arch/tile/mm/hugetlbpage.c

   1 /*
   2  * Copyright 2010 Tilera Corporation. All Rights Reserved.
   3  *
   4  *   This program is free software; you can redistribute it and/or
   5  *   modify it under the terms of the GNU General Public License
   6  *   as published by the Free Software Foundation, version 2.
   7  *
   8  *   This program is distributed in the hope that it will be useful, but
   9  *   WITHOUT ANY WARRANTY; without even the implied warranty of
  10  *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
  11  *   NON INFRINGEMENT.  See the GNU General Public License for
  12  *   more details.
  13  *
  14  * TILE Huge TLB Page Support for Kernel.
  15  * Taken from i386 hugetlb implementation:
  16  * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
  17  */
  18
  19 #include <linux/init.h>
  20 #include <linux/fs.h>
  21 #include <linux/mm.h>
  22 #include <linux/hugetlb.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/slab.h>
  25 #include <linux/err.h>
  26 #include <linux/sysctl.h>
  27 #include <linux/mman.h>
  28 #include <asm/tlb.h>
  29 #include <asm/tlbflush.h>
  30 #include <asm/setup.h>
  31
  32 #ifdef CONFIG_HUGETLB_SUPER_PAGES
  33
  34 /*
  35  * Provide an additional huge page size (in addition to the regular default
  36  * huge page size) if no "hugepagesz" arguments are specified.
  37  * Note that it must be smaller than the default huge page size so
  38  * that it's possible to allocate them on demand from the buddy allocator.
  39  * You can change this to 64K (on a 16K build), 256K, 1M, or 4M,
  40  * or not define it at all.
  41  */
  42 #define ADDITIONAL_HUGE_SIZE (1024 * 1024UL)
  43
  44 /* "Extra" page-size multipliers, one per level of the page table. */
  45 int huge_shift[HUGE_SHIFT_ENTRIES] = {
  46 #ifdef ADDITIONAL_HUGE_SIZE
  47 #define ADDITIONAL_HUGE_SHIFT __builtin_ctzl(ADDITIONAL_HUGE_SIZE / PAGE_SIZE)
  48         [HUGE_SHIFT_PAGE] = ADDITIONAL_HUGE_SHIFT
  49 #endif
  50 };
  51
  52 #endif
  53
  54 pte_t *huge_pte_alloc(struct mm_struct *mm,
  55                       unsigned long addr, unsigned long sz)
  56 {
  57         pgd_t *pgd;
  58         pud_t *pud;
  59
  60         addr &= -sz;   /* Mask off any low bits in the address. */
  61
  62         pgd = pgd_offset(mm, addr);
  63         pud = pud_alloc(mm, pgd, addr);
  64
  65 #ifdef CONFIG_HUGETLB_SUPER_PAGES
  66         if (sz >= PGDIR_SIZE) {
  67                 BUG_ON(sz != PGDIR_SIZE &&
  68                        sz != PGDIR_SIZE << huge_shift[HUGE_SHIFT_PGDIR]);
  69                 return (pte_t *)pud;
  70         } else {
  71                 pmd_t *pmd = pmd_alloc(mm, pud, addr);
  72                 if (sz >= PMD_SIZE) {
  73                         BUG_ON(sz != PMD_SIZE &&
  74                                sz != (PMD_SIZE << huge_shift[HUGE_SHIFT_PMD]));
  75                         return (pte_t *)pmd;
  76                 }
  77                 else {
  78                         if (sz != PAGE_SIZE << huge_shift[HUGE_SHIFT_PAGE])
  79                                 panic("Unexpected page size %#lx\n", sz);
  80                         return pte_alloc_map(mm, NULL, pmd, addr);
  81                 }
  82         }
  83 #else
  84         BUG_ON(sz != PMD_SIZE);
  85         return (pte_t *) pmd_alloc(mm, pud, addr);
  86 #endif
  87 }
  88
  89 static pte_t *get_pte(pte_t *base, int index, int level)
  90 {
  91         pte_t *ptep = base + index;
  92 #ifdef CONFIG_HUGETLB_SUPER_PAGES
  93         if (!pte_present(*ptep) && huge_shift[level] != 0) {
  94                 unsigned long mask = -1UL << huge_shift[level];
  95                 pte_t *super_ptep = base + (index & mask);
  96                 pte_t pte = *super_ptep;
  97                 if (pte_present(pte) && pte_super(pte))
  98                         ptep = super_ptep;
  99         }
 100 #endif
 101         return ptep;
 102 }
 103
 104 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
 105 {
 106         pgd_t *pgd;
 107         pud_t *pud;
 108         pmd_t *pmd;
 109 #ifdef CONFIG_HUGETLB_SUPER_PAGES
 110         pte_t *pte;
 111 #endif
 112
