] %2, %0\n\t"
3220					: "=r" (adi_tag)
3221					:  "r" (i), "i" (ASI_MCD_REAL));
3222			asm volatile("stxa %0, [%1] %2\n\t"
3223					:
3224					: "r" (adi_tag), "r" (pto),
3225					  "i" (ASI_MCD_REAL));
3226			pto += adi_blksize();
3227		}
3228		asm volatile("membar #Sync\n\t");
3229	}
3230}
3231EXPORT_SYMBOL(copy_highpage);
v5.9
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 *  arch/sparc64/mm/init.c
   4 *
   5 *  Copyright (C) 1996-1999 David S. Miller (davem@caip.rutgers.edu)
   6 *  Copyright (C) 1997-1999 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
   7 */
   8 
   9#include <linux/extable.h>
  10#include <linux/kernel.h>
  11#include <linux/sched.h>
  12#include <linux/string.h>
  13#include <linux/init.h>
  14#include <linux/memblock.h>
  15#include <linux/mm.h>
  16#include <linux/hugetlb.h>
  17#include <linux/initrd.h>
  18#include <linux/swap.h>
  19#include <linux/pagemap.h>
  20#include <linux/poison.h>
  21#include <linux/fs.h>
  22#include <linux/seq_file.h>
  23#include <linux/kprobes.h>
  24#include <linux/cache.h>
  25#include <linux/sort.h>
  26#include <linux/ioport.h>
  27#include <linux/percpu.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/oplib.h>
  35#include <asm/iommu.h>
  36#include <asm/io.h>
  37#include <linux/uaccess.h>
  38#include <asm/mmu_context.h>
  39#include <asm/tlbflush.h>
  40#include <asm/dma.h>
  41#include <asm/starfire.h>
  42#include <asm/tlb.h>
  43#include <asm/spitfire.h>
  44#include <asm/sections.h>
  45#include <asm/tsb.h>
  46#include <asm/hypervisor.h>
  47#include <asm/prom.h>
  48#include <asm/mdesc.h>
  49#include <asm/cpudata.h>
  50#include <asm/setup.h>
  51#include <asm/irq.h>
  52
  53#include "init_64.h"
  54
  55unsigned long kern_linear_pte_xor[4] __read_mostly;
  56static unsigned long page_cache4v_flag;
  57
  58/* A bitmap, two bits for every 256MB of physical memory.  These two
  59 * bits determine what page size we use for kernel linear
  60 * translations.  They form an index into kern_linear_pte_xor[].  The
  61 * value in the indexed slot is XOR'd with the TLB miss virtual
  62 * address to form the resulting TTE.  The mapping is:
  63 *
  64 *	0	==>	4MB
  65 *	1	==>	256MB
  66 *	2	==>	2GB
  67 *	3	==>	16GB
  68 *
  69 * All sun4v chips support 256MB pages.  Only SPARC-T4 and later
  70 * support 2GB pages, and hopefully future cpus will support the 16GB
  71 * pages as well.  For slots 2 and 3, we encode a 256MB TTE xor there
  72 * if these larger page sizes are not supported by the cpu.
  73 *
  74 * It would be nice to determine this from the machine description
  75 * 'cpu' properties, but we need to have this table setup before the
  76 * MDESC is initialized.
  77 */
  78
  79#ifndef CONFIG_DEBUG_PAGEALLOC
  80/* A special kernel TSB for 4MB, 256MB, 2GB and 16GB linear mappings.
  81 * Space is allocated for this right after the trap table in
  82 * arch/sparc64/kernel/head.S
  83 */
  84extern struct tsb swapper_4m_tsb[KERNEL_TSB4M_NENTRIES];
  85#endif
  86extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
  87
  88static unsigned long cpu_pgsz_mask;
  89
  90#define MAX_BANKS	1024
  91
  92static struct linux_prom64_registers pavail[MAX_BANKS];
  93static int pavail_ents;
  94
  95u64 numa_latency[MAX_NUMNODES][MAX_NUMNODES];
  96
  97static int cmp_p64(const void *a, const void *b)
  98{
  99	const struct linux_prom64_registers *x = a, *y = b;
 100
 101	if (x->phys_addr > y->phys_addr)
 102		return 1;
 103	if (x->phys_addr < y->phys_addr)
 104		return -1;
 105	return 0;
 106}
 107
 108static void __init read_obp_memory(const char *property,
 109				   struct linux_prom64_registers *regs,
 110				   int *num_ents)
 111{
 112	phandle node = prom_finddevice("/memory");
 113	int prop_size = prom_getproplen(node, property);
 114	int ents, ret, i;
 115
 116	ents = prop_size / sizeof(struct linux_prom64_registers);
 117	if (ents > MAX_BANKS) {
 118		prom_printf("The machine has more %s property entries than "
 119			    "this kernel can support (%d).\n",
 120			    property, MAX_BANKS);
 121		prom_halt();
 122	}
 123
 124	ret = prom_getproperty(node, property, (char *) regs, prop_size);
 125	if (ret == -1) {
 126		prom_printf("Couldn't get %s property from /memory.\n",
 127				property);
 128		prom_halt();
 129	}
 130
 131	/* Sanitize what we got from the firmware, by page aligning
 132	 * everything.
 133	 */
 134	for (i = 0; i < ents; i++) {
 135		unsigned long base, size;
 136
 137		base = regs[i].phys_addr;
 138		size = regs[i].reg_size;
 139
 140		size &= PAGE_MASK;
 141		if (base & ~PAGE_MASK) {
 142			unsigned long new_base = PAGE_ALIGN(base);
 143
 144			size -= new_base - base;
 145			if ((long) size < 0L)
 146				size = 0UL;
 147			base = new_base;
 148		}
 149		if (size == 0UL) {
 150			/* If it is empty, simply get rid of it.
 151			 * This simplifies the logic of the other
 152			 * functions that process these arrays.
 153			 */
 154			memmove(&regs[i], &regs[i + 1],
 155				(ents - i - 1) * sizeof(regs[0]));
 156			i--;
 157			ents--;
 158			continue;
 159		}
 160		regs[i].phys_addr = base;
 161		regs[i].reg_size = size;
 162	}
 163
 164	*num_ents = ents;
 165
 166	sort(regs, ents, sizeof(struct linux_prom64_registers),
 167	     cmp_p64, NULL);
 168}
 169
 170/* Kernel physical address base and size in bytes.  */
 171unsigned long kern_base __read_mostly;
 172unsigned long kern_size __read_mostly;
 173
 174/* Initial ramdisk setup */
 175extern unsigned long sparc_ramdisk_image64;
 176extern unsigned int sparc_ramdisk_image;
 177extern unsigned int sparc_ramdisk_size;
 178
 179struct page *mem_map_zero __read_mostly;
 180EXPORT_SYMBOL(mem_map_zero);
 181
 182unsigned int sparc64_highest_unlocked_tlb_ent __read_mostly;
 183
 184unsigned long sparc64_kern_pri_context __read_mostly;
 185unsigned long sparc64_kern_pri_nuc_bits __read_mostly;
 186unsigned long sparc64_kern_sec_context __read_mostly;
 187
 188int num_kernel_image_mappings;
 189
 190#ifdef CONFIG_DEBUG_DCFLUSH
 191atomic_t dcpage_flushes = ATOMIC_INIT(0);
 192#ifdef CONFIG_SMP
 193atomic_t dcpage_flushes_xcall = ATOMIC_INIT(0);
 194#endif
 195#endif
 196
 197inline void flush_dcache_page_impl(struct page *page)
 198{
 199	BUG_ON(tlb_type == hypervisor);
 200#ifdef CONFIG_DEBUG_DCFLUSH
 201	atomic_inc(&dcpage_flushes);
 202#endif
 203
 204#ifdef DCACHE_ALIASING_POSSIBLE
 205	__flush_dcache_page(page_address(page),
 206			    ((tlb_type == spitfire) &&
 207			     page_mapping_file(page) != NULL));
 208#else
 209	if (page_mapping_file(page) != NULL &&
 210	    tlb_type == spitfire)
 211		__flush_icache_page(__pa(page_address(page)));
 212#endif
 213}
 214
 215#define PG_dcache_dirty		PG_arch_1
 216#define PG_dcache_cpu_shift	32UL
 217#define PG_dcache_cpu_mask	\
 218	((1UL<<ilog2(roundup_pow_of_two(NR_CPUS)))-1UL)
 219
 220#define dcache_dirty_cpu(page) \
 221	(((page)->flags >> PG_dcache_cpu_shift) & PG_dcache_cpu_mask)
 222
 223static inline void set_dcache_dirty(struct page *page, int this_cpu)
 224{
 225	unsigned long mask = this_cpu;
 226	unsigned long non_cpu_bits;
 227
 228	non_cpu_bits = ~(PG_dcache_cpu_mask << PG_dcache_cpu_shift);
 229	mask = (mask << PG_dcache_cpu_shift) | (1UL << PG_dcache_dirty);
 230
 231	__asm__ __volatile__("1:\n\t"
 232			     "ldx	[%2], %%g7\n\t"
 233			     "and	%%g7, %1, %%g1\n\t"
 234			     "or	%%g1, %0, %%g1\n\t"
 235			     "casx	[%2], %%g7, %%g1\n\t"
 236			     "cmp	%%g7, %%g1\n\t"
 237			     "bne,pn	%%xcc, 1b\n\t"
 238			     " nop"
 239			     : /* no outputs */
 240			     : "r" (mask), "r" (non_cpu_bits), "r" (&page->flags)
 241			     : "g1", "g7");
 242}
 243
 244static inline void clear_dcache_dirty_cpu(struct page *page, unsigned long cpu)
 245{
 246	unsigned long mask = (1UL << PG_dcache_dirty);
 247
 248	__asm__ __volatile__("! test_and_clear_dcache_dirty\n"
 249			     "1:\n\t"
 250			     "ldx	[%2], %%g7\n\t"
 251			     "srlx	%%g7, %4, %%g1\n\t"
 252			     "and	%%g1, %3, %%g1\n\t"
 253			     "cmp	%%g1, %0\n\t"
 254			     "bne,pn	%%icc, 2f\n\t"
 255			     " andn	%%g7, %1, %%g1\n\t"
 256			     "casx	[%2], %%g7, %%g1\n\t"
 257			     "cmp	%%g7, %%g1\n\t"
 258			     "bne,pn	%%xcc, 1b\n\t"
 259			     " nop\n"
 260			     "2:"
 261			     : /* no outputs */
 262			     : "r" (cpu), "r" (mask), "r" (&page->flags),
 263			       "i" (PG_dcache_cpu_mask),
 264			       "i" (PG_dcache_cpu_shift)
 265			     : "g1", "g7");
 266}
 267
 268static inline void tsb_insert(struct tsb *ent, unsigned long tag, unsigned long pte)
 269{
 270	unsigned long tsb_addr = (unsigned long) ent;
 271
 272	if (tlb_type == cheetah_plus || tlb_type == hypervisor)
 273		tsb_addr = __pa(tsb_addr);
 274
 275	__tsb_insert(tsb_addr, tag, pte);
 276}
 277
 278unsigned long _PAGE_ALL_SZ_BITS __read_mostly;
 279
 280static void flush_dcache(unsigned long pfn)
 281{
 282	struct page *page;
 283
 284	page = pfn_to_page(pfn);
 285	if (page) {
 286		unsigned long pg_flags;
 287
 288		pg_flags = page->flags;
 289		if (pg_flags & (1UL << PG_dcache_dirty)) {
 290			int cpu = ((pg_flags >> PG_dcache_cpu_shift) &
 291				   PG_dcache_cpu_mask);
 292			int this_cpu = get_cpu();
 293
 294			/* This is just to optimize away some function calls
 295			 * in the SMP case.
 296			 */
 297			if (cpu == this_cpu)
 298				flush_dcache_page_impl(page);
 299			else
 300				smp_flush_dcache_page_impl(page, cpu);
 301
 302			clear_dcache_dirty_cpu(page, cpu);
 303
 304			put_cpu();
 305		}
 306	}
 307}
 308
 309/* mm->context.lock must be held */
 310static void __update_mmu_tsb_insert(struct mm_struct *mm, unsigned long tsb_index,
 311				    unsigned long tsb_hash_shift, unsigned long address,
 312				    unsigned long tte)
 313{
 314	struct tsb *tsb = mm->context.tsb_block[tsb_index].tsb;
 315	unsigned long tag;
 316
 317	if (unlikely(!tsb))
 318		return;
 319
 320	tsb += ((address >> tsb_hash_shift) &
 321		(mm->context.tsb_block[tsb_index].tsb_nentries - 1UL));
 322	tag = (address >> 22UL);
 323	tsb_insert(tsb, tag, tte);
 324}
 325
 326#ifdef CONFIG_HUGETLB_PAGE
 
 
 
 
 
 
 
 
 
 
 
 327static int __init hugetlbpage_init(void)
 328{
 329	hugetlb_add_hstate(HPAGE_64K_SHIFT - PAGE_SHIFT);
 330	hugetlb_add_hstate(HPAGE_SHIFT - PAGE_SHIFT);
 331	hugetlb_add_hstate(HPAGE_256MB_SHIFT - PAGE_SHIFT);
 332	hugetlb_add_hstate(HPAGE_2GB_SHIFT - PAGE_SHIFT);
 333
 334	return 0;
 335}
 336
 337arch_initcall(hugetlbpage_init);
 338
 339static void __init pud_huge_patch(void)
 340{
 341	struct pud_huge_patch_entry *p;
 342	unsigned long addr;
 343
 344	p = &__pud_huge_patch;
 345	addr = p->addr;
 346	*(unsigned int *)addr = p->insn;
 347
 348	__asm__ __volatile__("flush %0" : : "r" (addr));
 349}
 350
 351bool __init arch_hugetlb_valid_size(unsigned long size)
 352{
 353	unsigned int hugepage_shift = ilog2(size);
 
 354	unsigned short hv_pgsz_idx;
 355	unsigned int hv_pgsz_mask;
 
 
 
 
 356
 357	switch (hugepage_shift) {
 358	case HPAGE_16GB_SHIFT:
 359		hv_pgsz_mask = HV_PGSZ_MASK_16GB;
 360		hv_pgsz_idx = HV_PGSZ_IDX_16GB;
 361		pud_huge_patch();
 362		break;
 363	case HPAGE_2GB_SHIFT:
 364		hv_pgsz_mask = HV_PGSZ_MASK_2GB;
 365		hv_pgsz_idx = HV_PGSZ_IDX_2GB;
 366		break;
 367	case HPAGE_256MB_SHIFT:
 368		hv_pgsz_mask = HV_PGSZ_MASK_256MB;
 369		hv_pgsz_idx = HV_PGSZ_IDX_256MB;
 370		break;
 371	case HPAGE_SHIFT:
 372		hv_pgsz_mask = HV_PGSZ_MASK_4MB;
 373		hv_pgsz_idx = HV_PGSZ_IDX_4MB;
 374		break;
 375	case HPAGE_64K_SHIFT:
 376		hv_pgsz_mask = HV_PGSZ_MASK_64K;
 377		hv_pgsz_idx = HV_PGSZ_IDX_64K;
 378		break;
 379	default:
 380		hv_pgsz_mask = 0;
 381	}
 382
 383	if ((hv_pgsz_mask & cpu_pgsz_mask) == 0U)
 384		return false;
 
