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 NUM_LOW_AREAS   (0x100000000UL >> SID_SHIFT)
  31 #define NUM_HIGH_AREAS  (PGTABLE_RANGE >> HTLB_AREA_SHIFT)
  32
  33 /* Modelled after find_linux_pte() */
  34 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
  35 {
  36         pgd_t *pg;
  37         pud_t *pu;
  38         pmd_t *pm;
  39         pte_t *pt;
  40
  41         BUG_ON(! in_hugepage_area(mm->context, addr));
  42
  43         addr &= HPAGE_MASK;
  44
  45         pg = pgd_offset(mm, addr);
  46         if (!pgd_none(*pg)) {
  47                 pu = pud_offset(pg, addr);
  48                 if (!pud_none(*pu)) {
  49                         pm = pmd_offset(pu, addr);
  50                         pt = (pte_t *)pm;
  51                         BUG_ON(!pmd_none(*pm)
  52                                && !(pte_present(*pt) && pte_huge(*pt)));
  53                         return pt;
  54                 }
  55         }
  56
  57         return NULL;
  58 }
  59
  60 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
  61 {
  62         pgd_t *pg;
  63         pud_t *pu;
  64         pmd_t *pm;
  65         pte_t *pt;
  66
  67         BUG_ON(! in_hugepage_area(mm->context, addr));
  68
  69         addr &= HPAGE_MASK;
  70
  71         pg = pgd_offset(mm, addr);
  72         pu = pud_alloc(mm, pg, addr);
  73
  74         if (pu) {
  75                 pm = pmd_alloc(mm, pu, addr);
  76                 if (pm) {
  77                         pt = (pte_t *)pm;
  78                         BUG_ON(!pmd_none(*pm)
  79                                && !(pte_present(*pt) && pte_huge(*pt)));
  80                         return pt;
  81                 }
  82         }
  83
  84         return NULL;
  85 }
  86
  87 #define HUGEPTE_BATCH_SIZE      (HPAGE_SIZE / PMD_SIZE)
  88
  89 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
  90                      pte_t *ptep, pte_t pte)
  91 {
  92         int i;
  93
  94         if (pte_present(*ptep)) {
  95                 pte_clear(mm, addr, ptep);
  96                 flush_tlb_pending();
  97         }
  98
  99         for (i = 0; i < HUGEPTE_BATCH_SIZE; i++) {
 100                 *ptep = __pte(pte_val(pte) & ~_PAGE_HPTEFLAGS);
 101                 ptep++;
 102         }
 103 }
 104
 105 pte_t huge_ptep_get_and_clear(struct mm_struct *mm, unsigned long addr,
 106                               pte_t *ptep)
 107 {
 108         unsigned long old = pte_update(ptep, ~0UL);
 109         int i;
 110
 111         if (old & _PAGE_HASHPTE)
 112                 hpte_update(mm, addr, old, 0);
 113
 114         for (i = 1; i < HUGEPTE_BATCH_SIZE; i++)
 115                 ptep[i] = __pte(0);
 116
 117         return __pte(old);
 118 }
 119
 120 /*
 121  * This function checks for proper alignment of input addr and len parameters.
