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 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 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 /*
 151  * This function checks for proper alignment of input addr and len parameters.
 152  */
 153 int is_aligned_hugepage_range(unsigned long addr, unsigned long len)
 154 {
 155         if (len & ~HPAGE_MASK)
 156                 return -EINVAL;
 157         if (addr & ~HPAGE_MASK)
 158                 return -EINVAL;
 159         if (! (within_hugepage_low_range(addr, len)
 160                || within_hugepage_high_range(addr, len)) )
 161                 return -EINVAL;
 162         return 0;
 163 }
 164
 165 static void flush_segments(void *parm)
 166 {
 167         u16 segs = (unsigned long) parm;
 168         unsigned long i;
 169
 170         asm volatile("isync" : : : "memory");
 171
 172         for (i = 0; i < 16; i++) {
 173                 if (! (segs & (1U << i)))
 174                         continue;
 175                 asm volatile("slbie %0" : : "r" (i << SID_SHIFT));
 176         }
 177
 178         asm volatile("isync" : : : "memory");
 179 }
 180
 181 static int prepare_low_seg_for_htlb(struct mm_struct *mm, unsigned long seg)
 182 {
 183         unsigned long start = seg << SID_SHIFT;
 184         unsigned long end = (seg+1) << SID_SHIFT;
 185         struct vm_area_struct *vma;
 186
 187         BUG_ON(seg >= 16);
 188
 189         /* Check no VMAs are in the region */
 190         vma = find_vma(mm, start);
 191         if (vma && (vma->vm_start < end))
 192                 return -EBUSY;
 193
 194         return 0;
 195 }
 196
 197 static int open_low_hpage_segs(struct mm_struct *mm, u16 newsegs)
 198 {
 199         unsigned long i;
 200
 201         newsegs &= ~(mm->context.htlb_segs);
 202         if (! newsegs)
 203                 return 0; /* The segments we want are already open */
 204
 205         for (i = 0; i < 16; i++)
 206                 if ((1 << i) & newsegs)
 207                         if (prepare_low_seg_for_htlb(mm, i) != 0)
 208                                 return -EBUSY;
 209
 210         mm->context.htlb_segs |= newsegs;
 211
 212         /* update the paca copy of the context struct */
 213         get_paca()->context = mm->context;
 214
 215         /* the context change must make it to memory before the flush,
 216          * so that further SLB misses do the right thing. */
 217         mb();
 218         on_each_cpu(flush_segments, (void *)(unsigned long)newsegs, 0, 1);
 219
 220         return 0;
 221 }
 222
 223 int prepare_hugepage_range(unsigned long addr, unsigned long len)
 224 {
 225         if (within_hugepage_high_range(addr, len))
 226                 return 0;
 227         else if ((addr < 0x100000000UL) && ((addr+len) < 0x100000000UL)) {
 228                 int err;
 229                 /* Yes, we need both tests, in case addr+len overflows
 230                  * 64-bit arithmetic */
 231                 err = open_low_hpage_segs(current->mm,
 232                                           LOW_ESID_MASK(addr, len));
 233                 if (err)
 234                         printk(KERN_DEBUG "prepare_hugepage_range(%lx, %lx)"
 235                                " failed (segs: 0x%04hx)\n", addr, len,
 236                                LOW_ESID_MASK(addr, len));
 237                 return err;
 238         }
 239
 240         return -EINVAL;
 241 }
 242
 243 struct page *
 244 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 245 {
 246         pte_t *ptep;
 247         struct page *page;
 248
 249         if (! in_hugepage_area(mm->context, address))
 250                 return ERR_PTR(-EINVAL);
 251
 252         ptep = huge_pte_offset(mm, address);
 253         page = pte_page(*ptep);
 254         if (page)
 255                 page += (address % HPAGE_SIZE) / PAGE_SIZE;
 256
 257         return page;
 258 }
 259
 260 int pmd_huge(pmd_t pmd)
 261 {
 262         return 0;
 263 }
 264
 265 struct page *
 266 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 267                 pmd_t *pmd, int write)
 268 {
 269         BUG();
 270         return NULL;
 271 }
 272
 273 /* Because we have an exclusive hugepage region which lies within the
 274  * normal user address space, we have to take special measures to make
 275  * non-huge mmap()s evade the hugepage reserved regions. */
 276 unsigned long arch_get_unmapped_area(struct file *filp, unsigned long addr,
