arch/ppc64/mm/hugetlbpage.c

   1 /*
   2  * PPC64 (POWER4) Huge TLB Page Support for Kernel.
   3  *
   4  * Copyright (C) 2003 David Gibson, IBM Corporation.
   5  *
   6  * Based on the IA-32 version:
   7  * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
   8  */
   9
  10 #include <linux/init.h>
  11 #include <linux/fs.h>
  12 #include <linux/mm.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/pagemap.h>
  15 #include <linux/smp_lock.h>
  16 #include <linux/slab.h>
  17 #include <linux/err.h>
  18 #include <linux/sysctl.h>
  19 #include <asm/mman.h>
  20 #include <asm/pgalloc.h>
  21 #include <asm/tlb.h>
  22 #include <asm/tlbflush.h>
  23 #include <asm/mmu_context.h>
  24 #include <asm/machdep.h>
  25 #include <asm/cputable.h>
  26 #include <asm/tlb.h>
  27
  28 #include <linux/sysctl.h>
  29
  30 #define HUGEPGDIR_SHIFT         (HPAGE_SHIFT + PAGE_SHIFT - 3)
  31 #define HUGEPGDIR_SIZE          (1UL << HUGEPGDIR_SHIFT)
  32 #define HUGEPGDIR_MASK          (~(HUGEPGDIR_SIZE-1))
  33
  34 #define HUGEPTE_INDEX_SIZE      9
  35 #define HUGEPGD_INDEX_SIZE      10
  36
  37 #define PTRS_PER_HUGEPTE        (1 << HUGEPTE_INDEX_SIZE)
  38 #define PTRS_PER_HUGEPGD        (1 << HUGEPGD_INDEX_SIZE)
  39
  40 static inline int hugepgd_index(unsigned long addr)
  41 {
  42         return (addr & ~REGION_MASK) >> HUGEPGDIR_SHIFT;
  43 }
  44
  45 static pgd_t *hugepgd_offset(struct mm_struct *mm, unsigned long addr)
  46 {
  47         int index;
  48
  49         if (! mm->context.huge_pgdir)
  50                 return NULL;
  51
  52
  53         index = hugepgd_index(addr);
  54         BUG_ON(index >= PTRS_PER_HUGEPGD);
  55         return mm->context.huge_pgdir + index;
  56 }
  57
  58 static inline pte_t *hugepte_offset(pgd_t *dir, unsigned long addr)
  59 {
  60         int index;
  61
  62         if (pgd_none(*dir))
  63                 return NULL;
  64
  65         index = (addr >> HPAGE_SHIFT) % PTRS_PER_HUGEPTE;
  66         return (pte_t *)pgd_page(*dir) + index;
  67 }
  68
  69 static pgd_t *hugepgd_alloc(struct mm_struct *mm, unsigned long addr)
  70 {
  71         BUG_ON(! in_hugepage_area(mm->context, addr));
  72
  73         if (! mm->context.huge_pgdir) {
  74                 pgd_t *new;
  75                 spin_unlock(&mm->page_table_lock);
  76                 /* Don't use pgd_alloc(), because we want __GFP_REPEAT */
  77                 new = kmem_cache_alloc(zero_cache, GFP_KERNEL | __GFP_REPEAT);
  78                 BUG_ON(memcmp(new, empty_zero_page, PAGE_SIZE));
  79                 spin_lock(&mm->page_table_lock);
  80
  81                 /*
  82                  * Because we dropped the lock, we should re-check the
  83                  * entry, as somebody else could have populated it..
  84                  */
  85                 if (mm->context.huge_pgdir)
  86                         pgd_free(new);
  87                 else
  88                         mm->context.huge_pgdir = new;
  89         }
  90         return hugepgd_offset(mm, addr);
  91 }
  92
  93 static pte_t *hugepte_alloc(struct mm_struct *mm, pgd_t *dir,
  94                             unsigned long addr)
  95 {
  96         if (! pgd_present(*dir)) {
  97                 pte_t *new;
  98
  99                 spin_unlock(&mm->page_table_lock);
 100                 new = kmem_cache_alloc(zero_cache, GFP_KERNEL | __GFP_REPEAT);
 101                 BUG_ON(memcmp(new, empty_zero_page, PAGE_SIZE));
 102                 spin_lock(&mm->page_table_lock);
 103                 /*
 104                  * Because we dropped the lock, we should re-check the
 105                  * entry, as somebody else could have populated it..