 122  */
 123 int is_aligned_hugepage_range(unsigned long addr, unsigned long len)
 124 {
 125         if (len & ~HPAGE_MASK)
 126                 return -EINVAL;
 127         if (addr & ~HPAGE_MASK)
 128                 return -EINVAL;
 129         if (! (within_hugepage_low_range(addr, len)
 130                || within_hugepage_high_range(addr, len)) )
 131                 return -EINVAL;
 132         return 0;
 133 }
 134
 135 static void flush_low_segments(void *parm)
 136 {
 137         u16 areas = (unsigned long) parm;
 138         unsigned long i;
 139
 140         asm volatile("isync" : : : "memory");
 141
 142         BUILD_BUG_ON((sizeof(areas)*8) != NUM_LOW_AREAS);
 143
 144         for (i = 0; i < NUM_LOW_AREAS; i++) {
 145                 if (! (areas & (1U << i)))
 146                         continue;
 147                 asm volatile("slbie %0"
 148                              : : "r" ((i << SID_SHIFT) | SLBIE_C));
 149         }
 150
 151         asm volatile("isync" : : : "memory");
 152 }
 153
 154 static void flush_high_segments(void *parm)
 155 {
 156         u16 areas = (unsigned long) parm;
 157         unsigned long i, j;
 158
 159         asm volatile("isync" : : : "memory");
 160
 161         BUILD_BUG_ON((sizeof(areas)*8) != NUM_HIGH_AREAS);
 162
 163         for (i = 0; i < NUM_HIGH_AREAS; i++) {
 164                 if (! (areas & (1U << i)))
 165                         continue;
 166                 for (j = 0; j < (1UL << (HTLB_AREA_SHIFT-SID_SHIFT)); j++)
 167                         asm volatile("slbie %0"
 168                                      :: "r" (((i << HTLB_AREA_SHIFT)
 169                                              + (j << SID_SHIFT)) | SLBIE_C));
 170         }
 171
 172         asm volatile("isync" : : : "memory");
 173 }
 174
 175 static int prepare_low_area_for_htlb(struct mm_struct *mm, unsigned long area)
 176 {
 177         unsigned long start = area << SID_SHIFT;
 178         unsigned long end = (area+1) << SID_SHIFT;
 179         struct vm_area_struct *vma;
 180
 181         BUG_ON(area >= NUM_LOW_AREAS);
 182
 183         /* Check no VMAs are in the region */
 184         vma = find_vma(mm, start);
 185         if (vma && (vma->vm_start < end))
 186                 return -EBUSY;
 187
 188         return 0;
 189 }
 190
 191 static int prepare_high_area_for_htlb(struct mm_struct *mm, unsigned long area)
 192 {
 193         unsigned long start = area << HTLB_AREA_SHIFT;
 194         unsigned long end = (area+1) << HTLB_AREA_SHIFT;
 195         struct vm_area_struct *vma;
 196
 197         BUG_ON(area >= NUM_HIGH_AREAS);
 198
 199         /* Check no VMAs are in the region */
 200         vma = find_vma(mm, start);
 201         if (vma && (vma->vm_start < end))
 202                 return -EBUSY;
 203
 204         return 0;
 205 }
 206
 207 static int open_low_hpage_areas(struct mm_struct *mm, u16 newareas)
 208 {
 209         unsigned long i;
 210
 211         BUILD_BUG_ON((sizeof(newareas)*8) != NUM_LOW_AREAS);
 212         BUILD_BUG_ON((sizeof(mm->context.low_htlb_areas)*8) != NUM_LOW_AREAS);
 213
 214         newareas &= ~(mm->context.low_htlb_areas);
 215         if (! newareas)
 216                 return 0; /* The segments we want are already open */
 217
 218         for (i = 0; i < NUM_LOW_AREAS; i++)
 219                 if ((1 << i) & newareas)
 220                         if (prepare_low_area_for_htlb(mm, i) != 0)
 221                                 return -EBUSY;
 222
 223         mm->context.low_htlb_areas |= newareas;
 224
 225         /* update the paca copy of the context struct */
 226         get_paca()->context = mm->context;
 227
 228         /* the context change must make it to memory before the flush,
 229          * so that further SLB misses do the right thing. */
 230         mb();
 231         on_each_cpu(flush_low_segments, (void *)(unsigned long)newareas, 0, 1);
 232
 233         return 0;
 234 }
 235
 236 static int open_high_hpage_areas(struct mm_struct *mm, u16 newareas)
 237 {
 238         unsigned long i;
 239
 240         BUILD_BUG_ON((sizeof(newareas)*8) != NUM_HIGH_AREAS);
 241         BUILD_BUG_ON((sizeof(mm->context.high_htlb_areas)*8)
