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  *  Amiga/APUS changes by Jesper Skov (jskov@cygnus.co.uk).
   9  *  PPC44x/36-bit changes by Matt Porter (mporter@mvista.com)
  10  *
  11  *  Derived from "arch/i386/mm/init.c"
  12  *    Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
  13  *
  14  *  This program is free software; you can redistribute it and/or
  15  *  modify it under the terms of the GNU General Public License
  16  *  as published by the Free Software Foundation; either version
  17  *  2 of the License, or (at your option) any later version.
  18  *
  19  */
  20
  21 #include <linux/module.h>
  22 #include <linux/sched.h>
  23 #include <linux/kernel.h>
  24 #include <linux/errno.h>
  25 #include <linux/string.h>
  26 #include <linux/types.h>
  27 #include <linux/mm.h>
  28 #include <linux/stddef.h>
  29 #include <linux/init.h>
  30 #include <linux/bootmem.h>
  31 #include <linux/highmem.h>
  32 #include <linux/initrd.h>
  33 #include <linux/pagemap.h>
  34
  35 #include <asm/pgalloc.h>
  36 #include <asm/prom.h>
  37 #include <asm/io.h>
  38 #include <asm/mmu_context.h>
  39 #include <asm/pgtable.h>
  40 #include <asm/mmu.h>
  41 #include <asm/smp.h>
  42 #include <asm/machdep.h>
  43 #include <asm/btext.h>
  44 #include <asm/tlb.h>
  45 #include <asm/prom.h>
  46 #include <asm/lmb.h>
  47 #include <asm/sections.h>
  48 #include <asm/rtas.h>
  49 #include <asm/vdso.h>
  50
  51 #include "mmu_decl.h"
  52
  53 #ifndef CPU_FTR_COHERENT_ICACHE
  54 #define CPU_FTR_COHERENT_ICACHE 0       /* XXX for now */
  55 #define CPU_FTR_NOEXECUTE       0
  56 #endif
  57
  58 int init_bootmem_done;
  59 int mem_init_done;
  60 unsigned long memory_limit;
  61
  62 extern void hash_preload(struct mm_struct *mm, unsigned long ea,
  63                          unsigned long access, unsigned long trap);
  64
  65 /*
  66  * This is called by /dev/mem to know if a given address has to
  67  * be mapped non-cacheable or not
  68  */
  69 int page_is_ram(unsigned long pfn)
  70 {
  71         unsigned long paddr = (pfn << PAGE_SHIFT);
  72
  73 #ifndef CONFIG_PPC64    /* XXX for now */
  74         return paddr < __pa(high_memory);
  75 #else
  76         int i;
  77         for (i=0; i < lmb.memory.cnt; i++) {
  78                 unsigned long base;
  79
  80                 base = lmb.memory.region[i].base;
  81
  82                 if ((paddr >= base) &&
  83                         (paddr < (base + lmb.memory.region[i].size))) {
  84                         return 1;
  85                 }
  86         }
  87
  88         return 0;
  89 #endif
  90 }
  91 EXPORT_SYMBOL(page_is_ram);
  92
  93 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
  94                               unsigned long size, pgprot_t vma_prot)
  95 {
  96         if (ppc_md.phys_mem_access_prot)
  97                 return ppc_md.phys_mem_access_prot(file, pfn, size, vma_prot);
  98
  99         if (!page_is_ram(pfn))
 100                 vma_prot = __pgprot(pgprot_val(vma_prot)
 101                                     | _PAGE_GUARDED | _PAGE_NO_CACHE);
 102         return vma_prot;
 103 }
 104 EXPORT_SYMBOL(phys_mem_access_prot);
 105
 106 #ifdef CONFIG_MEMORY_HOTPLUG
 107
 108 void online_page(struct page *page)
 109 {
 110         ClearPageReserved(page);
 111         init_page_count(page);
 112         __free_page(page);
 113         totalram_pages++;
 114         num_physpages++;
 115 }
 116
 117 #ifdef CONFIG_NUMA
 118 int memory_add_physaddr_to_nid(u64 start)
 119 {
 120         return hot_add_scn_to_nid(start);
 121 }
 122 #endif
 123
 124 int __devinit arch_add_memory(int nid, u64 start, u64 size)
 125 {
 126         struct pglist_data *pgdata;
 127         struct zone *zone;
 128         unsigned long start_pfn = start >> PAGE_SHIFT;
 129         unsigned long nr_pages = size >> PAGE_SHIFT;
 130
 131         pgdata = NODE_DATA(nid);
 132
 133         start = (unsigned long)__va(start);
 134         create_section_mapping(start, start + size);
 135
 136         /* this should work for most non-highmem platforms */
 137         zone = pgdata->node_zones;
 138
 139         return __add_pages(zone, start_pfn, nr_pages);
 140
 141         return 0;
 142 }
 143
 144 /*
 145  * First pass at this code will check to determine if the remove
 146  * request is within the RMO.  Do not allow removal within the RMO.
