mm/hugetlb.c

   1 /*
   2  * Generic hugetlb support.
   3  * (C) William Irwin, April 2004
   4  */
   5 #include <linux/gfp.h>
   6 #include <linux/list.h>
   7 #include <linux/init.h>
   8 #include <linux/module.h>
   9 #include <linux/mm.h>
  10 #include <linux/sysctl.h>
  11 #include <linux/highmem.h>
  12 #include <linux/nodemask.h>
  13 #include <linux/pagemap.h>
  14 #include <asm/page.h>
  15 #include <asm/pgtable.h>
  16
  17 #include <linux/hugetlb.h>
  18
  19 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
  20 static unsigned long nr_huge_pages, free_huge_pages;
  21 unsigned long max_huge_pages;
  22 static struct list_head hugepage_freelists[MAX_NUMNODES];
  23 static unsigned int nr_huge_pages_node[MAX_NUMNODES];
  24 static unsigned int free_huge_pages_node[MAX_NUMNODES];
  25
  26 /*
  27  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
  28  */
  29 static DEFINE_SPINLOCK(hugetlb_lock);
  30
  31 static void enqueue_huge_page(struct page *page)
  32 {
  33         int nid = page_to_nid(page);
  34         list_add(&page->lru, &hugepage_freelists[nid]);
  35         free_huge_pages++;
  36         free_huge_pages_node[nid]++;
  37 }
  38
  39 static struct page *dequeue_huge_page(void)
  40 {
  41         int nid = numa_node_id();
  42         struct page *page = NULL;
  43
  44         if (list_empty(&hugepage_freelists[nid])) {
  45                 for (nid = 0; nid < MAX_NUMNODES; ++nid)
  46                         if (!list_empty(&hugepage_freelists[nid]))
  47                                 break;
  48         }
  49         if (nid >= 0 && nid < MAX_NUMNODES &&
  50             !list_empty(&hugepage_freelists[nid])) {
  51                 page = list_entry(hugepage_freelists[nid].next,
  52                                   struct page, lru);
  53                 list_del(&page->lru);
  54                 free_huge_pages--;
  55                 free_huge_pages_node[nid]--;
  56         }
  57         return page;
  58 }
  59
  60 static struct page *alloc_fresh_huge_page(void)
  61 {
  62         static int nid = 0;
  63         struct page *page;
  64         page = alloc_pages_node(nid, GFP_HIGHUSER|__GFP_COMP|__GFP_NOWARN,
  65                                         HUGETLB_PAGE_ORDER);
  66         nid = (nid + 1) % num_online_nodes();
  67         if (page) {
  68                 spin_lock(&hugetlb_lock);
  69                 nr_huge_pages++;
  70                 nr_huge_pages_node[page_to_nid(page)]++;
  71                 spin_unlock(&hugetlb_lock);
  72         }
  73         return page;
  74 }
  75
  76 void free_huge_page(struct page *page)
  77 {
  78         BUG_ON(page_count(page));
  79
  80         INIT_LIST_HEAD(&page->lru);
  81         page[1].mapping = NULL;
  82
  83         spin_lock(&hugetlb_lock);
  84         enqueue_huge_page(page);
  85         spin_unlock(&hugetlb_lock);
  86 }
  87
  88 struct page *alloc_huge_page(void)
  89 {
  90         struct page *page;
  91         int i;
  92
  93         spin_lock(&hugetlb_lock);
  94         page = dequeue_huge_page();
  95         if (!page) {
  96                 spin_unlock(&hugetlb_lock);
  97                 return NULL;
  98         }
  99         spin_unlock(&hugetlb_lock);
 100         set_page_count(page, 1);
 101         page[1].mapping = (void *)free_huge_page;
 102         for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); ++i)
 103                 clear_highpage(&page[i]);
 104         return page;
 105 }
 106
 107 static int __init hugetlb_init(void)
 108 {
 109         unsigned long i;
 110         struct page *page;
 111
 112         if (HPAGE_SHIFT == 0)
 113                 return 0;
 114
 115         for (i = 0; i < MAX_NUMNODES; ++i)
 116                 INIT_LIST_HEAD(&hugepage_freelists[i]);
 117
 118         for (i = 0; i < max_huge_pages; ++i) {
 119                 page = alloc_fresh_huge_page();
 120                 if (!page)
 121                         break;
 122                 spin_lock(&hugetlb_lock);
 123                 enqueue_huge_page(page);
 124                 spin_unlock(&hugetlb_lock);
 125         }
 126         max_huge_pages = free_huge_pages = nr_huge_pages = i;
 127         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
 128         return 0;
 129 }
 130 module_init(hugetlb_init);
 131
 132 static int __init hugetlb_setup(char *s)