 242                      != NUM_HIGH_AREAS);
 243
 244         newareas &= ~(mm->context.high_htlb_areas);
 245         if (! newareas)
 246                 return 0; /* The areas we want are already open */
 247
 248         for (i = 0; i < NUM_HIGH_AREAS; i++)
 249                 if ((1 << i) & newareas)
 250                         if (prepare_high_area_for_htlb(mm, i) != 0)
 251                                 return -EBUSY;
 252
 253         mm->context.high_htlb_areas |= newareas;
 254
 255         /* update the paca copy of the context struct */
 256         get_paca()->context = mm->context;
 257
 258         /* the context change must make it to memory before the flush,
 259          * so that further SLB misses do the right thing. */
 260         mb();
 261         on_each_cpu(flush_high_segments, (void *)(unsigned long)newareas, 0, 1);
 262
 263         return 0;
 264 }
 265
 266 int prepare_hugepage_range(unsigned long addr, unsigned long len)
 267 {
 268         int err;
 269
 270         if ( (addr+len) < addr )
 271                 return -EINVAL;
 272
 273         if ((addr + len) < 0x100000000UL)
 274                 err = open_low_hpage_areas(current->mm,
 275                                           LOW_ESID_MASK(addr, len));
 276         else
 277                 err = open_high_hpage_areas(current->mm,
 278                                             HTLB_AREA_MASK(addr, len));
 279         if (err) {
 280                 printk(KERN_DEBUG "prepare_hugepage_range(%lx, %lx)"
 281                        " failed (lowmask: 0x%04hx, highmask: 0x%04hx)\n",
 282                        addr, len,
 283                        LOW_ESID_MASK(addr, len), HTLB_AREA_MASK(addr, len));
 284                 return err;
 285         }
 286
 287         return 0;
 288 }
 289
 290 struct page *
 291 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 292 {
 293         pte_t *ptep;
 294         struct page *page;
 295
 296         if (! in_hugepage_area(mm->context, address))
 297                 return ERR_PTR(-EINVAL);
 298
 299         ptep = huge_pte_offset(mm, address);
 300         page = pte_page(*ptep);
 301         if (page)
 302                 page += (address % HPAGE_SIZE) / PAGE_SIZE;
 303
 304         return page;
 305 }
 306
 307 int pmd_huge(pmd_t pmd)
 308 {
 309         return 0;
 310 }
 311
 312 struct page *
 313 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 314                 pmd_t *pmd, int write)
 315 {
 316         BUG();
 317         return NULL;
 318 }
 319
 320 /* Because we have an exclusive hugepage region which lies within the
 321  * normal user address space, we have to take special measures to make
 322  * non-huge mmap()s evade the hugepage reserved regions. */
 323 unsigned long arch_get_unmapped_area(struct file *filp, unsigned long addr,
 324                                      unsigned long len, unsigned long pgoff,
 325                                      unsigned long flags)
 326 {
 327         struct mm_struct *mm = current->mm;
 328         struct vm_area_struct *vma;
 329         unsigned long start_addr;
 330
 331         if (len > TASK_SIZE)
 332                 return -ENOMEM;
 333
 334         if (addr) {
 335                 addr = PAGE_ALIGN(addr);
 336                 vma = find_vma(mm, addr);
 337                 if (((TASK_SIZE - len) >= addr)
 338                     && (!vma || (addr+len) <= vma->vm_start)
 339                     && !is_hugepage_only_range(mm, addr,len))
 340                         return addr;
 341         }
 342         if (len > mm->cached_hole_size) {
 343                 start_addr = addr = mm->free_area_cache;
 344         } else {
 345                 start_addr = addr = TASK_UNMAPPED_BASE;
 346                 mm->cached_hole_size = 0;
 347         }
 348
 349 full_search:
 350         vma = find_vma(mm, addr);
 351         while (TASK_SIZE - len >= addr) {
 352                 BUG_ON(vma && (addr >= vma->vm_end));
 353
 354                 if (touches_hugepage_low_range(mm, addr, len)) {
 355                         addr = ALIGN(addr+1, 1<<SID_SHIFT);
 356                         vma = find_vma(mm, addr);
 357                         continue;
 358                 }
