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         // mm->rss += (HPAGE_SIZE / PAGE_SIZE);
 157         vx_rsspages_add(mm, HPAGE_SIZE / PAGE_SIZE);
 158         if (write_access) {
 159                 entry =
 160                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
 161         } else {
 162                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
 163         }
 164         entry = pte_mkyoung(entry);
 165         entry = pte_mkhuge(entry);
 166
 167         set_pte(ptep, entry);
 168 }
 169
 170 /*
 171  * This function checks for proper alignment of input addr and len parameters.
 172  */
 173 int is_aligned_hugepage_range(unsigned long addr, unsigned long len)
 174 {
 175         if (len & ~HPAGE_MASK)
 176                 return -EINVAL;
 177         if (addr & ~HPAGE_MASK)
 178                 return -EINVAL;
 179         if (! (within_hugepage_low_range(addr, len)
 180                || within_hugepage_high_range(addr, len)) )
 181                 return -EINVAL;
 182         return 0;
 183 }
 184
 185 static void flush_segments(void *parm)
 186 {
 187         u16 segs = (unsigned long) parm;
 188         unsigned long i;
 189
 190         asm volatile("isync" : : : "memory");
 191
 192         for (i = 0; i < 16; i++) {
 193                 if (! (segs & (1U << i)))
 194                         continue;
 195                 asm volatile("slbie %0" : : "r" (i << SID_SHIFT));
 196         }
 197
 198         asm volatile("isync" : : : "memory");
 199 }
 200
 201 static int prepare_low_seg_for_htlb(struct mm_struct *mm, unsigned long seg)
 202 {
 203         unsigned long start = seg << SID_SHIFT;
 204         unsigned long end = (seg+1) << SID_SHIFT;
 205         struct vm_area_struct *vma;
 206         unsigned long addr;
 207         struct mmu_gather *tlb;
 208
 209         BUG_ON(seg >= 16);
 210
 211         /* Check no VMAs are in the region */
 212         vma = find_vma(mm, start);
 213         if (vma && (vma->vm_start < end))
 214                 return -EBUSY;
 215
 216         /* Clean up any leftover PTE pages in the region */
 217         spin_lock(&mm->page_table_lock);
 218         tlb = tlb_gather_mmu(mm, 0);
 219         for (addr = start; addr < end; addr += PMD_SIZE) {
 220                 pgd_t *pgd = pgd_offset(mm, addr);
 221                 pmd_t *pmd;
 222                 struct page *page;
 223                 pte_t *pte;
 224                 int i;
 225
 226                 if (pgd_none(*pgd))
 227                         continue;
 228                 pmd = pmd_offset(pgd, addr);
 229                 if (!pmd || pmd_none(*pmd))
 230                         continue;
 231                 if (pmd_bad(*pmd)) {
 232                         pmd_ERROR(*pmd);
 233                         pmd_clear(pmd);
 234                         continue;
 235                 }
 236                 pte = (pte_t *)pmd_page_kernel(*pmd);
 237                 /* No VMAs, so there should be no PTEs, check just in case. */
 238                 for (i = 0; i < PTRS_PER_PTE; i++) {
 239                         BUG_ON(!pte_none(*pte));
 240                         pte++;
 241                 }
 242                 page = pmd_page(*pmd);
 243                 pmd_clear(pmd);
 244                 mm->nr_ptes--;
 245                 dec_page_state(nr_page_table_pages);
 246                 pte_free_tlb(tlb, page);
 247         }
 248         tlb_finish_mmu(tlb, start, end);
 249         spin_unlock(&mm->page_table_lock);
 250
 251         return 0;
 252 }
 253
 254 static int open_low_hpage_segs(struct mm_struct *mm, u16 newsegs)
 255 {
 256         unsigned long i;
 257
 258         newsegs &= ~(mm->context.htlb_segs);
 259         if (! newsegs)
 260                 return 0; /* The segments we want are already open */
 261
 262         for (i = 0; i < 16; i++)
 263                 if ((1 << i) & newsegs)
 264                         if (prepare_low_seg_for_htlb(mm, i) != 0)
 265                                 return -EBUSY;
 266
 267         mm->context.htlb_segs |= newsegs;
 268
 269         /* update the paca copy of the context struct */
