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 pud_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 (pud_t *)(mm->context.huge_pgdir + index);
  56 }
  57
  58 static inline pte_t *hugepte_offset(pud_t *dir, unsigned long addr)
  59 {
  60         int index;
  61
  62         if (pud_none(*dir))
  63                 return NULL;
  64
  65         index = (addr >> HPAGE_SHIFT) % PTRS_PER_HUGEPTE;
  66         return (pte_t *)pud_page(*dir) + index;
  67 }
  68
  69 static pud_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, pud_t *dir, unsigned long addr)
  94 {
  95         if (! pud_present(*dir)) {
  96                 pte_t *new;
  97
  98                 spin_unlock(&mm->page_table_lock);
  99                 new = kmem_cache_alloc(zero_cache, GFP_KERNEL | __GFP_REPEAT);
 100                 BUG_ON(memcmp(new, empty_zero_page, PAGE_SIZE));
 101                 spin_lock(&mm->page_table_lock);
 102                 /*
 103                  * Because we dropped the lock, we should re-check the
 104                  * entry, as somebody else could have populated it..
 105                  */
 106                 if (pud_present(*dir)) {
 107                         if (new)
 108                                 kmem_cache_free(zero_cache, new);
 109                 } else {
 110                         struct page *ptepage;
 111
 112                         if (! new)
 113                                 return NULL;
 114                         ptepage = virt_to_page(new);
 115                         ptepage->mapping = (void *) mm;
 116                         ptepage->index = addr & HUGEPGDIR_MASK;
 117                         pud_populate(mm, dir, new);
 118                 }
 119         }
 120
 121         return hugepte_offset(dir, addr);
 122 }
 123
 124 static pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
 125 {
 126         pud_t *pud;
 127
 128         BUG_ON(! in_hugepage_area(mm->context, addr));
 129
 130         pud = hugepgd_offset(mm, addr);
 131         if (! pud)
 132                 return NULL;
 133
 134         return hugepte_offset(pud, addr);
 135 }
 136
 137 static pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
 138 {
 139         pud_t *pud;
 140
 141         BUG_ON(! in_hugepage_area(mm->context, addr));
 142
 143         pud = hugepgd_alloc(mm, addr);
 144         if (! pud)
 145                 return NULL;
 146
 147         return hugepte_alloc(mm, pud, addr);
 148 }
 149
 150 static void set_huge_pte(struct mm_struct *mm, struct vm_area_struct *vma,
 151                          unsigned long addr, struct page *page,
 152                          pte_t *ptep, int write_access)
 153 {
 154         pte_t entry;
 155
 156         add_mm_counter(mm, rss, 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_at(mm, addr, 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
 206         BUG_ON(seg >= 16);
 207
 208         /* Check no VMAs are in the region */
 209         vma = find_vma(mm, start);
 210         if (vma && (vma->vm_start < end))
 211                 return -EBUSY;
 212
 213         return 0;
 214 }
 215
 216 static int open_low_hpage_segs(struct mm_struct *mm, u16 newsegs)
 217 {
 218         unsigned long i;
 219
 220         newsegs &= ~(mm->context.htlb_segs);
 221         if (! newsegs)
 222                 return 0; /* The segments we want are already open */
 223
 224         for (i = 0; i < 16; i++)
 225                 if ((1 << i) & newsegs)
 226                         if (prepare_low_seg_for_htlb(mm, i) != 0)
 227                                 return -EBUSY;
 228
 229         mm->context.htlb_segs |= newsegs;
 230
 231         /* update the paca copy of the context struct */
 232         get_paca()->context = mm->context;
 233
 234         /* the context change must make it to memory before the flush,
 235          * so that further SLB misses do the right thing. */
 236         mb();
 237         on_each_cpu(flush_segments, (void *)(unsigned long)newsegs, 0, 1);
 238
 239         return 0;
 240 }
 241
 242 int prepare_hugepage_range(unsigned long addr, unsigned long len)
 243 {
 244         if (within_hugepage_high_range(addr, len))
 245                 return 0;
 246         else if ((addr < 0x100000000UL) && ((addr+len) < 0x100000000UL)) {
 247                 int err;
 248                 /* Yes, we need both tests, in case addr+len overflows
 249                  * 64-bit arithmetic */
 250                 err = open_low_hpage_segs(current->mm,
 251                                           LOW_ESID_MASK(addr, len));
 252                 if (err)
 253                         printk(KERN_DEBUG "prepare_hugepage_range(%lx, %lx)"