 359                 if (touches_hugepage_high_range(mm, addr, len)) {
 360                         addr = ALIGN(addr+1, 1UL<<HTLB_AREA_SHIFT);
 361                         vma = find_vma(mm, addr);
 362                         continue;
 363                 }
 364                 if (!vma || addr + len <= vma->vm_start) {
 365                         /*
 366                          * Remember the place where we stopped the search:
 367                          */
 368                         mm->free_area_cache = addr + len;
 369                         return addr;
 370                 }
 371                 if (addr + mm->cached_hole_size < vma->vm_start)
 372                         mm->cached_hole_size = vma->vm_start - addr;
 373                 addr = vma->vm_end;
 374                 vma = vma->vm_next;
 375         }
 376
 377         /* Make sure we didn't miss any holes */
 378         if (start_addr != TASK_UNMAPPED_BASE) {
 379                 start_addr = addr = TASK_UNMAPPED_BASE;
 380                 mm->cached_hole_size = 0;
 381                 goto full_search;
 382         }
 383         return -ENOMEM;
 384 }
 385
 386 /*
 387  * This mmap-allocator allocates new areas top-down from below the
 388  * stack's low limit (the base):
 389  *
 390  * Because we have an exclusive hugepage region which lies within the
 391  * normal user address space, we have to take special measures to make
 392  * non-huge mmap()s evade the hugepage reserved regions.
 393  */
 394 unsigned long
 395 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
 396                           const unsigned long len, const unsigned long pgoff,
 397                           const unsigned long flags)
 398 {
 399         struct vm_area_struct *vma, *prev_vma;
 400         struct mm_struct *mm = current->mm;
 401         unsigned long base = mm->mmap_base, addr = addr0;
 402         unsigned long largest_hole = mm->cached_hole_size;
 403         int first_time = 1;
 404
 405         /* requested length too big for entire address space */
 406         if (len > TASK_SIZE)
 407                 return -ENOMEM;
 408
 409         /* dont allow allocations above current base */
 410         if (mm->free_area_cache > base)
 411                 mm->free_area_cache = base;
 412
 413         /* requesting a specific address */
 414         if (addr) {
 415                 addr = PAGE_ALIGN(addr);
 416                 vma = find_vma(mm, addr);
 417                 if (TASK_SIZE - len >= addr &&
 418                                 (!vma || addr + len <= vma->vm_start)
 419                                 && !is_hugepage_only_range(mm, addr,len))
 420                         return addr;
 421         }
 422
 423         if (len <= largest_hole) {
 424                 largest_hole = 0;
 425                 mm->free_area_cache = base;
 426         }
 427 try_again:
 428         /* make sure it can fit in the remaining address space */
 429         if (mm->free_area_cache < len)
 430                 goto fail;
 431
 432         /* either no address requested or cant fit in requested address hole */
 433         addr = (mm->free_area_cache - len) & PAGE_MASK;
 434         do {
 435 hugepage_recheck:
 436                 if (touches_hugepage_low_range(mm, addr, len)) {
 437                         addr = (addr & ((~0) << SID_SHIFT)) - len;
 438                         goto hugepage_recheck;
 439                 } else if (touches_hugepage_high_range(mm, addr, len)) {
 440                         addr = (addr & ((~0UL) << HTLB_AREA_SHIFT)) - len;
 441                         goto hugepage_recheck;
 442                 }
 443
 444                 /*
 445                  * Lookup failure means no vma is above this address,
 446                  * i.e. return with success:
 447                  */
 448                 if (!(vma = find_vma_prev(mm, addr, &prev_vma)))
 449                         return addr;
 450
 451                 /*
 452                  * new region fits between prev_vma->vm_end and
 453                  * vma->vm_start, use it:
 454                  */
 455                 if (addr+len <= vma->vm_start &&
 456                           (!prev_vma || (addr >= prev_vma->vm_end))) {