 270         get_paca()->context = mm->context;
 271
 272         /* the context change must make it to memory before the flush,
 273          * so that further SLB misses do the right thing. */
 274         mb();
 275         on_each_cpu(flush_segments, (void *)(unsigned long)newsegs, 0, 1);
 276
 277         return 0;
 278 }
 279
 280 int prepare_hugepage_range(unsigned long addr, unsigned long len)
 281 {
 282         if (within_hugepage_high_range(addr, len))
 283                 return 0;
 284         else if ((addr < 0x100000000UL) && ((addr+len) < 0x100000000UL)) {
 285                 int err;
 286                 /* Yes, we need both tests, in case addr+len overflows
 287                  * 64-bit arithmetic */
 288                 err = open_low_hpage_segs(current->mm,
 289                                           LOW_ESID_MASK(addr, len));
 290                 if (err)
 291                         printk(KERN_DEBUG "prepare_hugepage_range(%lx, %lx)"
 292                                " failed (segs: 0x%04hx)\n", addr, len,
 293                                LOW_ESID_MASK(addr, len));
 294                 return err;
 295         }
 296
 297         return -EINVAL;
 298 }
 299
 300 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 301                         struct vm_area_struct *vma)
 302 {
 303         pte_t *src_pte, *dst_pte, entry;
 304         struct page *ptepage;
 305         unsigned long addr = vma->vm_start;
 306         unsigned long end = vma->vm_end;
 307         int err = -ENOMEM;
 308
 309         while (addr < end) {
 310                 dst_pte = huge_pte_alloc(dst, addr);
 311                 if (!dst_pte)
 312                         goto out;
 313
 314                 src_pte = huge_pte_offset(src, addr);
 315                 entry = *src_pte;
 316
 317                 ptepage = pte_page(entry);
 318                 get_page(ptepage);
 319                 // dst->rss += (HPAGE_SIZE / PAGE_SIZE);
 320                 vx_rsspages_add(dst, HPAGE_SIZE / PAGE_SIZE);
 321                 set_pte(dst_pte, entry);
 322
 323                 addr += HPAGE_SIZE;
 324         }
 325
 326         err = 0;
 327  out:
 328         return err;
 329 }
 330
 331 int
 332 follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
 333                     struct page **pages, struct vm_area_struct **vmas,
 334                     unsigned long *position, int *length, int i)
 335 {
 336         unsigned long vpfn, vaddr = *position;
 337         int remainder = *length;
 338
 339         WARN_ON(!is_vm_hugetlb_page(vma));
 340
 341         vpfn = vaddr/PAGE_SIZE;
 342         while (vaddr < vma->vm_end && remainder) {
 343                 if (pages) {
 344                         pte_t *pte;
 345                         struct page *page;
 346
 347                         pte = huge_pte_offset(mm, vaddr);
 348
 349                         /* hugetlb should be locked, and hence, prefaulted */
 350                         WARN_ON(!pte || pte_none(*pte));
 351
 352                         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
 353
 354                         WARN_ON(!PageCompound(page));
 355
 356                         get_page(page);
 357                         pages[i] = page;
 358                 }
 359
 360                 if (vmas)
 361                         vmas[i] = vma;
 362
 363                 vaddr += PAGE_SIZE;
 364                 ++vpfn;
 365                 --remainder;
 366                 ++i;
 367         }
 368
 369         *length = remainder;
 370         *position = vaddr;
 371
 372         return i;
 373 }
 374
 375 struct page *
 376 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 377 {
 378         pte_t *ptep;
 379         struct page *page;
 380
 381         if (! in_hugepage_area(mm->context, address))
 382                 return ERR_PTR(-EINVAL);
 383
 384         ptep = huge_pte_offset(mm, address);
 385         page = pte_page(*ptep);
 386         if (page)
 387                 page += (address % HPAGE_SIZE) / PAGE_SIZE;
 388
 389         return page;
 390 }
 391