 254                                " failed (segs: 0x%04hx)\n", addr, len,
 255                                LOW_ESID_MASK(addr, len));
 256                 return err;
 257         }
 258
 259         return -EINVAL;
 260 }
 261
 262 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 263                         struct vm_area_struct *vma)
 264 {
 265         pte_t *src_pte, *dst_pte, entry;
 266         struct page *ptepage;
 267         unsigned long addr = vma->vm_start;
 268         unsigned long end = vma->vm_end;
 269         int err = -ENOMEM;
 270
 271         while (addr < end) {
 272                 dst_pte = huge_pte_alloc(dst, addr);
 273                 if (!dst_pte)
 274                         goto out;
 275
 276                 src_pte = huge_pte_offset(src, addr);
 277                 entry = *src_pte;
 278
 279                 ptepage = pte_page(entry);
 280                 get_page(ptepage);
 281                 add_mm_counter(dst, rss, HPAGE_SIZE / PAGE_SIZE);
 282                 set_pte_at(dst, addr, dst_pte, entry);
 283
 284                 addr += HPAGE_SIZE;
 285         }
 286
 287         err = 0;
 288  out:
 289         return err;
 290 }
 291
 292 int
 293 follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
 294                     struct page **pages, struct vm_area_struct **vmas,
 295                     unsigned long *position, int *length, int i)
 296 {
 297         unsigned long vpfn, vaddr = *position;
 298         int remainder = *length;
 299
 300         WARN_ON(!is_vm_hugetlb_page(vma));
 301
 302         vpfn = vaddr/PAGE_SIZE;
 303         while (vaddr < vma->vm_end && remainder) {
 304                 if (pages) {
 305                         pte_t *pte;
 306                         struct page *page;
 307
 308                         pte = huge_pte_offset(mm, vaddr);
 309
 310                         /* hugetlb should be locked, and hence, prefaulted */
 311                         WARN_ON(!pte || pte_none(*pte));
 312
 313                         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
 314
 315                         WARN_ON(!PageCompound(page));
 316
 317                         get_page(page);
 318                         pages[i] = page;
 319                 }
 320
 321                 if (vmas)
 322                         vmas[i] = vma;
 323
 324                 vaddr += PAGE_SIZE;
 325                 ++vpfn;
 326                 --remainder;
 327                 ++i;
 328         }
 329
 330         *length = remainder;
 331         *position = vaddr;
 332
 333         return i;
 334 }
 335
 336 struct page *
 337 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 338 {
 339         pte_t *ptep;
 340         struct page *page;
 341
 342         if (! in_hugepage_area(mm->context, address))
 343                 return ERR_PTR(-EINVAL);
 344
 345         ptep = huge_pte_offset(mm, address);
 346         page = pte_page(*ptep);
 347         if (page)
 348                 page += (address % HPAGE_SIZE) / PAGE_SIZE;
 349
 350         return page;
 351 }
 352
 353 int pmd_huge(pmd_t pmd)
 354 {
 355         return 0;
 356 }
 357
 358 struct page *
 359 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 360                 pmd_t *pmd, int write)
 361 {
 362         BUG();
 363         return NULL;
 364 }
 365
 366 void unmap_hugepage_range(struct vm_area_struct *vma,
 367                           unsigned long start, unsigned long end)
 368 {
 369         struct mm_struct *mm = vma->vm_mm;
 370         unsigned long addr;
 371         pte_t *ptep;
 372         struct page *page;
 373
 374         WARN_ON(!is_vm_hugetlb_page(vma));
 375         BUG_ON((start % HPAGE_SIZE) != 0);
 376         BUG_ON((end % HPAGE_SIZE) != 0);
 377
 378         for (addr = start; addr < end; addr += HPAGE_SIZE) {
 379                 pte_t pte;
 380
 381                 ptep = huge_pte_offset(mm, addr);
 382                 if (!ptep || pte_none(*ptep))
 383                         continue;
 384
 385                 pte = *ptep;
 386                 page = pte_page(pte);
 387                 pte_clear(mm, addr, ptep);
 388
 389                 put_page(page);
 390         }
 391         add_mm_counter(mm, rss, -((end - start) >> PAGE_SHIFT));
 392         flush_tlb_pending();
 393 }
 394
 395 int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma)
 396 {
 397         struct mm_struct *mm = current->mm;
 398         unsigned long addr;
 399         int ret = 0;
 400
 401         WARN_ON(!is_vm_hugetlb_page(vma));
 402         BUG_ON((vma->vm_start % HPAGE_SIZE) != 0);
 403         BUG_ON((vma->vm_end % HPAGE_SIZE) != 0);
 404