 
 
 
 
 
 
 385
 386	return true;
 
 387}
 
 388#endif	/* CONFIG_HUGETLB_PAGE */
 389
 390void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
 391{
 392	struct mm_struct *mm;
 393	unsigned long flags;
 394	bool is_huge_tsb;
 395	pte_t pte = *ptep;
 396
 397	if (tlb_type != hypervisor) {
 398		unsigned long pfn = pte_pfn(pte);
 399
 400		if (pfn_valid(pfn))
 401			flush_dcache(pfn);
 402	}
 403
 404	mm = vma->vm_mm;
 405
 406	/* Don't insert a non-valid PTE into the TSB, we'll deadlock.  */
 407	if (!pte_accessible(mm, pte))
 408		return;
 409
 410	spin_lock_irqsave(&mm->context.lock, flags);
 411
 412	is_huge_tsb = false;
 413#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 414	if (mm->context.hugetlb_pte_count || mm->context.thp_pte_count) {
 415		unsigned long hugepage_size = PAGE_SIZE;
 416
 417		if (is_vm_hugetlb_page(vma))
 418			hugepage_size = huge_page_size(hstate_vma(vma));
 419
 420		if (hugepage_size >= PUD_SIZE) {
 421			unsigned long mask = 0x1ffc00000UL;
 422
 423			/* Transfer bits [32:22] from address to resolve
 424			 * at 4M granularity.
 425			 */
 426			pte_val(pte) &= ~mask;
 427			pte_val(pte) |= (address & mask);
 428		} else if (hugepage_size >= PMD_SIZE) {
 429			/* We are fabricating 8MB pages using 4MB
 430			 * real hw pages.
 431			 */
 432			pte_val(pte) |= (address & (1UL << REAL_HPAGE_SHIFT));
 433		}
 434
 435		if (hugepage_size >= PMD_SIZE) {
 436			__update_mmu_tsb_insert(mm, MM_TSB_HUGE,
 437				REAL_HPAGE_SHIFT, address, pte_val(pte));
 438			is_huge_tsb = true;
 439		}
 440	}
 441#endif
 442	if (!is_huge_tsb)
 443		__update_mmu_tsb_insert(mm, MM_TSB_BASE, PAGE_SHIFT,
 444					address, pte_val(pte));
 445
 446	spin_unlock_irqrestore(&mm->context.lock, flags);
 447}
 448
 449void flush_dcache_page(struct page *page)
 450{
 451	struct address_space *mapping;
 452	int this_cpu;
 453
 454	if (tlb_type == hypervisor)
 455		return;
 456
 457	/* Do not bother with the expensive D-cache flush if it
 458	 * is merely the zero page.  The 'bigcore' testcase in GDB
 459	 * causes this case to run millions of times.
 460	 */
 461	if (page == ZERO_PAGE(0))
 462		return;
 463
 464	this_cpu = get_cpu();
 465
 466	mapping = page_mapping_file(page);
 467	if (mapping && !mapping_mapped(mapping)) {
 468		int dirty = test_bit(PG_dcache_dirty, &page->flags);
 469		if (dirty) {
 470			int dirty_cpu = dcache_dirty_cpu(page);
 471
 472			if (dirty_cpu == this_cpu)
 473				goto out;
 474			smp_flush_dcache_page_impl(page, dirty_cpu);
 475		}
 476		set_dcache_dirty(page, this_cpu);
 477	} else {
 478		/* We could delay the flush for the !page_mapping
 479		 * case too.  But that case is for exec env/arg
 480		 * pages and those are %99 certainly going to get
 481		 * faulted into the tlb (and thus flushed) anyways.
 482		 */
 483		flush_dcache_page_impl(page);
 484	}
 485
 486out:
 487	put_cpu();
 488}
 489EXPORT_SYMBOL(flush_dcache_page);
 490
 491void __kprobes flush_icache_range(unsigned long start, unsigned long end)
 492{
 493	/* Cheetah and Hypervisor platform cpus have coherent I-cache. */
 494	if (tlb_type == spitfire) {
 495		unsigned long kaddr;
 496
 497		/* This code only runs on Spitfire cpus so this is
 498		 * why we can assume _PAGE_PADDR_4U.
 499		 */
 500		for (kaddr = start; kaddr < end; kaddr += PAGE_SIZE) {
 501			unsigned long paddr, mask = _PAGE_PADDR_4U;
 502
 503			if (kaddr >= PAGE_OFFSET)
 504				paddr = kaddr & mask;
 505			else {
 506				pte_t *ptep = virt_to_kpte(kaddr);
 
 
 