 147  */
 148 int __devinit remove_memory(u64 start, u64 size)
 149 {
 150         struct zone *zone;
 151         unsigned long start_pfn, end_pfn, nr_pages;
 152
 153         start_pfn = start >> PAGE_SHIFT;
 154         nr_pages = size >> PAGE_SHIFT;
 155         end_pfn = start_pfn + nr_pages;
 156
 157         printk("%s(): Attempting to remove memoy in range "
 158                         "%lx to %lx\n", __func__, start, start+size);
 159         /*
 160          * check for range within RMO
 161          */
 162         zone = page_zone(pfn_to_page(start_pfn));
 163
 164         printk("%s(): memory will be removed from "
 165                         "the %s zone\n", __func__, zone->name);
 166
 167         /*
 168          * not handling removing memory ranges that
 169          * overlap multiple zones yet
 170          */
 171         if (end_pfn > (zone->zone_start_pfn + zone->spanned_pages))
 172                 goto overlap;
 173
 174         /* make sure it is NOT in RMO */
 175         if ((start < lmb.rmo_size) || ((start+size) < lmb.rmo_size)) {
 176                 printk("%s(): range to be removed must NOT be in RMO!\n",
 177                         __func__);
 178                 goto in_rmo;
 179         }
 180
 181         return __remove_pages(zone, start_pfn, nr_pages);
 182
 183 overlap:
 184         printk("%s(): memory range to be removed overlaps "
 185                 "multiple zones!!!\n", __func__);
 186 in_rmo:
 187         return -1;
 188 }
 189 #endif /* CONFIG_MEMORY_HOTPLUG */
 190
 191 void show_mem(void)
 192 {
 193         unsigned long total = 0, reserved = 0;
 194         unsigned long shared = 0, cached = 0;
 195         unsigned long highmem = 0;
 196         struct page *page;
 197         pg_data_t *pgdat;
 198         unsigned long i;
 199
 200         printk("Mem-info:\n");
 201         show_free_areas();
 202         printk("Free swap:       %6ldkB\n", nr_swap_pages<<(PAGE_SHIFT-10));
 203         for_each_online_pgdat(pgdat) {
 204                 unsigned long flags;
 205                 pgdat_resize_lock(pgdat, &flags);
 206                 for (i = 0; i < pgdat->node_spanned_pages; i++) {
 207                         if (!pfn_valid(pgdat->node_start_pfn + i))
 208                                 continue;
 209                         page = pgdat_page_nr(pgdat, i);
 210                         total++;
 211                         if (PageHighMem(page))
 212                                 highmem++;
 213                         if (PageReserved(page))
 214                                 reserved++;
 215                         else if (PageSwapCache(page))
 216                                 cached++;
 217                         else if (page_count(page))
 218                                 shared += page_count(page) - 1;
 219                 }
 220                 pgdat_resize_unlock(pgdat, &flags);
 221         }
 222         printk("%ld pages of RAM\n", total);
 223 #ifdef CONFIG_HIGHMEM
 224         printk("%ld pages of HIGHMEM\n", highmem);
 225 #endif
 226         printk("%ld reserved pages\n", reserved);
 227         printk("%ld pages shared\n", shared);
 228         printk("%ld pages swap cached\n", cached);
 229 }
 230
 231 /*
 232  * Initialize the bootmem system and give it all the memory we
 233  * have available.  If we are using highmem, we only put the
 234  * lowmem into the bootmem system.
 235  */
 236 #ifndef CONFIG_NEED_MULTIPLE_NODES
 237 void __init do_init_bootmem(void)
 238 {
 239         unsigned long i;
 240         unsigned long start, bootmap_pages;
 241         unsigned long total_pages;
 242         int boot_mapsize;
 243
 244         max_pfn = total_pages = lmb_end_of_DRAM() >> PAGE_SHIFT;
 245 #ifdef CONFIG_HIGHMEM
 246         total_pages = total_lowmem >> PAGE_SHIFT;
 247 #endif
 248
 249         /*
 250          * Find an area to use for the bootmem bitmap.  Calculate the size of
 251          * bitmap required as (Total Memory) / PAGE_SIZE / BITS_PER_BYTE.
 252          * Add 1 additional page in case the address isn't page-aligned.
