arch/powerpc/mm/init_64.c

   1 /*
   2  *  PowerPC version
   3  *    Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org)
   4  *
   5  *  Modifications by Paul Mackerras (PowerMac) (paulus@cs.anu.edu.au)
   6  *  and Cort Dougan (PReP) (cort@cs.nmt.edu)
   7  *    Copyright (C) 1996 Paul Mackerras
   8  *
   9  *  Derived from "arch/i386/mm/init.c"
  10  *    Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
  11  *
  12  *  Dave Engebretsen <engebret@us.ibm.com>
  13  *      Rework for PPC64 port.
  14  *
  15  *  This program is free software; you can redistribute it and/or
  16  *  modify it under the terms of the GNU General Public License
  17  *  as published by the Free Software Foundation; either version
  18  *  2 of the License, or (at your option) any later version.
  19  *
  20  */
  21
  22 #undef DEBUG
  23
  24 #include <linux/signal.h>
  25 #include <linux/sched.h>
  26 #include <linux/kernel.h>
  27 #include <linux/errno.h>
  28 #include <linux/string.h>
  29 #include <linux/types.h>
  30 #include <linux/mman.h>
  31 #include <linux/mm.h>
  32 #include <linux/swap.h>
  33 #include <linux/stddef.h>
  34 #include <linux/vmalloc.h>
  35 #include <linux/init.h>
  36 #include <linux/delay.h>
  37 #include <linux/highmem.h>
  38 #include <linux/idr.h>
  39 #include <linux/nodemask.h>
  40 #include <linux/module.h>
  41 #include <linux/poison.h>
  42 #include <linux/memblock.h>
  43 #include <linux/hugetlb.h>
  44 #include <linux/slab.h>
  45
  46 #include <asm/pgalloc.h>
  47 #include <asm/page.h>
  48 #include <asm/prom.h>
  49 #include <asm/rtas.h>
  50 #include <asm/io.h>
  51 #include <asm/mmu_context.h>
  52 #include <asm/pgtable.h>
  53 #include <asm/mmu.h>
  54 #include <asm/uaccess.h>
  55 #include <asm/smp.h>
  56 #include <asm/machdep.h>
  57 #include <asm/tlb.h>
  58 #include <asm/eeh.h>
  59 #include <asm/processor.h>
  60 #include <asm/mmzone.h>
  61 #include <asm/cputable.h>
  62 #include <asm/sections.h>
  63 #include <asm/iommu.h>
  64 #include <asm/vdso.h>
  65
  66 #include "mmu_decl.h"
  67
  68 #ifdef CONFIG_PPC_STD_MMU_64
  69 #if PGTABLE_RANGE > USER_VSID_RANGE
  70 #warning Limited user VSID range means pagetable space is wasted
  71 #endif
  72
  73 #if (TASK_SIZE_USER64 < PGTABLE_RANGE) && (TASK_SIZE_USER64 < USER_VSID_RANGE)
  74 #warning TASK_SIZE is smaller than it needs to be.
  75 #endif
  76 #endif /* CONFIG_PPC_STD_MMU_64 */
  77
  78 phys_addr_t memstart_addr = ~0;
  79 EXPORT_SYMBOL_GPL(memstart_addr);
  80 phys_addr_t kernstart_addr;
  81 EXPORT_SYMBOL_GPL(kernstart_addr);
  82
  83 static void pgd_ctor(void *addr)
  84 {
  85         memset(addr, 0, PGD_TABLE_SIZE);
  86 }
  87
  88 static void pmd_ctor(void *addr)
  89 {
  90         memset(addr, 0, PMD_TABLE_SIZE);
  91 }
  92
  93 struct kmem_cache *pgtable_cache[MAX_PGTABLE_INDEX_SIZE];
  94
  95 /*
  96  * Create a kmem_cache() for pagetables.  This is not used for PTE
  97  * pages - they're linked to struct page, come from the normal free
  98  * pages pool and have a different entry size (see real_pte_t) to
  99  * everything else.  Caches created by this function are used for all
 100  * the higher level pagetables, and for hugepage pagetables.
 101  */
 102 void pgtable_cache_add(unsigned shift, void (*ctor)(void *))
 103 {
 104         char *name;
 105         unsigned long table_size = sizeof(void *) << shift;
 106         unsigned long align = table_size;
 107
 108         /* When batching pgtable pointers for RCU freeing, we store
 109          * the index size in the low bits.  Table alignment must be
 110          * big enough to fit it.