 106                  */
 107                 if (pgd_present(*dir)) {
 108                         if (new)
 109                                 kmem_cache_free(zero_cache, new);
 110                 } else {
 111                         struct page *ptepage;
 112
 113                         if (! new)
 114                                 return NULL;
 115                         ptepage = virt_to_page(new);
 116                         ptepage->mapping = (void *) mm;
 117                         ptepage->index = addr & HUGEPGDIR_MASK;
 118                         pgd_populate(mm, dir, new);
 119                 }
 120         }
 121
 122         return hugepte_offset(dir, addr);
 123 }
 124
 125 static pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
 126 {
 127         pgd_t *pgd;
 128
 129         BUG_ON(! in_hugepage_area(mm->context, addr));
 130
 131         pgd = hugepgd_offset(mm, addr);
 132         if (! pgd)
 133                 return NULL;
 134
 135         return hugepte_offset(pgd, addr);
 136 }
 137
 138 static pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
 139 {
 140         pgd_t *pgd;
 141
 142         BUG_ON(! in_hugepage_area(mm->context, addr));
 143
 144         pgd = hugepgd_alloc(mm, addr);
 145         if (! pgd)
 146                 return NULL;
 147
 148         return hugepte_alloc(mm, pgd, addr);
 149 }
 150
 151 static void set_huge_pte(struct mm_struct *mm, struct vm_area_struct *vma,
 152                          struct page *page, pte_t *ptep, int write_access)
 153 {
 154         pte_t entry;
 155
 156         vx_rsspages_add(mm, HPAGE_SIZE / PAGE_SIZE);
 157         if (write_access) {
 158                 entry =
 159                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
 160         } else {
 161                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
 162         }
 163         entry = pte_mkyoung(entry);
 164         entry = pte_mkhuge(entry);
 165
 166         set_pte(ptep, entry);
 167 }
 168
 169 /*
 170  * This function checks for proper alignment of input addr and len parameters.
 171  */
 172 int is_aligned_hugepage_range(unsigned long addr, unsigned long len)
 173 {
 174         if (len & ~HPAGE_MASK)
 175                 return -EINVAL;
 176         if (addr & ~HPAGE_MASK)
 177                 return -EINVAL;
 178         if (! (within_hugepage_low_range(addr, len)
 179                || within_hugepage_high_range(addr, len)) )
 180                 return -EINVAL;
 181         return 0;
 182 }
 183
 184 static void flush_segments(void *parm)
 185 {
 186         u16 segs = (unsigned long) parm;
 187         unsigned long i;
 188
 189         asm volatile("isync" : : : "memory");
 190
 191         for (i = 0; i < 16; i++) {
 192                 if (! (segs & (1U << i)))
 193                         continue;
 194                 asm volatile("slbie %0" : : "r" (i << SID_SHIFT));
 195         }
 196
 197         asm volatile("isync" : : : "memory");
 198 }
 199
 200 static int prepare_low_seg_for_htlb(struct mm_struct *mm, unsigned long seg)
 201 {
 202         unsigned long start = seg << SID_SHIFT;
 203         unsigned long end = (seg+1) << SID_SHIFT;
 204         struct vm_area_struct *vma;
 205         unsigned long addr;
 206         struct mmu_gather *tlb;
 207
 208         BUG_ON(seg >= 16);
 209
 210         /* Check no VMAs are in the region */
 211         vma = find_vma(mm, start);
 212         if (vma && (vma->vm_start < end))
 213                 return -EBUSY;
 214
 215         /* Clean up any leftover PTE pages in the region */
 216         spin_lock(&mm->page_table_lock);
 217         tlb = tlb_gather_mmu(mm, 0);
 218         for (addr = start; addr < end; addr += PMD_SIZE) {
 219                 pgd_t *pgd = pgd_offset(mm, addr);
 220                 pmd_t *pmd;
 221                 struct page *page;
 222                 pte_t *pte;
 223                 int i;
 224
 225                 if (pgd_none(*pgd))
 226                         continue;
 227                 pmd = pmd_offset(pgd, addr);
 228                 if (!pmd || pmd_none(*pmd))