 133 {
 134         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
 135                 max_huge_pages = 0;
 136         return 1;
 137 }
 138 __setup("hugepages=", hugetlb_setup);
 139
 140 #ifdef CONFIG_SYSCTL
 141 static void update_and_free_page(struct page *page)
 142 {
 143         int i;
 144         nr_huge_pages--;
 145         nr_huge_pages_node[page_zone(page)->zone_pgdat->node_id]--;
 146         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
 147                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
 148                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
 149                                 1 << PG_private | 1<< PG_writeback);
 150                 set_page_count(&page[i], 0);
 151         }
 152         set_page_count(page, 1);
 153         __free_pages(page, HUGETLB_PAGE_ORDER);
 154 }
 155
 156 #ifdef CONFIG_HIGHMEM
 157 static void try_to_free_low(unsigned long count)
 158 {
 159         int i, nid;
 160         for (i = 0; i < MAX_NUMNODES; ++i) {
 161                 struct page *page, *next;
 162                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
 163                         if (PageHighMem(page))
 164                                 continue;
 165                         list_del(&page->lru);
 166                         update_and_free_page(page);
 167                         nid = page_zone(page)->zone_pgdat->node_id;
 168                         free_huge_pages--;
 169                         free_huge_pages_node[nid]--;
 170                         if (count >= nr_huge_pages)
 171                                 return;
 172                 }
 173         }
 174 }
 175 #else
 176 static inline void try_to_free_low(unsigned long count)
 177 {
 178 }
 179 #endif
 180
 181 static unsigned long set_max_huge_pages(unsigned long count)
 182 {
 183         while (count > nr_huge_pages) {
 184                 struct page *page = alloc_fresh_huge_page();
 185                 if (!page)
 186                         return nr_huge_pages;
 187                 spin_lock(&hugetlb_lock);
 188                 enqueue_huge_page(page);
 189                 spin_unlock(&hugetlb_lock);
 190         }
 191         if (count >= nr_huge_pages)
 192                 return nr_huge_pages;
 193
 194         spin_lock(&hugetlb_lock);
 195         try_to_free_low(count);
 196         while (count < nr_huge_pages) {
 197                 struct page *page = dequeue_huge_page();
 198                 if (!page)
 199                         break;
 200                 update_and_free_page(page);
 201         }
 202         spin_unlock(&hugetlb_lock);
 203         return nr_huge_pages;
 204 }
 205
 206 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
 207                            struct file *file, void __user *buffer,
 208                            size_t *length, loff_t *ppos)
 209 {
 210         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
 211         max_huge_pages = set_max_huge_pages(max_huge_pages);
 212         return 0;
 213 }
 214 #endif /* CONFIG_SYSCTL */
 215
 216 int hugetlb_report_meminfo(char *buf)
 217 {
 218         return sprintf(buf,
 219                         "HugePages_Total: %5lu\n"
 220                         "HugePages_Free:  %5lu\n"
 221                         "Hugepagesize:    %5lu kB\n",
 222                         nr_huge_pages,
 223                         free_huge_pages,
 224                         HPAGE_SIZE/1024);
 225 }
 226
 227 int hugetlb_report_node_meminfo(int nid, char *buf)
 228 {
 229         return sprintf(buf,
 230                 "Node %d HugePages_Total: %5u\n"
 231                 "Node %d HugePages_Free:  %5u\n",
 232                 nid, nr_huge_pages_node[nid],
 233                 nid, free_huge_pages_node[nid]);
 234 }
 235
 236 int is_hugepage_mem_enough(size_t size)
 237 {
 238         return (size + ~HPAGE_MASK)/HPAGE_SIZE <= free_huge_pages;
 239 }
 240
 241 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
 242 unsigned long hugetlb_total_pages(void)
 243 {
 244         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
 245 }
 246
 247 /*
 248  * We cannot handle pagefaults against hugetlb pages at all.  They cause
 249  * handle_mm_fault() to try to instantiate regular-sized pages in the
 250  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
 251  * this far.