 392 int pmd_huge(pmd_t pmd)
 393 {
 394         return 0;
 395 }
 396
 397 struct page *
 398 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 399                 pmd_t *pmd, int write)
 400 {
 401         BUG();
 402         return NULL;
 403 }
 404
 405 void unmap_hugepage_range(struct vm_area_struct *vma,
 406                           unsigned long start, unsigned long end)
 407 {
 408         struct mm_struct *mm = vma->vm_mm;
 409         unsigned long addr;
 410         pte_t *ptep;
 411         struct page *page;
 412
 413         WARN_ON(!is_vm_hugetlb_page(vma));
 414         BUG_ON((start % HPAGE_SIZE) != 0);
 415         BUG_ON((end % HPAGE_SIZE) != 0);
 416
 417         for (addr = start; addr < end; addr += HPAGE_SIZE) {
 418                 pte_t pte;
 419
 420                 ptep = huge_pte_offset(mm, addr);
 421                 if (!ptep || pte_none(*ptep))
 422                         continue;
 423
 424                 pte = *ptep;
 425                 page = pte_page(pte);
 426                 pte_clear(ptep);
 427
 428                 put_page(page);
 429         }
 430         // mm->rss -= (end - start) >> PAGE_SHIFT;
 431         vx_rsspages_sub(mm, (end - start) >> PAGE_SHIFT);
 432         flush_tlb_pending();
 433 }
 434
 435 void hugetlb_free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *prev,
 436                            unsigned long start, unsigned long end)
 437 {
 438         /* Because the huge pgtables are only 2 level, they can take
 439          * at most around 4M, much less than one hugepage which the
 440          * process is presumably entitled to use.  So we don't bother
 441          * freeing up the pagetables on unmap, and wait until
 442          * destroy_context() to clean up the lot. */
 443 }
 444
 445 int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma)
 446 {
 447         struct mm_struct *mm = current->mm;
 448         unsigned long addr;
 449         int ret = 0;
 450
 451         WARN_ON(!is_vm_hugetlb_page(vma));
 452         BUG_ON((vma->vm_start % HPAGE_SIZE) != 0);
 453         BUG_ON((vma->vm_end % HPAGE_SIZE) != 0);
 454
 455         spin_lock(&mm->page_table_lock);
 456         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 457                 unsigned long idx;
 458                 pte_t *pte = huge_pte_alloc(mm, addr);
 459                 struct page *page;
 460
 461                 if (!pte) {
 462                         ret = -ENOMEM;
 463                         goto out;
 464                 }
 465                 if (! pte_none(*pte))
 466                         continue;
 467
 468                 idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
 469                         + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 470                 page = find_get_page(mapping, idx);
 471                 if (!page) {
 472                         /* charge the fs quota first */
 473                         if (hugetlb_get_quota(mapping)) {
 474                                 ret = -ENOMEM;
 475                                 goto out;
 476                         }
 477                         page = alloc_huge_page();
 478                         if (!page) {
 479                                 hugetlb_put_quota(mapping);
 480                                 ret = -ENOMEM;
 481                                 goto out;
 482                         }
 483                         ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC);
 484                         if (! ret) {
 485                                 unlock_page(page);
 486                         } else {
 487                                 hugetlb_put_quota(mapping);
 488                                 free_huge_page(page);
 489                                 goto out;
 490                         }
 491                 }
 492                 set_huge_pte(mm, vma, page, pte, vma->vm_flags & VM_WRITE);
 493         }
 494 out:
 495         spin_unlock(&mm->page_table_lock);
 496         return ret;
 497 }
 498
 499 /* Because we have an exclusive hugepage region which lies within the
