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 H_PGTABLE_RANGE > USER_VSID_RANGE
  70 #warning Limited user VSID range means pagetable space is wasted
  71 #endif
  72
  73 #if (TASK_SIZE_USER64 < H_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 pud_ctor(void *addr)
  89 {
  90         memset(addr, 0, PUD_TABLE_SIZE);
  91 }
  92
  93 static void pmd_ctor(void *addr)
  94 {
  95         memset(addr, 0, PMD_TABLE_SIZE);
  96 }
  97
  98 struct kmem_cache *pgtable_cache[MAX_PGTABLE_INDEX_SIZE];
  99
 100 /*
 101  * Create a kmem_cache() for pagetables.  This is not used for PTE
 102  * pages - they're linked to struct page, come from the normal free
 103  * pages pool and have a different entry size (see real_pte_t) to
 104  * everything else.  Caches created by this function are used for all
 105  * the higher level pagetables, and for hugepage pagetables.
 106  */
 107 void pgtable_cache_add(unsigned shift, void (*ctor)(void *))
 108 {
 109         char *name;
 110         unsigned long table_size = sizeof(void *) << shift;
 111         unsigned long align = table_size;
 112
 113         /* When batching pgtable pointers for RCU freeing, we store
 114          * the index size in the low bits.  Table alignment must be
 115          * big enough to fit it.
 116          *
 117          * Likewise, hugeapge pagetable pointers contain a (different)
 118          * shift value in the low bits.  All tables must be aligned so
 119          * as to leave enough 0 bits in the address to contain it. */
 120         unsigned long minalign = max(MAX_PGTABLE_INDEX_SIZE + 1,
 121                                      HUGEPD_SHIFT_MASK + 1);
 122         struct kmem_cache *new;
 123
 124         /* It would be nice if this was a BUILD_BUG_ON(), but at the
 125          * moment, gcc doesn't seem to recognize is_power_of_2 as a
 126          * constant expression, so so much for that. */
 127         BUG_ON(!is_power_of_2(minalign));
 128         BUG_ON((shift < 1) || (shift > MAX_PGTABLE_INDEX_SIZE));
 129
 130         if (PGT_CACHE(shift))
 131                 return; /* Already have a cache of this size */
 132
 133         align = max_t(unsigned long, align, minalign);
 134         name = kasprintf(GFP_KERNEL, "pgtable-2^%d", shift);
 135         new = kmem_cache_create(name, table_size, align, 0, ctor);
 136         kfree(name);
 137         pgtable_cache[shift - 1] = new;
 138         pr_debug("Allocated pgtable cache for order %d\n", shift);
 139 }
 140
 141
 142 void pgtable_cache_init(void)
 143 {
 144         pgtable_cache_add(PGD_INDEX_SIZE, pgd_ctor);
 145         pgtable_cache_add(PMD_CACHE_INDEX, pmd_ctor);
 146         /*
 147          * In all current configs, when the PUD index exists it's the
 148          * same size as either the pgd or pmd index except with THP enabled
 149          * on book3s 64
 150          */
 151         if (PUD_INDEX_SIZE && !PGT_CACHE(PUD_INDEX_SIZE))
 152                 pgtable_cache_add(PUD_INDEX_SIZE, pud_ctor);
 153
 154         if (!PGT_CACHE(PGD_INDEX_SIZE) || !PGT_CACHE(PMD_CACHE_INDEX))
 155                 panic("Couldn't allocate pgtable caches");
 156         if (PUD_INDEX_SIZE && !PGT_CACHE(PUD_INDEX_SIZE))
 157                 panic("Couldn't allocate pud pgtable caches");
 158 }
 159
 160 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 161 /*
 162  * Given an address within the vmemmap, determine the pfn of the page that
 163  * represents the start of the section it is within.  Note that we have to
 164  * do this by hand as the proffered address may not be correctly aligned.
 165  * Subtraction of non-aligned pointers produces undefined results.
 166  */
 167 static unsigned long __meminit vmemmap_section_start(unsigned long page)
 168 {
 169         unsigned long offset = page - ((unsigned long)(vmemmap));
 170
 171         /* Return the pfn of the start of the section. */
 172         return (offset / sizeof(struct page)) & PAGE_SECTION_MASK;
 173 }
 174
 175 /*
 176  * Check if this vmemmap page is already initialised.  If any section
 177  * which overlaps this vmemmap page is initialised then this page is
 178  * initialised already.
