arch/powerpc/mm/mem.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  *  PPC44x/36-bit changes by Matt Porter (mporter@mvista.com)
   9  *
  10  *  Derived from "arch/i386/mm/init.c"
  11  *    Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
  12  *
  13  *  This program is free software; you can redistribute it and/or
  14  *  modify it under the terms of the GNU General Public License
  15  *  as published by the Free Software Foundation; either version
  16  *  2 of the License, or (at your option) any later version.
  17  *
  18  */
  19
  20 #include <linux/module.h>
  21 #include <linux/sched.h>
  22 #include <linux/kernel.h>
  23 #include <linux/errno.h>
  24 #include <linux/string.h>
  25 #include <linux/types.h>
  26 #include <linux/mm.h>
  27 #include <linux/stddef.h>
  28 #include <linux/init.h>
  29 #include <linux/bootmem.h>
  30 #include <linux/highmem.h>
  31 #include <linux/initrd.h>
  32 #include <linux/pagemap.h>
  33 #include <linux/suspend.h>
  34 #include <linux/lmb.h>
  35
  36 #include <asm/pgalloc.h>
  37 #include <asm/prom.h>
  38 #include <asm/io.h>
  39 #include <asm/mmu_context.h>
  40 #include <asm/pgtable.h>
  41 #include <asm/mmu.h>
  42 #include <asm/smp.h>
  43 #include <asm/machdep.h>
  44 #include <asm/btext.h>
  45 #include <asm/tlb.h>
  46 #include <asm/sections.h>
  47 #include <asm/vdso.h>
  48 #include <asm/fixmap.h>
  49
  50 #include "mmu_decl.h"
  51
  52 #ifndef CPU_FTR_COHERENT_ICACHE
  53 #define CPU_FTR_COHERENT_ICACHE 0       /* XXX for now */
  54 #define CPU_FTR_NOEXECUTE       0
  55 #endif
  56
  57 int init_bootmem_done;
  58 int mem_init_done;
  59 unsigned long memory_limit;
  60
  61 #ifdef CONFIG_HIGHMEM
  62 pte_t *kmap_pte;
  63 pgprot_t kmap_prot;
  64
  65 EXPORT_SYMBOL(kmap_prot);
  66 EXPORT_SYMBOL(kmap_pte);
  67
  68 static inline pte_t *virt_to_kpte(unsigned long vaddr)
  69 {
  70         return pte_offset_kernel(pmd_offset(pud_offset(pgd_offset_k(vaddr),
  71                         vaddr), vaddr), vaddr);
  72 }
  73 #endif
  74
  75 int page_is_ram(unsigned long pfn)
  76 {
  77         unsigned long paddr = (pfn << PAGE_SHIFT);
  78
  79 #ifndef CONFIG_PPC64    /* XXX for now */
  80         return paddr < __pa(high_memory);
  81 #else
  82         int i;
  83         for (i=0; i < lmb.memory.cnt; i++) {
  84                 unsigned long base;
  85
  86                 base = lmb.memory.region[i].base;
  87
  88                 if ((paddr >= base) &&
  89                         (paddr < (base + lmb.memory.region[i].size))) {
  90                         return 1;
  91                 }
  92         }
  93
  94         return 0;
  95 #endif
  96 }
  97
  98 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
  99                               unsigned long size, pgprot_t vma_prot)
 100 {
 101         if (ppc_md.phys_mem_access_prot)
 102                 return ppc_md.phys_mem_access_prot(file, pfn, size, vma_prot);
 103
 104         if (!page_is_ram(pfn))
 105                 vma_prot = __pgprot(pgprot_val(vma_prot)
 106                                     | _PAGE_GUARDED | _PAGE_NO_CACHE);
 107         return vma_prot;
 108 }
 109 EXPORT_SYMBOL(phys_mem_access_prot);
 110
 111 #ifdef CONFIG_MEMORY_HOTPLUG
 112
 113 void online_page(struct page *page)
 114 {
 115         ClearPageReserved(page);
 116         init_page_count(page);
 117         __free_page(page);
 118         totalram_pages++;
 119         num_physpages++;
 120 }
 121
 122 #ifdef CONFIG_NUMA
 123 int memory_add_physaddr_to_nid(u64 start)
 124 {
 125         return hot_add_scn_to_nid(start);
 126 }
 127 #endif
 128
 129 int arch_add_memory(int nid, u64 start, u64 size)
 130 {