 405         spin_lock(&mm->page_table_lock);
 406         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 407                 unsigned long idx;
 408                 pte_t *pte = huge_pte_alloc(mm, addr);
 409                 struct page *page;
 410
 411                 if (!pte) {
 412                         ret = -ENOMEM;
 413                         goto out;
 414                 }
 415                 if (! pte_none(*pte))
 416                         continue;
 417
 418                 idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
 419                         + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 420                 page = find_get_page(mapping, idx);
 421                 if (!page) {
 422                         /* charge the fs quota first */
 423                         if (hugetlb_get_quota(mapping)) {
 424                                 ret = -ENOMEM;
 425                                 goto out;
 426                         }
 427                         page = alloc_huge_page();
 428                         if (!page) {
 429                                 hugetlb_put_quota(mapping);
 430                                 ret = -ENOMEM;
 431                                 goto out;
 432                         }
 433                         ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC);
 434                         if (! ret) {
 435                                 unlock_page(page);
 436                         } else {
 437                                 hugetlb_put_quota(mapping);
 438                                 free_huge_page(page);
 439                                 goto out;
 440                         }
 441                 }
 442                 set_huge_pte(mm, vma, addr, page, pte, vma->vm_flags & VM_WRITE);
 443         }
 444 out:
 445         spin_unlock(&mm->page_table_lock);
 446         return ret;
 447 }
 448
 449 /* Because we have an exclusive hugepage region which lies within the
 450  * normal user address space, we have to take special measures to make
 451  * non-huge mmap()s evade the hugepage reserved regions. */
 452 unsigned long arch_get_unmapped_area(struct file *filp, unsigned long addr,
 453                                      unsigned long len, unsigned long pgoff,
 454                                      unsigned long flags)
 455 {
 456         struct mm_struct *mm = current->mm;
 457         struct vm_area_struct *vma;
 458         unsigned long start_addr;
 459
 460         if (len > TASK_SIZE)
 461                 return -ENOMEM;
 462
 463         if (addr) {
 464                 addr = PAGE_ALIGN(addr);
 465                 vma = find_vma(mm, addr);
 466                 if (((TASK_SIZE - len) >= addr)
 467                     && (!vma || (addr+len) <= vma->vm_start)
 468                     && !is_hugepage_only_range(mm, addr,len))
 469                         return addr;
 470         }
 471         start_addr = addr = mm->free_area_cache;
 472
 473 full_search:
 474         vma = find_vma(mm, addr);
 475         while (TASK_SIZE - len >= addr) {
 476                 BUG_ON(vma && (addr >= vma->vm_end));
 477
 478                 if (touches_hugepage_low_range(mm, addr, len)) {
 479                         addr = ALIGN(addr+1, 1<<SID_SHIFT);
 480                         vma = find_vma(mm, addr);
 481                         continue;
 482                 }
 483                 if (touches_hugepage_high_range(addr, len)) {
 484                         addr = TASK_HPAGE_END;
 485                         vma = find_vma(mm, addr);
 486                         continue;
 487                 }
 488                 if (!vma || addr + len <= vma->vm_start) {
 489                         /*
 490                          * Remember the place where we stopped the search:
 491                          */
 492                         mm->free_area_cache = addr + len;
 493                         return addr;
 494                 }
 495                 addr = vma->vm_end;
 496                 vma = vma->vm_next;
 497         }
 498
 499         /* Make sure we didn't miss any holes */
 500         if (start_addr != TASK_UNMAPPED_BASE) {
 501                 start_addr = addr = TASK_UNMAPPED_BASE;
 502                 goto full_search;
 503         }
 504         return -ENOMEM;
 505 }
 506
 507 /*
 508  * This mmap-allocator allocates new areas top-down from below the
 509  * stack's low limit (the base):
 510  *
 511  * Because we have an exclusive hugepage region which lies within the
 512  * normal user address space, we have to take special measures to make
 513  * non-huge mmap()s evade the hugepage reserved regions.