 229                         continue;
 230                 if (pmd_bad(*pmd)) {
 231                         pmd_ERROR(*pmd);
 232                         pmd_clear(pmd);
 233                         continue;
 234                 }
 235                 pte = (pte_t *)pmd_page_kernel(*pmd);
 236                 /* No VMAs, so there should be no PTEs, check just in case. */
 237                 for (i = 0; i < PTRS_PER_PTE; i++) {
 238                         BUG_ON(!pte_none(*pte));
 239                         pte++;
 240                 }
 241                 page = pmd_page(*pmd);
 242                 pmd_clear(pmd);
 243                 mm->nr_ptes--;
 244                 dec_page_state(nr_page_table_pages);
 245                 pte_free_tlb(tlb, page);
 246         }
 247         tlb_finish_mmu(tlb, start, end);
 248         spin_unlock(&mm->page_table_lock);
 249
 250         return 0;
 251 }
 252
 253 static int open_low_hpage_segs(struct mm_struct *mm, u16 newsegs)
 254 {
 255         unsigned long i;
 256
 257         newsegs &= ~(mm->context.htlb_segs);
 258         if (! newsegs)
 259                 return 0; /* The segments we want are already open */
 260
 261         for (i = 0; i < 16; i++)
 262                 if ((1 << i) & newsegs)
 263                         if (prepare_low_seg_for_htlb(mm, i) != 0)
 264                                 return -EBUSY;
 265
 266         mm->context.htlb_segs |= newsegs;
 267
 268         /* update the paca copy of the context struct */
 269         get_paca()->context = mm->context;
 270
 271         /* the context change must make it to memory before the flush,
 272          * so that further SLB misses do the right thing. */
 273         mb();
 274         on_each_cpu(flush_segments, (void *)(unsigned long)newsegs, 0, 1);
 275
 276         return 0;
 277 }
 278
 279 int prepare_hugepage_range(unsigned long addr, unsigned long len)
 280 {
 281         if (within_hugepage_high_range(addr, len))
 282                 return 0;
 283         else if ((addr < 0x100000000UL) && ((addr+len) < 0x100000000UL)) {
 284                 int err;
 285                 /* Yes, we need both tests, in case addr+len overflows
 286                  * 64-bit arithmetic */
 287                 err = open_low_hpage_segs(current->mm,
 288                                           LOW_ESID_MASK(addr, len));
 289                 if (err)
 290                         printk(KERN_DEBUG "prepare_hugepage_range(%lx, %lx)"
 291                                " failed (segs: 0x%04hx)\n", addr, len,
 292                                LOW_ESID_MASK(addr, len));
 293                 return err;
 294         }
 295
 296         return -EINVAL;
 297 }
 298
 299 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 300                         struct vm_area_struct *vma)
 301 {
 302         pte_t *src_pte, *dst_pte, entry;
 303         struct page *ptepage;
 304         unsigned long addr = vma->vm_start;
 305         unsigned long end = vma->vm_end;
 306         int err = -ENOMEM;
 307
 308         while (addr < end) {
 309                 dst_pte = huge_pte_alloc(dst, addr);
 310                 if (!dst_pte)
 311                         goto out;
 312
 313                 src_pte = huge_pte_offset(src, addr);
 314                 entry = *src_pte;
 315
 316                 ptepage = pte_page(entry);
 317                 get_page(ptepage);
 318                 vx_rsspages_add(dst, HPAGE_SIZE / PAGE_SIZE);
 319                 set_pte(dst_pte, entry);
 320
 321                 addr += HPAGE_SIZE;
 322         }
 323
 324         err = 0;
 325  out:
 326         return err;
 327 }
 328
 329 int
 330 follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
 331                     struct page **pages, struct vm_area_struct **vmas,
 332                     unsigned long *position, int *length, int i)
 333 {
 334         unsigned long vpfn, vaddr = *position;
 335         int remainder = *length;
 336
 337         WARN_ON(!is_vm_hugetlb_page(vma));
 338
 339         vpfn = vaddr/PAGE_SIZE;