 113         /* Get the top-level page table entry. */
 114         pgd = (pgd_t *)get_pte((pte_t *)mm->pgd, pgd_index(addr), 0);
 115
 116         /* We don't have four levels. */
 117         pud = pud_offset(pgd, addr);
 118 #ifndef __PAGETABLE_PUD_FOLDED
 119 # error support fourth page table level
 120 #endif
 121         if (!pud_present(*pud))
 122                 return NULL;
 123
 124         /* Check for an L0 huge PTE, if we have three levels. */
 125 #ifndef __PAGETABLE_PMD_FOLDED
 126         if (pud_huge(*pud))
 127                 return (pte_t *)pud;
 128
 129         pmd = (pmd_t *)get_pte((pte_t *)pud_page_vaddr(*pud),
 130                                pmd_index(addr), 1);
 131         if (!pmd_present(*pmd))
 132                 return NULL;
 133 #else
 134         pmd = pmd_offset(pud, addr);
 135 #endif
 136
 137         /* Check for an L1 huge PTE. */
 138         if (pmd_huge(*pmd))
 139                 return (pte_t *)pmd;
 140
 141 #ifdef CONFIG_HUGETLB_SUPER_PAGES
 142         /* Check for an L2 huge PTE. */
 143         pte = get_pte((pte_t *)pmd_page_vaddr(*pmd), pte_index(addr), 2);
 144         if (!pte_present(*pte))
 145                 return NULL;
 146         if (pte_super(*pte))
 147                 return pte;
 148 #endif
 149
 150         return NULL;
 151 }
 152
 153 struct page *follow_huge_addr(struct mm_struct *mm, unsigned long address,
 154                               int write)
 155 {
 156         return ERR_PTR(-EINVAL);
 157 }
 158
 159 int pmd_huge(pmd_t pmd)
 160 {
 161         return !!(pmd_val(pmd) & _PAGE_HUGE_PAGE);
 162 }
 163
 164 int pud_huge(pud_t pud)
 165 {
 166         return !!(pud_val(pud) & _PAGE_HUGE_PAGE);
 167 }
 168
 169 struct page *follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 170                              pmd_t *pmd, int write)
 171 {
 172         struct page *page;
 173
 174         page = pte_page(*(pte_t *)pmd);
 175         if (page)
 176                 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
 177         return page;
 178 }
 179
 180 struct page *follow_huge_pud(struct mm_struct *mm, unsigned long address,
 181                              pud_t *pud, int write)
 182 {
 183         struct page *page;
 184
 185         page = pte_page(*(pte_t *)pud);
 186         if (page)
 187                 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
 188         return page;
 189 }
 190
 191 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
 192 {
 193         return 0;
 194 }
 195
 196 #ifdef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 197 static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *file,
 198                 unsigned long addr, unsigned long len,
 199                 unsigned long pgoff, unsigned long flags)
 200 {
 201         struct hstate *h = hstate_file(file);
 202         struct vm_unmapped_area_info info;
 203
 204         info.flags = 0;
 205         info.length = len;
 206         info.low_limit = TASK_UNMAPPED_BASE;
 207         info.high_limit = TASK_SIZE;
 208         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 209         info.align_offset = 0;
 210         return vm_unmapped_area(&info);
 211 }
 212
 213 static unsigned long hugetlb_get_unmapped_area_topdown(struct file *file,
 214                 unsigned long addr0, unsigned long len,
 215                 unsigned long pgoff, unsigned long flags)
 216 {
 217         struct hstate *h = hstate_file(file);
 218         struct vm_unmapped_area_info info;
 219         unsigned long addr;
 220
 221         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
 222         info.length = len;
 223         info.low_limit = PAGE_SIZE;
 224         info.high_limit = current->mm->mmap_base;
 225         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 226         info.align_offset = 0;
 227         addr = vm_unmapped_area(&info);
 228
 229         /*
 230          * A failed mmap() very likely causes application failure,
 231          * so fall back to the bottom-up function here. This scenario
 232          * can happen with large stack limits and large mmap()
 233          * allocations.