 457                         /* remember the address as a hint for next time */
 458                         mm->cached_hole_size = largest_hole;
 459                         return (mm->free_area_cache = addr);
 460                 } else {
 461                         /* pull free_area_cache down to the first hole */
 462                         if (mm->free_area_cache == vma->vm_end) {
 463                                 mm->free_area_cache = vma->vm_start;
 464                                 mm->cached_hole_size = largest_hole;
 465                         }
 466                 }
 467
 468                 /* remember the largest hole we saw so far */
 469                 if (addr + largest_hole < vma->vm_start)
 470                         largest_hole = vma->vm_start - addr;
 471
 472                 /* try just below the current vma->vm_start */
 473                 addr = vma->vm_start-len;
 474         } while (len <= vma->vm_start);
 475
 476 fail:
 477         /*
 478          * if hint left us with no space for the requested
 479          * mapping then try again:
 480          */
 481         if (first_time) {
 482                 mm->free_area_cache = base;
 483                 largest_hole = 0;
 484                 first_time = 0;
 485                 goto try_again;
 486         }
 487         /*
 488          * A failed mmap() very likely causes application failure,
 489          * so fall back to the bottom-up function here. This scenario
 490          * can happen with large stack limits and large mmap()
 491          * allocations.
 492          */
 493         mm->free_area_cache = TASK_UNMAPPED_BASE;
 494         mm->cached_hole_size = ~0UL;
 495         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
 496         /*
 497          * Restore the topdown base:
 498          */
 499         mm->free_area_cache = base;
 500         mm->cached_hole_size = ~0UL;
 501
 502         return addr;
 503 }
 504
 505 static unsigned long htlb_get_low_area(unsigned long len, u16 segmask)
 506 {
 507         unsigned long addr = 0;
 508         struct vm_area_struct *vma;
 509
 510         vma = find_vma(current->mm, addr);
 511         while (addr + len <= 0x100000000UL) {
 512                 BUG_ON(vma && (addr >= vma->vm_end)); /* invariant */
 513
 514                 if (! __within_hugepage_low_range(addr, len, segmask)) {
 515                         addr = ALIGN(addr+1, 1<<SID_SHIFT);
 516                         vma = find_vma(current->mm, addr);
 517                         continue;
 518                 }
 519
 520                 if (!vma || (addr + len) <= vma->vm_start)
 521                         return addr;
 522                 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
 523                 /* Depending on segmask this might not be a confirmed
 524                  * hugepage region, so the ALIGN could have skipped
 525                  * some VMAs */
 526                 vma = find_vma(current->mm, addr);
 527         }
 528
 529         return -ENOMEM;
 530 }
 531
 532 static unsigned long htlb_get_high_area(unsigned long len, u16 areamask)
 533 {
 534         unsigned long addr = 0x100000000UL;
 535         struct vm_area_struct *vma;
 536
 537         vma = find_vma(current->mm, addr);
 538         while (addr + len <= TASK_SIZE_USER64) {
 539                 BUG_ON(vma && (addr >= vma->vm_end)); /* invariant */
 540
 541                 if (! __within_hugepage_high_range(addr, len, areamask)) {
 542                         addr = ALIGN(addr+1, 1UL<<HTLB_AREA_SHIFT);
 543                         vma = find_vma(current->mm, addr);
 544                         continue;
 545                 }
 546
 547                 if (!vma || (addr + len) <= vma->vm_start)
 548                         return addr;
 549                 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
 550                 /* Depending on segmask this might not be a confirmed
 551                  * hugepage region, so the ALIGN could have skipped
 552                  * some VMAs */
 553                 vma = find_vma(current->mm, addr);
 554         }
 555
 556         return -ENOMEM;
 557 }
 558