 500  * normal user address space, we have to take special measures to make
 501  * non-huge mmap()s evade the hugepage reserved regions. */
 502 unsigned long arch_get_unmapped_area(struct file *filp, unsigned long addr,
 503                                      unsigned long len, unsigned long pgoff,
 504                                      unsigned long flags)
 505 {
 506         struct mm_struct *mm = current->mm;
 507         struct vm_area_struct *vma;
 508         unsigned long start_addr;
 509
 510         if (len > TASK_SIZE)
 511                 return -ENOMEM;
 512
 513         if (addr) {
 514                 addr = PAGE_ALIGN(addr);
 515                 vma = find_vma(mm, addr);
 516                 if (((TASK_SIZE - len) >= addr)
 517                     && (!vma || (addr+len) <= vma->vm_start)
 518                     && !is_hugepage_only_range(addr,len))
 519                         return addr;
 520         }
 521         start_addr = addr = mm->free_area_cache;
 522
 523 full_search:
 524         vma = find_vma(mm, addr);
 525         while (TASK_SIZE - len >= addr) {
 526                 BUG_ON(vma && (addr >= vma->vm_end));
 527
 528                 if (touches_hugepage_low_range(addr, len)) {
 529                         addr = ALIGN(addr+1, 1<<SID_SHIFT);
 530                         vma = find_vma(mm, addr);
 531                         continue;
 532                 }
 533                 if (touches_hugepage_high_range(addr, len)) {
 534                         addr = TASK_HPAGE_END;
 535                         vma = find_vma(mm, addr);
 536                         continue;
 537                 }
 538                 if (!vma || addr + len <= vma->vm_start) {
 539                         /*
 540                          * Remember the place where we stopped the search:
 541                          */
 542                         mm->free_area_cache = addr + len;
 543                         return addr;
 544                 }
 545                 addr = vma->vm_end;
 546                 vma = vma->vm_next;
 547         }
 548
 549         /* Make sure we didn't miss any holes */
 550         if (start_addr != TASK_UNMAPPED_BASE) {
 551                 start_addr = addr = TASK_UNMAPPED_BASE;
 552                 goto full_search;
 553         }
 554         return -ENOMEM;
 555 }
 556
 557 /*
 558  * This mmap-allocator allocates new areas top-down from below the
 559  * stack's low limit (the base):
 560  *
 561  * Because we have an exclusive hugepage region which lies within the
 562  * normal user address space, we have to take special measures to make
 563  * non-huge mmap()s evade the hugepage reserved regions.
 564  */
 565 unsigned long
 566 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
 567                           const unsigned long len, const unsigned long pgoff,
 568                           const unsigned long flags)
 569 {
 570         struct vm_area_struct *vma, *prev_vma;
 571         struct mm_struct *mm = current->mm;
 572         unsigned long base = mm->mmap_base, addr = addr0;
 573         int first_time = 1;
 574
 575         /* requested length too big for entire address space */
 576         if (len > TASK_SIZE)
 577                 return -ENOMEM;
 578
 579         /* dont allow allocations above current base */
 580         if (mm->free_area_cache > base)
 581                 mm->free_area_cache = base;
 582
 583         /* requesting a specific address */
 584         if (addr) {
 585                 addr = PAGE_ALIGN(addr);
 586                 vma = find_vma(mm, addr);
 587                 if (TASK_SIZE - len >= addr &&
 588                                 (!vma || addr + len <= vma->vm_start)
 589                                 && !is_hugepage_only_range(addr,len))
 590                         return addr;
 591         }
 592
 593 try_again:
 594         /* make sure it can fit in the remaining address space */
 595         if (mm->free_area_cache < len)
 596                 goto fail;
 597
 598         /* either no address requested or cant fit in requested address hole */