 507
 508				paddr = pte_val(*ptep) & mask;
 509			}
 510			__flush_icache_page(paddr);
 511		}
 512	}
 513}
 514EXPORT_SYMBOL(flush_icache_range);
 515
 516void mmu_info(struct seq_file *m)
 517{
 518	static const char *pgsz_strings[] = {
 519		"8K", "64K", "512K", "4MB", "32MB",
 520		"256MB", "2GB", "16GB",
 521	};
 522	int i, printed;
 523
 524	if (tlb_type == cheetah)
 525		seq_printf(m, "MMU Type\t: Cheetah\n");
 526	else if (tlb_type == cheetah_plus)
 527		seq_printf(m, "MMU Type\t: Cheetah+\n");
 528	else if (tlb_type == spitfire)
 529		seq_printf(m, "MMU Type\t: Spitfire\n");
 530	else if (tlb_type == hypervisor)
 531		seq_printf(m, "MMU Type\t: Hypervisor (sun4v)\n");
 532	else
 533		seq_printf(m, "MMU Type\t: ???\n");
 534
 535	seq_printf(m, "MMU PGSZs\t: ");
 536	printed = 0;
 537	for (i = 0; i < ARRAY_SIZE(pgsz_strings); i++) {
 538		if (cpu_pgsz_mask & (1UL << i)) {
 539			seq_printf(m, "%s%s",
 540				   printed ? "," : "", pgsz_strings[i]);
 541			printed++;
 542		}
 543	}
 544	seq_putc(m, '\n');
 545
 546#ifdef CONFIG_DEBUG_DCFLUSH
 547	seq_printf(m, "DCPageFlushes\t: %d\n",
 548		   atomic_read(&dcpage_flushes));
 549#ifdef CONFIG_SMP
 550	seq_printf(m, "DCPageFlushesXC\t: %d\n",
 551		   atomic_read(&dcpage_flushes_xcall));
 552#endif /* CONFIG_SMP */
 553#endif /* CONFIG_DEBUG_DCFLUSH */
 554}
 555
 556struct linux_prom_translation prom_trans[512] __read_mostly;
 557unsigned int prom_trans_ents __read_mostly;
 558
 559unsigned long kern_locked_tte_data;
 560
 561/* The obp translations are saved based on 8k pagesize, since obp can
 562 * use a mixture of pagesizes. Misses to the LOW_OBP_ADDRESS ->
 563 * HI_OBP_ADDRESS range are handled in ktlb.S.
 564 */
 565static inline int in_obp_range(unsigned long vaddr)
 566{
 567	return (vaddr >= LOW_OBP_ADDRESS &&
 568		vaddr < HI_OBP_ADDRESS);
 569}
 570
 571static int cmp_ptrans(const void *a, const void *b)
 572{
 573	const struct linux_prom_translation *x = a, *y = b;
 574
 575	if (x->virt > y->virt)
 576		return 1;
 577	if (x->virt < y->virt)
 578		return -1;
 579	return 0;
 580}
 581
 582/* Read OBP translations property into 'prom_trans[]'.  */
 583static void __init read_obp_translations(void)
 584{
 585	int n, node, ents, first, last, i;
 586
 587	node = prom_finddevice("/virtual-memory");
 588	n = prom_getproplen(node, "translations");
 589	if (unlikely(n == 0 || n == -1)) {
 590		prom_printf("prom_mappings: Couldn't get size.\n");
 591		prom_halt();
 592	}
 593	if (unlikely(n > sizeof(prom_trans))) {
 594		prom_printf("prom_mappings: Size %d is too big.\n", n);
 595		prom_halt();
 596	}
 597
 598	if ((n = prom_getproperty(node, "translations",
 599				  (char *)&prom_trans[0],
 600				  sizeof(prom_trans))) == -1) {
 601		prom_printf("prom_mappings: Couldn't get property.\n");
 602		prom_halt();
 603	}
 604
 605	n = n / sizeof(struct linux_prom_translation);
 606
 607	ents = n;
 608
 609	sort(prom_trans, ents, sizeof(struct linux_prom_translation),
 610	     cmp_ptrans, NULL);
 611
 612	/* Now kick out all the non-OBP entries.  */
 613	for (i = 0; i < ents; i++) {
 614		if (in_obp_range(prom_trans[i].virt))
 615			break;
 616	}
 617	first = i;
 618	for (; i < ents; i++) {
 619		if (!in_obp_range(prom_trans[i].virt))
 620			break;
 621	}
 622	last = i;
 623
 624	for (i = 0; i < (last - first); i++) {
 625		struct linux_prom_translation *src = &prom_trans[i + first];
 626		struct linux_prom_translation *dest = &prom_trans[i];
 627
 628		*dest = *src;
 629	}
 630	for (; i < ents; i++) {
 631		struct linux_prom_translation *dest = &prom_trans[i];
 632		dest->virt = dest->size = dest->data = 0x0UL;
 633	}
 634
 635	prom_trans_ents = last - first;
 636
 637	if (tlb_type == spitfire) {
 638		/* Clear diag TTE bits. */
 639		for (i = 0; i < prom_trans_ents; i++)
 640			prom_trans[i].data &= ~0x0003fe0000000000UL;
 641	}
 642
 643	/* Force execute bit on.  */
 644	for (i = 0; i < prom_trans_ents; i++)
 645		prom_trans[i].data |= (tlb_type == hypervisor ?
 646				       _PAGE_EXEC_4V : _PAGE_EXEC_4U);
 647}
 648
 649static void __init hypervisor_tlb_lock(unsigned long vaddr,
 650				       unsigned long pte,
 651				       unsigned long mmu)
 652{
 653	unsigned long ret = sun4v_mmu_map_perm_addr(vaddr, 0, pte, mmu);
 654
 655	if (ret != 0) {
 656		prom_printf("hypervisor_tlb_lock[%lx:%x:%lx:%lx]: "
 657			    "errors with %lx\n", vaddr, 0, pte, mmu, ret);
 658		prom_halt();
 659	}
 660}
 661
 662static unsigned long kern_large_tte(unsigned long paddr);
 663
 664static void __init remap_kernel(void)
 665{
 666	unsigned long phys_page, tte_vaddr, tte_data;
 667	int i, tlb_ent = sparc64_highest_locked_tlbent();
 668
 669	tte_vaddr = (unsigned long) KERNBASE;
 670	phys_page = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
 671	tte_data = kern_large_tte(phys_page);
 672
 673	kern_locked_tte_data = tte_data;
 674
 675	/* Now lock us into the TLBs via Hypervisor or OBP. */
 676	if (tlb_type == hypervisor) {
 677		for (i = 0; i < num_kernel_image_mappings; i++) {
 678			hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_DMMU);
 679			hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_IMMU);
 680			tte_vaddr += 0x400000;
 681			tte_data += 0x400000;
 682		}
 683	} else {
 684		for (i = 0; i < num_kernel_image_mappings; i++) {
 685			prom_dtlb_load(tlb_ent - i, tte_data, tte_vaddr);
 686			prom_itlb_load(tlb_ent - i, tte_data, tte_vaddr);
 687			tte_vaddr += 0x400000;
 688			tte_data += 0x400000;
 689		}
 690		sparc64_highest_unlocked_tlb_ent = tlb_ent - i;
 691	}
 692	if (tlb_type == cheetah_plus) {
 693		sparc64_kern_pri_context = (CTX_CHEETAH_PLUS_CTX0 |
 694					    CTX_CHEETAH_PLUS_NUC);
 695		sparc64_kern_pri_nuc_bits = CTX_CHEETAH_PLUS_NUC;
 696		sparc64_kern_sec_context = CTX_CHEETAH_PLUS_CTX0;
 697	}
 698}
 699
 700
 701static void __init inherit_prom_mappings(void)
 702{
 703	/* Now fixup OBP's idea about where we really are mapped. */
 704	printk("Remapping the kernel... ");
 705	remap_kernel();
 706	printk("done.\n");
 707}
 708
 709void prom_world(int enter)
 710{
 711	if (!enter)
 712		set_fs(get_fs());
 713
 714	__asm__ __volatile__("flushw");
 715}
 716
 717void __flush_dcache_range(unsigned long start, unsigned long end)
 718{
 719	unsigned long va;
 720
 721	if (tlb_type == spitfire) {
 722		int n = 0;
 723
 724		for (va = start; va < end; va += 32) {
 725			spitfire_put_dcache_tag(va & 0x3fe0, 0x0);
 726			if (++n >= 512)
 727				break;
 728		}
 729	} else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
 730		start = __pa(start);
 731		end = __pa(end);
 732		for (va = start; va < end; va += 32)
 733			__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
 734					     "membar #Sync"
 735					     : /* no outputs */
 736					     : "r" (va),
 737					       "i" (ASI_DCACHE_INVALIDATE));
 738	}
 739}
 740EXPORT_SYMBOL(__flush_dcache_range);
 741
 742/* get_new_mmu_context() uses "cache + 1".  */
 743DEFINE_SPINLOCK(ctx_alloc_lock);
 744unsigned long tlb_context_cache = CTX_FIRST_VERSION;
 745#define MAX_CTX_NR	(1UL << CTX_NR_BITS)
 746#define CTX_BMAP_SLOTS	BITS_TO_LONGS(MAX_CTX_NR)
 747DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR);
 748DEFINE_PER_CPU(struct mm_struct *, per_cpu_secondary_mm) = {0};
 749
 750static void mmu_context_wrap(void)
 751{
 752	unsigned long old_ver = tlb_context_cache & CTX_VERSION_MASK;
 753	unsigned long new_ver, new_ctx, old_ctx;
 754	struct mm_struct *mm;
 755	int cpu;
 756
 757	bitmap_zero(mmu_context_bmap, 1 << CTX_NR_BITS);
 758
 759	/* Reserve kernel context */
 760	set_bit(0, mmu_context_bmap);
 761
 762	new_ver = (tlb_context_cache & CTX_VERSION_MASK) + CTX_FIRST_VERSION;
 763	if (unlikely(new_ver == 0))
 764		new_ver = CTX_FIRST_VERSION;
 765	tlb_context_cache = new_ver;
 766
 767	/*
 768	 * Make sure that any new mm that are added into per_cpu_secondary_mm,
 769	 * are going to go through get_new_mmu_context() path.
 770	 */
 771	mb();
 772
 773	/*
 774	 * Updated versions to current on those CPUs that had valid secondary
 775	 * contexts
 776	 */
 777	for_each_online_cpu(cpu) {
 778		/*
 779		 * If a new mm is stored after we took this mm from the array,
 780		 * it will go into get_new_mmu_context() path, because we
 781		 * already bumped the version in tlb_context_cache.
 782		 */
 783		mm = per_cpu(per_cpu_secondary_mm, cpu);
 784
 785		if (unlikely(!mm || mm == &init_mm))
 786			continue;
 787
 788		old_ctx = mm->context.sparc64_ctx_val;
 789		if (likely((old_ctx & CTX_VERSION_MASK) == old_ver)) {
 790			new_ctx = (old_ctx & ~CTX_VERSION_MASK) | new_ver;
 791			set_bit(new_ctx & CTX_NR_MASK, mmu_context_bmap);
 792			mm->context.sparc64_ctx_val = new_ctx;
 793		}
 794	}
 795}
 796
 797/* Caller does TLB context flushing on local CPU if necessary.
 798 * The caller also ensures that CTX_VALID(mm->context) is false.
 799 *
 800 * We must be careful about boundary cases so that we never
 801 * let the user have CTX 0 (nucleus) or we ever use a CTX
 802 * version of zero (and thus NO_CONTEXT would not be caught
 803 * by version mis-match tests in mmu_context.h).
 804 *
 805 * Always invoked with interrupts disabled.
 806 */
 807void get_new_mmu_context(struct mm_struct *mm)
 808{
 809	unsigned long ctx, new_ctx;
 810	unsigned long orig_pgsz_bits;
 811
 812	spin_lock(&ctx_alloc_lock);
 813retry:
 814	/* wrap might have happened, test again if our context became valid */
 815	if (unlikely(CTX_VALID(mm->context)))
 816		goto out;
 817	orig_pgsz_bits = (mm->context.sparc64_ctx_val & CTX_PGSZ_MASK);
 818	ctx = (tlb_context_cache + 1) & CTX_NR_MASK;
 819	new_ctx = find_next_zero_bit(mmu_context_bmap, 1 << CTX_NR_BITS, ctx);
 820	if (new_ctx >= (1 << CTX_NR_BITS)) {
 821		new_ctx = find_next_zero_bit(mmu_context_bmap, ctx, 1);
 822		if (new_ctx >= ctx) {
 823			mmu_context_wrap();
 824			goto retry;
 825		}
 826	}
 827	if (mm->context.sparc64_ctx_val)
 828		cpumask_clear(mm_cpumask(mm));
 829	mmu_context_bmap[new_ctx>>6] |= (1UL << (new_ctx & 63));
 830	new_ctx |= (tlb_context_cache & CTX_VERSION_MASK);
 831	tlb_context_cache = new_ctx;
 832	mm->context.sparc64_ctx_val = new_ctx | orig_pgsz_bits;
 833out:
 834	spin_unlock(&ctx_alloc_lock);
 835}
 836
 837static int numa_enabled = 1;
 838static int numa_debug;
 839
 840static int __init early_numa(char *p)
 841{
 842	if (!p)
 843		return 0;
 844
 845	if (strstr(p, "off"))
 846		numa_enabled = 0;
 847
 848	if (strstr(p, "debug"))
 849		numa_debug = 1;
 850
 851	return 0;
 852}
 853early_param("numa", early_numa);
 854
 855#define numadbg(f, a...) \
 856do {	if (numa_debug) \
 857		printk(KERN_INFO f, ## a); \
 858} while (0)
 859
 860static void __init find_ramdisk(unsigned long phys_base)
 861{
 862#ifdef CONFIG_BLK_DEV_INITRD
 863	if (sparc_ramdisk_image || sparc_ramdisk_image64) {
 864		unsigned long ramdisk_image;
 865
 866		/* Older versions of the bootloader only supported a
 867		 * 32-bit physical address for the ramdisk image
 868		 * location, stored at sparc_ramdisk_image.  Newer
 869		 * SILO versions set sparc_ramdisk_image to zero and
 870		 * provide a full 64-bit physical address at
 871		 * sparc_ramdisk_image64.
 872		 */
 873		ramdisk_image = sparc_ramdisk_image;
 874		if (!ramdisk_image)
 875			ramdisk_image = sparc_ramdisk_image64;
 876
 877		/* Another bootloader quirk.  The bootloader normalizes
 878		 * the physical address to KERNBASE, so we have to
 879		 * factor that back out and add in the lowest valid
 880		 * physical page address to get the true physical address.
 881		 */
 882		ramdisk_image -= KERNBASE;
 883		ramdisk_image += phys_base;
 884
 885		numadbg("Found ramdisk at physical address 0x%lx, size %u\n",
 886			ramdisk_image, sparc_ramdisk_size);
 887
 888		initrd_start = ramdisk_image;
 889		initrd_end = ramdisk_image + sparc_ramdisk_size;
 890
 891		memblock_reserve(initrd_start, sparc_ramdisk_size);
 892
 893		initrd_start += PAGE_OFFSET;
 894		initrd_end += PAGE_OFFSET;
 895	}
 896#endif
 897}
 898
 899struct node_mem_mask {
 900	unsigned long mask;
 901	unsigned long match;
 902};
 903static struct node_mem_mask node_masks[MAX_NUMNODES];
 904static int num_node_masks;
 905
 906#ifdef CONFIG_NEED_MULTIPLE_NODES
 907
 908struct mdesc_mlgroup {
 909	u64	node;
 910	u64	latency;
 911	u64	match;
 912	u64	mask;
 913};
 914
 915static struct mdesc_mlgroup *mlgroups;
 916static int num_mlgroups;
 917
 918int numa_cpu_lookup_table[NR_CPUS];
 919cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
 920
 921struct mdesc_mblock {
 922	u64	base;
 923	u64	size;
 924	u64	offset; /* RA-to-PA */
 925};
 926static struct mdesc_mblock *mblocks;
 927static int num_mblocks;
 928
 929static struct mdesc_mblock * __init addr_to_mblock(unsigned long addr)
 930{
 931	struct mdesc_mblock *m = NULL;
 932	int i;
 933
 934	for (i = 0; i < num_mblocks; i++) {
 935		m = &mblocks[i];
 936
 937		if (addr >= m->base &&
 938		    addr < (m->base + m->size)) {
 939			break;
 940		}
 941	}
 942
 943	return m;
 944}
 945
 946static u64 __init memblock_nid_range_sun4u(u64 start, u64 end, int *nid)
 947{
 948	int prev_nid, new_nid;
 949
 950	prev_nid = NUMA_NO_NODE;
 951	for ( ; start < end; start += PAGE_SIZE) {
 952		for (new_nid = 0; new_nid < num_node_masks; new_nid++) {
 953			struct node_mem_mask *p = &node_masks[new_nid];
 954
 955			if ((start & p->mask) == p->match) {
 956				if (prev_nid == NUMA_NO_NODE)
 957					prev_nid = new_nid;
 958				break;
 959			}
 960		}
 961
 962		if (new_nid == num_node_masks) {
 963			prev_nid = 0;
 964			WARN_ONCE(1, "addr[%Lx] doesn't match a NUMA node rule. Some memory will be owned by node 0.",
 965				  start);
 966			break;
 967		}
 968
 969		if (prev_nid != new_nid)
 970			break;
 971	}
 972	*nid = prev_nid;
 973
 974	return start > end ? end : start;
 975}
 976
 977static u64 __init memblock_nid_range(u64 start, u64 end, int *nid)
 978{
 979	u64 ret_end, pa_start, m_mask, m_match, m_end;
 980	struct mdesc_mblock *mblock;
 981	int _nid, i;
 982
 983	if (tlb_type != hypervisor)
 984		return memblock_nid_range_sun4u(start, end, nid);
 985
 986	mblock = addr_to_mblock(start);
 987	if (!mblock) {
 988		WARN_ONCE(1, "memblock_nid_range: Can't find mblock addr[%Lx]",
 989			  start);
 990
 991		_nid = 0;
 992		ret_end = end;
 993		goto done;
 994	}
 995
 996	pa_start = start + mblock->offset;
 997	m_match = 0;
 998	m_mask = 0;
 999
1000	for (_nid = 0; _nid < num_node_masks; _nid++) {
1001		struct node_mem_mask *const m = &node_masks[_nid];
1002
1003		if ((pa_start & m->mask) == m->match) {
1004			m_match = m->match;
1005			m_mask = m->mask;
1006			break;
1007		}
1008	}
1009
1010	if (num_node_masks == _nid) {
1011		/* We could not find NUMA group, so default to 0, but lets
1012		 * search for latency group, so we could calculate the correct
1013		 * end address that we return
1014		 */
1015		_nid = 0;
1016
1017		for (i = 0; i < num_mlgroups; i++) {
1018			struct mdesc_mlgroup *const m = &mlgroups[i];
1019
1020			if ((pa_start & m->mask) == m->match) {
1021				m_match = m->match;
1022				m_mask = m->mask;
1023				break;
1024			}
1025		}
1026
1027		if (i == num_mlgroups) {
1028			WARN_ONCE(1, "memblock_nid_range: Can't find latency group addr[%Lx]",
1029				  start);
1030
1031			ret_end = end;
1032			goto done;
1033		}
1034	}
1035
1036	/*
1037	 * Each latency group has match and mask, and each memory block has an
1038	 * offset.  An address belongs to a latency group if its address matches
1039	 * the following formula: ((addr + offset) & mask) == match
1040	 * It is, however, slow to check every single page if it matches a
1041	 * particular latency group. As optimization we calculate end value by
1042	 * using bit arithmetics.
1043	 */
1044	m_end = m_match + (1ul << __ffs(m_mask)) - mblock->offset;
1045	m_end += pa_start & ~((1ul << fls64(m_mask)) - 1);
1046	ret_end = m_end > end ? end : m_end;
1047
1048done:
1049	*nid = _nid;
1050	return ret_end;
1051}
1052#endif
1053
1054/* This must be invoked after performing all of the necessary
1055 * memblock_set_node() calls for 'nid'.  We need to be able to get
1056 * correct data from get_pfn_range_for_nid().
1057 */
1058static void __init allocate_node_data(int nid)
1059{
1060	struct pglist_data *p;
1061	unsigned long start_pfn, end_pfn;
1062#ifdef CONFIG_NEED_MULTIPLE_NODES
 
1063
1064	NODE_DATA(nid) = memblock_alloc_node(sizeof(struct pglist_data),
1065					     SMP_CACHE_BYTES, nid);
1066	if (!NODE_DATA(nid)) {
1067		prom_printf("Cannot allocate pglist_data for nid[%d]\n", nid);
1068		prom_halt();
1069	}
 