 277                                      unsigned long len, unsigned long pgoff,
 278                                      unsigned long flags)
 279 {
 280         struct mm_struct *mm = current->mm;
 281         struct vm_area_struct *vma;
 282         unsigned long start_addr;
 283
 284         if (len > TASK_SIZE)
 285                 return -ENOMEM;
 286
 287         if (addr) {
 288                 addr = PAGE_ALIGN(addr);
 289                 vma = find_vma(mm, addr);
 290                 if (((TASK_SIZE - len) >= addr)
 291                     && (!vma || (addr+len) <= vma->vm_start)
 292                     && !is_hugepage_only_range(mm, addr,len))
 293                         return addr;
 294         }
 295         if (len > mm->cached_hole_size) {
 296                 start_addr = addr = mm->free_area_cache;
 297         } else {
 298                 start_addr = addr = TASK_UNMAPPED_BASE;
 299                 mm->cached_hole_size = 0;
 300         }
 301
 302 full_search:
 303         vma = find_vma(mm, addr);
 304         while (TASK_SIZE - len >= addr) {
 305                 BUG_ON(vma && (addr >= vma->vm_end));
 306
 307                 if (touches_hugepage_low_range(mm, addr, len)) {
 308                         addr = ALIGN(addr+1, 1<<SID_SHIFT);
 309                         vma = find_vma(mm, addr);
 310                         continue;
 311                 }
 312                 if (touches_hugepage_high_range(addr, len)) {
 313                         addr = TASK_HPAGE_END;
 314                         vma = find_vma(mm, addr);
 315                         continue;
 316                 }
 317                 if (!vma || addr + len <= vma->vm_start) {
 318                         /*
 319                          * Remember the place where we stopped the search:
 320                          */
 321                         mm->free_area_cache = addr + len;
 322                         return addr;
 323                 }
 324                 if (addr + mm->cached_hole_size < vma->vm_start)
 325                         mm->cached_hole_size = vma->vm_start - addr;
 326                 addr = vma->vm_end;
 327                 vma = vma->vm_next;
 328         }
 329
 330         /* Make sure we didn't miss any holes */
 331         if (start_addr != TASK_UNMAPPED_BASE) {
 332                 start_addr = addr = TASK_UNMAPPED_BASE;
 333                 mm->cached_hole_size = 0;
 334                 goto full_search;
 335         }
 336         return -ENOMEM;
 337 }
 338
 339 /*
 340  * This mmap-allocator allocates new areas top-down from below the
 341  * stack's low limit (the base):
 342  *
 343  * Because we have an exclusive hugepage region which lies within the
 344  * normal user address space, we have to take special measures to make
 345  * non-huge mmap()s evade the hugepage reserved regions.
 346  */
 347 unsigned long
 348 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
 349                           const unsigned long len, const unsigned long pgoff,
 350                           const unsigned long flags)
 351 {
 352         struct vm_area_struct *vma, *prev_vma;
 353         struct mm_struct *mm = current->mm;
 354         unsigned long base = mm->mmap_base, addr = addr0;
 355         unsigned long largest_hole = mm->cached_hole_size;
 356         int first_time = 1;
 357
 358         /* requested length too big for entire address space */
 359         if (len > TASK_SIZE)
 360                 return -ENOMEM;
 361
 362         /* dont allow allocations above current base */
 363         if (mm->free_area_cache > base)
 364                 mm->free_area_cache = base;
 365
 366         /* requesting a specific address */
 367         if (addr) {
 368                 addr = PAGE_ALIGN(addr);
 369                 vma = find_vma(mm, addr);
 370                 if (TASK_SIZE - len >= addr &&
 371                                 (!vma || addr + len <= vma->vm_start)
 372                                 && !is_hugepage_only_range(mm, addr,len))
 373                         return addr;
 374         }
 375
 376         if (len <= largest_hole) {
 377                 largest_hole = 0;
 378                 mm->free_area_cache = base;
 379         }
 380 try_again:
 381         /* make sure it can fit in the remaining address space */
 382         if (mm->free_area_cache < len)
 383                 goto fail;