 599         addr = (mm->free_area_cache - len) & PAGE_MASK;
 600         do {
 601 hugepage_recheck:
 602                 if (touches_hugepage_low_range(addr, len)) {
 603                         addr = (addr & ((~0) << SID_SHIFT)) - len;
 604                         goto hugepage_recheck;
 605                 } else if (touches_hugepage_high_range(addr, len)) {
 606                         addr = TASK_HPAGE_BASE - len;
 607                 }
 608
 609                 /*
 610                  * Lookup failure means no vma is above this address,
 611                  * i.e. return with success:
 612                  */
 613                 if (!(vma = find_vma_prev(mm, addr, &prev_vma)))
 614                         return addr;
 615
 616                 /*
 617                  * new region fits between prev_vma->vm_end and
 618                  * vma->vm_start, use it:
 619                  */
 620                 if (addr+len <= vma->vm_start &&
 621                                 (!prev_vma || (addr >= prev_vma->vm_end)))
 622                         /* remember the address as a hint for next time */
 623                         return (mm->free_area_cache = addr);
 624                 else
 625                         /* pull free_area_cache down to the first hole */
 626                         if (mm->free_area_cache == vma->vm_end)
 627                                 mm->free_area_cache = vma->vm_start;
 628
 629                 /* try just below the current vma->vm_start */
 630                 addr = vma->vm_start-len;
 631         } while (len <= vma->vm_start);
 632
 633 fail:
 634         /*
 635          * if hint left us with no space for the requested
 636          * mapping then try again:
 637          */
 638         if (first_time) {
 639                 mm->free_area_cache = base;
 640                 first_time = 0;
 641                 goto try_again;
 642         }
 643         /*
 644          * A failed mmap() very likely causes application failure,
 645          * so fall back to the bottom-up function here. This scenario
 646          * can happen with large stack limits and large mmap()
 647          * allocations.
 648          */
 649         mm->free_area_cache = TASK_UNMAPPED_BASE;
 650         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
 651         /*
 652          * Restore the topdown base:
 653          */
 654         mm->free_area_cache = base;
 655
 656         return addr;
 657 }
 658
 659 static unsigned long htlb_get_low_area(unsigned long len, u16 segmask)
 660 {
 661         unsigned long addr = 0;
 662         struct vm_area_struct *vma;
 663
 664         vma = find_vma(current->mm, addr);
 665         while (addr + len <= 0x100000000UL) {
 666                 BUG_ON(vma && (addr >= vma->vm_end)); /* invariant */
 667
 668                 if (! __within_hugepage_low_range(addr, len, segmask)) {
 669                         addr = ALIGN(addr+1, 1<<SID_SHIFT);
 670                         vma = find_vma(current->mm, addr);
 671                         continue;
 672                 }
 673
 674                 if (!vma || (addr + len) <= vma->vm_start)
 675                         return addr;
 676                 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
 677                 /* Depending on segmask this might not be a confirmed
 678                  * hugepage region, so the ALIGN could have skipped
 679                  * some VMAs */
 680                 vma = find_vma(current->mm, addr);
 681         }
 682
 683         return -ENOMEM;
 684 }
 685
 686 static unsigned long htlb_get_high_area(unsigned long len)
 687 {
 688         unsigned long addr = TASK_HPAGE_BASE;
 689         struct vm_area_struct *vma;
 690
 691         vma = find_vma(current->mm, addr);
 692         for (vma = find_vma(current->mm, addr);
 693              addr + len <= TASK_HPAGE_END;
 694              vma = vma->vm_next) {
 695                 BUG_ON(vma && (addr >= vma->vm_end)); /* invariant */
 696                 BUG_ON(! within_hugepage_high_range(addr, len));
 697
 698                 if (!vma || (addr + len) <= vma->vm_start)