 131         struct pglist_data *pgdata;
 132         struct zone *zone;
 133         unsigned long start_pfn = start >> PAGE_SHIFT;
 134         unsigned long nr_pages = size >> PAGE_SHIFT;
 135
 136         pgdata = NODE_DATA(nid);
 137
 138         start = (unsigned long)__va(start);
 139         create_section_mapping(start, start + size);
 140
 141         /* this should work for most non-highmem platforms */
 142         zone = pgdata->node_zones;
 143
 144         return __add_pages(zone, start_pfn, nr_pages);
 145 }
 146
 147 #ifdef CONFIG_MEMORY_HOTREMOVE
 148 int remove_memory(u64 start, u64 size)
 149 {
 150         unsigned long start_pfn, end_pfn;
 151         int ret;
 152
 153         start_pfn = start >> PAGE_SHIFT;
 154         end_pfn = start_pfn + (size >> PAGE_SHIFT);
 155         ret = offline_pages(start_pfn, end_pfn, 120 * HZ);
 156         if (ret)
 157                 goto out;
 158         /* Arch-specific calls go here - next patch */
 159 out:
 160         return ret;
 161 }
 162 #endif /* CONFIG_MEMORY_HOTREMOVE */
 163
 164 /*
 165  * walk_memory_resource() needs to make sure there is no holes in a given
 166  * memory range. On PPC64, since this range comes from /sysfs, the range
 167  * is guaranteed to be valid, non-overlapping and can not contain any
 168  * holes. By the time we get here (memory add or remove), /proc/device-tree
 169  * is updated and correct. Only reason we need to check against device-tree
 170  * would be if we allow user-land to specify a memory range through a
 171  * system call/ioctl etc. instead of doing offline/online through /sysfs.
 172  */
 173 int
 174 walk_memory_resource(unsigned long start_pfn, unsigned long nr_pages, void *arg,
 175                         int (*func)(unsigned long, unsigned long, void *))
 176 {
 177         return  (*func)(start_pfn, nr_pages, arg);
 178 }
 179
 180 #endif /* CONFIG_MEMORY_HOTPLUG */
 181
 182 void show_mem(void)
 183 {
 184         unsigned long total = 0, reserved = 0;
 185         unsigned long shared = 0, cached = 0;
 186         unsigned long highmem = 0;
 187         struct page *page;
 188         pg_data_t *pgdat;
 189         unsigned long i;
 190
 191         printk("Mem-info:\n");
 192         show_free_areas();
 193         for_each_online_pgdat(pgdat) {
 194                 unsigned long flags;
 195                 pgdat_resize_lock(pgdat, &flags);
 196                 for (i = 0; i < pgdat->node_spanned_pages; i++) {
 197                         if (!pfn_valid(pgdat->node_start_pfn + i))
 198                                 continue;
 199                         page = pgdat_page_nr(pgdat, i);
 200                         total++;
 201                         if (PageHighMem(page))
 202                                 highmem++;
 203                         if (PageReserved(page))
 204                                 reserved++;
 205                         else if (PageSwapCache(page))
 206                                 cached++;
 207                         else if (page_count(page))
 208                                 shared += page_count(page) - 1;
 209                 }
 210                 pgdat_resize_unlock(pgdat, &flags);
 211         }
 212         printk("%ld pages of RAM\n", total);
 213 #ifdef CONFIG_HIGHMEM
 214         printk("%ld pages of HIGHMEM\n", highmem);
 215 #endif
 216         printk("%ld reserved pages\n", reserved);
 217         printk("%ld pages shared\n", shared);
 218         printk("%ld pages swap cached\n", cached);
 219 }
 220
 221 /*
 222  * Initialize the bootmem system and give it all the memory we
 223  * have available.  If we are using highmem, we only put the
 224  * lowmem into the bootmem system.