 384
 385         /* either no address requested or cant fit in requested address hole */
 386         addr = (mm->free_area_cache - len) & PAGE_MASK;
 387         do {
 388 hugepage_recheck:
 389                 if (touches_hugepage_low_range(mm, addr, len)) {
 390                         addr = (addr & ((~0) << SID_SHIFT)) - len;
 391                         goto hugepage_recheck;
 392                 } else if (touches_hugepage_high_range(addr, len)) {
 393                         addr = TASK_HPAGE_BASE - len;
 394                 }
 395
 396                 /*
 397                  * Lookup failure means no vma is above this address,
 398                  * i.e. return with success:
 399                  */
 400                 if (!(vma = find_vma_prev(mm, addr, &prev_vma)))
 401                         return addr;
 402
 403                 /*
 404                  * new region fits between prev_vma->vm_end and
 405                  * vma->vm_start, use it:
 406                  */
 407                 if (addr+len <= vma->vm_start &&
 408                           (!prev_vma || (addr >= prev_vma->vm_end))) {
 409                         /* remember the address as a hint for next time */
 410                         mm->cached_hole_size = largest_hole;
 411                         return (mm->free_area_cache = addr);
 412                 } else {
 413                         /* pull free_area_cache down to the first hole */
 414                         if (mm->free_area_cache == vma->vm_end) {
 415                                 mm->free_area_cache = vma->vm_start;
 416                                 mm->cached_hole_size = largest_hole;
 417                         }
 418                 }
 419
 420                 /* remember the largest hole we saw so far */
 421                 if (addr + largest_hole < vma->vm_start)
 422                         largest_hole = vma->vm_start - addr;
 423
 424                 /* try just below the current vma->vm_start */
 425                 addr = vma->vm_start-len;
 426         } while (len <= vma->vm_start);
 427
 428 fail:
 429         /*
 430          * if hint left us with no space for the requested
 431          * mapping then try again:
 432          */
 433         if (first_time) {
 434                 mm->free_area_cache = base;
 435                 largest_hole = 0;
 436                 first_time = 0;
 437                 goto try_again;
 438         }
 439         /*
 440          * A failed mmap() very likely causes application failure,
 441          * so fall back to the bottom-up function here. This scenario
 442          * can happen with large stack limits and large mmap()
 443          * allocations.
 444          */
 445         mm->free_area_cache = TASK_UNMAPPED_BASE;
 446         mm->cached_hole_size = ~0UL;
 447         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
 448         /*
 449          * Restore the topdown base:
 450          */
 451         mm->free_area_cache = base;
 452         mm->cached_hole_size = ~0UL;
 453
 454         return addr;
 455 }
 456
 457 static unsigned long htlb_get_low_area(unsigned long len, u16 segmask)
 458 {
 459         unsigned long addr = 0;
 460         struct vm_area_struct *vma;
 461
 462         vma = find_vma(current->mm, addr);
 463         while (addr + len <= 0x100000000UL) {
 464                 BUG_ON(vma && (addr >= vma->vm_end)); /* invariant */
 465
 466                 if (! __within_hugepage_low_range(addr, len, segmask)) {
 467                         addr = ALIGN(addr+1, 1<<SID_SHIFT);
 468                         vma = find_vma(current->mm, addr);
 469                         continue;
 470                 }
 471
 472                 if (!vma || (addr + len) <= vma->vm_start)
 473                         return addr;
 474                 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
 475                 /* Depending on segmask this might not be a confirmed
 476                  * hugepage region, so the ALIGN could have skipped
 477                  * some VMAs */
 478                 vma = find_vma(current->mm, addr);
 479         }
 480
 481         return -ENOMEM;
 482 }
 483
 484 static unsigned long htlb_get_high_area(unsigned long len)
 485 {
 486         unsigned long addr = TASK_HPAGE_BASE;
 487         struct vm_area_struct *vma;
 488
 489         vma = find_vma(current->mm, addr);