 111          *
 112          * Likewise, hugeapge pagetable pointers contain a (different)
 113          * shift value in the low bits.  All tables must be aligned so
 114          * as to leave enough 0 bits in the address to contain it. */
 115         unsigned long minalign = max(MAX_PGTABLE_INDEX_SIZE + 1,
 116                                      HUGEPD_SHIFT_MASK + 1);
 117         struct kmem_cache *new;
 118
 119         /* It would be nice if this was a BUILD_BUG_ON(), but at the
 120          * moment, gcc doesn't seem to recognize is_power_of_2 as a
 121          * constant expression, so so much for that. */
 122         BUG_ON(!is_power_of_2(minalign));
 123         BUG_ON((shift < 1) || (shift > MAX_PGTABLE_INDEX_SIZE));
 124
 125         if (PGT_CACHE(shift))
 126                 return; /* Already have a cache of this size */
 127
 128         align = max_t(unsigned long, align, minalign);
 129         name = kasprintf(GFP_KERNEL, "pgtable-2^%d", shift);
 130         new = kmem_cache_create(name, table_size, align, 0, ctor);
 131         kfree(name);
 132         pgtable_cache[shift - 1] = new;
 133         pr_debug("Allocated pgtable cache for order %d\n", shift);
 134 }
 135
 136
 137 void pgtable_cache_init(void)
 138 {
 139         pgtable_cache_add(PGD_INDEX_SIZE, pgd_ctor);
 140         pgtable_cache_add(PMD_CACHE_INDEX, pmd_ctor);
 141         if (!PGT_CACHE(PGD_INDEX_SIZE) || !PGT_CACHE(PMD_CACHE_INDEX))
 142                 panic("Couldn't allocate pgtable caches");
 143         /* In all current configs, when the PUD index exists it's the
 144          * same size as either the pgd or pmd index.  Verify that the
 145          * initialization above has also created a PUD cache.  This
 146          * will need re-examiniation if we add new possibilities for
 147          * the pagetable layout. */
 148         BUG_ON(PUD_INDEX_SIZE && !PGT_CACHE(PUD_INDEX_SIZE));
 149 }
 150
 151 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 152 /*
 153  * Given an address within the vmemmap, determine the pfn of the page that
 154  * represents the start of the section it is within.  Note that we have to
 155  * do this by hand as the proffered address may not be correctly aligned.
 156  * Subtraction of non-aligned pointers produces undefined results.
 157  */
 158 static unsigned long __meminit vmemmap_section_start(unsigned long page)
 159 {
 160         unsigned long offset = page - ((unsigned long)(vmemmap));
 161
 162         /* Return the pfn of the start of the section. */
 163         return (offset / sizeof(struct page)) & PAGE_SECTION_MASK;
 164 }
 165
 166 /*
 167  * Check if this vmemmap page is already initialised.  If any section
 168  * which overlaps this vmemmap page is initialised then this page is
 169  * initialised already.
 170  */
 171 static int __meminit vmemmap_populated(unsigned long start, int page_size)
 172 {
 173         unsigned long end = start + page_size;
 174         start = (unsigned long)(pfn_to_page(vmemmap_section_start(start)));
 175
 176         for (; start < end; start += (PAGES_PER_SECTION * sizeof(struct page)))
 177                 if (pfn_valid(page_to_pfn((struct page *)start)))
 178                         return 1;
 179
 180         return 0;
 181 }
 182
 183 /* On hash-based CPUs, the vmemmap is bolted in the hash table.