 225  */
 226 #ifndef CONFIG_NEED_MULTIPLE_NODES
 227 void __init do_init_bootmem(void)
 228 {
 229         unsigned long i;
 230         unsigned long start, bootmap_pages;
 231         unsigned long total_pages;
 232         int boot_mapsize;
 233
 234         max_low_pfn = max_pfn = lmb_end_of_DRAM() >> PAGE_SHIFT;
 235         total_pages = (lmb_end_of_DRAM() - memstart_addr) >> PAGE_SHIFT;
 236 #ifdef CONFIG_HIGHMEM
 237         total_pages = total_lowmem >> PAGE_SHIFT;
 238         max_low_pfn = lowmem_end_addr >> PAGE_SHIFT;
 239 #endif
 240
 241         /*
 242          * Find an area to use for the bootmem bitmap.  Calculate the size of
 243          * bitmap required as (Total Memory) / PAGE_SIZE / BITS_PER_BYTE.
 244          * Add 1 additional page in case the address isn't page-aligned.
 245          */
 246         bootmap_pages = bootmem_bootmap_pages(total_pages);
 247
 248         start = lmb_alloc(bootmap_pages << PAGE_SHIFT, PAGE_SIZE);
 249
 250         min_low_pfn = MEMORY_START >> PAGE_SHIFT;
 251         boot_mapsize = init_bootmem_node(NODE_DATA(0), start >> PAGE_SHIFT, min_low_pfn, max_low_pfn);
 252
 253         /* Add active regions with valid PFNs */
 254         for (i = 0; i < lmb.memory.cnt; i++) {
 255                 unsigned long start_pfn, end_pfn;
 256                 start_pfn = lmb.memory.region[i].base >> PAGE_SHIFT;
 257                 end_pfn = start_pfn + lmb_size_pages(&lmb.memory, i);
 258                 add_active_range(0, start_pfn, end_pfn);
 259         }
 260
 261         /* Add all physical memory to the bootmem map, mark each area
 262          * present.
 263          */
 264 #ifdef CONFIG_HIGHMEM
 265         free_bootmem_with_active_regions(0, lowmem_end_addr >> PAGE_SHIFT);
 266
 267         /* reserve the sections we're already using */
 268         for (i = 0; i < lmb.reserved.cnt; i++) {
 269                 unsigned long addr = lmb.reserved.region[i].base +
 270                                      lmb_size_bytes(&lmb.reserved, i) - 1;
 271                 if (addr < lowmem_end_addr)
 272                         reserve_bootmem(lmb.reserved.region[i].base,
 273                                         lmb_size_bytes(&lmb.reserved, i),
 274                                         BOOTMEM_DEFAULT);
 275                 else if (lmb.reserved.region[i].base < lowmem_end_addr) {
 276                         unsigned long adjusted_size = lowmem_end_addr -
 277                                       lmb.reserved.region[i].base;
 278                         reserve_bootmem(lmb.reserved.region[i].base,
 279                                         adjusted_size, BOOTMEM_DEFAULT);
 280                 }
 281         }
 282 #else
 283         free_bootmem_with_active_regions(0, max_pfn);
 284
 285         /* reserve the sections we're already using */
 286         for (i = 0; i < lmb.reserved.cnt; i++)
 287                 reserve_bootmem(lmb.reserved.region[i].base,
 288                                 lmb_size_bytes(&lmb.reserved, i),
 289                                 BOOTMEM_DEFAULT);
 290
 291 #endif
 292         /* XXX need to clip this if using highmem? */
 293         sparse_memory_present_with_active_regions(0);
 294
 295         init_bootmem_done = 1;
 296 }
 297
 298 /* mark pages that don't exist as nosave */
 299 static int __init mark_nonram_nosave(void)
 300 {
 301         unsigned long lmb_next_region_start_pfn,
 302                       lmb_region_max_pfn;
 303         int i;
 304
 305         for (i = 0; i < lmb.memory.cnt - 1; i++) {
 306                 lmb_region_max_pfn =
 307                         (lmb.memory.region[i].base >> PAGE_SHIFT) +
 308                         (lmb.memory.region[i].size >> PAGE_SHIFT);
 309                 lmb_next_region_start_pfn =
 310                         lmb.memory.region[i+1].base >> PAGE_SHIFT;
 311
 312                 if (lmb_region_max_pfn < lmb_next_region_start_pfn)
 313                         register_nosave_region(lmb_region_max_pfn,
 314                                                lmb_next_region_start_pfn);
 315         }
 316
 317         return 0;
 318 }
 319
 320 /*
 321  * paging_init() sets up the page tables - in fact we've already done this.