 253          */
 254         bootmap_pages = bootmem_bootmap_pages(total_pages);
 255
 256         start = lmb_alloc(bootmap_pages << PAGE_SHIFT, PAGE_SIZE);
 257
 258         boot_mapsize = init_bootmem(start >> PAGE_SHIFT, total_pages);
 259
 260         /* Add all physical memory to the bootmem map, mark each area
 261          * present.
 262          */
 263         for (i = 0; i < lmb.memory.cnt; i++) {
 264                 unsigned long base = lmb.memory.region[i].base;
 265                 unsigned long size = lmb_size_bytes(&lmb.memory, i);
 266 #ifdef CONFIG_HIGHMEM
 267                 if (base >= total_lowmem)
 268                         continue;
 269                 if (base + size > total_lowmem)
 270                         size = total_lowmem - base;
 271 #endif
 272                 free_bootmem(base, size);
 273         }
 274
 275         /* reserve the sections we're already using */
 276         for (i = 0; i < lmb.reserved.cnt; i++)
 277                 reserve_bootmem(lmb.reserved.region[i].base,
 278                                 lmb_size_bytes(&lmb.reserved, i));
 279
 280         /* XXX need to clip this if using highmem? */
 281         for (i = 0; i < lmb.memory.cnt; i++)
 282                 memory_present(0, lmb_start_pfn(&lmb.memory, i),
 283                                lmb_end_pfn(&lmb.memory, i));
 284         init_bootmem_done = 1;
 285 }
 286
 287 /*
 288  * paging_init() sets up the page tables - in fact we've already done this.
 289  */
 290 void __init paging_init(void)
 291 {
 292         unsigned long zones_size[MAX_NR_ZONES];
 293         unsigned long zholes_size[MAX_NR_ZONES];
 294         unsigned long total_ram = lmb_phys_mem_size();
 295         unsigned long top_of_ram = lmb_end_of_DRAM();
 296
 297 #ifdef CONFIG_HIGHMEM
 298         map_page(PKMAP_BASE, 0, 0);     /* XXX gross */
 299         pkmap_page_table = pte_offset_kernel(pmd_offset(pgd_offset_k
 300                         (PKMAP_BASE), PKMAP_BASE), PKMAP_BASE);
 301         map_page(KMAP_FIX_BEGIN, 0, 0); /* XXX gross */
 302         kmap_pte = pte_offset_kernel(pmd_offset(pgd_offset_k
 303                         (KMAP_FIX_BEGIN), KMAP_FIX_BEGIN), KMAP_FIX_BEGIN);
 304         kmap_prot = PAGE_KERNEL;
 305 #endif /* CONFIG_HIGHMEM */
 306
 307         printk(KERN_DEBUG "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
 308                top_of_ram, total_ram);
 309         printk(KERN_DEBUG "Memory hole size: %ldMB\n",
 310                (top_of_ram - total_ram) >> 20);
 311         /*
 312          * All pages are DMA-able so we put them all in the DMA zone.
 313          */
 314         memset(zones_size, 0, sizeof(zones_size));
 315         memset(zholes_size, 0, sizeof(zholes_size));
 316
 317         zones_size[ZONE_DMA] = top_of_ram >> PAGE_SHIFT;
 318         zholes_size[ZONE_DMA] = (top_of_ram - total_ram) >> PAGE_SHIFT;
 319
 320 #ifdef CONFIG_HIGHMEM
 321         zones_size[ZONE_DMA] = total_lowmem >> PAGE_SHIFT;
 322         zones_size[ZONE_HIGHMEM] = (total_memory - total_lowmem) >> PAGE_SHIFT;
 323         zholes_size[ZONE_HIGHMEM] = (top_of_ram - total_ram) >> PAGE_SHIFT;
 324 #else
 325         zones_size[ZONE_DMA] = top_of_ram >> PAGE_SHIFT;
 326         zholes_size[ZONE_DMA] = (top_of_ram - total_ram) >> PAGE_SHIFT;
 327 #endif /* CONFIG_HIGHMEM */
 328
 329         free_area_init_node(0, NODE_DATA(0), zones_size,
 330                             __pa(PAGE_OFFSET) >> PAGE_SHIFT, zholes_size);
 331 }
 332 #endif /* ! CONFIG_NEED_MULTIPLE_NODES */
 333
 334 void __init mem_init(void)
 335 {
 336 #ifdef CONFIG_NEED_MULTIPLE_NODES
 337         int nid;
 338 #endif