 179  */
 180 static int __meminit vmemmap_populated(unsigned long start, int page_size)
 181 {
 182         unsigned long end = start + page_size;
 183         start = (unsigned long)(pfn_to_page(vmemmap_section_start(start)));
 184
 185         for (; start < end; start += (PAGES_PER_SECTION * sizeof(struct page)))
 186                 if (pfn_valid(page_to_pfn((struct page *)start)))
 187                         return 1;
 188
 189         return 0;
 190 }
 191
 192 struct vmemmap_backing *vmemmap_list;
 193 static struct vmemmap_backing *next;
 194 static int num_left;
 195 static int num_freed;
 196
 197 static __meminit struct vmemmap_backing * vmemmap_list_alloc(int node)
 198 {
 199         struct vmemmap_backing *vmem_back;
 200         /* get from freed entries first */
 201         if (num_freed) {
 202                 num_freed--;
 203                 vmem_back = next;
 204                 next = next->list;
 205
 206                 return vmem_back;
 207         }
 208
 209         /* allocate a page when required and hand out chunks */
 210         if (!num_left) {
 211                 next = vmemmap_alloc_block(PAGE_SIZE, node);
 212                 if (unlikely(!next)) {
 213                         WARN_ON(1);
 214                         return NULL;
 215                 }
 216                 num_left = PAGE_SIZE / sizeof(struct vmemmap_backing);
 217         }
 218
 219         num_left--;
 220
 221         return next++;
 222 }
 223
 224 static __meminit void vmemmap_list_populate(unsigned long phys,
 225                                             unsigned long start,
 226                                             int node)
 227 {
 228         struct vmemmap_backing *vmem_back;
 229
 230         vmem_back = vmemmap_list_alloc(node);
 231         if (unlikely(!vmem_back)) {
 232                 WARN_ON(1);
 233                 return;
 234         }
 235
 236         vmem_back->phys = phys;
 237         vmem_back->virt_addr = start;
 238         vmem_back->list = vmemmap_list;
 239
 240         vmemmap_list = vmem_back;
 241 }
 242
 243 int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node)
 244 {
 245         unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
 246
 247         /* Align to the page size of the linear mapping. */
 248         start = _ALIGN_DOWN(start, page_size);
 249
 250         pr_debug("vmemmap_populate %lx..%lx, node %d\n", start, end, node);
 251
 252         for (; start < end; start += page_size) {
 253                 void *p;
 254                 int rc;
 255
 256                 if (vmemmap_populated(start, page_size))
 257                         continue;
 258
 259                 p = vmemmap_alloc_block(page_size, node);
 260                 if (!p)
 261                         return -ENOMEM;
 262
 263                 vmemmap_list_populate(__pa(p), start, node);
 264
 265                 pr_debug("      * %016lx..%016lx allocated at %p\n",
 266                          start, start + page_size, p);
 267
 268                 rc = vmemmap_create_mapping(start, page_size, __pa(p));
 269                 if (rc < 0) {
 270                         pr_warning(
 271                                 "vmemmap_populate: Unable to create vmemmap mapping: %d\n",
 272                                 rc);
 273                         return -EFAULT;
 274                 }
 275         }
 276
 277         return 0;
 278 }
 279
 280 #ifdef CONFIG_MEMORY_HOTPLUG
 281 static unsigned long vmemmap_list_free(unsigned long start)
 282 {
 283         struct vmemmap_backing *vmem_back, *vmem_back_prev;
 284
 285         vmem_back_prev = vmem_back = vmemmap_list;
 286
 287         /* look for it with prev pointer recorded */
 288         for (; vmem_back; vmem_back = vmem_back->list) {
 289                 if (vmem_back->virt_addr == start)
 290                         break;
 291                 vmem_back_prev = vmem_back;
 292         }
 293
 294         if (unlikely(!vmem_back)) {
 295                 WARN_ON(1);
 296                 return 0;
 297         }
 298
 299         /* remove it from vmemmap_list */
 300         if (vmem_back == vmemmap_list) /* remove head */
 301                 vmemmap_list = vmem_back->list;
 302         else
 303                 vmem_back_prev->list = vmem_back->list;
 304
 305         /* next point to this freed entry */
 306         vmem_back->list = next;
 307         next = vmem_back;
 308         num_freed++;
 309
 310         return vmem_back->phys;
 311 }
 312
 313 void __ref vmemmap_free(unsigned long start, unsigned long end)
 314 {
 315         unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
 316
 317         start = _ALIGN_DOWN(start, page_size);
 318
 319         pr_debug("vmemmap_free %lx...%lx\n", start, end);
 320
 321         for (; start < end; start += page_size) {
 322                 unsigned long addr;
 323
 324                 /*
 325                  * the section has already be marked as invalid, so
 326                  * vmemmap_populated() true means some other sections still
 327                  * in this page, so skip it.