 
1070
1071	NODE_DATA(nid)->node_id = nid;
1072#endif
1073
1074	p = NODE_DATA(nid);
1075
1076	get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
1077	p->node_start_pfn = start_pfn;
1078	p->node_spanned_pages = end_pfn - start_pfn;
1079}
1080
1081static void init_node_masks_nonnuma(void)
1082{
1083#ifdef CONFIG_NEED_MULTIPLE_NODES
1084	int i;
1085#endif
1086
1087	numadbg("Initializing tables for non-numa.\n");
1088
1089	node_masks[0].mask = 0;
1090	node_masks[0].match = 0;
1091	num_node_masks = 1;
1092
1093#ifdef CONFIG_NEED_MULTIPLE_NODES
1094	for (i = 0; i < NR_CPUS; i++)
1095		numa_cpu_lookup_table[i] = 0;
1096
1097	cpumask_setall(&numa_cpumask_lookup_table[0]);
1098#endif
1099}
1100
1101#ifdef CONFIG_NEED_MULTIPLE_NODES
1102struct pglist_data *node_data[MAX_NUMNODES];
1103
1104EXPORT_SYMBOL(numa_cpu_lookup_table);
1105EXPORT_SYMBOL(numa_cpumask_lookup_table);
1106EXPORT_SYMBOL(node_data);
1107
1108static int scan_pio_for_cfg_handle(struct mdesc_handle *md, u64 pio,
1109				   u32 cfg_handle)
1110{
1111	u64 arc;
1112
1113	mdesc_for_each_arc(arc, md, pio, MDESC_ARC_TYPE_FWD) {
1114		u64 target = mdesc_arc_target(md, arc);
1115		const u64 *val;
1116
1117		val = mdesc_get_property(md, target,
1118					 "cfg-handle", NULL);
1119		if (val && *val == cfg_handle)
1120			return 0;
1121	}
1122	return -ENODEV;
1123}
1124
1125static int scan_arcs_for_cfg_handle(struct mdesc_handle *md, u64 grp,
1126				    u32 cfg_handle)
1127{
1128	u64 arc, candidate, best_latency = ~(u64)0;
1129
1130	candidate = MDESC_NODE_NULL;
1131	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1132		u64 target = mdesc_arc_target(md, arc);
1133		const char *name = mdesc_node_name(md, target);
1134		const u64 *val;
1135
1136		if (strcmp(name, "pio-latency-group"))
1137			continue;
1138
1139		val = mdesc_get_property(md, target, "latency", NULL);
1140		if (!val)
1141			continue;
1142
1143		if (*val < best_latency) {
1144			candidate = target;
1145			best_latency = *val;
1146		}
1147	}
1148
1149	if (candidate == MDESC_NODE_NULL)
1150		return -ENODEV;
1151
1152	return scan_pio_for_cfg_handle(md, candidate, cfg_handle);
1153}
1154
1155int of_node_to_nid(struct device_node *dp)
1156{
1157	const struct linux_prom64_registers *regs;
1158	struct mdesc_handle *md;
1159	u32 cfg_handle;
1160	int count, nid;
1161	u64 grp;
1162
1163	/* This is the right thing to do on currently supported
1164	 * SUN4U NUMA platforms as well, as the PCI controller does
1165	 * not sit behind any particular memory controller.
1166	 */
1167	if (!mlgroups)
1168		return -1;
1169
1170	regs = of_get_property(dp, "reg", NULL);
1171	if (!regs)
1172		return -1;
1173
1174	cfg_handle = (regs->phys_addr >> 32UL) & 0x0fffffff;
1175
1176	md = mdesc_grab();
1177
1178	count = 0;
1179	nid = NUMA_NO_NODE;
1180	mdesc_for_each_node_by_name(md, grp, "group") {
1181		if (!scan_arcs_for_cfg_handle(md, grp, cfg_handle)) {
1182			nid = count;
1183			break;
1184		}
1185		count++;
1186	}
1187
1188	mdesc_release(md);
1189
1190	return nid;
1191}
1192
1193static void __init add_node_ranges(void)
1194{
1195	struct memblock_region *reg;
1196	unsigned long prev_max;
1197
1198memblock_resized:
1199	prev_max = memblock.memory.max;
1200
1201	for_each_memblock(memory, reg) {
1202		unsigned long size = reg->size;
1203		unsigned long start, end;
1204
1205		start = reg->base;
1206		end = start + size;
1207		while (start < end) {
1208			unsigned long this_end;
1209			int nid;
1210
1211			this_end = memblock_nid_range(start, end, &nid);
1212
1213			numadbg("Setting memblock NUMA node nid[%d] "
1214				"start[%lx] end[%lx]\n",
1215				nid, start, this_end);
1216
1217			memblock_set_node(start, this_end - start,
1218					  &memblock.memory, nid);
1219			if (memblock.memory.max != prev_max)
1220				goto memblock_resized;
1221			start = this_end;
1222		}
1223	}
1224}
1225
1226static int __init grab_mlgroups(struct mdesc_handle *md)
1227{
1228	unsigned long paddr;
1229	int count = 0;
1230	u64 node;
1231
1232	mdesc_for_each_node_by_name(md, node, "memory-latency-group")
1233		count++;
1234	if (!count)
1235		return -ENOENT;
1236
1237	paddr = memblock_phys_alloc(count * sizeof(struct mdesc_mlgroup),
1238				    SMP_CACHE_BYTES);
1239	if (!paddr)
1240		return -ENOMEM;
1241
1242	mlgroups = __va(paddr);
1243	num_mlgroups = count;
1244
1245	count = 0;
1246	mdesc_for_each_node_by_name(md, node, "memory-latency-group") {
1247		struct mdesc_mlgroup *m = &mlgroups[count++];
1248		const u64 *val;
1249
1250		m->node = node;
1251
1252		val = mdesc_get_property(md, node, "latency", NULL);
1253		m->latency = *val;
1254		val = mdesc_get_property(md, node, "address-match", NULL);
1255		m->match = *val;
1256		val = mdesc_get_property(md, node, "address-mask", NULL);
1257		m->mask = *val;
1258
1259		numadbg("MLGROUP[%d]: node[%llx] latency[%llx] "
1260			"match[%llx] mask[%llx]\n",
1261			count - 1, m->node, m->latency, m->match, m->mask);
1262	}
1263
1264	return 0;
1265}
1266
1267static int __init grab_mblocks(struct mdesc_handle *md)
1268{
1269	unsigned long paddr;
1270	int count = 0;
1271	u64 node;
1272
1273	mdesc_for_each_node_by_name(md, node, "mblock")
1274		count++;
1275	if (!count)
1276		return -ENOENT;
1277
1278	paddr = memblock_phys_alloc(count * sizeof(struct mdesc_mblock),
1279				    SMP_CACHE_BYTES);
1280	if (!paddr)
1281		return -ENOMEM;
1282
1283	mblocks = __va(paddr);
1284	num_mblocks = count;
1285
1286	count = 0;
1287	mdesc_for_each_node_by_name(md, node, "mblock") {
1288		struct mdesc_mblock *m = &mblocks[count++];
1289		const u64 *val;
1290
1291		val = mdesc_get_property(md, node, "base", NULL);
1292		m->base = *val;
1293		val = mdesc_get_property(md, node, "size", NULL);
1294		m->size = *val;
1295		val = mdesc_get_property(md, node,
1296					 "address-congruence-offset", NULL);
1297
1298		/* The address-congruence-offset property is optional.
1299		 * Explicity zero it be identifty this.
1300		 */
1301		if (val)
1302			m->offset = *val;
1303		else
1304			m->offset = 0UL;
1305
1306		numadbg("MBLOCK[%d]: base[%llx] size[%llx] offset[%llx]\n",
1307			count - 1, m->base, m->size, m->offset);
1308	}
1309
1310	return 0;
1311}
1312
1313static void __init numa_parse_mdesc_group_cpus(struct mdesc_handle *md,
1314					       u64 grp, cpumask_t *mask)
1315{
1316	u64 arc;
1317
1318	cpumask_clear(mask);
1319
1320	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_BACK) {
1321		u64 target = mdesc_arc_target(md, arc);
1322		const char *name = mdesc_node_name(md, target);
1323		const u64 *id;
1324
1325		if (strcmp(name, "cpu"))
1326			continue;
1327		id = mdesc_get_property(md, target, "id", NULL);
1328		if (*id < nr_cpu_ids)
1329			cpumask_set_cpu(*id, mask);
1330	}
1331}
1332
1333static struct mdesc_mlgroup * __init find_mlgroup(u64 node)
1334{
1335	int i;
1336
1337	for (i = 0; i < num_mlgroups; i++) {
1338		struct mdesc_mlgroup *m = &mlgroups[i];
1339		if (m->node == node)
1340			return m;
1341	}
1342	return NULL;
1343}
1344
1345int __node_distance(int from, int to)
1346{
1347	if ((from >= MAX_NUMNODES) || (to >= MAX_NUMNODES)) {
1348		pr_warn("Returning default NUMA distance value for %d->%d\n",
1349			from, to);
1350		return (from == to) ? LOCAL_DISTANCE : REMOTE_DISTANCE;
1351	}
1352	return numa_latency[from][to];
1353}
1354EXPORT_SYMBOL(__node_distance);
1355
1356static int __init find_best_numa_node_for_mlgroup(struct mdesc_mlgroup *grp)
1357{
1358	int i;
1359
1360	for (i = 0; i < MAX_NUMNODES; i++) {
1361		struct node_mem_mask *n = &node_masks[i];
1362
1363		if ((grp->mask == n->mask) && (grp->match == n->match))
1364			break;
1365	}
1366	return i;
1367}
1368
1369static void __init find_numa_latencies_for_group(struct mdesc_handle *md,
1370						 u64 grp, int index)
1371{
1372	u64 arc;
1373
1374	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1375		int tnode;
1376		u64 target = mdesc_arc_target(md, arc);
1377		struct mdesc_mlgroup *m = find_mlgroup(target);
1378
1379		if (!m)
1380			continue;
1381		tnode = find_best_numa_node_for_mlgroup(m);
1382		if (tnode == MAX_NUMNODES)
1383			continue;
1384		numa_latency[index][tnode] = m->latency;
1385	}
1386}
1387
1388static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp,
1389				      int index)
1390{
1391	struct mdesc_mlgroup *candidate = NULL;
1392	u64 arc, best_latency = ~(u64)0;
1393	struct node_mem_mask *n;
1394
1395	mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1396		u64 target = mdesc_arc_target(md, arc);
1397		struct mdesc_mlgroup *m = find_mlgroup(target);
1398		if (!m)
1399			continue;
1400		if (m->latency < best_latency) {
1401			candidate = m;
1402			best_latency = m->latency;
1403		}
1404	}
1405	if (!candidate)
1406		return -ENOENT;
1407
1408	if (num_node_masks != index) {
1409		printk(KERN_ERR "Inconsistent NUMA state, "
1410		       "index[%d] != num_node_masks[%d]\n",
1411		       index, num_node_masks);
1412		return -EINVAL;
1413	}
1414
1415	n = &node_masks[num_node_masks++];
1416
1417	n->mask = candidate->mask;
1418	n->match = candidate->match;
1419
1420	numadbg("NUMA NODE[%d]: mask[%lx] match[%lx] (latency[%llx])\n",
1421		index, n->mask, n->match, candidate->latency);
1422
1423	return 0;
1424}
1425
1426static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp,
1427					 int index)
1428{
1429	cpumask_t mask;
1430	int cpu;
1431
1432	numa_parse_mdesc_group_cpus(md, grp, &mask);
1433
1434	for_each_cpu(cpu, &mask)
1435		numa_cpu_lookup_table[cpu] = index;
1436	cpumask_copy(&numa_cpumask_lookup_table[index], &mask);
1437
1438	if (numa_debug) {
1439		printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index);
1440		for_each_cpu(cpu, &mask)
1441			printk("%d ", cpu);
1442		printk("]\n");
1443	}
1444
1445	return numa_attach_mlgroup(md, grp, index);
1446}
1447
1448static int __init numa_parse_mdesc(void)
1449{
1450	struct mdesc_handle *md = mdesc_grab();
1451	int i, j, err, count;
1452	u64 node;
1453
1454	node = mdesc_node_by_name(md, MDESC_NODE_NULL, "latency-groups");
1455	if (node == MDESC_NODE_NULL) {
1456		mdesc_release(md);
1457		return -ENOENT;
1458	}
1459
1460	err = grab_mblocks(md);
1461	if (err < 0)
1462		goto out;
1463
1464	err = grab_mlgroups(md);
1465	if (err < 0)
1466		goto out;
1467
1468	count = 0;
1469	mdesc_for_each_node_by_name(md, node, "group") {
1470		err = numa_parse_mdesc_group(md, node, count);
1471		if (err < 0)
1472			break;
1473		count++;
1474	}
1475
1476	count = 0;
1477	mdesc_for_each_node_by_name(md, node, "group") {
1478		find_numa_latencies_for_group(md, node, count);
1479		count++;
1480	}
1481
1482	/* Normalize numa latency matrix according to ACPI SLIT spec. */
1483	for (i = 0; i < MAX_NUMNODES; i++) {
1484		u64 self_latency = numa_latency[i][i];
1485
1486		for (j = 0; j < MAX_NUMNODES; j++) {
1487			numa_latency[i][j] =
1488				(numa_latency[i][j] * LOCAL_DISTANCE) /
1489				self_latency;
1490		}
1491	}
1492
1493	add_node_ranges();
1494
1495	for (i = 0; i < num_node_masks; i++) {
1496		allocate_node_data(i);
1497		node_set_online(i);
1498	}
1499
1500	err = 0;
1501out:
1502	mdesc_release(md);
1503	return err;
1504}
1505
1506static int __init numa_parse_jbus(void)
1507{
1508	unsigned long cpu, index;
1509
1510	/* NUMA node id is encoded in bits 36 and higher, and there is
1511	 * a 1-to-1 mapping from CPU ID to NUMA node ID.
1512	 */
1513	index = 0;
1514	for_each_present_cpu(cpu) {
1515		numa_cpu_lookup_table[cpu] = index;
1516		cpumask_copy(&numa_cpumask_lookup_table[index], cpumask_of(cpu));
1517		node_masks[index].mask = ~((1UL << 36UL) - 1UL);
1518		node_masks[index].match = cpu << 36UL;
1519
1520		index++;
1521	}
1522	num_node_masks = index;
1523
1524	add_node_ranges();
1525
1526	for (index = 0; index < num_node_masks; index++) {
1527		allocate_node_data(index);
1528		node_set_online(index);
1529	}
1530
1531	return 0;
1532}
1533
1534static int __init numa_parse_sun4u(void)
1535{
1536	if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1537		unsigned long ver;
1538
1539		__asm__ ("rdpr %%ver, %0" : "=r" (ver));
1540		if ((ver >> 32UL) == __JALAPENO_ID ||
1541		    (ver >> 32UL) == __SERRANO_ID)
1542			return numa_parse_jbus();
1543	}
1544	return -1;
1545}
1546
1547static int __init bootmem_init_numa(void)
1548{
1549	int i, j;
1550	int err = -1;
1551
1552	numadbg("bootmem_init_numa()\n");
1553
1554	/* Some sane defaults for numa latency values */
1555	for (i = 0; i < MAX_NUMNODES; i++) {
1556		for (j = 0; j < MAX_NUMNODES; j++)
1557			numa_latency[i][j] = (i == j) ?
1558				LOCAL_DISTANCE : REMOTE_DISTANCE;
1559	}
1560
1561	if (numa_enabled) {
1562		if (tlb_type == hypervisor)
1563			err = numa_parse_mdesc();
1564		else
1565			err = numa_parse_sun4u();
1566	}
1567	return err;
1568}
1569
1570#else
1571
1572static int bootmem_init_numa(void)
1573{
1574	return -1;
1575}
1576
1577#endif
1578
1579static void __init bootmem_init_nonnuma(void)
1580{
1581	unsigned long top_of_ram = memblock_end_of_DRAM();
1582	unsigned long total_ram = memblock_phys_mem_size();
1583
1584	numadbg("bootmem_init_nonnuma()\n");
1585
1586	printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
1587	       top_of_ram, total_ram);
1588	printk(KERN_INFO "Memory hole size: %ldMB\n",
1589	       (top_of_ram - total_ram) >> 20);
1590
1591	init_node_masks_nonnuma();
1592	memblock_set_node(0, PHYS_ADDR_MAX, &memblock.memory, 0);
1593	allocate_node_data(0);
1594	node_set_online(0);
1595}
1596
1597static unsigned long __init bootmem_init(unsigned long phys_base)
1598{
1599	unsigned long end_pfn;
1600
1601	end_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1602	max_pfn = max_low_pfn = end_pfn;
1603	min_low_pfn = (phys_base >> PAGE_SHIFT);
1604
1605	if (bootmem_init_numa() < 0)
1606		bootmem_init_nonnuma();
1607
1608	/* Dump memblock with node info. */
1609	memblock_dump_all();
1610
1611	/* XXX cpu notifier XXX */
1612
 