 184  *
 185  * On Book3E CPUs, the vmemmap is currently mapped in the top half of
 186  * the vmalloc space using normal page tables, though the size of
 187  * pages encoded in the PTEs can be different
 188  */
 189
 190 #ifdef CONFIG_PPC_BOOK3E
 191 static void __meminit vmemmap_create_mapping(unsigned long start,
 192                                              unsigned long page_size,
 193                                              unsigned long phys)
 194 {
 195         /* Create a PTE encoding without page size */
 196         unsigned long i, flags = _PAGE_PRESENT | _PAGE_ACCESSED |
 197                 _PAGE_KERNEL_RW;
 198
 199         /* PTEs only contain page size encodings up to 32M */
 200         BUG_ON(mmu_psize_defs[mmu_vmemmap_psize].enc > 0xf);
 201
 202         /* Encode the size in the PTE */
 203         flags |= mmu_psize_defs[mmu_vmemmap_psize].enc << 8;
 204
 205         /* For each PTE for that area, map things. Note that we don't
 206          * increment phys because all PTEs are of the large size and
 207          * thus must have the low bits clear
 208          */
 209         for (i = 0; i < page_size; i += PAGE_SIZE)
 210                 BUG_ON(map_kernel_page(start + i, phys, flags));
 211 }
 212
 213 #ifdef CONFIG_MEMORY_HOTPLUG
 214 static void vmemmap_remove_mapping(unsigned long start,
 215                                    unsigned long page_size)
 216 {
 217 }
 218 #endif
 219 #else /* CONFIG_PPC_BOOK3E */
 220 static void __meminit vmemmap_create_mapping(unsigned long start,
 221                                              unsigned long page_size,
 222                                              unsigned long phys)
 223 {
 224         int  mapped = htab_bolt_mapping(start, start + page_size, phys,
 225                                         pgprot_val(PAGE_KERNEL),
 226                                         mmu_vmemmap_psize,
 227                                         mmu_kernel_ssize);
 228         BUG_ON(mapped < 0);
 229 }
 230
 231 #ifdef CONFIG_MEMORY_HOTPLUG
 232 static void vmemmap_remove_mapping(unsigned long start,
 233                                    unsigned long page_size)
 234 {
 235         int mapped = htab_remove_mapping(start, start + page_size,
 236                                          mmu_vmemmap_psize,
 237                                          mmu_kernel_ssize);
 238         BUG_ON(mapped < 0);
 239 }
 240 #endif
 241
 242 #endif /* CONFIG_PPC_BOOK3E */
 243
 244 struct vmemmap_backing *vmemmap_list;
 245 static struct vmemmap_backing *next;
 246 static int num_left;
 247 static int num_freed;
 248
 249 static __meminit struct vmemmap_backing * vmemmap_list_alloc(int node)
 250 {
 251         struct vmemmap_backing *vmem_back;
 252         /* get from freed entries first */
 253         if (num_freed) {
 254                 num_freed--;
 255                 vmem_back = next;
 256                 next = next->list;
 257
 258                 return vmem_back;
 259         }
 260
 261         /* allocate a page when required and hand out chunks */
 262         if (!num_left) {
 263                 next = vmemmap_alloc_block(PAGE_SIZE, node);
 264                 if (unlikely(!next)) {
 265                         WARN_ON(1);
 266                         return NULL;
 267                 }
 268                 num_left = PAGE_SIZE / sizeof(struct vmemmap_backing);
 269         }
 270
 271         num_left--;
 272
 273         return next++;
 274 }
 275
 276 static __meminit void vmemmap_list_populate(unsigned long phys,
 277                                             unsigned long start,
 278                                             int node)
 279 {
 280         struct vmemmap_backing *vmem_back;
 281
 282         vmem_back = vmemmap_list_alloc(node);
 283         if (unlikely(!vmem_back)) {
 284                 WARN_ON(1);
 285                 return;
 286         }
 287
 288         vmem_back->phys = phys;
 289         vmem_back->virt_addr = start;
 290         vmem_back->list = vmemmap_list;
 291
 292         vmemmap_list = vmem_back;
 293 }
 294
 295 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
 296 {
 297         unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
 298
 299         /* Align to the page size of the linear mapping. */
 300         start = _ALIGN_DOWN(start, page_size);
 301
 302         pr_debug("vmemmap_populate %lx..%lx, node %d\n", start, end, node);
 303
 304         for (; start < end; start += page_size) {
 305                 void *p;
 306
 307                 if (vmemmap_populated(start, page_size))
 308                         continue;
 309
 310                 p = vmemmap_alloc_block(page_size, node);
 311                 if (!p)
 312                         return -ENOMEM;
 313
 314                 vmemmap_list_populate(__pa(p), start, node);
 315
 316                 pr_debug("      * %016lx..%016lx allocated at %p\n",
 317                          start, start + page_size, p);
 318
 319                 vmemmap_create_mapping(start, page_size, __pa(p));
 320         }
 321
 322         return 0;
 323 }
 324
 325 #ifdef CONFIG_MEMORY_HOTPLUG
 326 static unsigned long vmemmap_list_free(unsigned long start)
 327 {
 328         struct vmemmap_backing *vmem_back, *vmem_back_prev;
 329
 330         vmem_back_prev = vmem_back = vmemmap_list;
 331
 332         /* look for it with prev pointer recorded */
 333         for (; vmem_back; vmem_back = vmem_back->list) {
 334                 if (vmem_back->virt_addr == start)
 335                         break;
 336                 vmem_back_prev = vmem_back;
 337         }
 338
 339         if (unlikely(!vmem_back)) {
 340                 WARN_ON(1);
 341                 return 0;
 342         }
 343
 344         /* remove it from vmemmap_list */
 345         if (vmem_back == vmemmap_list) /* remove head */
 346                 vmemmap_list = vmem_back->list;
 347         else
 348                 vmem_back_prev->list = vmem_back->list;
 349
 350         /* next point to this freed entry */
 351         vmem_back->list = next;
 352         next = vmem_back;
 353         num_freed++;
 354
 355         return vmem_back->phys;
 356 }
 357
 358 void __ref vmemmap_free(unsigned long start, unsigned long end)
 359 {
 360         unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
 361
 362         start = _ALIGN_DOWN(start, page_size);
 363
 364         pr_debug("vmemmap_free %lx...%lx\n", start, end);
 365
 366         for (; start < end; start += page_size) {
 367                 unsigned long addr;
 368
 369                 /*
 370                  * the section has already be marked as invalid, so
 371                  * vmemmap_populated() true means some other sections still
 372                  * in this page, so skip it.