 340         while (vaddr < vma->vm_end && remainder) {
 341                 if (pages) {
 342                         pte_t *pte;
 343                         struct page *page;
 344
 345                         pte = huge_pte_offset(mm, vaddr);
 346
 347                         /* hugetlb should be locked, and hence, prefaulted */
 348                         WARN_ON(!pte || pte_none(*pte));
 349
 350                         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
 351
 352                         WARN_ON(!PageCompound(page));
 353
 354                         get_page(page);
 355                         pages[i] = page;
 356                 }
 357
 358                 if (vmas)
 359                         vmas[i] = vma;
 360
 361                 vaddr += PAGE_SIZE;
 362                 ++vpfn;
 363                 --remainder;
 364                 ++i;
 365         }
 366
 367         *length = remainder;
 368         *position = vaddr;
 369
 370         return i;
 371 }
 372
 373 struct page *
 374 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 375 {
 376         pte_t *ptep;
 377         struct page *page;
 378
 379         if (! in_hugepage_area(mm->context, address))
 380                 return ERR_PTR(-EINVAL);
 381
 382         ptep = huge_pte_offset(mm, address);
 383         page = pte_page(*ptep);
 384         if (page)
 385                 page += (address % HPAGE_SIZE) / PAGE_SIZE;
 386
 387         return page;
 388 }
 389
 390 int pmd_huge(pmd_t pmd)
 391 {
 392         return 0;
 393 }
 394
 395 struct page *
 396 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 397                 pmd_t *pmd, int write)
 398 {
 399         BUG();
 400         return NULL;
 401 }
 402
 403 void unmap_hugepage_range(struct vm_area_struct *vma,
 404                           unsigned long start, unsigned long end)
 405 {
 406         struct mm_struct *mm = vma->vm_mm;
 407         unsigned long addr;
 408         pte_t *ptep;
 409         struct page *page;
 410
 411         WARN_ON(!is_vm_hugetlb_page(vma));
 412         BUG_ON((start % HPAGE_SIZE) != 0);
 413         BUG_ON((end % HPAGE_SIZE) != 0);
 414
 415         for (addr = start; addr < end; addr += HPAGE_SIZE) {
 416                 pte_t pte;
 417
 418                 ptep = huge_pte_offset(mm, addr);
 419                 if (!ptep || pte_none(*ptep))
 420                         continue;
 421
 422                 pte = *ptep;
 423                 page = pte_page(pte);
 424                 pte_clear(ptep);
 425
 426                 put_page(page);
 427         }
 428         vx_rsspages_sub(mm, (end - start) >> PAGE_SHIFT);
 429         flush_tlb_pending();
 430 }
 431
 432 void hugetlb_free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *prev,
 433                            unsigned long start, unsigned long end)
 434 {
 435         /* Because the huge pgtables are only 2 level, they can take
 436          * at most around 4M, much less than one hugepage which the
 437          * process is presumably entitled to use.  So we don't bother
 438          * freeing up the pagetables on unmap, and wait until
 439          * destroy_context() to clean up the lot. */
 440 }
 441
 442 int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma)
 443 {
 444         struct mm_struct *mm = current->mm;
 445         unsigned long addr;
 446         int ret = 0;
 447
 448         WARN_ON(!is_vm_hugetlb_page(vma));
 449         BUG_ON((vma->vm_start % HPAGE_SIZE) != 0);
 450         BUG_ON((vma->vm_end % HPAGE_SIZE) != 0);
 451
 452         spin_lock(&mm->page_table_lock);
 453         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 454                 unsigned long idx;
 455                 pte_t *pte = huge_pte_alloc(mm, addr);
 456                 struct page *page;
 457
 458                 if (!pte) {
 459                         ret = -ENOMEM;
 460                         goto out;
 461                 }
 462                 if (! pte_none(*pte))
 463                         continue;
 464
 465                 idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