1613	sparse_init();
1614
1615	return end_pfn;
1616}
1617
1618static struct linux_prom64_registers pall[MAX_BANKS] __initdata;
1619static int pall_ents __initdata;
1620
1621static unsigned long max_phys_bits = 40;
1622
1623bool kern_addr_valid(unsigned long addr)
1624{
1625	pgd_t *pgd;
1626	p4d_t *p4d;
1627	pud_t *pud;
1628	pmd_t *pmd;
1629	pte_t *pte;
1630
1631	if ((long)addr < 0L) {
1632		unsigned long pa = __pa(addr);
1633
1634		if ((pa >> max_phys_bits) != 0UL)
1635			return false;
1636
1637		return pfn_valid(pa >> PAGE_SHIFT);
1638	}
1639
1640	if (addr >= (unsigned long) KERNBASE &&
1641	    addr < (unsigned long)&_end)
1642		return true;
1643
1644	pgd = pgd_offset_k(addr);
1645	if (pgd_none(*pgd))
1646		return false;
1647
1648	p4d = p4d_offset(pgd, addr);
1649	if (p4d_none(*p4d))
1650		return false;
1651
1652	pud = pud_offset(p4d, addr);
1653	if (pud_none(*pud))
1654		return false;
1655
1656	if (pud_large(*pud))
1657		return pfn_valid(pud_pfn(*pud));
1658
1659	pmd = pmd_offset(pud, addr);
1660	if (pmd_none(*pmd))
1661		return false;
1662
1663	if (pmd_large(*pmd))
1664		return pfn_valid(pmd_pfn(*pmd));
1665
1666	pte = pte_offset_kernel(pmd, addr);
1667	if (pte_none(*pte))
1668		return false;
1669
1670	return pfn_valid(pte_pfn(*pte));
1671}
1672EXPORT_SYMBOL(kern_addr_valid);
1673
1674static unsigned long __ref kernel_map_hugepud(unsigned long vstart,
1675					      unsigned long vend,
1676					      pud_t *pud)
1677{
1678	const unsigned long mask16gb = (1UL << 34) - 1UL;
1679	u64 pte_val = vstart;
1680
1681	/* Each PUD is 8GB */
1682	if ((vstart & mask16gb) ||
1683	    (vend - vstart <= mask16gb)) {
1684		pte_val ^= kern_linear_pte_xor[2];
1685		pud_val(*pud) = pte_val | _PAGE_PUD_HUGE;
1686
1687		return vstart + PUD_SIZE;
1688	}
1689
1690	pte_val ^= kern_linear_pte_xor[3];
1691	pte_val |= _PAGE_PUD_HUGE;
1692
1693	vend = vstart + mask16gb + 1UL;
1694	while (vstart < vend) {
1695		pud_val(*pud) = pte_val;
1696
1697		pte_val += PUD_SIZE;
1698		vstart += PUD_SIZE;
1699		pud++;
1700	}
1701	return vstart;
1702}
1703
1704static bool kernel_can_map_hugepud(unsigned long vstart, unsigned long vend,
1705				   bool guard)
1706{
1707	if (guard && !(vstart & ~PUD_MASK) && (vend - vstart) >= PUD_SIZE)
1708		return true;
1709
1710	return false;
1711}
1712
1713static unsigned long __ref kernel_map_hugepmd(unsigned long vstart,
1714					      unsigned long vend,
1715					      pmd_t *pmd)
1716{
1717	const unsigned long mask256mb = (1UL << 28) - 1UL;
1718	const unsigned long mask2gb = (1UL << 31) - 1UL;
1719	u64 pte_val = vstart;
1720
1721	/* Each PMD is 8MB */
1722	if ((vstart & mask256mb) ||
1723	    (vend - vstart <= mask256mb)) {
1724		pte_val ^= kern_linear_pte_xor[0];
1725		pmd_val(*pmd) = pte_val | _PAGE_PMD_HUGE;
1726
1727		return vstart + PMD_SIZE;
1728	}
1729
1730	if ((vstart & mask2gb) ||
1731	    (vend - vstart <= mask2gb)) {
1732		pte_val ^= kern_linear_pte_xor[1];
1733		pte_val |= _PAGE_PMD_HUGE;
1734		vend = vstart + mask256mb + 1UL;
1735	} else {
1736		pte_val ^= kern_linear_pte_xor[2];
1737		pte_val |= _PAGE_PMD_HUGE;
1738		vend = vstart + mask2gb + 1UL;
1739	}
1740
1741	while (vstart < vend) {
1742		pmd_val(*pmd) = pte_val;
1743
1744		pte_val += PMD_SIZE;
1745		vstart += PMD_SIZE;
1746		pmd++;
1747	}
1748
1749	return vstart;
1750}
1751
1752static bool kernel_can_map_hugepmd(unsigned long vstart, unsigned long vend,
1753				   bool guard)
1754{
1755	if (guard && !(vstart & ~PMD_MASK) && (vend - vstart) >= PMD_SIZE)
1756		return true;
1757
1758	return false;
1759}
1760
1761static unsigned long __ref kernel_map_range(unsigned long pstart,
1762					    unsigned long pend, pgprot_t prot,
1763					    bool use_huge)
1764{
1765	unsigned long vstart = PAGE_OFFSET + pstart;
1766	unsigned long vend = PAGE_OFFSET + pend;
1767	unsigned long alloc_bytes = 0UL;
1768
1769	if ((vstart & ~PAGE_MASK) || (vend & ~PAGE_MASK)) {
1770		prom_printf("kernel_map: Unaligned physmem[%lx:%lx]\n",
1771			    vstart, vend);
1772		prom_halt();
1773	}
1774
1775	while (vstart < vend) {
1776		unsigned long this_end, paddr = __pa(vstart);
1777		pgd_t *pgd = pgd_offset_k(vstart);
1778		p4d_t *p4d;
1779		pud_t *pud;
1780		pmd_t *pmd;
1781		pte_t *pte;
1782
1783		if (pgd_none(*pgd)) {
1784			pud_t *new;
1785
1786			new = memblock_alloc_from(PAGE_SIZE, PAGE_SIZE,
1787						  PAGE_SIZE);
1788			if (!new)
1789				goto err_alloc;
1790			alloc_bytes += PAGE_SIZE;
1791			pgd_populate(&init_mm, pgd, new);
1792		}
1793
1794		p4d = p4d_offset(pgd, vstart);
1795		if (p4d_none(*p4d)) {
1796			pud_t *new;
1797
1798			new = memblock_alloc_from(PAGE_SIZE, PAGE_SIZE,
1799						  PAGE_SIZE);
1800			if (!new)
1801				goto err_alloc;
1802			alloc_bytes += PAGE_SIZE;
1803			p4d_populate(&init_mm, p4d, new);
1804		}
1805
1806		pud = pud_offset(p4d, vstart);
1807		if (pud_none(*pud)) {
1808			pmd_t *new;
1809
1810			if (kernel_can_map_hugepud(vstart, vend, use_huge)) {
1811				vstart = kernel_map_hugepud(vstart, vend, pud);
1812				continue;
1813			}
1814			new = memblock_alloc_from(PAGE_SIZE, PAGE_SIZE,
1815						  PAGE_SIZE);
1816			if (!new)
1817				goto err_alloc;
1818			alloc_bytes += PAGE_SIZE;
1819			pud_populate(&init_mm, pud, new);
1820		}
1821
1822		pmd = pmd_offset(pud, vstart);
1823		if (pmd_none(*pmd)) {
1824			pte_t *new;
1825
1826			if (kernel_can_map_hugepmd(vstart, vend, use_huge)) {
1827				vstart = kernel_map_hugepmd(vstart, vend, pmd);
1828				continue;
1829			}
1830			new = memblock_alloc_from(PAGE_SIZE, PAGE_SIZE,
1831						  PAGE_SIZE);
1832			if (!new)
1833				goto err_alloc;
1834			alloc_bytes += PAGE_SIZE;
1835			pmd_populate_kernel(&init_mm, pmd, new);
1836		}
1837
1838		pte = pte_offset_kernel(pmd, vstart);
1839		this_end = (vstart + PMD_SIZE) & PMD_MASK;
1840		if (this_end > vend)
1841			this_end = vend;
1842
1843		while (vstart < this_end) {
1844			pte_val(*pte) = (paddr | pgprot_val(prot));
1845
1846			vstart += PAGE_SIZE;
1847			paddr += PAGE_SIZE;
1848			pte++;
1849		}
1850	}
1851
1852	return alloc_bytes;
1853
1854err_alloc:
1855	panic("%s: Failed to allocate %lu bytes align=%lx from=%lx\n",
1856	      __func__, PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1857	return -ENOMEM;
1858}
1859
1860static void __init flush_all_kernel_tsbs(void)
1861{
1862	int i;
1863
1864	for (i = 0; i < KERNEL_TSB_NENTRIES; i++) {
1865		struct tsb *ent = &swapper_tsb[i];
1866
1867		ent->tag = (1UL << TSB_TAG_INVALID_BIT);
1868	}
1869#ifndef CONFIG_DEBUG_PAGEALLOC
1870	for (i = 0; i < KERNEL_TSB4M_NENTRIES; i++) {
1871		struct tsb *ent = &swapper_4m_tsb[i];
1872
1873		ent->tag = (1UL << TSB_TAG_INVALID_BIT);
1874	}
1875#endif
1876}
1877
1878extern unsigned int kvmap_linear_patch[1];
1879
1880static void __init kernel_physical_mapping_init(void)
1881{
1882	unsigned long i, mem_alloced = 0UL;
1883	bool use_huge = true;
1884
1885#ifdef CONFIG_DEBUG_PAGEALLOC
1886	use_huge = false;
1887#endif
1888	for (i = 0; i < pall_ents; i++) {
1889		unsigned long phys_start, phys_end;
1890
1891		phys_start = pall[i].phys_addr;
1892		phys_end = phys_start + pall[i].reg_size;
1893
1894		mem_alloced += kernel_map_range(phys_start, phys_end,
1895						PAGE_KERNEL, use_huge);
1896	}
1897
1898	printk("Allocated %ld bytes for kernel page tables.\n",
1899	       mem_alloced);
1900
1901	kvmap_linear_patch[0] = 0x01000000; /* nop */
1902	flushi(&kvmap_linear_patch[0]);
1903
1904	flush_all_kernel_tsbs();
1905
1906	__flush_tlb_all();
1907}
1908
1909#ifdef CONFIG_DEBUG_PAGEALLOC
1910void __kernel_map_pages(struct page *page, int numpages, int enable)
1911{
1912	unsigned long phys_start = page_to_pfn(page) << PAGE_SHIFT;
1913	unsigned long phys_end = phys_start + (numpages * PAGE_SIZE);
1914
1915	kernel_map_range(phys_start, phys_end,
1916			 (enable ? PAGE_KERNEL : __pgprot(0)), false);
1917
1918	flush_tsb_kernel_range(PAGE_OFFSET + phys_start,
1919			       PAGE_OFFSET + phys_end);
1920
1921	/* we should perform an IPI and flush all tlbs,
1922	 * but that can deadlock->flush only current cpu.
1923	 */
1924	__flush_tlb_kernel_range(PAGE_OFFSET + phys_start,
1925				 PAGE_OFFSET + phys_end);
1926}
1927#endif
1928
1929unsigned long __init find_ecache_flush_span(unsigned long size)
1930{
1931	int i;
1932
1933	for (i = 0; i < pavail_ents; i++) {
1934		if (pavail[i].reg_size >= size)
1935			return pavail[i].phys_addr;
1936	}
1937
1938	return ~0UL;
1939}
1940
1941unsigned long PAGE_OFFSET;
1942EXPORT_SYMBOL(PAGE_OFFSET);
1943
1944unsigned long VMALLOC_END   = 0x0000010000000000UL;
1945EXPORT_SYMBOL(VMALLOC_END);
1946
1947unsigned long sparc64_va_hole_top =    0xfffff80000000000UL;
1948unsigned long sparc64_va_hole_bottom = 0x0000080000000000UL;
1949
1950static void __init setup_page_offset(void)
1951{
1952	if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1953		/* Cheetah/Panther support a full 64-bit virtual
1954		 * address, so we can use all that our page tables
1955		 * support.
1956		 */
1957		sparc64_va_hole_top =    0xfff0000000000000UL;
1958		sparc64_va_hole_bottom = 0x0010000000000000UL;
1959
1960		max_phys_bits = 42;
1961	} else if (tlb_type == hypervisor) {
1962		switch (sun4v_chip_type) {
1963		case SUN4V_CHIP_NIAGARA1:
1964		case SUN4V_CHIP_NIAGARA2:
1965			/* T1 and T2 support 48-bit virtual addresses.  */
1966			sparc64_va_hole_top =    0xffff800000000000UL;
1967			sparc64_va_hole_bottom = 0x0000800000000000UL;
1968
1969			max_phys_bits = 39;
1970			break;
1971		case SUN4V_CHIP_NIAGARA3:
1972			/* T3 supports 48-bit virtual addresses.  */
1973			sparc64_va_hole_top =    0xffff800000000000UL;
1974			sparc64_va_hole_bottom = 0x0000800000000000UL;
1975
1976			max_phys_bits = 43;
1977			break;
1978		case SUN4V_CHIP_NIAGARA4:
1979		case SUN4V_CHIP_NIAGARA5:
1980		case SUN4V_CHIP_SPARC64X:
1981		case SUN4V_CHIP_SPARC_M6:
1982			/* T4 and later support 52-bit virtual addresses.  */
1983			sparc64_va_hole_top =    0xfff8000000000000UL;
1984			sparc64_va_hole_bottom = 0x0008000000000000UL;
1985			max_phys_bits = 47;
1986			break;
1987		case SUN4V_CHIP_SPARC_M7:
1988		case SUN4V_CHIP_SPARC_SN:
1989			/* M7 and later support 52-bit virtual addresses.  */
1990			sparc64_va_hole_top =    0xfff8000000000000UL;
1991			sparc64_va_hole_bottom = 0x0008000000000000UL;
1992			max_phys_bits = 49;
1993			break;
1994		case SUN4V_CHIP_SPARC_M8:
1995		default:
1996			/* M8 and later support 54-bit virtual addresses.
1997			 * However, restricting M8 and above VA bits to 53
1998			 * as 4-level page table cannot support more than
1999			 * 53 VA bits.
2000			 */
2001			sparc64_va_hole_top =    0xfff0000000000000UL;
2002			sparc64_va_hole_bottom = 0x0010000000000000UL;
2003			max_phys_bits = 51;
2004			break;
2005		}
2006	}
2007
2008	if (max_phys_bits > MAX_PHYS_ADDRESS_BITS) {
2009		prom_printf("MAX_PHYS_ADDRESS_BITS is too small, need %lu\n",
2010			    max_phys_bits);
2011		prom_halt();
2012	}
2013
2014	PAGE_OFFSET = sparc64_va_hole_top;
2015	VMALLOC_END = ((sparc64_va_hole_bottom >> 1) +
2016		       (sparc64_va_hole_bottom >> 2));
2017
2018	pr_info("MM: PAGE_OFFSET is 0x%016lx (max_phys_bits == %lu)\n",
2019		PAGE_OFFSET, max_phys_bits);
2020	pr_info("MM: VMALLOC [0x%016lx --> 0x%016lx]\n",
2021		VMALLOC_START, VMALLOC_END);
2022	pr_info("MM: VMEMMAP [0x%016lx --> 0x%016lx]\n",
2023		VMEMMAP_BASE, VMEMMAP_BASE << 1);
2024}
2025
2026static void __init tsb_phys_patch(void)
2027{
2028	struct tsb_ldquad_phys_patch_entry *pquad;
2029	struct tsb_phys_patch_entry *p;
2030
2031	pquad = &__tsb_ldquad_phys_patch;
2032	while (pquad < &__tsb_ldquad_phys_patch_end) {
2033		unsigned long addr = pquad->addr;
2034
2035		if (tlb_type == hypervisor)
2036			*(unsigned int *) addr = pquad->sun4v_insn;
2037		else
2038			*(unsigned int *) addr = pquad->sun4u_insn;
2039		wmb();
2040		__asm__ __volatile__("flush	%0"
2041				     : /* no outputs */
2042				     : "r" (addr));
2043
2044		pquad++;
2045	}
2046
2047	p = &__tsb_phys_patch;
2048	while (p < &__tsb_phys_patch_end) {
2049		unsigned long addr = p->addr;
2050
2051		*(unsigned int *) addr = p->insn;
2052		wmb();
2053		__asm__ __volatile__("flush	%0"
2054				     : /* no outputs */
2055				     : "r" (addr));
2056
2057		p++;
2058	}
2059}
2060
2061/* Don't mark as init, we give this to the Hypervisor.  */
2062#ifndef CONFIG_DEBUG_PAGEALLOC
2063#define NUM_KTSB_DESCR	2
2064#else
2065#define NUM_KTSB_DESCR	1
2066#endif
2067static struct hv_tsb_descr ktsb_descr[NUM_KTSB_DESCR];
2068
2069/* The swapper TSBs are loaded with a base sequence of:
2070 *
2071 *	sethi	%uhi(SYMBOL), REG1
2072 *	sethi	%hi(SYMBOL), REG2
2073 *	or	REG1, %ulo(SYMBOL), REG1
2074 *	or	REG2, %lo(SYMBOL), REG2
2075 *	sllx	REG1, 32, REG1
2076 *	or	REG1, REG2, REG1
2077 *
2078 * When we use physical addressing for the TSB accesses, we patch the
2079 * first four instructions in the above sequence.
2080 */
2081
2082static void patch_one_ktsb_phys(unsigned int *start, unsigned int *end, unsigned long pa)
2083{
2084	unsigned long high_bits, low_bits;
2085
2086	high_bits = (pa >> 32) & 0xffffffff;
2087	low_bits = (pa >> 0) & 0xffffffff;
2088
2089	while (start < end) {
2090		unsigned int *ia = (unsigned int *)(unsigned long)*start;
2091
2092		ia[0] = (ia[0] & ~0x3fffff) | (high_bits >> 10);
2093		__asm__ __volatile__("flush	%0" : : "r" (ia));
2094
2095		ia[1] = (ia[1] & ~0x3fffff) | (low_bits >> 10);
2096		__asm__ __volatile__("flush	%0" : : "r" (ia + 1));
2097
2098		ia[2] = (ia[2] & ~0x1fff) | (high_bits & 0x3ff);
2099		__asm__ __volatile__("flush	%0" : : "r" (ia + 2));
2100
2101		ia[3] = (ia[3] & ~0x1fff) | (low_bits & 0x3ff);
2102		__asm__ __volatile__("flush	%0" : : "r" (ia + 3));
2103
2104		start++;
2105	}
2106}
2107
2108static void ktsb_phys_patch(void)
2109{
2110	extern unsigned int __swapper_tsb_phys_patch;
2111	extern unsigned int __swapper_tsb_phys_patch_end;
2112	unsigned long ktsb_pa;
2113
2114	ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
2115	patch_one_ktsb_phys(&__swapper_tsb_phys_patch,
2116			    &__swapper_tsb_phys_patch_end, ktsb_pa);
2117#ifndef CONFIG_DEBUG_PAGEALLOC
2118	{
2119	extern unsigned int __swapper_4m_tsb_phys_patch;
2120	extern unsigned int __swapper_4m_tsb_phys_patch_end;
2121	ktsb_pa = (kern_base +
2122		   ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
2123	patch_one_ktsb_phys(&__swapper_4m_tsb_phys_patch,
2124			    &__swapper_4m_tsb_phys_patch_end, ktsb_pa);
2125	}
2126#endif
2127}
2128
2129static void __init sun4v_ktsb_init(void)
2130{
2131	unsigned long ktsb_pa;
2132
2133	/* First KTSB for PAGE_SIZE mappings.  */
2134	ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
2135
2136	switch (PAGE_SIZE) {
2137	case 8 * 1024:
2138	default:
2139		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_8K;
2140		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_8K;
2141		break;
2142
2143	case 64 * 1024:
2144		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_64K;
2145		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_64K;
2146		break;
2147
2148	case 512 * 1024:
2149		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_512K;
2150		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_512K;
2151		break;
2152
2153	case 4 * 1024 * 1024:
2154		ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_4MB;
2155		ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_4MB;
2156		break;
2157	}
2158
2159	ktsb_descr[0].assoc = 1;
2160	ktsb_descr[0].num_ttes = KERNEL_TSB_NENTRIES;
2161	ktsb_descr[0].ctx_idx = 0;
2162	ktsb_descr[0].tsb_base = ktsb_pa;
2163	ktsb_descr[0].resv = 0;
2164
2165#ifndef CONFIG_DEBUG_PAGEALLOC
2166	/* Second KTSB for 4MB/256MB/2GB/16GB mappings.  */
2167	ktsb_pa = (kern_base +
2168		   ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
2169
2170	ktsb_descr[1].pgsz_idx = HV_PGSZ_IDX_4MB;
2171	ktsb_descr[1].pgsz_mask = ((HV_PGSZ_MASK_4MB |
2172				    HV_PGSZ_MASK_256MB |
2173				    HV_PGSZ_MASK_2GB |
2174				    HV_PGSZ_MASK_16GB) &
2175				   cpu_pgsz_mask);
2176	ktsb_descr[1].assoc = 1;
2177	ktsb_descr[1].num_ttes = KERNEL_TSB4M_NENTRIES;
2178	ktsb_descr[1].ctx_idx = 0;
2179	ktsb_descr[1].tsb_base = ktsb_pa;
2180	ktsb_descr[1].resv = 0;
2181#endif
2182}
2183
2184void sun4v_ktsb_register(void)
2185{
2186	unsigned long pa, ret;
2187
2188	pa = kern_base + ((unsigned long)&ktsb_descr[0] - KERNBASE);
2189
2190	ret = sun4v_mmu_tsb_ctx0(NUM_KTSB_DESCR, pa);
2191	if (ret != 0) {
2192		prom_printf("hypervisor_mmu_tsb_ctx0[%lx]: "
2193			    "errors with %lx\n", pa, ret);
2194		prom_halt();
2195	}
2196}
2197
2198static void __init sun4u_linear_pte_xor_finalize(void)
2199{
2200#ifndef CONFIG_DEBUG_PAGEALLOC
2201	/* This is where we would add Panther support for
2202	 * 32MB and 256MB pages.
2203	 */
2204#endif
2205}
2206
2207static void __init sun4v_linear_pte_xor_finalize(void)
2208{
2209	unsigned long pagecv_flag;
2210
2211	/* Bit 9 of TTE is no longer CV bit on M7 processor and it instead
2212	 * enables MCD error. Do not set bit 9 on M7 processor.
2213	 */
2214	switch (sun4v_chip_type) {
2215	case SUN4V_CHIP_SPARC_M7:
2216	case SUN4V_CHIP_SPARC_M8:
2217	case SUN4V_CHIP_SPARC_SN:
2218		pagecv_flag = 0x00;
2219		break;
2220	default:
2221		pagecv_flag = _PAGE_CV_4V;
2222		break;
2223	}
2224#ifndef CONFIG_DEBUG_PAGEALLOC
2225	if (cpu_pgsz_mask & HV_PGSZ_MASK_256MB) {
2226		kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZ256MB_4V) ^
2227			PAGE_OFFSET;
2228		kern_linear_pte_xor[1] |= (_PAGE_CP_4V | pagecv_flag |
2229					   _PAGE_P_4V | _PAGE_W_4V);
2230	} else {
2231		kern_linear_pte_xor[1] = kern_linear_pte_xor[0];
2232	}
2233
2234	if (cpu_pgsz_mask & HV_PGSZ_MASK_2GB) {
2235		kern_linear_pte_xor[2] = (_PAGE_VALID | _PAGE_SZ2GB_4V) ^
2236			PAGE_OFFSET;
2237		kern_linear_pte_xor[2] |= (_PAGE_CP_4V | pagecv_flag |
2238					   _PAGE_P_4V | _PAGE_W_4V);
2239	} else {
2240		kern_linear_pte_xor[2] = kern_linear_pte_xor[1];
2241	}
2242
2243	if (cpu_pgsz_mask & HV_PGSZ_MASK_16GB) {
2244		kern_linear_pte_xor[3] = (_PAGE_VALID | _PAGE_SZ16GB_4V) ^
2245			PAGE_OFFSET;
2246		kern_linear_pte_xor[3] |= (_PAGE_CP_4V | pagecv_flag |
2247					   _PAGE_P_4V | _PAGE_W_4V);
2248	} else {
2249		kern_linear_pte_xor[3] = kern_linear_pte_xor[2];
2250	}
2251#endif
2252}
2253
2254/* paging_init() sets up the page tables */
2255
2256static unsigned long last_valid_pfn;
2257
2258static void sun4u_pgprot_init(void);
2259static void sun4v_pgprot_init(void);
2260
 