 699                         return addr;
 700                 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
 701                 /* Because we're in a hugepage region, this alignment
 702                  * should not skip us over any VMAs */
 703         }
 704
 705         return -ENOMEM;
 706 }
 707
 708 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 709                                         unsigned long len, unsigned long pgoff,
 710                                         unsigned long flags)
 711 {
 712         if (len & ~HPAGE_MASK)
 713                 return -EINVAL;
 714
 715         if (!(cur_cpu_spec->cpu_features & CPU_FTR_16M_PAGE))
 716                 return -EINVAL;
 717
 718         if (test_thread_flag(TIF_32BIT)) {
 719                 int lastshift = 0;
 720                 u16 segmask, cursegs = current->mm->context.htlb_segs;
 721
 722                 /* First see if we can do the mapping in the existing
 723                  * low hpage segments */
 724                 addr = htlb_get_low_area(len, cursegs);
 725                 if (addr != -ENOMEM)
 726                         return addr;
 727
 728                 for (segmask = LOW_ESID_MASK(0x100000000UL-len, len);
 729                      ! lastshift; segmask >>=1) {
 730                         if (segmask & 1)
 731                                 lastshift = 1;
 732
 733                         addr = htlb_get_low_area(len, cursegs | segmask);
 734                         if ((addr != -ENOMEM)
 735                             && open_low_hpage_segs(current->mm, segmask) == 0)
 736                                 return addr;
 737                 }
 738                 printk(KERN_DEBUG "hugetlb_get_unmapped_area() unable to open"
 739                        " enough segments\n");
 740                 return -ENOMEM;
 741         } else {
 742                 return htlb_get_high_area(len);
 743         }
 744 }
 745
 746 void hugetlb_mm_free_pgd(struct mm_struct *mm)
 747 {
 748         int i;
 749         pgd_t *pgdir;
 750
 751         spin_lock(&mm->page_table_lock);
 752
 753         pgdir = mm->context.huge_pgdir;
 754         if (! pgdir)
 755                 goto out;
 756
 757         mm->context.huge_pgdir = NULL;
 758
 759         /* cleanup any hugepte pages leftover */
 760         for (i = 0; i < PTRS_PER_HUGEPGD; i++) {
 761                 pgd_t *pgd = pgdir + i;
 762
 763                 if (! pgd_none(*pgd)) {
 764                         pte_t *pte = (pte_t *)pgd_page(*pgd);
 765                         struct page *ptepage = virt_to_page(pte);
 766
 767                         ptepage->mapping = NULL;
 768
 769                         BUG_ON(memcmp(pte, empty_zero_page, PAGE_SIZE));
 770                         kmem_cache_free(zero_cache, pte);
 771                 }
 772                 pgd_clear(pgd);
 773         }
 774
 775         BUG_ON(memcmp(pgdir, empty_zero_page, PAGE_SIZE));
 776         kmem_cache_free(zero_cache, pgdir);
 777
 778  out:
 779         spin_unlock(&mm->page_table_lock);
 780 }
 781
 782 int hash_huge_page(struct mm_struct *mm, unsigned long access,
 783                    unsigned long ea, unsigned long vsid, int local)
 784 {
 785         pte_t *ptep;
 786         unsigned long va, vpn;
 787         int is_write;
 788         pte_t old_pte, new_pte;
 789         unsigned long hpteflags, prpn;
 790         long slot;
 791         int err = 1;
 792
 793         spin_lock(&mm->page_table_lock);
 794
 795         ptep = huge_pte_offset(mm, ea);
 796
 797         /* Search the Linux page table for a match with va */
 798         va = (vsid << 28) | (ea & 0x0fffffff);
 799         vpn = va >> HPAGE_SHIFT;
 800
 801         /*
 802          * If no pte found or not present, send the problem up to
 803          * do_page_fault
 804          */
 805         if (unlikely(!ptep || pte_none(*ptep)))
 806                 goto out;
 807
 808 /*      BUG_ON(pte_bad(*ptep)); */
 809
 810         /*
 811          * Check the user's access rights to the page.  If access should be
 812          * prevented then send the problem up to do_page_fault.