 234          */
 235         if (addr & ~PAGE_MASK) {
 236                 VM_BUG_ON(addr != -ENOMEM);
 237                 info.flags = 0;
 238                 info.low_limit = TASK_UNMAPPED_BASE;
 239                 info.high_limit = TASK_SIZE;
 240                 addr = vm_unmapped_area(&info);
 241         }
 242
 243         return addr;
 244 }
 245
 246 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 247                 unsigned long len, unsigned long pgoff, unsigned long flags)
 248 {
 249         struct hstate *h = hstate_file(file);
 250         struct mm_struct *mm = current->mm;
 251         struct vm_area_struct *vma;
 252
 253         if (len & ~huge_page_mask(h))
 254                 return -EINVAL;
 255         if (len > TASK_SIZE)
 256                 return -ENOMEM;
 257
 258         if (flags & MAP_FIXED) {
 259                 if (prepare_hugepage_range(file, addr, len))
 260                         return -EINVAL;
 261                 return addr;
 262         }
 263
 264         if (addr) {
 265                 addr = ALIGN(addr, huge_page_size(h));
 266                 vma = find_vma(mm, addr);
 267                 if (TASK_SIZE - len >= addr &&
 268                     (!vma || addr + len <= vma->vm_start))
 269                         return addr;
 270         }
 271         if (current->mm->get_unmapped_area == arch_get_unmapped_area)
 272                 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
 273                                 pgoff, flags);
 274         else
 275                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
 276                                 pgoff, flags);
 277 }
 278 #endif /* HAVE_ARCH_HUGETLB_UNMAPPED_AREA */
 279
 280 #ifdef CONFIG_HUGETLB_SUPER_PAGES
 281 static __init int __setup_hugepagesz(unsigned long ps)
 282 {
 283         int log_ps = __builtin_ctzl(ps);
 284         int level, base_shift;
 285
 286         if ((1UL << log_ps) != ps || (log_ps & 1) != 0) {
 287                 pr_warn("Not enabling %ld byte huge pages; must be a power of four\n",
 288                         ps);
 289                 return -EINVAL;
 290         }
 291
 292         if (ps > 64*1024*1024*1024UL) {
 293                 pr_warn("Not enabling %ld MB huge pages; largest legal value is 64 GB\n",
 294                         ps >> 20);
 295                 return -EINVAL;
 296         } else if (ps >= PUD_SIZE) {
 297                 static long hv_jpage_size;
 298                 if (hv_jpage_size == 0)
 299                         hv_jpage_size = hv_sysconf(HV_SYSCONF_PAGE_SIZE_JUMBO);
 300                 if (hv_jpage_size != PUD_SIZE) {
 301                         pr_warn("Not enabling >= %ld MB huge pages: hypervisor reports size %ld\n",
 302                                 PUD_SIZE >> 20, hv_jpage_size);
 303                         return -EINVAL;
 304                 }
 305                 level = 0;
 306                 base_shift = PUD_SHIFT;
 307         } else if (ps >= PMD_SIZE) {
 308                 level = 1;
 309                 base_shift = PMD_SHIFT;
 310         } else if (ps > PAGE_SIZE) {
 311                 level = 2;
 312                 base_shift = PAGE_SHIFT;
 313         } else {
 314                 pr_err("hugepagesz: huge page size %ld too small\n", ps);
 315                 return -EINVAL;
 316         }
 317
 318         if (log_ps != base_shift) {
 319                 int shift_val = log_ps - base_shift;
 320                 if (huge_shift[level] != 0) {
 321                         int old_shift = base_shift + huge_shift[level];
 322                         pr_warn("Not enabling %ld MB huge pages; already have size %ld MB\n",
 323                                 ps >> 20, (1UL << old_shift) >> 20);
 324                         return -EINVAL;
 325                 }
 326                 if (hv_set_pte_super_shift(level, shift_val) != 0) {
 327                         pr_warn("Not enabling %ld MB huge pages; no hypervisor support\n",
 328                                 ps >> 20);
 329                         return -EINVAL;
 330                 }
 331                 printk(KERN_DEBUG "Enabled %ld MB huge pages\n", ps >> 20);
 332                 huge_shift[level] = shift_val;
 333         }
 334
 335         hugetlb_add_hstate(log_ps - PAGE_SHIFT);
 336
 337         return 0;
 338 }
 339
 340 static bool saw_hugepagesz;
 341
 342 static __init int setup_hugepagesz(char *opt)
 343 {
 344         if (!saw_hugepagesz) {
 345                 saw_hugepagesz = true;
 346                 memset(huge_shift, 0, sizeof(huge_shift));
 347         }
 348         return __setup_hugepagesz(memparse(opt, NULL));
 349 }
 350 __setup("hugepagesz=", setup_hugepagesz);
 351
 352 #ifdef ADDITIONAL_HUGE_SIZE
 353 /*
 354  * Provide an additional huge page size if no "hugepagesz" args are given.
 355  * In that case, all the cores have properly set up their hv super_shift
 356  * already, but we need to notify the hugetlb code to enable the
 357  * new huge page size from the Linux point of view.
 358  */
 359 static __init int add_default_hugepagesz(void)
 360 {
 361         if (!saw_hugepagesz) {
 362                 BUILD_BUG_ON(ADDITIONAL_HUGE_SIZE >= PMD_SIZE ||
 363                              ADDITIONAL_HUGE_SIZE <= PAGE_SIZE);
 364                 BUILD_BUG_ON((PAGE_SIZE << ADDITIONAL_HUGE_SHIFT) !=
 365                              ADDITIONAL_HUGE_SIZE);
 366                 BUILD_BUG_ON(ADDITIONAL_HUGE_SHIFT & 1);
 367                 hugetlb_add_hstate(ADDITIONAL_HUGE_SHIFT);
 368         }
 369         return 0;
 370 }
 371 arch_initcall(add_default_hugepagesz);
 372 #endif
 373
 374 #endif /* CONFIG_HUGETLB_SUPER_PAGES */