 
 
 
 
 
 
 
 
 
 
 
 
2261#define _PAGE_CACHE_4U	(_PAGE_CP_4U | _PAGE_CV_4U)
2262#define _PAGE_CACHE_4V	(_PAGE_CP_4V | _PAGE_CV_4V)
2263#define __DIRTY_BITS_4U	 (_PAGE_MODIFIED_4U | _PAGE_WRITE_4U | _PAGE_W_4U)
2264#define __DIRTY_BITS_4V	 (_PAGE_MODIFIED_4V | _PAGE_WRITE_4V | _PAGE_W_4V)
2265#define __ACCESS_BITS_4U (_PAGE_ACCESSED_4U | _PAGE_READ_4U | _PAGE_R)
2266#define __ACCESS_BITS_4V (_PAGE_ACCESSED_4V | _PAGE_READ_4V | _PAGE_R)
2267
2268/* We need to exclude reserved regions. This exclusion will include
2269 * vmlinux and initrd. To be more precise the initrd size could be used to
2270 * compute a new lower limit because it is freed later during initialization.
2271 */
2272static void __init reduce_memory(phys_addr_t limit_ram)
2273{
2274	limit_ram += memblock_reserved_size();
2275	memblock_enforce_memory_limit(limit_ram);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2276}
2277
2278void __init paging_init(void)
2279{
2280	unsigned long end_pfn, shift, phys_base;
2281	unsigned long real_end, i;
2282
2283	setup_page_offset();
2284
2285	/* These build time checkes make sure that the dcache_dirty_cpu()
2286	 * page->flags usage will work.
2287	 *
2288	 * When a page gets marked as dcache-dirty, we store the
2289	 * cpu number starting at bit 32 in the page->flags.  Also,
2290	 * functions like clear_dcache_dirty_cpu use the cpu mask
2291	 * in 13-bit signed-immediate instruction fields.
2292	 */
2293
2294	/*
2295	 * Page flags must not reach into upper 32 bits that are used
2296	 * for the cpu number
2297	 */
2298	BUILD_BUG_ON(NR_PAGEFLAGS > 32);
2299
2300	/*
2301	 * The bit fields placed in the high range must not reach below
2302	 * the 32 bit boundary. Otherwise we cannot place the cpu field
2303	 * at the 32 bit boundary.
2304	 */
2305	BUILD_BUG_ON(SECTIONS_WIDTH + NODES_WIDTH + ZONES_WIDTH +
2306		ilog2(roundup_pow_of_two(NR_CPUS)) > 32);
2307
2308	BUILD_BUG_ON(NR_CPUS > 4096);
2309
2310	kern_base = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
2311	kern_size = (unsigned long)&_end - (unsigned long)KERNBASE;
2312
2313	/* Invalidate both kernel TSBs.  */
2314	memset(swapper_tsb, 0x40, sizeof(swapper_tsb));
2315#ifndef CONFIG_DEBUG_PAGEALLOC
2316	memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2317#endif
2318
2319	/* TTE.cv bit on sparc v9 occupies the same position as TTE.mcde
2320	 * bit on M7 processor. This is a conflicting usage of the same
2321	 * bit. Enabling TTE.cv on M7 would turn on Memory Corruption
2322	 * Detection error on all pages and this will lead to problems
2323	 * later. Kernel does not run with MCD enabled and hence rest
2324	 * of the required steps to fully configure memory corruption
2325	 * detection are not taken. We need to ensure TTE.mcde is not
2326	 * set on M7 processor. Compute the value of cacheability
2327	 * flag for use later taking this into consideration.
2328	 */
2329	switch (sun4v_chip_type) {
2330	case SUN4V_CHIP_SPARC_M7:
2331	case SUN4V_CHIP_SPARC_M8:
2332	case SUN4V_CHIP_SPARC_SN:
2333		page_cache4v_flag = _PAGE_CP_4V;
2334		break;
2335	default:
2336		page_cache4v_flag = _PAGE_CACHE_4V;
2337		break;
2338	}
2339
2340	if (tlb_type == hypervisor)
2341		sun4v_pgprot_init();
2342	else
2343		sun4u_pgprot_init();
2344
2345	if (tlb_type == cheetah_plus ||
2346	    tlb_type == hypervisor) {
2347		tsb_phys_patch();
2348		ktsb_phys_patch();
2349	}
2350
2351	if (tlb_type == hypervisor)
2352		sun4v_patch_tlb_handlers();
2353
2354	/* Find available physical memory...
2355	 *
2356	 * Read it twice in order to work around a bug in openfirmware.
2357	 * The call to grab this table itself can cause openfirmware to
2358	 * allocate memory, which in turn can take away some space from
2359	 * the list of available memory.  Reading it twice makes sure
2360	 * we really do get the final value.
2361	 */
2362	read_obp_translations();
2363	read_obp_memory("reg", &pall[0], &pall_ents);
2364	read_obp_memory("available", &pavail[0], &pavail_ents);
2365	read_obp_memory("available", &pavail[0], &pavail_ents);
2366
2367	phys_base = 0xffffffffffffffffUL;
2368	for (i = 0; i < pavail_ents; i++) {
2369		phys_base = min(phys_base, pavail[i].phys_addr);
2370		memblock_add(pavail[i].phys_addr, pavail[i].reg_size);
2371	}
2372
2373	memblock_reserve(kern_base, kern_size);
2374
2375	find_ramdisk(phys_base);
2376
2377	if (cmdline_memory_size)
2378		reduce_memory(cmdline_memory_size);
2379
2380	memblock_allow_resize();
2381	memblock_dump_all();
2382
2383	set_bit(0, mmu_context_bmap);
2384
2385	shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE);
2386
2387	real_end = (unsigned long)_end;
2388	num_kernel_image_mappings = DIV_ROUND_UP(real_end - KERNBASE, 1 << ILOG2_4MB);
2389	printk("Kernel: Using %d locked TLB entries for main kernel image.\n",
2390	       num_kernel_image_mappings);
2391
2392	/* Set kernel pgd to upper alias so physical page computations
2393	 * work.
2394	 */
2395	init_mm.pgd += ((shift) / (sizeof(pgd_t)));
2396	
2397	memset(swapper_pg_dir, 0, sizeof(swapper_pg_dir));
2398
2399	inherit_prom_mappings();
2400	
2401	/* Ok, we can use our TLB miss and window trap handlers safely.  */
2402	setup_tba();
2403
2404	__flush_tlb_all();
2405
2406	prom_build_devicetree();
2407	of_populate_present_mask();
2408#ifndef CONFIG_SMP
2409	of_fill_in_cpu_data();
2410#endif
2411
2412	if (tlb_type == hypervisor) {
2413		sun4v_mdesc_init();
2414		mdesc_populate_present_mask(cpu_all_mask);
2415#ifndef CONFIG_SMP
2416		mdesc_fill_in_cpu_data(cpu_all_mask);
2417#endif
2418		mdesc_get_page_sizes(cpu_all_mask, &cpu_pgsz_mask);
2419
2420		sun4v_linear_pte_xor_finalize();
2421
2422		sun4v_ktsb_init();
2423		sun4v_ktsb_register();
2424	} else {
2425		unsigned long impl, ver;
2426
2427		cpu_pgsz_mask = (HV_PGSZ_MASK_8K | HV_PGSZ_MASK_64K |
2428				 HV_PGSZ_MASK_512K | HV_PGSZ_MASK_4MB);
2429
2430		__asm__ __volatile__("rdpr %%ver, %0" : "=r" (ver));
2431		impl = ((ver >> 32) & 0xffff);
2432		if (impl == PANTHER_IMPL)
2433			cpu_pgsz_mask |= (HV_PGSZ_MASK_32MB |
2434					  HV_PGSZ_MASK_256MB);
2435
2436		sun4u_linear_pte_xor_finalize();
2437	}
2438
2439	/* Flush the TLBs and the 4M TSB so that the updated linear
2440	 * pte XOR settings are realized for all mappings.
2441	 */
2442	__flush_tlb_all();
2443#ifndef CONFIG_DEBUG_PAGEALLOC
2444	memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2445#endif
2446	__flush_tlb_all();
2447
2448	/* Setup bootmem... */
2449	last_valid_pfn = end_pfn = bootmem_init(phys_base);
2450
2451	kernel_physical_mapping_init();
2452
2453	{
2454		unsigned long max_zone_pfns[MAX_NR_ZONES];
2455
2456		memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
2457
2458		max_zone_pfns[ZONE_NORMAL] = end_pfn;
2459
2460		free_area_init(max_zone_pfns);
2461	}
2462
2463	printk("Booting Linux...\n");
2464}
2465
2466int page_in_phys_avail(unsigned long paddr)
2467{
2468	int i;
2469
2470	paddr &= PAGE_MASK;
2471
2472	for (i = 0; i < pavail_ents; i++) {
2473		unsigned long start, end;
2474
2475		start = pavail[i].phys_addr;
2476		end = start + pavail[i].reg_size;
2477
2478		if (paddr >= start && paddr < end)
2479			return 1;
2480	}
2481	if (paddr >= kern_base && paddr < (kern_base + kern_size))
2482		return 1;
2483#ifdef CONFIG_BLK_DEV_INITRD
2484	if (paddr >= __pa(initrd_start) &&
2485	    paddr < __pa(PAGE_ALIGN(initrd_end)))
2486		return 1;
2487#endif
2488
2489	return 0;
2490}
2491
2492static void __init register_page_bootmem_info(void)
2493{
2494#ifdef CONFIG_NEED_MULTIPLE_NODES
2495	int i;
2496
2497	for_each_online_node(i)
2498		if (NODE_DATA(i)->node_spanned_pages)
2499			register_page_bootmem_info_node(NODE_DATA(i));
2500#endif
2501}
2502void __init mem_init(void)
2503{
2504	high_memory = __va(last_valid_pfn << PAGE_SHIFT);
2505
2506	memblock_free_all();
2507
2508	/*
2509	 * Must be done after boot memory is put on freelist, because here we
2510	 * might set fields in deferred struct pages that have not yet been
2511	 * initialized, and memblock_free_all() initializes all the reserved
2512	 * deferred pages for us.
2513	 */
2514	register_page_bootmem_info();
2515
2516	/*
2517	 * Set up the zero page, mark it reserved, so that page count
2518	 * is not manipulated when freeing the page from user ptes.
2519	 */
2520	mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0);
2521	if (mem_map_zero == NULL) {
2522		prom_printf("paging_init: Cannot alloc zero page.\n");
2523		prom_halt();
2524	}
2525	mark_page_reserved(mem_map_zero);
2526
2527	mem_init_print_info(NULL);
2528
2529	if (tlb_type == cheetah || tlb_type == cheetah_plus)
2530		cheetah_ecache_flush_init();
2531}
2532
2533void free_initmem(void)
2534{
2535	unsigned long addr, initend;
2536	int do_free = 1;
2537
2538	/* If the physical memory maps were trimmed by kernel command
2539	 * line options, don't even try freeing this initmem stuff up.
2540	 * The kernel image could have been in the trimmed out region
2541	 * and if so the freeing below will free invalid page structs.
2542	 */
2543	if (cmdline_memory_size)
2544		do_free = 0;
2545
2546	/*
2547	 * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes.
2548	 */
2549	addr = PAGE_ALIGN((unsigned long)(__init_begin));
2550	initend = (unsigned long)(__init_end) & PAGE_MASK;
2551	for (; addr < initend; addr += PAGE_SIZE) {
2552		unsigned long page;
2553
2554		page = (addr +
2555			((unsigned long) __va(kern_base)) -
2556			((unsigned long) KERNBASE));
2557		memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
2558
2559		if (do_free)
2560			free_reserved_page(virt_to_page(page));
2561	}
2562}
2563
 
 
 
 
 
 
 
 
2564pgprot_t PAGE_KERNEL __read_mostly;
2565EXPORT_SYMBOL(PAGE_KERNEL);
2566
2567pgprot_t PAGE_KERNEL_LOCKED __read_mostly;
2568pgprot_t PAGE_COPY __read_mostly;
2569
2570pgprot_t PAGE_SHARED __read_mostly;
2571EXPORT_SYMBOL(PAGE_SHARED);
2572
2573unsigned long pg_iobits __read_mostly;
2574
2575unsigned long _PAGE_IE __read_mostly;
2576EXPORT_SYMBOL(_PAGE_IE);
2577
2578unsigned long _PAGE_E __read_mostly;
2579EXPORT_SYMBOL(_PAGE_E);
2580
2581unsigned long _PAGE_CACHE __read_mostly;
2582EXPORT_SYMBOL(_PAGE_CACHE);
2583
2584#ifdef CONFIG_SPARSEMEM_VMEMMAP
2585int __meminit vmemmap_populate(unsigned long vstart, unsigned long vend,
2586			       int node, struct vmem_altmap *altmap)
2587{
2588	unsigned long pte_base;
2589
2590	pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2591		    _PAGE_CP_4U | _PAGE_CV_4U |
2592		    _PAGE_P_4U | _PAGE_W_4U);
2593	if (tlb_type == hypervisor)
2594		pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2595			    page_cache4v_flag | _PAGE_P_4V | _PAGE_W_4V);
2596
2597	pte_base |= _PAGE_PMD_HUGE;
2598
2599	vstart = vstart & PMD_MASK;
2600	vend = ALIGN(vend, PMD_SIZE);
2601	for (; vstart < vend; vstart += PMD_SIZE) {
2602		pgd_t *pgd = vmemmap_pgd_populate(vstart, node);
2603		unsigned long pte;
2604		p4d_t *p4d;
2605		pud_t *pud;
2606		pmd_t *pmd;
2607
2608		if (!pgd)
2609			return -ENOMEM;
2610
2611		p4d = vmemmap_p4d_populate(pgd, vstart, node);
2612		if (!p4d)
2613			return -ENOMEM;
2614
2615		pud = vmemmap_pud_populate(p4d, vstart, node);
2616		if (!pud)
2617			return -ENOMEM;
2618
2619		pmd = pmd_offset(pud, vstart);
2620		pte = pmd_val(*pmd);
2621		if (!(pte & _PAGE_VALID)) {
2622			void *block = vmemmap_alloc_block(PMD_SIZE, node);
2623
2624			if (!block)
2625				return -ENOMEM;
2626
2627			pmd_val(*pmd) = pte_base | __pa(block);
2628		}
2629	}
2630
2631	return 0;
2632}
2633
2634void vmemmap_free(unsigned long start, unsigned long end,
2635		struct vmem_altmap *altmap)
2636{
2637}
2638#endif /* CONFIG_SPARSEMEM_VMEMMAP */
2639
2640static void prot_init_common(unsigned long page_none,
2641			     unsigned long page_shared,
2642			     unsigned long page_copy,
2643			     unsigned long page_readonly,
2644			     unsigned long page_exec_bit)
2645{
2646	PAGE_COPY = __pgprot(page_copy);
2647	PAGE_SHARED = __pgprot(page_shared);
2648
2649	protection_map[0x0] = __pgprot(page_none);
2650	protection_map[0x1] = __pgprot(page_readonly & ~page_exec_bit);
2651	protection_map[0x2] = __pgprot(page_copy & ~page_exec_bit);
2652	protection_map[0x3] = __pgprot(page_copy & ~page_exec_bit);
2653	protection_map[0x4] = __pgprot(page_readonly);
2654	protection_map[0x5] = __pgprot(page_readonly);
2655	protection_map[0x6] = __pgprot(page_copy);
2656	protection_map[0x7] = __pgprot(page_copy);
2657	protection_map[0x8] = __pgprot(page_none);
2658	protection_map[0x9] = __pgprot(page_readonly & ~page_exec_bit);
2659	protection_map[0xa] = __pgprot(page_shared & ~page_exec_bit);
2660	protection_map[0xb] = __pgprot(page_shared & ~page_exec_bit);
2661	protection_map[0xc] = __pgprot(page_readonly);
2662	protection_map[0xd] = __pgprot(page_readonly);
2663	protection_map[0xe] = __pgprot(page_shared);
2664	protection_map[0xf] = __pgprot(page_shared);
2665}
2666
2667static void __init sun4u_pgprot_init(void)
2668{
2669	unsigned long page_none, page_shared, page_copy, page_readonly;
2670	unsigned long page_exec_bit;
2671	int i;
2672
2673	PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2674				_PAGE_CACHE_4U | _PAGE_P_4U |
2675				__ACCESS_BITS_4U | __DIRTY_BITS_4U |
2676				_PAGE_EXEC_4U);
2677	PAGE_KERNEL_LOCKED = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2678				       _PAGE_CACHE_4U | _PAGE_P_4U |
2679				       __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2680				       _PAGE_EXEC_4U | _PAGE_L_4U);
2681
2682	_PAGE_IE = _PAGE_IE_4U;
2683	_PAGE_E = _PAGE_E_4U;
2684	_PAGE_CACHE = _PAGE_CACHE_4U;
2685
2686	pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4U | __DIRTY_BITS_4U |
2687		     __ACCESS_BITS_4U | _PAGE_E_4U);
2688
2689#ifdef CONFIG_DEBUG_PAGEALLOC
2690	kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2691#else
2692	kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4U) ^
2693		PAGE_OFFSET;
2694#endif
2695	kern_linear_pte_xor[0] |= (_PAGE_CP_4U | _PAGE_CV_4U |
2696				   _PAGE_P_4U | _PAGE_W_4U);
2697
2698	for (i = 1; i < 4; i++)
2699		kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2700
2701	_PAGE_ALL_SZ_BITS =  (_PAGE_SZ4MB_4U | _PAGE_SZ512K_4U |
2702			      _PAGE_SZ64K_4U | _PAGE_SZ8K_4U |
2703			      _PAGE_SZ32MB_4U | _PAGE_SZ256MB_4U);
2704
2705
2706	page_none = _PAGE_PRESENT_4U | _PAGE_ACCESSED_4U | _PAGE_CACHE_4U;
2707	page_shared = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2708		       __ACCESS_BITS_4U | _PAGE_WRITE_4U | _PAGE_EXEC_4U);
2709	page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2710		       __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2711	page_readonly   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2712			   __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2713
2714	page_exec_bit = _PAGE_EXEC_4U;
2715
2716	prot_init_common(page_none, page_shared, page_copy, page_readonly,
2717			 page_exec_bit);
2718}
2719
2720static void __init sun4v_pgprot_init(void)
2721{
2722	unsigned long page_none, page_shared, page_copy, page_readonly;
2723	unsigned long page_exec_bit;
2724	int i;
2725
2726	PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4V | _PAGE_VALID |
2727				page_cache4v_flag | _PAGE_P_4V |
2728				__ACCESS_BITS_4V | __DIRTY_BITS_4V |
2729				_PAGE_EXEC_4V);
2730	PAGE_KERNEL_LOCKED = PAGE_KERNEL;
2731
2732	_PAGE_IE = _PAGE_IE_4V;
2733	_PAGE_E = _PAGE_E_4V;
2734	_PAGE_CACHE = page_cache4v_flag;
2735
2736#ifdef CONFIG_DEBUG_PAGEALLOC
2737	kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2738#else
2739	kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4V) ^
2740		PAGE_OFFSET;
2741#endif
2742	kern_linear_pte_xor[0] |= (page_cache4v_flag | _PAGE_P_4V |
2743				   _PAGE_W_4V);
2744
2745	for (i = 1; i < 4; i++)
2746		kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2747
2748	pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4V | __DIRTY_BITS_4V |
2749		     __ACCESS_BITS_4V | _PAGE_E_4V);
2750
2751	_PAGE_ALL_SZ_BITS = (_PAGE_SZ16GB_4V | _PAGE_SZ2GB_4V |
2752			     _PAGE_SZ256MB_4V | _PAGE_SZ32MB_4V |
2753			     _PAGE_SZ4MB_4V | _PAGE_SZ512K_4V |
2754			     _PAGE_SZ64K_4V | _PAGE_SZ8K_4V);
2755
2756	page_none = _PAGE_PRESENT_4V | _PAGE_ACCESSED_4V | page_cache4v_flag;
2757	page_shared = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2758		       __ACCESS_BITS_4V | _PAGE_WRITE_4V | _PAGE_EXEC_4V);
2759	page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2760		       __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2761	page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2762			 __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2763
2764	page_exec_bit = _PAGE_EXEC_4V;
2765
2766	prot_init_common(page_none, page_shared, page_copy, page_readonly,
2767			 page_exec_bit);
2768}
2769
2770unsigned long pte_sz_bits(unsigned long sz)
2771{
2772	if (tlb_type == hypervisor) {
2773		switch (sz) {
2774		case 8 * 1024:
2775		default:
2776			return _PAGE_SZ8K_4V;
2777		case 64 * 1024:
2778			return _PAGE_SZ64K_4V;
2779		case 512 * 1024:
2780			return _PAGE_SZ512K_4V;
2781		case 4 * 1024 * 1024:
2782			return _PAGE_SZ4MB_4V;
2783		}
2784	} else {
2785		switch (sz) {
2786		case 8 * 1024:
2787		default:
2788			return _PAGE_SZ8K_4U;
2789		case 64 * 1024:
2790			return _PAGE_SZ64K_4U;
2791		case 512 * 1024:
2792			return _PAGE_SZ512K_4U;
2793		case 4 * 1024 * 1024:
2794			return _PAGE_SZ4MB_4U;
2795		}
2796	}
2797}
2798
2799pte_t mk_pte_io(unsigned long page, pgprot_t prot, int space, unsigned long page_size)
2800{
2801	pte_t pte;
2802
2803	pte_val(pte)  = page | pgprot_val(pgprot_noncached(prot));
2804	pte_val(pte) |= (((unsigned long)space) << 32);
2805	pte_val(pte) |= pte_sz_bits(page_size);
2806
2807	return pte;
2808}
2809
2810static unsigned long kern_large_tte(unsigned long paddr)
2811{
2812	unsigned long val;
2813
2814	val = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2815	       _PAGE_CP_4U | _PAGE_CV_4U | _PAGE_P_4U |
2816	       _PAGE_EXEC_4U | _PAGE_L_4U | _PAGE_W_4U);
2817	if (tlb_type == hypervisor)
2818		val = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2819		       page_cache4v_flag | _PAGE_P_4V |
2820		       _PAGE_EXEC_4V | _PAGE_W_4V);
2821
2822	return val | paddr;
2823}
2824
2825/* If not locked, zap it. */
2826void __flush_tlb_all(void)
2827{
2828	unsigned long pstate;
2829	int i;
2830
2831	__asm__ __volatile__("flushw\n\t"
2832			     "rdpr	%%pstate, %0\n\t"
2833			     "wrpr	%0, %1, %%pstate"
2834			     : "=r" (pstate)
2835			     : "i" (PSTATE_IE));
2836	if (tlb_type == hypervisor) {
2837		sun4v_mmu_demap_all();
2838	} else if (tlb_type == spitfire) {
2839		for (i = 0; i < 64; i++) {
2840			/* Spitfire Errata #32 workaround */
2841			/* NOTE: Always runs on spitfire, so no
2842			 *       cheetah+ page size encodings.
2843			 */
2844			__asm__ __volatile__("stxa	%0, [%1] %2\n\t"
2845					     "flush	%%g6"
2846					     : /* No outputs */
2847					     : "r" (0),
2848					     "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2849
2850			if (!(spitfire_get_dtlb_data(i) & _PAGE_L_4U)) {
2851				__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2852						     "membar #Sync"
2853						     : /* no outputs */
2854						     : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
2855				spitfire_put_dtlb_data(i, 0x0UL);
2856			}
2857
2858			/* Spitfire Errata #32 workaround */
2859			/* NOTE: Always runs on spitfire, so no
2860			 *       cheetah+ page size encodings.
2861			 */
2862			__asm__ __volatile__("stxa	%0, [%1] %2\n\t"
2863					     "flush	%%g6"
2864					     : /* No outputs */
2865					     : "r" (0),
2866					     "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2867
2868			if (!(spitfire_get_itlb_data(i) & _PAGE_L_4U)) {
2869				__asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2870						     "membar #Sync"
2871						     : /* no outputs */
2872						     : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
2873				spitfire_put_itlb_data(i, 0x0UL);
2874			}
2875		}
2876	} else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
2877		cheetah_flush_dtlb_all();
2878		cheetah_flush_itlb_all();
2879	}
2880	__asm__ __volatile__("wrpr	%0, 0, %%pstate"
2881			     : : "r" (pstate));
2882}
2883
2884pte_t *pte_alloc_one_kernel(struct mm_struct *mm)
 