 490         for (vma = find_vma(current->mm, addr);
 491              addr + len <= TASK_HPAGE_END;
 492              vma = vma->vm_next) {
 493                 BUG_ON(vma && (addr >= vma->vm_end)); /* invariant */
 494                 BUG_ON(! within_hugepage_high_range(addr, len));
 495
 496                 if (!vma || (addr + len) <= vma->vm_start)
 497                         return addr;
 498                 addr = ALIGN(vma->vm_end, HPAGE_SIZE);
 499                 /* Because we're in a hugepage region, this alignment
 500                  * should not skip us over any VMAs */
 501         }
 502
 503         return -ENOMEM;
 504 }
 505
 506 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 507                                         unsigned long len, unsigned long pgoff,
 508                                         unsigned long flags)
 509 {
 510         if (len & ~HPAGE_MASK)
 511                 return -EINVAL;
 512
 513         if (!cpu_has_feature(CPU_FTR_16M_PAGE))
 514                 return -EINVAL;
 515
 516         if (test_thread_flag(TIF_32BIT)) {
 517                 int lastshift = 0;
 518                 u16 segmask, cursegs = current->mm->context.htlb_segs;
 519
 520                 /* First see if we can do the mapping in the existing
 521                  * low hpage segments */
 522                 addr = htlb_get_low_area(len, cursegs);
 523                 if (addr != -ENOMEM)
 524                         return addr;
 525
 526                 for (segmask = LOW_ESID_MASK(0x100000000UL-len, len);
 527                      ! lastshift; segmask >>=1) {
 528                         if (segmask & 1)
 529                                 lastshift = 1;
 530
 531                         addr = htlb_get_low_area(len, cursegs | segmask);
 532                         if ((addr != -ENOMEM)
 533                             && open_low_hpage_segs(current->mm, segmask) == 0)
 534                                 return addr;
 535                 }
 536                 printk(KERN_DEBUG "hugetlb_get_unmapped_area() unable to open"
 537                        " enough segments\n");
 538                 return -ENOMEM;
 539         } else {
 540                 return htlb_get_high_area(len);
 541         }
 542 }
 543
 544 void hugetlb_mm_free_pgd(struct mm_struct *mm)
 545 {
 546         int i;
 547         pgd_t *pgdir;
 548
 549         spin_lock(&mm->page_table_lock);
 550
 551         pgdir = mm->context.huge_pgdir;
 552         if (! pgdir)
 553                 goto out;
 554
 555         mm->context.huge_pgdir = NULL;
 556
 557         /* cleanup any hugepte pages leftover */
 558         for (i = 0; i < PTRS_PER_HUGEPGD; i++) {
 559                 pud_t *pud = (pud_t *)(pgdir + i);
 560
 561                 if (! pud_none(*pud)) {
 562                         pte_t *pte = (pte_t *)pud_page(*pud);
 563                         struct page *ptepage = virt_to_page(pte);
 564
 565                         ptepage->mapping = NULL;
 566
 567                         BUG_ON(memcmp(pte, empty_zero_page, PAGE_SIZE));
 568                         kmem_cache_free(zero_cache, pte);
 569                 }
 570                 pud_clear(pud);
 571         }
 572
 573         BUG_ON(memcmp(pgdir, empty_zero_page, PAGE_SIZE));
 574         kmem_cache_free(zero_cache, pgdir);
 575
 576  out:
 577         spin_unlock(&mm->page_table_lock);
 578 }
 579
 580 int hash_huge_page(struct mm_struct *mm, unsigned long access,
 581                    unsigned long ea, unsigned long vsid, int local)
 582 {
 583         pte_t *ptep;
 584         unsigned long va, vpn;
 585         pte_t old_pte, new_pte;
 586         unsigned long rflags, prpn;
 587         long slot;
 588         int err = 1;
 589
 590         spin_lock(&mm->page_table_lock);
 591
 592         ptep = huge_pte_offset(mm, ea);
 593
 594         /* Search the Linux page table for a match with va */
 595         va = (vsid << 28) | (ea & 0x0fffffff);
 596         vpn = va >> HPAGE_SHIFT;
 597
 598         /*
 599          * If no pte found or not present, send the problem up to
 600          * do_page_fault
 601          */
 602         if (unlikely(!ptep || pte_none(*ptep)))
 603                 goto out;
 604
 605 /*      BUG_ON(pte_bad(*ptep)); */
 606
 607         /*
 608          * Check the user's access rights to the page.  If access should be
 609          * prevented then send the problem up to do_page_fault.