 328                  */
 329                 if (vmemmap_populated(start, page_size))
 330                         continue;
 331
 332                 addr = vmemmap_list_free(start);
 333                 if (addr) {
 334                         struct page *page = pfn_to_page(addr >> PAGE_SHIFT);
 335
 336                         if (PageReserved(page)) {
 337                                 /* allocated from bootmem */
 338                                 if (page_size < PAGE_SIZE) {
 339                                         /*
 340                                          * this shouldn't happen, but if it is
 341                                          * the case, leave the memory there
 342                                          */
 343                                         WARN_ON_ONCE(1);
 344                                 } else {
 345                                         unsigned int nr_pages =
 346                                                 1 << get_order(page_size);
 347                                         while (nr_pages--)
 348                                                 free_reserved_page(page++);
 349                                 }
 350                         } else
 351                                 free_pages((unsigned long)(__va(addr)),
 352                                                         get_order(page_size));
 353
 354                         vmemmap_remove_mapping(start, page_size);
 355                 }
 356         }
 357 }
 358 #endif
 359 void register_page_bootmem_memmap(unsigned long section_nr,
 360                                   struct page *start_page, unsigned long size)
 361 {
 362 }
 363
 364 /*
 365  * We do not have access to the sparsemem vmemmap, so we fallback to
 366  * walking the list of sparsemem blocks which we already maintain for
 367  * the sake of crashdump. In the long run, we might want to maintain
 368  * a tree if performance of that linear walk becomes a problem.
 369  *
 370  * realmode_pfn_to_page functions can fail due to:
 371  * 1) As real sparsemem blocks do not lay in RAM continously (they
 372  * are in virtual address space which is not available in the real mode),
 373  * the requested page struct can be split between blocks so get_page/put_page
 374  * may fail.
 375  * 2) When huge pages are used, the get_page/put_page API will fail
 376  * in real mode as the linked addresses in the page struct are virtual
 377  * too.
 378  */
 379 struct page *realmode_pfn_to_page(unsigned long pfn)
 380 {
 381         struct vmemmap_backing *vmem_back;
 382         struct page *page;
 383         unsigned long page_size = 1 << mmu_psize_defs[mmu_vmemmap_psize].shift;
 384         unsigned long pg_va = (unsigned long) pfn_to_page(pfn);
 385
 386         for (vmem_back = vmemmap_list; vmem_back; vmem_back = vmem_back->list) {
 387                 if (pg_va < vmem_back->virt_addr)
 388                         continue;
 389
 390                 /* After vmemmap_list entry free is possible, need check all */
 391                 if ((pg_va + sizeof(struct page)) <=
 392                                 (vmem_back->virt_addr + page_size)) {
 393                         page = (struct page *) (vmem_back->phys + pg_va -
 394                                 vmem_back->virt_addr);
 395                         return page;
 396                 }
 397         }
 398
 399         /* Probably that page struct is split between real pages */
 400         return NULL;
 401 }
 402 EXPORT_SYMBOL_GPL(realmode_pfn_to_page);
 403
 404 #elif defined(CONFIG_FLATMEM)
 405
 406 struct page *realmode_pfn_to_page(unsigned long pfn)
 407 {
 408         struct page *page = pfn_to_page(pfn);
 409         return page;
 410 }
 411 EXPORT_SYMBOL_GPL(realmode_pfn_to_page);
 412
 413 #endif /* CONFIG_SPARSEMEM_VMEMMAP/CONFIG_FLATMEM */