 466                         + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 467                 page = find_get_page(mapping, idx);
 468                 if (!page) {
 469                         /* charge the fs quota first */
 470                         if (hugetlb_get_quota(mapping)) {
 471                                 ret = -ENOMEM;
 472                                 goto out;
 473                         }
 474                         page = alloc_huge_page();
 475                         if (!page) {
 476                                 hugetlb_put_quota(mapping);
 477                                 ret = -ENOMEM;
 478                                 goto out;
 479                         }
 480                         ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC);
 481                         if (! ret) {
 482                                 unlock_page(page);
 483                         } else {
 484                                 hugetlb_put_quota(mapping);
 485                                 free_huge_page(page);
 486                                 goto out;
 487                         }
 488                 }
 489                 set_huge_pte(mm, vma, page, pte, vma->vm_flags & VM_WRITE);
 490         }
 491 out:
 492         spin_unlock(&mm->page_table_lock);
 493         return ret;
 494 }
 495
 496 /* Because we have an exclusive hugepage region which lies within the
 497  * normal user address space, we have to take special measures to make
 498  * non-huge mmap()s evade the hugepage reserved regions. */
 499 unsigned long arch_get_unmapped_area(struct file *filp, unsigned long addr,
 500                                      unsigned long len, unsigned long pgoff,
 501                                      unsigned long flags)
 502 {
 503         struct mm_struct *mm = current->mm;
 504         struct vm_area_struct *vma;
 505         unsigned long start_addr;
 506
 507         if (len > TASK_SIZE)
 508                 return -ENOMEM;
 509
 510         if (addr) {
 511                 addr = PAGE_ALIGN(addr);
 512                 vma = find_vma(mm, addr);
 513                 if (((TASK_SIZE - len) >= addr)
 514                     && (!vma || (addr+len) <= vma->vm_start)
 515                     && !is_hugepage_only_range(addr,len))
 516                         return addr;
 517         }
 518         start_addr = addr = mm->free_area_cache;
 519
 520 full_search:
 521         vma = find_vma(mm, addr);
 522         while (TASK_SIZE - len >= addr) {
 523                 BUG_ON(vma && (addr >= vma->vm_end));
 524
 525                 if (touches_hugepage_low_range(addr, len)) {
 526                         addr = ALIGN(addr+1, 1<<SID_SHIFT);
 527                         vma = find_vma(mm, addr);
 528                         continue;
 529                 }
 530                 if (touches_hugepage_high_range(addr, len)) {
 531                         addr = TASK_HPAGE_END;
 532                         vma = find_vma(mm, addr);
 533                         continue;
 534                 }
 535                 if (!vma || addr + len <= vma->vm_start) {
 536                         /*
 537                          * Remember the place where we stopped the search:
 538                          */
 539                         mm->free_area_cache = addr + len;
 540                         return addr;
 541                 }
 542                 addr = vma->vm_end;
 543                 vma = vma->vm_next;
 544         }
 545
 546         /* Make sure we didn't miss any holes */
 547         if (start_addr != TASK_UNMAPPED_BASE) {
 548                 start_addr = addr = TASK_UNMAPPED_BASE;
 549                 goto full_search;
 550         }
 551         return -ENOMEM;
 552 }
 553
 554 /*
 555  * This mmap-allocator allocates new areas top-down from below the
 556  * stack's low limit (the base):
 557  *
 558  * Because we have an exclusive hugepage region which lies within the
 559  * normal user address space, we have to take special measures to make
 560  * non-huge mmap()s evade the hugepage reserved regions.