 339         pg_data_t *pgdat;
 340         unsigned long i;
 341         struct page *page;
 342         unsigned long reservedpages = 0, codesize, initsize, datasize, bsssize;
 343
 344         num_physpages = lmb.memory.size >> PAGE_SHIFT;
 345         high_memory = (void *) __va(max_low_pfn * PAGE_SIZE);
 346
 347 #ifdef CONFIG_NEED_MULTIPLE_NODES
 348         for_each_online_node(nid) {
 349                 if (NODE_DATA(nid)->node_spanned_pages != 0) {
 350                         printk("freeing bootmem node %d\n", nid);
 351                         totalram_pages +=
 352                                 free_all_bootmem_node(NODE_DATA(nid));
 353                 }
 354         }
 355 #else
 356         max_mapnr = max_pfn;
 357         totalram_pages += free_all_bootmem();
 358 #endif
 359
 360 #ifdef CONFIG_PPC_PSERIES
 361         /* Mark the RTAS pages as PG_reserved so userspace can mmap them */
 362         if (rtas_rmo_buf) {
 363                 unsigned long pfn, start_pfn, end_pfn;
 364
 365                 start_pfn = rtas_rmo_buf >> PAGE_SHIFT;
 366                 end_pfn = (rtas_rmo_buf + RTAS_RMOBUF_MAX) >>  PAGE_SHIFT;
 367                 for (pfn = start_pfn; pfn < end_pfn; pfn++)
 368                         SetPageReserved(pfn_to_page(pfn));
 369         }
 370 #endif
 371
 372         for_each_online_pgdat(pgdat) {
 373                 for (i = 0; i < pgdat->node_spanned_pages; i++) {
 374                         if (!pfn_valid(pgdat->node_start_pfn + i))
 375                                 continue;
 376                         page = pgdat_page_nr(pgdat, i);
 377                         if (PageReserved(page))
 378                                 reservedpages++;
 379                 }
 380         }
 381
 382         codesize = (unsigned long)&_sdata - (unsigned long)&_stext;
 383         datasize = (unsigned long)&_edata - (unsigned long)&_sdata;
 384         initsize = (unsigned long)&__init_end - (unsigned long)&__init_begin;
 385         bsssize = (unsigned long)&__bss_stop - (unsigned long)&__bss_start;
 386
 387 #ifdef CONFIG_HIGHMEM
 388         {
 389                 unsigned long pfn, highmem_mapnr;
 390
 391                 highmem_mapnr = total_lowmem >> PAGE_SHIFT;
 392                 for (pfn = highmem_mapnr; pfn < max_mapnr; ++pfn) {
 393                         struct page *page = pfn_to_page(pfn);
 394
 395                         ClearPageReserved(page);
 396                         init_page_count(page);
 397                         __free_page(page);
 398                         totalhigh_pages++;
 399                 }
 400                 totalram_pages += totalhigh_pages;
 401                 printk(KERN_DEBUG "High memory: %luk\n",
 402                        totalhigh_pages << (PAGE_SHIFT-10));
 403         }
 404 #endif /* CONFIG_HIGHMEM */
 405
 406         printk(KERN_INFO "Memory: %luk/%luk available (%luk kernel code, "
 407                "%luk reserved, %luk data, %luk bss, %luk init)\n",
 408                 (unsigned long)nr_free_pages() << (PAGE_SHIFT-10),
 409                 num_physpages << (PAGE_SHIFT-10),
 410                 codesize >> 10,
 411                 reservedpages << (PAGE_SHIFT-10),
 412                 datasize >> 10,
 413                 bsssize >> 10,
 414                 initsize >> 10);
 415
 416         mem_init_done = 1;
 417
 418         /* Initialize the vDSO */
 419         vdso_init();
 420 }
 421
 422 /*
 423  * This is called when a page has been modified by the kernel.
 424  * It just marks the page as not i-cache clean.  We do the i-cache
 425  * flush later when the page is given to a user process, if necessary.