 322  */
 323 void __init paging_init(void)
 324 {
 325         unsigned long total_ram = lmb_phys_mem_size();
 326         unsigned long top_of_ram = lmb_end_of_DRAM();
 327         unsigned long max_zone_pfns[MAX_NR_ZONES];
 328
 329 #ifdef CONFIG_PPC32
 330         unsigned long v = __fix_to_virt(__end_of_fixed_addresses - 1);
 331         unsigned long end = __fix_to_virt(FIX_HOLE);
 332
 333         for (; v < end; v += PAGE_SIZE)
 334                 map_page(v, 0, 0); /* XXX gross */
 335 #endif
 336
 337 #ifdef CONFIG_HIGHMEM
 338         map_page(PKMAP_BASE, 0, 0);     /* XXX gross */
 339         pkmap_page_table = virt_to_kpte(PKMAP_BASE);
 340
 341         kmap_pte = virt_to_kpte(__fix_to_virt(FIX_KMAP_BEGIN));
 342         kmap_prot = PAGE_KERNEL;
 343 #endif /* CONFIG_HIGHMEM */
 344
 345         printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
 346                top_of_ram, total_ram);
 347         printk(KERN_DEBUG "Memory hole size: %ldMB\n",
 348                (top_of_ram - total_ram) >> 20);
 349         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 350 #ifdef CONFIG_HIGHMEM
 351         max_zone_pfns[ZONE_DMA] = lowmem_end_addr >> PAGE_SHIFT;
 352         max_zone_pfns[ZONE_HIGHMEM] = top_of_ram >> PAGE_SHIFT;
 353 #else
 354         max_zone_pfns[ZONE_DMA] = top_of_ram >> PAGE_SHIFT;
 355 #endif
 356         free_area_init_nodes(max_zone_pfns);
 357
 358         mark_nonram_nosave();
 359 }
 360 #endif /* ! CONFIG_NEED_MULTIPLE_NODES */
 361
 362 void __init mem_init(void)
 363 {
 364 #ifdef CONFIG_NEED_MULTIPLE_NODES
 365         int nid;
 366 #endif
 367         pg_data_t *pgdat;
 368         unsigned long i;
 369         struct page *page;
 370         unsigned long reservedpages = 0, codesize, initsize, datasize, bsssize;
 371
 372         num_physpages = lmb.memory.size >> PAGE_SHIFT;
 373         high_memory = (void *) __va(max_low_pfn * PAGE_SIZE);
 374
 375 #ifdef CONFIG_NEED_MULTIPLE_NODES
 376         for_each_online_node(nid) {
 377                 if (NODE_DATA(nid)->node_spanned_pages != 0) {
 378                         printk("freeing bootmem node %d\n", nid);
 379                         totalram_pages +=
 380                                 free_all_bootmem_node(NODE_DATA(nid));
 381                 }
 382         }
 383 #else
 384         max_mapnr = max_pfn;
 385         totalram_pages += free_all_bootmem();
 386 #endif
 387         for_each_online_pgdat(pgdat) {
 388                 for (i = 0; i < pgdat->node_spanned_pages; i++) {
 389                         if (!pfn_valid(pgdat->node_start_pfn + i))
 390                                 continue;
 391                         page = pgdat_page_nr(pgdat, i);
 392                         if (PageReserved(page))
 393                                 reservedpages++;
 394                 }
 395         }
 396
 397         codesize = (unsigned long)&_sdata - (unsigned long)&_stext;
 398         datasize = (unsigned long)&_edata - (unsigned long)&_sdata;
 399         initsize = (unsigned long)&__init_end - (unsigned long)&__init_begin;
 400         bsssize = (unsigned long)&__bss_stop - (unsigned long)&__bss_start;
 401
 402 #ifdef CONFIG_HIGHMEM
 403         {