 561  */
 562 unsigned long
 563 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
 564                           const unsigned long len, const unsigned long pgoff,
 565                           const unsigned long flags)
 566 {
 567         struct vm_area_struct *vma, *prev_vma;
 568         struct mm_struct *mm = current->mm;
 569         unsigned long base = mm->mmap_base, addr = addr0;
 570         int first_time = 1;
 571
 572         /* requested length too big for entire address space */
 573         if (len > TASK_SIZE)
 574                 return -ENOMEM;
 575
 576         /* dont allow allocations above current base */
 577         if (mm->free_area_cache > base)
 578                 mm->free_area_cache = base;
 579
 580         /* requesting a specific address */
 581         if (addr) {
 582                 addr = PAGE_ALIGN(addr);
 583                 vma = find_vma(mm, addr);
 584                 if (TASK_SIZE - len >= addr &&
 585                                 (!vma || addr + len <= vma->vm_start)
 586                                 && !is_hugepage_only_range(addr,len))
 587                         return addr;
 588         }
 589
 590 try_again:
 591         /* make sure it can fit in the remaining address space */
 592         if (mm->free_area_cache < len)
 593                 goto fail;
 594
 595         /* either no address requested or cant fit in requested address hole */
 596         addr = (mm->free_area_cache - len) & PAGE_MASK;
 597         do {
 598 hugepage_recheck:
 599                 if (touches_hugepage_low_range(addr, len)) {
 600                         addr = (addr & ((~0) << SID_SHIFT)) - len;
 601                         goto hugepage_recheck;
 602                 } else if (touches_hugepage_high_range(addr, len)) {
 603                         addr = TASK_HPAGE_BASE - len;
 604                 }
 605
 606                 /*
 607                  * Lookup failure means no vma is above this address,
 608                  * i.e. return with success:
 609                  */
 610                 if (!(vma = find_vma_prev(mm, addr, &prev_vma)))
 611                         return addr;
 612
 613                 /*
 614                  * new region fits between prev_vma->vm_end and
 615                  * vma->vm_start, use it:
 616                  */
 617                 if (addr+len <= vma->vm_start &&
 618                                 (!prev_vma || (addr >= prev_vma->vm_end)))
 619                         /* remember the address as a hint for next time */
 620                         return (mm->free_area_cache = addr);
 621                 else
 622                         /* pull free_area_cache down to the first hole */
 623                         if (mm->free_area_cache == vma->vm_end)
 624                                 mm->free_area_cache = vma->vm_start;
 625
 626                 /* try just below the current vma->vm_start */
 627                 addr = vma->vm_start-len;
 628         } while (len <= vma->vm_start);
 629
 630 fail:
 631         /*
 632          * if hint left us with no space for the requested
 633          * mapping then try again:
 634          */
 635         if (first_time) {
 636                 mm->free_area_cache = base;
 637                 first_time = 0;
 638                 goto try_again;
 639         }
 640         /*
 641          * A failed mmap() very likely causes application failure,
 642          * so fall back to the bottom-up function here. This scenario
 643          * can happen with large stack limits and large mmap()
 644          * allocations.
 645          */
 646         mm->free_area_cache = TASK_UNMAPPED_BASE;
 647         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
 648         /*
 649          * Restore the topdown base:
 650          */
 651         mm->free_area_cache = base;
 652
 653         return addr;
 654 }
 655
 656 static unsigned long htlb_get_low_area(unsigned long len, u16 segmask)
 657 {
 658         unsigned long addr = 0;
 659         struct vm_area_struct *vma;
 660
 661         vma = find_vma(current->mm, addr);
 662         while (addr + len <= 0x100000000UL) {
 663                 BUG_ON(vma && (addr >= vma->vm_end)); /* invariant */
 664
 665                 if (! __within_hugepage_low_range(addr, len, segmask)) {
 666                         addr = ALIGN(addr+1, 1<<SID_SHIFT);
 667                         vma = find_vma(current->mm, addr);
 668                         continue;
 669                 }
 670
 671                 if (!vma || (addr + len) <= vma->vm_start)
 672                         return addr;
 673                 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
 674                 /* Depending on segmask this might not be a confirmed
 675                  * hugepage region, so the ALIGN could have skipped