2885{
2886	struct page *page = alloc_page(GFP_KERNEL | __GFP_ZERO);
2887	pte_t *pte = NULL;
2888
2889	if (page)
2890		pte = (pte_t *) page_address(page);
2891
2892	return pte;
2893}
2894
2895pgtable_t pte_alloc_one(struct mm_struct *mm)
 
2896{
2897	struct page *page = alloc_page(GFP_KERNEL | __GFP_ZERO);
2898	if (!page)
2899		return NULL;
2900	if (!pgtable_pte_page_ctor(page)) {
2901		free_unref_page(page);
2902		return NULL;
2903	}
2904	return (pte_t *) page_address(page);
2905}
2906
2907void pte_free_kernel(struct mm_struct *mm, pte_t *pte)
2908{
2909	free_page((unsigned long)pte);
2910}
2911
2912static void __pte_free(pgtable_t pte)
2913{
2914	struct page *page = virt_to_page(pte);
2915
2916	pgtable_pte_page_dtor(page);
2917	__free_page(page);
2918}
2919
2920void pte_free(struct mm_struct *mm, pgtable_t pte)
2921{
2922	__pte_free(pte);
2923}
2924
2925void pgtable_free(void *table, bool is_page)
2926{
2927	if (is_page)
2928		__pte_free(table);
2929	else
2930		kmem_cache_free(pgtable_cache, table);
2931}
2932
2933#ifdef CONFIG_TRANSPARENT_HUGEPAGE
2934void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
2935			  pmd_t *pmd)
2936{
2937	unsigned long pte, flags;
2938	struct mm_struct *mm;
2939	pmd_t entry = *pmd;
2940
2941	if (!pmd_large(entry) || !pmd_young(entry))
2942		return;
2943
2944	pte = pmd_val(entry);
2945
2946	/* Don't insert a non-valid PMD into the TSB, we'll deadlock.  */
2947	if (!(pte & _PAGE_VALID))
2948		return;
2949
2950	/* We are fabricating 8MB pages using 4MB real hw pages.  */
2951	pte |= (addr & (1UL << REAL_HPAGE_SHIFT));
2952
2953	mm = vma->vm_mm;
2954
2955	spin_lock_irqsave(&mm->context.lock, flags);
2956
2957	if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL)
2958		__update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
2959					addr, pte);
2960
2961	spin_unlock_irqrestore(&mm->context.lock, flags);
2962}
2963#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2964
2965#if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
2966static void context_reload(void *__data)
2967{
2968	struct mm_struct *mm = __data;
2969
2970	if (mm == current->mm)
2971		load_secondary_context(mm);
2972}
2973
2974void hugetlb_setup(struct pt_regs *regs)
2975{
2976	struct mm_struct *mm = current->mm;
2977	struct tsb_config *tp;
2978
2979	if (faulthandler_disabled() || !mm) {
2980		const struct exception_table_entry *entry;
2981
2982		entry = search_exception_tables(regs->tpc);
2983		if (entry) {
2984			regs->tpc = entry->fixup;
2985			regs->tnpc = regs->tpc + 4;
2986			return;
2987		}
2988		pr_alert("Unexpected HugeTLB setup in atomic context.\n");
2989		die_if_kernel("HugeTSB in atomic", regs);
2990	}
2991
2992	tp = &mm->context.tsb_block[MM_TSB_HUGE];
2993	if (likely(tp->tsb == NULL))
2994		tsb_grow(mm, MM_TSB_HUGE, 0);
2995
2996	tsb_context_switch(mm);
2997	smp_tsb_sync(mm);
2998
2999	/* On UltraSPARC-III+ and later, configure the second half of
3000	 * the Data-TLB for huge pages.
3001	 */
3002	if (tlb_type == cheetah_plus) {
3003		bool need_context_reload = false;
3004		unsigned long ctx;
3005
3006		spin_lock_irq(&ctx_alloc_lock);
3007		ctx = mm->context.sparc64_ctx_val;
3008		ctx &= ~CTX_PGSZ_MASK;
3009		ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT;
3010		ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT;
3011
3012		if (ctx != mm->context.sparc64_ctx_val) {
3013			/* When changing the page size fields, we
3014			 * must perform a context flush so that no
3015			 * stale entries match.  This flush must
3016			 * occur with the original context register
3017			 * settings.
3018			 */
3019			do_flush_tlb_mm(mm);
3020
3021			/* Reload the context register of all processors
3022			 * also executing in this address space.
3023			 */
3024			mm->context.sparc64_ctx_val = ctx;
3025			need_context_reload = true;
3026		}
3027		spin_unlock_irq(&ctx_alloc_lock);
3028
3029		if (need_context_reload)
3030			on_each_cpu(context_reload, mm, 0);
3031	}
3032}
3033#endif
3034
3035static struct resource code_resource = {
3036	.name	= "Kernel code",
3037	.flags	= IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
3038};
3039
3040static struct resource data_resource = {
3041	.name	= "Kernel data",
3042	.flags	= IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
3043};
3044
3045static struct resource bss_resource = {
3046	.name	= "Kernel bss",
3047	.flags	= IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
3048};
3049
3050static inline resource_size_t compute_kern_paddr(void *addr)
3051{
3052	return (resource_size_t) (addr - KERNBASE + kern_base);
3053}
3054
3055static void __init kernel_lds_init(void)
3056{
3057	code_resource.start = compute_kern_paddr(_text);
3058	code_resource.end   = compute_kern_paddr(_etext - 1);
3059	data_resource.start = compute_kern_paddr(_etext);
3060	data_resource.end   = compute_kern_paddr(_edata - 1);
3061	bss_resource.start  = compute_kern_paddr(__bss_start);
3062	bss_resource.end    = compute_kern_paddr(_end - 1);
3063}
3064
3065static int __init report_memory(void)
3066{
3067	int i;
3068	struct resource *res;
3069
3070	kernel_lds_init();
3071
3072	for (i = 0; i < pavail_ents; i++) {
3073		res = kzalloc(sizeof(struct resource), GFP_KERNEL);
3074
3075		if (!res) {
3076			pr_warn("Failed to allocate source.\n");
3077			break;
3078		}
3079
3080		res->name = "System RAM";
3081		res->start = pavail[i].phys_addr;
3082		res->end = pavail[i].phys_addr + pavail[i].reg_size - 1;
3083		res->flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM;
3084
3085		if (insert_resource(&iomem_resource, res) < 0) {
3086			pr_warn("Resource insertion failed.\n");
3087			break;
3088		}
3089
3090		insert_resource(res, &code_resource);
3091		insert_resource(res, &data_resource);
3092		insert_resource(res, &bss_resource);
3093	}
3094
3095	return 0;
3096}
3097arch_initcall(report_memory);
3098
3099#ifdef CONFIG_SMP
3100#define do_flush_tlb_kernel_range	smp_flush_tlb_kernel_range
3101#else
3102#define do_flush_tlb_kernel_range	__flush_tlb_kernel_range
3103#endif
3104
3105void flush_tlb_kernel_range(unsigned long start, unsigned long end)
3106{
3107	if (start < HI_OBP_ADDRESS && end > LOW_OBP_ADDRESS) {
3108		if (start < LOW_OBP_ADDRESS) {
3109			flush_tsb_kernel_range(start, LOW_OBP_ADDRESS);
3110			do_flush_tlb_kernel_range(start, LOW_OBP_ADDRESS);
3111		}
3112		if (end > HI_OBP_ADDRESS) {
3113			flush_tsb_kernel_range(HI_OBP_ADDRESS, end);
3114			do_flush_tlb_kernel_range(HI_OBP_ADDRESS, end);
3115		}
3116	} else {
3117		flush_tsb_kernel_range(start, end);
3118		do_flush_tlb_kernel_range(start, end);
3119	}
3120}
3121
3122void copy_user_highpage(struct page *to, struct page *from,
3123	unsigned long vaddr, struct vm_area_struct *vma)
3124{
3125	char *vfrom, *vto;
3126
3127	vfrom = kmap_atomic(from);
3128	vto = kmap_atomic(to);
3129	copy_user_page(vto, vfrom, vaddr, to);
3130	kunmap_atomic(vto);
3131	kunmap_atomic(vfrom);
3132
3133	/* If this page has ADI enabled, copy over any ADI tags
3134	 * as well
3135	 */
3136	if (vma->vm_flags & VM_SPARC_ADI) {
3137		unsigned long pfrom, pto, i, adi_tag;
3138
3139		pfrom = page_to_phys(from);
3140		pto = page_to_phys(to);
3141
3142		for (i = pfrom; i < (pfrom + PAGE_SIZE); i += adi_blksize()) {
3143			asm volatile("ldxa [%1] %2, %0\n\t"
3144					: "=r" (adi_tag)
3145					:  "r" (i), "i" (ASI_MCD_REAL));
3146			asm volatile("stxa %0, [%1] %2\n\t"
3147					:
3148					: "r" (adi_tag), "r" (pto),
3149					  "i" (ASI_MCD_REAL));
3150			pto += adi_blksize();
3151		}
3152		asm volatile("membar #Sync\n\t");
3153	}
3154}
3155EXPORT_SYMBOL(copy_user_highpage);
3156
3157void copy_highpage(struct page *to, struct page *from)
3158{
3159	char *vfrom, *vto;
3160
3161	vfrom = kmap_atomic(from);
3162	vto = kmap_atomic(to);
3163	copy_page(vto, vfrom);
3164	kunmap_atomic(vto);
3165	kunmap_atomic(vfrom);
3166
3167	/* If this platform is ADI enabled, copy any ADI tags
3168	 * as well
3169	 */
3170	if (adi_capable()) {
3171		unsigned long pfrom, pto, i, adi_tag;
3172
3173		pfrom = page_to_phys(from);
3174		pto = page_to_phys(to);
3175
3176		for (i = pfrom; i < (pfrom + PAGE_SIZE); i += adi_blksize()) {
3177			asm volatile("ldxa [%1] %2, %0\n\t"
3178					: "=r" (adi_tag)
3179					:  "r" (i), "i" (ASI_MCD_REAL));
3180			asm volatile("stxa %0, [%1] %2\n\t"
3181					:
3182					: "r" (adi_tag), "r" (pto),
3183					  "i" (ASI_MCD_REAL));
3184			pto += adi_blksize();
3185		}
3186		asm volatile("membar #Sync\n\t");
3187	}
3188}
3189EXPORT_SYMBOL(copy_highpage);