 813          */
 814         is_write = access & _PAGE_RW;
 815         if (unlikely(is_write && !(pte_val(*ptep) & _PAGE_RW)))
 816                 goto out;
 817         /*
 818          * At this point, we have a pte (old_pte) which can be used to build
 819          * or update an HPTE. There are 2 cases:
 820          *
 821          * 1. There is a valid (present) pte with no associated HPTE (this is
 822          *      the most common case)
 823          * 2. There is a valid (present) pte with an associated HPTE. The
 824          *      current values of the pp bits in the HPTE prevent access
 825          *      because we are doing software DIRTY bit management and the
 826          *      page is currently not DIRTY.
 827          */
 828
 829
 830         old_pte = *ptep;
 831         new_pte = old_pte;
 832
 833         hpteflags = 0x2 | (! (pte_val(new_pte) & _PAGE_RW));
 834
 835         /* Check if pte already has an hpte (case 2) */
 836         if (unlikely(pte_val(old_pte) & _PAGE_HASHPTE)) {
 837                 /* There MIGHT be an HPTE for this pte */
 838                 unsigned long hash, slot;
 839
 840                 hash = hpt_hash(vpn, 1);
 841                 if (pte_val(old_pte) & _PAGE_SECONDARY)
 842                         hash = ~hash;
 843                 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
 844                 slot += (pte_val(old_pte) & _PAGE_GROUP_IX) >> 12;
 845
 846                 if (ppc_md.hpte_updatepp(slot, hpteflags, va, 1, local) == -1)
 847                         pte_val(old_pte) &= ~_PAGE_HPTEFLAGS;
 848         }
 849
 850         if (likely(!(pte_val(old_pte) & _PAGE_HASHPTE))) {
 851                 unsigned long hash = hpt_hash(vpn, 1);
 852                 unsigned long hpte_group;
 853
 854                 prpn = pte_pfn(old_pte);
 855
 856 repeat:
 857                 hpte_group = ((hash & htab_hash_mask) *
 858                               HPTES_PER_GROUP) & ~0x7UL;
 859
 860                 /* Update the linux pte with the HPTE slot */
 861                 pte_val(new_pte) &= ~_PAGE_HPTEFLAGS;
 862                 pte_val(new_pte) |= _PAGE_HASHPTE;
 863
 864                 /* Add in WIMG bits */
 865                 /* XXX We should store these in the pte */
 866                 hpteflags |= _PAGE_COHERENT;
 867
 868                 slot = ppc_md.hpte_insert(hpte_group, va, prpn, 0,
 869                                           hpteflags, 0, 1);
 870
 871                 /* Primary is full, try the secondary */
 872                 if (unlikely(slot == -1)) {
 873                         pte_val(new_pte) |= _PAGE_SECONDARY;
 874                         hpte_group = ((~hash & htab_hash_mask) *
 875                                       HPTES_PER_GROUP) & ~0x7UL;
 876                         slot = ppc_md.hpte_insert(hpte_group, va, prpn,
 877                                                   1, hpteflags, 0, 1);
 878                         if (slot == -1) {
 879                                 if (mftb() & 0x1)
 880                                         hpte_group = ((hash & htab_hash_mask) * HPTES_PER_GROUP) & ~0x7UL;
 881
 882                                 ppc_md.hpte_remove(hpte_group);
 883                                 goto repeat;
 884                         }
 885                 }
 886
 887                 if (unlikely(slot == -2))
 888                         panic("hash_huge_page: pte_insert failed\n");
 889
 890                 pte_val(new_pte) |= (slot<<12) & _PAGE_GROUP_IX;
 891
 892                 /*
 893                  * No need to use ldarx/stdcx here because all who
 894                  * might be updating the pte will hold the
 895                  * page_table_lock
 896                  */
 897                 *ptep = new_pte;
 898         }
 899
 900         err = 0;
 901
 902  out:
 903         spin_unlock(&mm->page_table_lock);
 904
 905         return err;
 906 }