 676                  * some VMAs */
 677                 vma = find_vma(current->mm, addr);
 678         }
 679
 680         return -ENOMEM;
 681 }
 682
 683 static unsigned long htlb_get_high_area(unsigned long len)
 684 {
 685         unsigned long addr = TASK_HPAGE_BASE;
 686         struct vm_area_struct *vma;
 687
 688         vma = find_vma(current->mm, addr);
 689         for (vma = find_vma(current->mm, addr);
 690              addr + len <= TASK_HPAGE_END;
 691              vma = vma->vm_next) {
 692                 BUG_ON(vma && (addr >= vma->vm_end)); /* invariant */
 693                 BUG_ON(! within_hugepage_high_range(addr, len));
 694
 695                 if (!vma || (addr + len) <= vma->vm_start)
 696                         return addr;
 697                 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
 698                 /* Because we're in a hugepage region, this alignment
 699                  * should not skip us over any VMAs */
 700         }
 701
 702         return -ENOMEM;
 703 }
 704
 705 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 706                                         unsigned long len, unsigned long pgoff,
 707                                         unsigned long flags)
 708 {
 709         if (len & ~HPAGE_MASK)
 710                 return -EINVAL;
 711
 712         if (!(cur_cpu_spec->cpu_features & CPU_FTR_16M_PAGE))
 713                 return -EINVAL;
 714
 715         if (test_thread_flag(TIF_32BIT)) {
 716                 int lastshift = 0;
 717                 u16 segmask, cursegs = current->mm->context.htlb_segs;
 718
 719                 /* First see if we can do the mapping in the existing
 720                  * low hpage segments */
 721                 addr = htlb_get_low_area(len, cursegs);
 722                 if (addr != -ENOMEM)
 723                         return addr;
 724
 725                 for (segmask = LOW_ESID_MASK(0x100000000UL-len, len);
 726                      ! lastshift; segmask >>=1) {
 727                         if (segmask & 1)
 728                                 lastshift = 1;
 729
 730                         addr = htlb_get_low_area(len, cursegs | segmask);
 731                         if ((addr != -ENOMEM)
 732                             && open_low_hpage_segs(current->mm, segmask) == 0)
 733                                 return addr;
 734                 }
 735                 printk(KERN_DEBUG "hugetlb_get_unmapped_area() unable to open"
 736                        " enough segments\n");
 737                 return -ENOMEM;
 738         } else {
 739                 return htlb_get_high_area(len);
 740         }
 741 }
 742
 743 void hugetlb_mm_free_pgd(struct mm_struct *mm)
 744 {
 745         int i;
 746         pgd_t *pgdir;
 747
 748         spin_lock(&mm->page_table_lock);
 749
 750         pgdir = mm->context.huge_pgdir;
 751         if (! pgdir)
 752                 goto out;
 753
 754         mm->context.huge_pgdir = NULL;
 755
 756         /* cleanup any hugepte pages leftover */
 757         for (i = 0; i < PTRS_PER_HUGEPGD; i++) {
 758                 pgd_t *pgd = pgdir + i;
 759
 760                 if (! pgd_none(*pgd)) {
 761                         pte_t *pte = (pte_t *)pgd_page(*pgd);
 762                         struct page *ptepage = virt_to_page(pte);
 763
 764                         ptepage->mapping = NULL;
 765
 766                         BUG_ON(memcmp(pte, empty_zero_page, PAGE_SIZE));
 767                         kmem_cache_free(zero_cache, pte);
 768                 }
 769                 pgd_clear(pgd);
 770         }
 771
 772         BUG_ON(memcmp(pgdir, empty_zero_page, PAGE_SIZE));
 773         kmem_cache_free(zero_cache, pgdir);
 774
 775  out:
 776         spin_unlock(&mm->page_table_lock);
 777 }
 778
 779 int hash_huge_page(struct mm_struct *mm, unsigned long access,
 780                    unsigned long ea, unsigned long vsid, int local)
 781 {
 782         pte_t *ptep;
 783         unsigned long va, vpn;
 784         int is_write;
 785         pte_t old_pte, new_pte;
 786         unsigned long hpteflags, prpn;
 787         long slot;
 788         int err = 1;
 789
 790         spin_lock(&mm->page_table_lock);
 791
 792         ptep = huge_pte_offset(mm, ea);
 793
 794         /* Search the Linux page table for a match with va */
 795         va = (vsid << 28) | (ea & 0x0fffffff);
 796         vpn = va >> HPAGE_SHIFT;
 797
 798         /*
 799          * If no pte found or not present, send the problem up to
 800          * do_page_fault
 801          */
 802         if (unlikely(!ptep || pte_none(*ptep)))
 803                 goto out;
 804
 805 /*      BUG_ON(pte_bad(*ptep)); */
 806
 807         /*
 808          * Check the user's access rights to the page.  If access should be
 809          * prevented then send the problem up to do_page_fault.