 559 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 560                                         unsigned long len, unsigned long pgoff,
 561                                         unsigned long flags)
 562 {
 563         int lastshift;
 564         u16 areamask, curareas;
 565
 566         if (len & ~HPAGE_MASK)
 567                 return -EINVAL;
 568
 569         if (!cpu_has_feature(CPU_FTR_16M_PAGE))
 570                 return -EINVAL;
 571
 572         if (test_thread_flag(TIF_32BIT)) {
 573                 curareas = current->mm->context.low_htlb_areas;
 574
 575                 /* First see if we can do the mapping in the existing
 576                  * low areas */
 577                 addr = htlb_get_low_area(len, curareas);
 578                 if (addr != -ENOMEM)
 579                         return addr;
 580
 581                 lastshift = 0;
 582                 for (areamask = LOW_ESID_MASK(0x100000000UL-len, len);
 583                      ! lastshift; areamask >>=1) {
 584                         if (areamask & 1)
 585                                 lastshift = 1;
 586
 587                         addr = htlb_get_low_area(len, curareas | areamask);
 588                         if ((addr != -ENOMEM)
 589                             && open_low_hpage_areas(current->mm, areamask) == 0)
 590                                 return addr;
 591                 }
 592         } else {
 593                 curareas = current->mm->context.high_htlb_areas;
 594
 595                 /* First see if we can do the mapping in the existing
 596                  * high areas */
 597                 addr = htlb_get_high_area(len, curareas);
 598                 if (addr != -ENOMEM)
 599                         return addr;
 600
 601                 lastshift = 0;
 602                 for (areamask = HTLB_AREA_MASK(TASK_SIZE_USER64-len, len);
 603                      ! lastshift; areamask >>=1) {
 604                         if (areamask & 1)
 605                                 lastshift = 1;
 606
 607                         addr = htlb_get_high_area(len, curareas | areamask);
 608                         if ((addr != -ENOMEM)
 609                             && open_high_hpage_areas(current->mm, areamask) == 0)
 610                                 return addr;
 611                 }
 612         }
 613         printk(KERN_DEBUG "hugetlb_get_unmapped_area() unable to open"
 614                " enough areas\n");
 615         return -ENOMEM;
 616 }
 617
 618 int hash_huge_page(struct mm_struct *mm, unsigned long access,
 619                    unsigned long ea, unsigned long vsid, int local)
 620 {
 621         pte_t *ptep;
 622         unsigned long va, vpn;
 623         pte_t old_pte, new_pte;
 624         unsigned long rflags, prpn;
 625         long slot;
 626         int err = 1;
 627
 628         spin_lock(&mm->page_table_lock);
 629
 630         ptep = huge_pte_offset(mm, ea);
 631
 632         /* Search the Linux page table for a match with va */
 633         va = (vsid << 28) | (ea & 0x0fffffff);
 634         vpn = va >> HPAGE_SHIFT;
 635
 636         /*
 637          * If no pte found or not present, send the problem up to
 638          * do_page_fault
 639          */
 640         if (unlikely(!ptep || pte_none(*ptep)))
 641                 goto out;
 642
 643 /*      BUG_ON(pte_bad(*ptep)); */
 644
 645         /*
 646          * Check the user's access rights to the page.  If access should be
 647          * prevented then send the problem up to do_page_fault.
 648          */
 649         if (unlikely(access & ~pte_val(*ptep)))
 650                 goto out;
 651         /*
 652          * At this point, we have a pte (old_pte) which can be used to build
 653          * or update an HPTE. There are 2 cases:
 654          *
 655          * 1. There is a valid (present) pte with no associated HPTE (this is
 656          *      the most common case)
 657          * 2. There is a valid (present) pte with an associated HPTE. The
 658          *      current values of the pp bits in the HPTE prevent access
 659          *      because we are doing software DIRTY bit management and the
 660          *      page is currently not DIRTY.