 404                 unsigned long pfn, highmem_mapnr;
 405
 406                 highmem_mapnr = lowmem_end_addr >> PAGE_SHIFT;
 407                 for (pfn = highmem_mapnr; pfn < max_mapnr; ++pfn) {
 408                         struct page *page = pfn_to_page(pfn);
 409                         if (lmb_is_reserved(pfn << PAGE_SHIFT))
 410                                 continue;
 411                         ClearPageReserved(page);
 412                         init_page_count(page);
 413                         __free_page(page);
 414                         totalhigh_pages++;
 415                         reservedpages--;
 416                 }
 417                 totalram_pages += totalhigh_pages;
 418                 printk(KERN_DEBUG "High memory: %luk\n",
 419                        totalhigh_pages << (PAGE_SHIFT-10));
 420         }
 421 #endif /* CONFIG_HIGHMEM */
 422
 423         printk(KERN_INFO "Memory: %luk/%luk available (%luk kernel code, "
 424                "%luk reserved, %luk data, %luk bss, %luk init)\n",
 425                 (unsigned long)nr_free_pages() << (PAGE_SHIFT-10),
 426                 num_physpages << (PAGE_SHIFT-10),
 427                 codesize >> 10,
 428                 reservedpages << (PAGE_SHIFT-10),
 429                 datasize >> 10,
 430                 bsssize >> 10,
 431                 initsize >> 10);
 432
 433         mem_init_done = 1;
 434 }
 435
 436 /*
 437  * This is called when a page has been modified by the kernel.
 438  * It just marks the page as not i-cache clean.  We do the i-cache
 439  * flush later when the page is given to a user process, if necessary.
 440  */
 441 void flush_dcache_page(struct page *page)
 442 {
 443         if (cpu_has_feature(CPU_FTR_COHERENT_ICACHE))
 444                 return;
 445         /* avoid an atomic op if possible */
 446         if (test_bit(PG_arch_1, &page->flags))
 447                 clear_bit(PG_arch_1, &page->flags);
 448 }
 449 EXPORT_SYMBOL(flush_dcache_page);
 450
 451 void flush_dcache_icache_page(struct page *page)
 452 {
 453 #ifdef CONFIG_BOOKE
 454         void *start = kmap_atomic(page, KM_PPC_SYNC_ICACHE);
 455         __flush_dcache_icache(start);
 456         kunmap_atomic(start, KM_PPC_SYNC_ICACHE);
 457 #elif defined(CONFIG_8xx) || defined(CONFIG_PPC64)
 458         /* On 8xx there is no need to kmap since highmem is not supported */
 459         __flush_dcache_icache(page_address(page));
 460 #else
 461         __flush_dcache_icache_phys(page_to_pfn(page) << PAGE_SHIFT);
 462 #endif
 463
 464 }
 465 void clear_user_page(void *page, unsigned long vaddr, struct page *pg)
 466 {
 467         clear_page(page);
 468
 469         /*
 470          * We shouldnt have to do this, but some versions of glibc
 471          * require it (ld.so assumes zero filled pages are icache clean)
 472          * - Anton
 473          */
 474         flush_dcache_page(pg);
 475 }
 476 EXPORT_SYMBOL(clear_user_page);
 477
 478 void copy_user_page(void *vto, void *vfrom, unsigned long vaddr,
 479                     struct page *pg)
 480 {
 481         copy_page(vto, vfrom);
 482
 483         /*
 484          * We should be able to use the following optimisation, however
 485          * there are two problems.
