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 int pmd_huge_support(void)
 170 {
 171         return 1;
 172 }
 173
 174 struct page *follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 175                              pmd_t *pmd, int write)
 176 {
 177         struct page *page;
 178
 179         page = pte_page(*(pte_t *)pmd);
 180         if (page)
 181                 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
 182         return page;
 183 }
 184
 185 struct page *follow_huge_pud(struct mm_struct *mm, unsigned long address,
 186                              pud_t *pud, int write)
 187 {
 188         struct page *page;
 189
 190         page = pte_page(*(pte_t *)pud);
 191         if (page)
 192                 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
 193         return page;
 194 }
 195
 196 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
 197 {
 198         return 0;
 199 }
 200
 201 #ifdef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 202 static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *file,
 203                 unsigned long addr, unsigned long len,
 204                 unsigned long pgoff, unsigned long flags)
 205 {
 206         struct hstate *h = hstate_file(file);
 207         struct vm_unmapped_area_info info;
 208
 209         info.flags = 0;
 210         info.length = len;
 211         info.low_limit = TASK_UNMAPPED_BASE;
 212         info.high_limit = TASK_SIZE;
 213         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 214         info.align_offset = 0;
 215         return vm_unmapped_area(&info);
 216 }
 217
 218 static unsigned long hugetlb_get_unmapped_area_topdown(struct file *file,
 219                 unsigned long addr0, unsigned long len,
 220                 unsigned long pgoff, unsigned long flags)
 221 {
 222         struct hstate *h = hstate_file(file);
 223         struct vm_unmapped_area_info info;
 224         unsigned long addr;
 225
 226         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
 227         info.length = len;
 228         info.low_limit = PAGE_SIZE;
 229         info.high_limit = current->mm->mmap_base;
 230         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 231         info.align_offset = 0;
 232         addr = vm_unmapped_area(&info);
 233
 234         /*
 235          * A failed mmap() very likely causes application failure,
 236          * so fall back to the bottom-up function here. This scenario
 237          * can happen with large stack limits and large mmap()
 238          * allocations.
 239          */
 240         if (addr & ~PAGE_MASK) {
 241                 VM_BUG_ON(addr != -ENOMEM);
 242                 info.flags = 0;
 243                 info.low_limit = TASK_UNMAPPED_BASE;
 244                 info.high_limit = TASK_SIZE;
 245                 addr = vm_unmapped_area(&info);
 246         }
 247
 248         return addr;
 249 }
 250
 251 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 252                 unsigned long len, unsigned long pgoff, unsigned long flags)
 253 {
 254         struct hstate *h = hstate_file(file);
 255         struct mm_struct *mm = current->mm;
 256         struct vm_area_struct *vma;
 257
 258         if (len & ~huge_page_mask(h))
 259                 return -EINVAL;
 260         if (len > TASK_SIZE)
 261                 return -ENOMEM;
 262
 263         if (flags & MAP_FIXED) {
 264                 if (prepare_hugepage_range(file, addr, len))
 265                         return -EINVAL;
 266                 return addr;
 267         }
 268
 269         if (addr) {
 270                 addr = ALIGN(addr, huge_page_size(h));
 271                 vma = find_vma(mm, addr);
 272                 if (TASK_SIZE - len >= addr &&
 273                     (!vma || addr + len <= vma->vm_start))
 274                         return addr;
 275         }
 276         if (current->mm->get_unmapped_area == arch_get_unmapped_area)
 277                 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
 278                                 pgoff, flags);
 279         else
 280                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