 514  */
 515 unsigned long
 516 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
 517                           const unsigned long len, const unsigned long pgoff,
 518                           const unsigned long flags)
 519 {
 520         struct vm_area_struct *vma, *prev_vma;
 521         struct mm_struct *mm = current->mm;
 522         unsigned long base = mm->mmap_base, addr = addr0;
 523         int first_time = 1;
 524
 525         /* requested length too big for entire address space */
 526         if (len > TASK_SIZE)
 527                 return -ENOMEM;
 528
 529         /* dont allow allocations above current base */
 530         if (mm->free_area_cache > base)
 531                 mm->free_area_cache = base;
 532
 533         /* requesting a specific address */
 534         if (addr) {
 535                 addr = PAGE_ALIGN(addr);
 536                 vma = find_vma(mm, addr);
 537                 if (TASK_SIZE - len >= addr &&
 538                                 (!vma || addr + len <= vma->vm_start)
 539                                 && !is_hugepage_only_range(mm, addr,len))
 540                         return addr;
 541         }
 542
 543 try_again:
 544         /* make sure it can fit in the remaining address space */
 545         if (mm->free_area_cache < len)
 546                 goto fail;
 547
 548         /* either no address requested or cant fit in requested address hole */
 549         addr = (mm->free_area_cache - len) & PAGE_MASK;
 550         do {
 551 hugepage_recheck:
 552                 if (touches_hugepage_low_range(mm, addr, len)) {
 553                         addr = (addr & ((~0) << SID_SHIFT)) - len;
 554                         goto hugepage_recheck;
 555                 } else if (touches_hugepage_high_range(addr, len)) {
 556                         addr = TASK_HPAGE_BASE - len;
 557                 }
 558
 559                 /*
 560                  * Lookup failure means no vma is above this address,
 561                  * i.e. return with success:
 562                  */
 563                 if (!(vma = find_vma_prev(mm, addr, &prev_vma)))
 564                         return addr;
 565
 566                 /*
 567                  * new region fits between prev_vma->vm_end and
 568                  * vma->vm_start, use it:
 569                  */
 570                 if (addr+len <= vma->vm_start &&
 571                                 (!prev_vma || (addr >= prev_vma->vm_end)))
 572                         /* remember the address as a hint for next time */
 573                         return (mm->free_area_cache = addr);
 574                 else
 575                         /* pull free_area_cache down to the first hole */
 576                         if (mm->free_area_cache == vma->vm_end)
 577                                 mm->free_area_cache = vma->vm_start;
 578
 579                 /* try just below the current vma->vm_start */
 580                 addr = vma->vm_start-len;
 581         } while (len <= vma->vm_start);
 582
 583 fail:
 584         /*
 585          * if hint left us with no space for the requested
 586          * mapping then try again:
 587          */
 588         if (first_time) {
 589                 mm->free_area_cache = base;
 590                 first_time = 0;
 591                 goto try_again;
 592         }
 593         /*
 594          * A failed mmap() very likely causes application failure,
 595          * so fall back to the bottom-up function here. This scenario
 596          * can happen with large stack limits and large mmap()
 597          * allocations.
 598          */
 599         mm->free_area_cache = TASK_UNMAPPED_BASE;
 600         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
 601         /*
 602          * Restore the topdown base:
 603          */
 604         mm->free_area_cache = base;
 605
 606         return addr;
 607 }
 608
 609 static unsigned long htlb_get_low_area(unsigned long len, u16 segmask)
 610 {
 611         unsigned long addr = 0;
 612         struct vm_area_struct *vma;
 613
 614         vma = find_vma(current->mm, addr);
 615         while (addr + len <= 0x100000000UL) {
 616                 BUG_ON(vma && (addr >= vma->vm_end)); /* invariant */
 617
 618                 if (! __within_hugepage_low_range(addr, len, segmask)) {
 619                         addr = ALIGN(addr+1, 1<<SID_SHIFT);
 620                         vma = find_vma(current->mm, addr);
 621                         continue;
 622                 }
 623
 624                 if (!vma || (addr + len) <= vma->vm_start)
 625                         return addr;
 626                 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
 627                 /* Depending on segmask this might not be a confirmed
 628                  * hugepage region, so the ALIGN could have skipped