 486          * Firstly a bug in some versions of binutils meant PLT sections
 487          * were not marked executable.
 488          * Secondly the first word in the GOT section is blrl, used
 489          * to establish the GOT address. Until recently the GOT was
 490          * not marked executable.
 491          * - Anton
 492          */
 493 #if 0
 494         if (!vma->vm_file && ((vma->vm_flags & VM_EXEC) == 0))
 495                 return;
 496 #endif
 497
 498         flush_dcache_page(pg);
 499 }
 500
 501 void flush_icache_user_range(struct vm_area_struct *vma, struct page *page,
 502                              unsigned long addr, int len)
 503 {
 504         unsigned long maddr;
 505
 506         maddr = (unsigned long) kmap(page) + (addr & ~PAGE_MASK);
 507         flush_icache_range(maddr, maddr + len);
 508         kunmap(page);
 509 }
 510 EXPORT_SYMBOL(flush_icache_user_range);
 511
 512 /*
 513  * This is called at the end of handling a user page fault, when the
 514  * fault has been handled by updating a PTE in the linux page tables.
 515  * We use it to preload an HPTE into the hash table corresponding to
 516  * the updated linux PTE.
 517  *
 518  * This must always be called with the pte lock held.
 519  */
 520 void update_mmu_cache(struct vm_area_struct *vma, unsigned long address,
 521                       pte_t pte)
 522 {
 523 #ifdef CONFIG_PPC_STD_MMU
 524         unsigned long access = 0, trap;
 525 #endif
 526         unsigned long pfn = pte_pfn(pte);
 527
 528         /* handle i-cache coherency */
 529         if (!cpu_has_feature(CPU_FTR_COHERENT_ICACHE) &&
 530             !cpu_has_feature(CPU_FTR_NOEXECUTE) &&
 531             pfn_valid(pfn)) {
 532                 struct page *page = pfn_to_page(pfn);
 533 #ifdef CONFIG_8xx
 534                 /* On 8xx, cache control instructions (particularly
 535                  * "dcbst" from flush_dcache_icache) fault as write
 536                  * operation if there is an unpopulated TLB entry
 537                  * for the address in question. To workaround that,
 538                  * we invalidate the TLB here, thus avoiding dcbst
 539                  * misbehaviour.
 540                  */
 541                 _tlbie(address, 0 /* 8xx doesn't care about PID */);
 542 #endif
 543                 /* The _PAGE_USER test should really be _PAGE_EXEC, but
 544                  * older glibc versions execute some code from no-exec
 545                  * pages, which for now we are supporting.  If exec-only
 546                  * pages are ever implemented, this will have to change.
 547                  */
 548                 if (!PageReserved(page) && (pte_val(pte) & _PAGE_USER)
 549                     && !test_bit(PG_arch_1, &page->flags)) {
 550                         if (vma->vm_mm == current->active_mm) {
 551                                 __flush_dcache_icache((void *) address);
 552                         } else
 553                                 flush_dcache_icache_page(page);
 554                         set_bit(PG_arch_1, &page->flags);
 555                 }
 556         }
 557
 558 #ifdef CONFIG_PPC_STD_MMU
 559         /* We only want HPTEs for linux PTEs that have _PAGE_ACCESSED set */
 560         if (!pte_young(pte) || address >= TASK_SIZE)
 561                 return;
 562
 563         /* We try to figure out if we are coming from an instruction
 564          * access fault and pass that down to __hash_page so we avoid
 565          * double-faulting on execution of fresh text. We have to test
 566          * for regs NULL since init will get here first thing at boot
 567          *
 568          * We also avoid filling the hash if not coming from a fault
 569          */
 570         if (current->thread.regs == NULL)
 571                 return;
 572         trap = TRAP(current->thread.regs);
 573         if (trap == 0x400)
 574                 access |= _PAGE_EXEC;
 575         else if (trap != 0x300)
 576                 return;
 577         hash_preload(vma->vm_mm, address, access, trap);
 578 #endif /* CONFIG_PPC_STD_MMU */
 579 }