 373                  */
 374                 if (vmemmap_populated(start, page_size))
 375                         continue;
 376
 377                 addr = vmemmap_list_free(start);
 378                 if (addr) {
 379                         struct page *page = pfn_to_page(addr >> PAGE_SHIFT);
 380
 381                         if (PageReserved(page)) {
 382                                 /* allocated from bootmem */
 383                                 if (page_size < PAGE_SIZE) {
 384                                         /*
 385                                          * this shouldn't happen, but if it is
 386                                          * the case, leave the memory there
 387                                          */
 388                                         WARN_ON_ONCE(1);
 389                                 } else {
 390                                         unsigned int nr_pages =
 391                                                 1 << get_order(page_size);
 392                                         while (nr_pages--)
 393                                                 free_reserved_page(page++);
 394                                 }
 395                         } else
 396                                 free_pages((unsigned long)(__va(addr)),
 397                                                         get_order(page_size));
 398
 399                         vmemmap_remove_mapping(start, page_size);
 400                 }
 401         }
 402 }
 403 #endif
 404 void register_page_bootmem_memmap(unsigned long section_nr,
 405                                   struct page *start_page, unsigned long size)
 406 {
 407 }
 408
 409 /*
 410  * We do not have access to the sparsemem vmemmap, so we fallback to
 411  * walking the list of sparsemem blocks which we already maintain for
 412  * the sake of crashdump. In the long run, we might want to maintain
 413  * a tree if performance of that linear walk becomes a problem.
 414  *
 415  * realmode_pfn_to_page functions can fail due to:
 416  * 1) As real sparsemem blocks do not lay in RAM continously (they
 417  * are in virtual address space which is not available in the real mode),
 418  * the requested page struct can be split between blocks so get_page/put_page
 419  * may fail.
 420  * 2) When huge pages are used, the get_page/put_page API will fail
 421  * in real mode as the linked addresses in the page struct are virtual
 422  * too.
 423  */
 424 struct page *realmode_pfn_to_page(unsigned long pfn)
 425 {
 426         struct vmemmap_backing *vmem_back;
 427         struct page *page;
 428         unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
 429         unsigned long pg_va = (unsigned long) pfn_to_page(pfn);
 430
 431         for (vmem_back = vmemmap_list; vmem_back; vmem_back = vmem_back->list) {
 432                 if (pg_va < vmem_back->virt_addr)
 433                         continue;
 434
 435                 /* After vmemmap_list entry free is possible, need check all */
 436                 if ((pg_va + sizeof(struct page)) <=
 437                                 (vmem_back->virt_addr + page_size)) {
 438                         page = (struct page *) (vmem_back->phys + pg_va -
 439                                 vmem_back->virt_addr);
 440                         return page;
 441                 }
 442         }
 443
 444         /* Probably that page struct is split between real pages */
 445         return NULL;
 446 }
 447 EXPORT_SYMBOL_GPL(realmode_pfn_to_page);
 448
 449 #elif defined(CONFIG_FLATMEM)
 450
 451 struct page *realmode_pfn_to_page(unsigned long pfn)
 452 {
 453         struct page *page = pfn_to_page(pfn);
 454         return page;
 455 }
 456 EXPORT_SYMBOL_GPL(realmode_pfn_to_page);
 457
 458 #endif /* CONFIG_SPARSEMEM_VMEMMAP/CONFIG_FLATMEM */