 629                  * some VMAs */
 630                 vma = find_vma(current->mm, addr);
 631         }
 632
 633         return -ENOMEM;
 634 }
 635
 636 static unsigned long htlb_get_high_area(unsigned long len)
 637 {
 638         unsigned long addr = TASK_HPAGE_BASE;
 639         struct vm_area_struct *vma;
 640
 641         vma = find_vma(current->mm, addr);
 642         for (vma = find_vma(current->mm, addr);
 643              addr + len <= TASK_HPAGE_END;
 644              vma = vma->vm_next) {
 645                 BUG_ON(vma && (addr >= vma->vm_end)); /* invariant */
 646                 BUG_ON(! within_hugepage_high_range(addr, len));
 647
 648                 if (!vma || (addr + len) <= vma->vm_start)
 649                         return addr;
 650                 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
 651                 /* Because we're in a hugepage region, this alignment
 652                  * should not skip us over any VMAs */
 653         }
 654
 655         return -ENOMEM;
 656 }
 657
 658 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 659                                         unsigned long len, unsigned long pgoff,
 660                                         unsigned long flags)
 661 {
 662         if (len & ~HPAGE_MASK)
 663                 return -EINVAL;
 664
 665         if (!cpu_has_feature(CPU_FTR_16M_PAGE))
 666                 return -EINVAL;
 667
 668         if (test_thread_flag(TIF_32BIT)) {
 669                 int lastshift = 0;
 670                 u16 segmask, cursegs = current->mm->context.htlb_segs;
 671
 672                 /* First see if we can do the mapping in the existing
 673                  * low hpage segments */
 674                 addr = htlb_get_low_area(len, cursegs);
 675                 if (addr != -ENOMEM)
 676                         return addr;
 677
 678                 for (segmask = LOW_ESID_MASK(0x100000000UL-len, len);
 679                      ! lastshift; segmask >>=1) {
 680                         if (segmask & 1)
 681                                 lastshift = 1;
 682
 683                         addr = htlb_get_low_area(len, cursegs | segmask);
 684                         if ((addr != -ENOMEM)
 685                             && open_low_hpage_segs(current->mm, segmask) == 0)
 686                                 return addr;
 687                 }
 688                 printk(KERN_DEBUG "hugetlb_get_unmapped_area() unable to open"
 689                        " enough segments\n");
 690                 return -ENOMEM;
 691         } else {
 692                 return htlb_get_high_area(len);
 693         }
 694 }
 695
 696 void hugetlb_mm_free_pgd(struct mm_struct *mm)
 697 {
 698         int i;
 699         pgd_t *pgdir;
 700
 701         spin_lock(&mm->page_table_lock);
 702
 703         pgdir = mm->context.huge_pgdir;
 704         if (! pgdir)
 705                 goto out;
 706
 707         mm->context.huge_pgdir = NULL;
 708
 709         /* cleanup any hugepte pages leftover */
 710         for (i = 0; i < PTRS_PER_HUGEPGD; i++) {
 711                 pud_t *pud = (pud_t *)(pgdir + i);
 712
 713                 if (! pud_none(*pud)) {
 714                         pte_t *pte = (pte_t *)pud_page(*pud);
 715                         struct page *ptepage = virt_to_page(pte);
 716
 717                         ptepage->mapping = NULL;
 718
 719                         BUG_ON(memcmp(pte, empty_zero_page, PAGE_SIZE));
 720                         kmem_cache_free(zero_cache, pte);
 721                 }
 722                 pud_clear(pud);
 723         }
 724
 725         BUG_ON(memcmp(pgdir, empty_zero_page, PAGE_SIZE));
 726         kmem_cache_free(zero_cache, pgdir);
 727
 728  out:
 729         spin_unlock(&mm->page_table_lock);
 730 }
 731
 732 int hash_huge_page(struct mm_struct *mm, unsigned long access,
 733                    unsigned long ea, unsigned long vsid, int local)
 734 {
 735         pte_t *ptep;
 736         unsigned long va, vpn;
 737         pte_t old_pte, new_pte;
 738         unsigned long hpteflags, prpn;
 739         long slot;
 740         int err = 1;
 741
 742         spin_lock(&mm->page_table_lock);
 743
 744         ptep = huge_pte_offset(mm, ea);
 745
 746         /* Search the Linux page table for a match with va */
 747         va = (vsid << 28) | (ea & 0x0fffffff);
 748         vpn = va >> HPAGE_SHIFT;
 749
 750         /*
 751          * If no pte found or not present, send the problem up to
 752          * do_page_fault
 753          */
 754         if (unlikely(!ptep || pte_none(*ptep)))
 755                 goto out;
 756
 757 /*      BUG_ON(pte_bad(*ptep)); */
 758
 759         /*
 760          * Check the user's access rights to the page.  If access should be
 761          * prevented then send the problem up to do_page_fault.