 252  */
 253 static struct page *hugetlb_nopage(struct vm_area_struct *vma,
 254                                 unsigned long address, int *unused)
 255 {
 256         BUG();
 257         return NULL;
 258 }
 259
 260 struct vm_operations_struct hugetlb_vm_ops = {
 261         .nopage = hugetlb_nopage,
 262 };
 263
 264 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page)
 265 {
 266         pte_t entry;
 267
 268         if (vma->vm_flags & VM_WRITE) {
 269                 entry =
 270                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
 271         } else {
 272                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
 273         }
 274         entry = pte_mkyoung(entry);
 275         entry = pte_mkhuge(entry);
 276
 277         return entry;
 278 }
 279
 280 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 281                             struct vm_area_struct *vma)
 282 {
 283         pte_t *src_pte, *dst_pte, entry;
 284         struct page *ptepage;
 285         unsigned long addr;
 286
 287         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 288                 src_pte = huge_pte_offset(src, addr);
 289                 if (!src_pte)
 290                         continue;
 291                 dst_pte = huge_pte_alloc(dst, addr);
 292                 if (!dst_pte)
 293                         goto nomem;
 294                 spin_lock(&dst->page_table_lock);
 295                 spin_lock(&src->page_table_lock);
 296                 if (!pte_none(*src_pte)) {
 297                         entry = *src_pte;
 298                         ptepage = pte_page(entry);
 299                         get_page(ptepage);
 300                         add_mm_counter(dst, file_rss, HPAGE_SIZE / PAGE_SIZE);
 301                         set_huge_pte_at(dst, addr, dst_pte, entry);
 302                 }
 303                 spin_unlock(&src->page_table_lock);
 304                 spin_unlock(&dst->page_table_lock);
 305         }
 306         return 0;
 307
 308 nomem:
 309         return -ENOMEM;
 310 }
 311
 312 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 313                           unsigned long end)
 314 {
 315         struct mm_struct *mm = vma->vm_mm;
 316         unsigned long address;
 317         pte_t *ptep;
 318         pte_t pte;
 319         struct page *page;
 320
 321         WARN_ON(!is_vm_hugetlb_page(vma));
 322         BUG_ON(start & ~HPAGE_MASK);
 323         BUG_ON(end & ~HPAGE_MASK);
 324
 325         spin_lock(&mm->page_table_lock);
 326
 327         /* Update high watermark before we lower rss */
 328         update_hiwater_rss(mm);
 329
 330         for (address = start; address < end; address += HPAGE_SIZE) {
 331                 ptep = huge_pte_offset(mm, address);
 332                 if (!ptep)
 333                         continue;
 334
 335                 pte = huge_ptep_get_and_clear(mm, address, ptep);
 336                 if (pte_none(pte))
 337                         continue;
 338
 339                 page = pte_page(pte);
 340                 put_page(page);
 341                 add_mm_counter(mm, file_rss, (int) -(HPAGE_SIZE / PAGE_SIZE));
 342         }
 343
 344         spin_unlock(&mm->page_table_lock);
 345         flush_tlb_range(vma, start, end);
 346 }
 347
 348 static struct page *find_or_alloc_huge_page(struct address_space *mapping,
 349                                                 unsigned long idx)
 350 {
 351         struct page *page;
 352         int err;
 353
 354 retry:
 355         page = find_lock_page(mapping, idx);
 356         if (page)
 357                 goto out;
 358
 359         if (hugetlb_get_quota(mapping))
 360                 goto out;
 361         page = alloc_huge_page();
 362         if (!page) {
 363                 hugetlb_put_quota(mapping);
 364                 goto out;
 365         }
 366
 367         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
 368         if (err) {