 610          */
 611         if (unlikely(access & ~pte_val(*ptep)))
 612                 goto out;
 613         /*
 614          * At this point, we have a pte (old_pte) which can be used to build
 615          * or update an HPTE. There are 2 cases:
 616          *
 617          * 1. There is a valid (present) pte with no associated HPTE (this is
 618          *      the most common case)
 619          * 2. There is a valid (present) pte with an associated HPTE. The
 620          *      current values of the pp bits in the HPTE prevent access
 621          *      because we are doing software DIRTY bit management and the
 622          *      page is currently not DIRTY.
 623          */
 624
 625
 626         old_pte = *ptep;
 627         new_pte = old_pte;
 628
 629         rflags = 0x2 | (! (pte_val(new_pte) & _PAGE_RW));
 630         /* _PAGE_EXEC -> HW_NO_EXEC since it's inverted */
 631         rflags |= ((pte_val(new_pte) & _PAGE_EXEC) ? 0 : HW_NO_EXEC);
 632
 633         /* Check if pte already has an hpte (case 2) */
 634         if (unlikely(pte_val(old_pte) & _PAGE_HASHPTE)) {
 635                 /* There MIGHT be an HPTE for this pte */
 636                 unsigned long hash, slot;
 637
 638                 hash = hpt_hash(vpn, 1);
 639                 if (pte_val(old_pte) & _PAGE_SECONDARY)
 640                         hash = ~hash;
 641                 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
 642                 slot += (pte_val(old_pte) & _PAGE_GROUP_IX) >> 12;
 643
 644                 if (ppc_md.hpte_updatepp(slot, rflags, va, 1, local) == -1)
 645                         pte_val(old_pte) &= ~_PAGE_HPTEFLAGS;
 646         }
 647
 648         if (likely(!(pte_val(old_pte) & _PAGE_HASHPTE))) {
 649                 unsigned long hash = hpt_hash(vpn, 1);
 650                 unsigned long hpte_group;
 651
 652                 prpn = pte_pfn(old_pte);
 653
 654 repeat:
 655                 hpte_group = ((hash & htab_hash_mask) *
 656                               HPTES_PER_GROUP) & ~0x7UL;
 657
 658                 /* Update the linux pte with the HPTE slot */
 659                 pte_val(new_pte) &= ~_PAGE_HPTEFLAGS;
 660                 pte_val(new_pte) |= _PAGE_HASHPTE;
 661
 662                 /* Add in WIMG bits */
 663                 /* XXX We should store these in the pte */
 664                 rflags |= _PAGE_COHERENT;
 665
 666                 slot = ppc_md.hpte_insert(hpte_group, va, prpn,
 667                                           HPTE_V_LARGE, rflags);
 668
 669                 /* Primary is full, try the secondary */
 670                 if (unlikely(slot == -1)) {
 671                         pte_val(new_pte) |= _PAGE_SECONDARY;
 672                         hpte_group = ((~hash & htab_hash_mask) *
 673                                       HPTES_PER_GROUP) & ~0x7UL;
 674                         slot = ppc_md.hpte_insert(hpte_group, va, prpn,
 675                                                   HPTE_V_LARGE, rflags);
 676                         if (slot == -1) {
 677                                 if (mftb() & 0x1)
 678                                         hpte_group = ((hash & htab_hash_mask) * HPTES_PER_GROUP) & ~0x7UL;
 679
 680                                 ppc_md.hpte_remove(hpte_group);
 681                                 goto repeat;
 682                         }
 683                 }
 684
 685                 if (unlikely(slot == -2))
 686                         panic("hash_huge_page: pte_insert failed\n");
 687
 688                 pte_val(new_pte) |= (slot<<12) & _PAGE_GROUP_IX;
 689
 690                 /*
 691                  * No need to use ldarx/stdcx here because all who
 692                  * might be updating the pte will hold the
 693                  * page_table_lock
 694                  */
 695                 *ptep = new_pte;
 696         }
 697
 698         err = 0;
 699
 700  out:
 701         spin_unlock(&mm->page_table_lock);
 702
 703         return err;
 704 }