 810          */
 811         is_write = access & _PAGE_RW;
 812         if (unlikely(is_write && !(pte_val(*ptep) & _PAGE_RW)))
 813                 goto out;
 814         /*
 815          * At this point, we have a pte (old_pte) which can be used to build
 816          * or update an HPTE. There are 2 cases:
 817          *
 818          * 1. There is a valid (present) pte with no associated HPTE (this is
 819          *      the most common case)
 820          * 2. There is a valid (present) pte with an associated HPTE. The
 821          *      current values of the pp bits in the HPTE prevent access
 822          *      because we are doing software DIRTY bit management and the
 823          *      page is currently not DIRTY.
 824          */
 825
 826
 827         old_pte = *ptep;
 828         new_pte = old_pte;
 829
 830         hpteflags = 0x2 | (! (pte_val(new_pte) & _PAGE_RW));
 831
 832         /* Check if pte already has an hpte (case 2) */
 833         if (unlikely(pte_val(old_pte) & _PAGE_HASHPTE)) {
 834                 /* There MIGHT be an HPTE for this pte */
 835                 unsigned long hash, slot;
 836
 837                 hash = hpt_hash(vpn, 1);
 838                 if (pte_val(old_pte) & _PAGE_SECONDARY)
 839                         hash = ~hash;
 840                 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
 841                 slot += (pte_val(old_pte) & _PAGE_GROUP_IX) >> 12;
 842
 843                 if (ppc_md.hpte_updatepp(slot, hpteflags, va, 1, local) == -1)
 844                         pte_val(old_pte) &= ~_PAGE_HPTEFLAGS;
 845         }
 846
 847         if (likely(!(pte_val(old_pte) & _PAGE_HASHPTE))) {
 848                 unsigned long hash = hpt_hash(vpn, 1);
 849                 unsigned long hpte_group;
 850
 851                 prpn = pte_pfn(old_pte);
 852
 853 repeat:
 854                 hpte_group = ((hash & htab_hash_mask) *
 855                               HPTES_PER_GROUP) & ~0x7UL;
 856
 857                 /* Update the linux pte with the HPTE slot */
 858                 pte_val(new_pte) &= ~_PAGE_HPTEFLAGS;
 859                 pte_val(new_pte) |= _PAGE_HASHPTE;
 860
 861                 /* Add in WIMG bits */
 862                 /* XXX We should store these in the pte */
 863                 hpteflags |= _PAGE_COHERENT;
 864
 865                 slot = ppc_md.hpte_insert(hpte_group, va, prpn, 0,
 866                                           hpteflags, 0, 1);
 867
 868                 /* Primary is full, try the secondary */
 869                 if (unlikely(slot == -1)) {
 870                         pte_val(new_pte) |= _PAGE_SECONDARY;
 871                         hpte_group = ((~hash & htab_hash_mask) *
 872                                       HPTES_PER_GROUP) & ~0x7UL;
 873                         slot = ppc_md.hpte_insert(hpte_group, va, prpn,
 874                                                   1, hpteflags, 0, 1);
 875                         if (slot == -1) {
 876                                 if (mftb() & 0x1)
 877                                         hpte_group = ((hash & htab_hash_mask) * HPTES_PER_GROUP) & ~0x7UL;
 878
 879                                 ppc_md.hpte_remove(hpte_group);
 880                                 goto repeat;
 881                         }
 882                 }
 883
 884                 if (unlikely(slot == -2))
 885                         panic("hash_huge_page: pte_insert failed\n");
 886
 887                 pte_val(new_pte) |= (slot<<12) & _PAGE_GROUP_IX;
 888
 889                 /*
 890                  * No need to use ldarx/stdcx here because all who
 891                  * might be updating the pte will hold the
 892                  * page_table_lock
 893                  */
 894                 *ptep = new_pte;
 895         }
 896
 897         err = 0;
 898
 899  out:
 900         spin_unlock(&mm->page_table_lock);
 901
 902         return err;
 903 }