 281                                 pgoff, flags);
 282 }
 283 #endif /* HAVE_ARCH_HUGETLB_UNMAPPED_AREA */
 284
 285 #ifdef CONFIG_HUGETLB_SUPER_PAGES
 286 static __init int __setup_hugepagesz(unsigned long ps)
 287 {
 288         int log_ps = __builtin_ctzl(ps);
 289         int level, base_shift;
 290
 291         if ((1UL << log_ps) != ps || (log_ps & 1) != 0) {
 292                 pr_warn("Not enabling %ld byte huge pages;"
 293                         " must be a power of four.\n", ps);
 294                 return -EINVAL;
 295         }
 296
 297         if (ps > 64*1024*1024*1024UL) {
 298                 pr_warn("Not enabling %ld MB huge pages;"
 299                         " largest legal value is 64 GB .\n", ps >> 20);
 300                 return -EINVAL;
 301         } else if (ps >= PUD_SIZE) {
 302                 static long hv_jpage_size;
 303                 if (hv_jpage_size == 0)
 304                         hv_jpage_size = hv_sysconf(HV_SYSCONF_PAGE_SIZE_JUMBO);
 305                 if (hv_jpage_size != PUD_SIZE) {
 306                         pr_warn("Not enabling >= %ld MB huge pages:"
 307                                 " hypervisor reports size %ld\n",
 308                                 PUD_SIZE >> 20, hv_jpage_size);
 309                         return -EINVAL;
 310                 }
 311                 level = 0;
 312                 base_shift = PUD_SHIFT;
 313         } else if (ps >= PMD_SIZE) {
 314                 level = 1;
 315                 base_shift = PMD_SHIFT;
 316         } else if (ps > PAGE_SIZE) {
 317                 level = 2;
 318                 base_shift = PAGE_SHIFT;
 319         } else {
 320                 pr_err("hugepagesz: huge page size %ld too small\n", ps);
 321                 return -EINVAL;
 322         }
 323
 324         if (log_ps != base_shift) {
 325                 int shift_val = log_ps - base_shift;
 326                 if (huge_shift[level] != 0) {
 327                         int old_shift = base_shift + huge_shift[level];
 328                         pr_warn("Not enabling %ld MB huge pages;"
 329                                 " already have size %ld MB.\n",
 330                                 ps >> 20, (1UL << old_shift) >> 20);
 331                         return -EINVAL;
 332                 }
 333                 if (hv_set_pte_super_shift(level, shift_val) != 0) {
 334                         pr_warn("Not enabling %ld MB huge pages;"
 335                                 " no hypervisor support.\n", ps >> 20);
 336                         return -EINVAL;
 337                 }
 338                 printk(KERN_DEBUG "Enabled %ld MB huge pages\n", ps >> 20);
 339                 huge_shift[level] = shift_val;
 340         }
 341
 342         hugetlb_add_hstate(log_ps - PAGE_SHIFT);
 343
 344         return 0;
 345 }
 346
 347 static bool saw_hugepagesz;
 348
 349 static __init int setup_hugepagesz(char *opt)
 350 {
 351         if (!saw_hugepagesz) {
 352                 saw_hugepagesz = true;
 353                 memset(huge_shift, 0, sizeof(huge_shift));
 354         }
 355         return __setup_hugepagesz(memparse(opt, NULL));
 356 }
 357 __setup("hugepagesz=", setup_hugepagesz);
 358
 359 #ifdef ADDITIONAL_HUGE_SIZE
 360 /*
 361  * Provide an additional huge page size if no "hugepagesz" args are given.
 362  * In that case, all the cores have properly set up their hv super_shift
 363  * already, but we need to notify the hugetlb code to enable the
 364  * new huge page size from the Linux point of view.
 365  */
 366 static __init int add_default_hugepagesz(void)
 367 {
 368         if (!saw_hugepagesz) {
 369                 BUILD_BUG_ON(ADDITIONAL_HUGE_SIZE >= PMD_SIZE ||
 370                              ADDITIONAL_HUGE_SIZE <= PAGE_SIZE);
 371                 BUILD_BUG_ON((PAGE_SIZE << ADDITIONAL_HUGE_SHIFT) !=
 372                              ADDITIONAL_HUGE_SIZE);
 373                 BUILD_BUG_ON(ADDITIONAL_HUGE_SHIFT & 1);
 374                 hugetlb_add_hstate(ADDITIONAL_HUGE_SHIFT);
 375         }
 376         return 0;
 377 }
 378 arch_initcall(add_default_hugepagesz);
 379 #endif
 380
 381 #endif /* CONFIG_HUGETLB_SUPER_PAGES */