 369                 put_page(page);
 370                 hugetlb_put_quota(mapping);
 371                 if (err == -EEXIST)
 372                         goto retry;
 373                 page = NULL;
 374         }
 375 out:
 376         return page;
 377 }
 378
 379 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 380                         unsigned long address, int write_access)
 381 {
 382         int ret = VM_FAULT_SIGBUS;
 383         unsigned long idx;
 384         unsigned long size;
 385         pte_t *pte;
 386         struct page *page;
 387         struct address_space *mapping;
 388
 389         pte = huge_pte_alloc(mm, address);
 390         if (!pte)
 391                 goto out;
 392
 393         mapping = vma->vm_file->f_mapping;
 394         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
 395                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 396
 397         /*
 398          * Use page lock to guard against racing truncation
 399          * before we get page_table_lock.
 400          */
 401         page = find_or_alloc_huge_page(mapping, idx);
 402         if (!page)
 403                 goto out;
 404
 405         spin_lock(&mm->page_table_lock);
 406         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 407         if (idx >= size)
 408                 goto backout;
 409
 410         ret = VM_FAULT_MINOR;
 411         if (!pte_none(*pte))
 412                 goto backout;
 413
 414         add_mm_counter(mm, file_rss, HPAGE_SIZE / PAGE_SIZE);
 415         set_huge_pte_at(mm, address, pte, make_huge_pte(vma, page));
 416         spin_unlock(&mm->page_table_lock);
 417         unlock_page(page);
 418 out:
 419         return ret;
 420
 421 backout:
 422         spin_unlock(&mm->page_table_lock);
 423         hugetlb_put_quota(mapping);
 424         unlock_page(page);
 425         put_page(page);
 426         goto out;
 427 }
 428
 429 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
 430                         struct page **pages, struct vm_area_struct **vmas,
 431                         unsigned long *position, int *length, int i)
 432 {
 433         unsigned long vpfn, vaddr = *position;
 434         int remainder = *length;
 435
 436         vpfn = vaddr/PAGE_SIZE;
 437         spin_lock(&mm->page_table_lock);
 438         while (vaddr < vma->vm_end && remainder) {
 439                 pte_t *pte;
 440                 struct page *page;
 441
 442                 /*
 443                  * Some archs (sparc64, sh*) have multiple pte_ts to
 444                  * each hugepage.  We have to make * sure we get the
 445                  * first, for the page indexing below to work.
 446                  */
 447                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
 448
 449                 if (!pte || pte_none(*pte)) {
 450                         int ret;
 451
 452                         spin_unlock(&mm->page_table_lock);
 453                         ret = hugetlb_fault(mm, vma, vaddr, 0);
 454                         spin_lock(&mm->page_table_lock);
 455                         if (ret == VM_FAULT_MINOR)
 456                                 continue;
 457
 458                         remainder = 0;
 459                         if (!i)
 460                                 i = -EFAULT;
 461                         break;
 462                 }
 463
 464                 if (pages) {
 465                         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
 466                         get_page(page);
 467                         pages[i] = page;
 468                 }
 469
 470                 if (vmas)
 471                         vmas[i] = vma;
 472
 473                 vaddr += PAGE_SIZE;
 474                 ++vpfn;
 475                 --remainder;
 476                 ++i;
 477         }
 478         spin_unlock(&mm->page_table_lock);
 479         *length = remainder;
 480         *position = vaddr;
 481
 482         return i;
 483 }