 762          */
 763         if (unlikely(access & ~pte_val(*ptep)))
 764                 goto out;
 765         /*
 766          * At this point, we have a pte (old_pte) which can be used to build
 767          * or update an HPTE. There are 2 cases:
 768          *
 769          * 1. There is a valid (present) pte with no associated HPTE (this is
 770          *      the most common case)
 771          * 2. There is a valid (present) pte with an associated HPTE. The
 772          *      current values of the pp bits in the HPTE prevent access
 773          *      because we are doing software DIRTY bit management and the
 774          *      page is currently not DIRTY.
 775          */
 776
 777
 778         old_pte = *ptep;
 779         new_pte = old_pte;
 780
 781         hpteflags = 0x2 | (! (pte_val(new_pte) & _PAGE_RW));
 782         /* _PAGE_EXEC -> HW_NO_EXEC since it's inverted */
 783         hpteflags |= ((pte_val(new_pte) & _PAGE_EXEC) ? 0 : HW_NO_EXEC);
 784
 785         /* Check if pte already has an hpte (case 2) */
 786         if (unlikely(pte_val(old_pte) & _PAGE_HASHPTE)) {
 787                 /* There MIGHT be an HPTE for this pte */
 788                 unsigned long hash, slot;
 789
 790                 hash = hpt_hash(vpn, 1);
 791                 if (pte_val(old_pte) & _PAGE_SECONDARY)
 792                         hash = ~hash;
 793                 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
 794                 slot += (pte_val(old_pte) & _PAGE_GROUP_IX) >> 12;
 795
 796                 if (ppc_md.hpte_updatepp(slot, hpteflags, va, 1, local) == -1)
 797                         pte_val(old_pte) &= ~_PAGE_HPTEFLAGS;
 798         }
 799
 800         if (likely(!(pte_val(old_pte) & _PAGE_HASHPTE))) {
 801                 unsigned long hash = hpt_hash(vpn, 1);
 802                 unsigned long hpte_group;
 803
 804                 prpn = pte_pfn(old_pte);
 805
 806 repeat:
 807                 hpte_group = ((hash & htab_hash_mask) *
 808                               HPTES_PER_GROUP) & ~0x7UL;
 809
 810                 /* Update the linux pte with the HPTE slot */
 811                 pte_val(new_pte) &= ~_PAGE_HPTEFLAGS;
 812                 pte_val(new_pte) |= _PAGE_HASHPTE;
 813
 814                 /* Add in WIMG bits */
 815                 /* XXX We should store these in the pte */
 816                 hpteflags |= _PAGE_COHERENT;
 817
 818                 slot = ppc_md.hpte_insert(hpte_group, va, prpn, 0,
 819                                           hpteflags, 0, 1);
 820
 821                 /* Primary is full, try the secondary */
 822                 if (unlikely(slot == -1)) {
 823                         pte_val(new_pte) |= _PAGE_SECONDARY;
 824                         hpte_group = ((~hash & htab_hash_mask) *
 825                                       HPTES_PER_GROUP) & ~0x7UL;
 826                         slot = ppc_md.hpte_insert(hpte_group, va, prpn,
 827                                                   1, hpteflags, 0, 1);
 828                         if (slot == -1) {
 829                                 if (mftb() & 0x1)
 830                                         hpte_group = ((hash & htab_hash_mask) * HPTES_PER_GROUP) & ~0x7UL;
 831
 832                                 ppc_md.hpte_remove(hpte_group);
 833                                 goto repeat;
 834                         }
 835                 }
 836
 837                 if (unlikely(slot == -2))
 838                         panic("hash_huge_page: pte_insert failed\n");
 839
 840                 pte_val(new_pte) |= (slot<<12) & _PAGE_GROUP_IX;
 841
 842                 /*
 843                  * No need to use ldarx/stdcx here because all who
 844                  * might be updating the pte will hold the
 845                  * page_table_lock
 846                  */
 847                 *ptep = new_pte;
 848         }
 849
 850         err = 0;
 851
 852  out:
 853         spin_unlock(&mm->page_table_lock);
 854
 855         return err;
 856 }