 661          */
 662
 663
 664         old_pte = *ptep;
 665         new_pte = old_pte;
 666
 667         rflags = 0x2 | (! (pte_val(new_pte) & _PAGE_RW));
 668         /* _PAGE_EXEC -> HW_NO_EXEC since it's inverted */
 669         rflags |= ((pte_val(new_pte) & _PAGE_EXEC) ? 0 : HW_NO_EXEC);
 670
 671         /* Check if pte already has an hpte (case 2) */
 672         if (unlikely(pte_val(old_pte) & _PAGE_HASHPTE)) {
 673                 /* There MIGHT be an HPTE for this pte */
 674                 unsigned long hash, slot;
 675
 676                 hash = hpt_hash(vpn, 1);
 677                 if (pte_val(old_pte) & _PAGE_SECONDARY)
 678                         hash = ~hash;
 679                 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
 680                 slot += (pte_val(old_pte) & _PAGE_GROUP_IX) >> 12;
 681
 682                 if (ppc_md.hpte_updatepp(slot, rflags, va, 1, local) == -1)
 683                         pte_val(old_pte) &= ~_PAGE_HPTEFLAGS;
 684         }
 685
 686         if (likely(!(pte_val(old_pte) & _PAGE_HASHPTE))) {
 687                 unsigned long hash = hpt_hash(vpn, 1);
 688                 unsigned long hpte_group;
 689
 690                 prpn = pte_pfn(old_pte);
 691
 692 repeat:
 693                 hpte_group = ((hash & htab_hash_mask) *
 694                               HPTES_PER_GROUP) & ~0x7UL;
 695
 696                 /* Update the linux pte with the HPTE slot */
 697                 pte_val(new_pte) &= ~_PAGE_HPTEFLAGS;
 698                 pte_val(new_pte) |= _PAGE_HASHPTE;
 699
 700                 /* Add in WIMG bits */
 701                 /* XXX We should store these in the pte */
 702                 rflags |= _PAGE_COHERENT;
 703
 704                 slot = ppc_md.hpte_insert(hpte_group, va, prpn,
 705                                           HPTE_V_LARGE, rflags);
 706
 707                 /* Primary is full, try the secondary */
 708                 if (unlikely(slot == -1)) {
 709                         pte_val(new_pte) |= _PAGE_SECONDARY;
 710                         hpte_group = ((~hash & htab_hash_mask) *
 711                                       HPTES_PER_GROUP) & ~0x7UL;
 712                         slot = ppc_md.hpte_insert(hpte_group, va, prpn,
 713                                                   HPTE_V_LARGE |
 714                                                   HPTE_V_SECONDARY,
 715                                                   rflags);
 716                         if (slot == -1) {
 717                                 if (mftb() & 0x1)
 718                                         hpte_group = ((hash & htab_hash_mask) *
 719                                                       HPTES_PER_GROUP)&~0x7UL;
 720
 721                                 ppc_md.hpte_remove(hpte_group);
 722                                 goto repeat;
 723                         }
 724                 }
 725
 726                 if (unlikely(slot == -2))
 727                         panic("hash_huge_page: pte_insert failed\n");
 728
 729                 pte_val(new_pte) |= (slot<<12) & _PAGE_GROUP_IX;
 730
 731                 /*
 732                  * No need to use ldarx/stdcx here because all who
 733                  * might be updating the pte will hold the
 734                  * page_table_lock
 735                  */
 736                 *ptep = new_pte;
 737         }
 738
 739         err = 0;
 740
 741  out:
 742         spin_unlock(&mm->page_table_lock);
 743
 744         return err;
 745 }