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