 426  */
 427 void flush_dcache_page(struct page *page)
 428 {
 429         if (cpu_has_feature(CPU_FTR_COHERENT_ICACHE))
 430                 return;
 431         /* avoid an atomic op if possible */
 432         if (test_bit(PG_arch_1, &page->flags))
 433                 clear_bit(PG_arch_1, &page->flags);
 434 }
 435 EXPORT_SYMBOL(flush_dcache_page);
 436
 437 void flush_dcache_icache_page(struct page *page)
 438 {
 439 #ifdef CONFIG_BOOKE
 440         void *start = kmap_atomic(page, KM_PPC_SYNC_ICACHE);
 441         __flush_dcache_icache(start);
 442         kunmap_atomic(start, KM_PPC_SYNC_ICACHE);
 443 #elif defined(CONFIG_8xx) || defined(CONFIG_PPC64)
 444         /* On 8xx there is no need to kmap since highmem is not supported */
 445         __flush_dcache_icache(page_address(page));
 446 #else
 447         __flush_dcache_icache_phys(page_to_pfn(page) << PAGE_SHIFT);
 448 #endif
 449
 450 }
 451 void clear_user_page(void *page, unsigned long vaddr, struct page *pg)
 452 {
 453         clear_page(page);
 454
 455         /*
 456          * We shouldnt have to do this, but some versions of glibc
 457          * require it (ld.so assumes zero filled pages are icache clean)
 458          * - Anton
 459          */
 460         flush_dcache_page(pg);
 461 }
 462 EXPORT_SYMBOL(clear_user_page);
 463
 464 void copy_user_page(void *vto, void *vfrom, unsigned long vaddr,
 465                     struct page *pg)
 466 {
 467         copy_page(vto, vfrom);
 468
 469         /*
 470          * We should be able to use the following optimisation, however
 471          * there are two problems.
 472          * Firstly a bug in some versions of binutils meant PLT sections
 473          * were not marked executable.
 474          * Secondly the first word in the GOT section is blrl, used
 475          * to establish the GOT address. Until recently the GOT was
 476          * not marked executable.
 477          * - Anton
 478          */
 479 #if 0
 480         if (!vma->vm_file && ((vma->vm_flags & VM_EXEC) == 0))
 481                 return;
 482 #endif
 483
 484         flush_dcache_page(pg);
 485 }
 486
 487 void flush_icache_user_range(struct vm_area_struct *vma, struct page *page,
 488                              unsigned long addr, int len)
 489 {
 490         unsigned long maddr;
 491
 492         maddr = (unsigned long) kmap(page) + (addr & ~PAGE_MASK);
 493         flush_icache_range(maddr, maddr + len);
 494         kunmap(page);
 495 }
 496 EXPORT_SYMBOL(flush_icache_user_range);
 497
 498 /*
 499  * This is called at the end of handling a user page fault, when the
 500  * fault has been handled by updating a PTE in the linux page tables.
 501  * We use it to preload an HPTE into the hash table corresponding to
 502  * the updated linux PTE.
 503  *
 504  * This must always be called with the pte lock held.
 505  */
 506 void update_mmu_cache(struct vm_area_struct *vma, unsigned long address,
 507                       pte_t pte)
 508 {
 509 #ifdef CONFIG_PPC_STD_MMU
 510         unsigned long access = 0, trap;
 511 #endif
 512         unsigned long pfn = pte_pfn(pte);
 513
 514         /* handle i-cache coherency */
 515         if (!cpu_has_feature(CPU_FTR_COHERENT_ICACHE) &&
 516             !cpu_has_feature(CPU_FTR_NOEXECUTE) &&
 517             pfn_valid(pfn)) {
 518                 struct page *page = pfn_to_page(pfn);
 519                 if (!PageReserved(page)
 520                     && !test_bit(PG_arch_1, &page->flags)) {
 521                         if (vma->vm_mm == current->active_mm) {
 522 #ifdef CONFIG_8xx
 523                         /* On 8xx, cache control instructions (particularly
 524                          * "dcbst" from flush_dcache_icache) fault as write
 525                          * operation if there is an unpopulated TLB entry
 526                          * for the address in question. To workaround that,
 527                          * we invalidate the TLB here, thus avoiding dcbst
 528                          * misbehaviour.
 529                          */
 530                                 _tlbie(address);
 531 #endif
 532                                 __flush_dcache_icache((void *) address);
 533                         } else
 534                                 flush_dcache_icache_page(page);
 535                         set_bit(PG_arch_1, &page->flags);
 536                 }
 537         }
 538
 539 #ifdef CONFIG_PPC_STD_MMU
 540         /* We only want HPTEs for linux PTEs that have _PAGE_ACCESSED set */
 541         if (!pte_young(pte) || address >= TASK_SIZE)
 542                 return;
 543
 544         /* We try to figure out if we are coming from an instruction
 545          * access fault and pass that down to __hash_page so we avoid
 546          * double-faulting on execution of fresh text. We have to test
 547          * for regs NULL since init will get here first thing at boot
 548          *
 549          * We also avoid filling the hash if not coming from a fault
 550          */
 551         if (current->thread.regs == NULL)
 552                 return;
 553         trap = TRAP(current->thread.regs);
 554         if (trap == 0x400)
 555                 access |= _PAGE_EXEC;
 556         else if (trap != 0x300)
 557                 return;
 558         hash_preload(vma->vm_mm, address, access, trap);
 559 #endif /* CONFIG_PPC_STD_MMU */
 560 }