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 static DEFINE_SPINLOCK(hugetlb_lock);
  26
  27 static void enqueue_huge_page(struct page *page)
  28 {
  29         int nid = page_to_nid(page);
  30         list_add(&page->lru, &hugepage_freelists[nid]);
  31         free_huge_pages++;
  32         free_huge_pages_node[nid]++;
  33 }
  34
  35 static struct page *dequeue_huge_page(void)
  36 {
  37         int nid = numa_node_id();
  38         struct page *page = NULL;
  39
  40         if (list_empty(&hugepage_freelists[nid])) {
  41                 for (nid = 0; nid < MAX_NUMNODES; ++nid)
  42                         if (!list_empty(&hugepage_freelists[nid]))
  43                                 break;
  44         }
  45         if (nid >= 0 && nid < MAX_NUMNODES &&
  46             !list_empty(&hugepage_freelists[nid])) {
  47                 page = list_entry(hugepage_freelists[nid].next,
  48                                   struct page, lru);
  49                 list_del(&page->lru);
  50                 free_huge_pages--;
  51                 free_huge_pages_node[nid]--;
  52         }
  53         return page;
  54 }
  55
  56 static struct page *alloc_fresh_huge_page(void)
  57 {
  58         static int nid = 0;
  59         struct page *page;
  60         page = alloc_pages_node(nid, GFP_HIGHUSER|__GFP_COMP|__GFP_NOWARN,
  61                                         HUGETLB_PAGE_ORDER);
  62         nid = (nid + 1) % num_online_nodes();
  63         if (page) {
  64                 nr_huge_pages++;
  65                 nr_huge_pages_node[page_to_nid(page)]++;
  66         }
  67         return page;
  68 }
  69
  70 void free_huge_page(struct page *page)
  71 {
  72         BUG_ON(page_count(page));
  73
  74         INIT_LIST_HEAD(&page->lru);
  75         page[1].mapping = NULL;
  76
  77         spin_lock(&hugetlb_lock);
  78         enqueue_huge_page(page);
  79         spin_unlock(&hugetlb_lock);
  80 }
  81
  82 struct page *alloc_huge_page(void)
  83 {
  84         struct page *page;
  85         int i;
  86
  87         spin_lock(&hugetlb_lock);
  88         page = dequeue_huge_page();
  89         if (!page) {
  90                 spin_unlock(&hugetlb_lock);
  91                 return NULL;
  92         }
  93         spin_unlock(&hugetlb_lock);
  94         set_page_count(page, 1);
  95         page[1].mapping = (void *)free_huge_page;
  96         for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); ++i)
  97                 clear_highpage(&page[i]);
  98         return page;
  99 }
 100
 101 static int __init hugetlb_init(void)
 102 {
 103         unsigned long i;
 104         struct page *page;
 105
 106         if (HPAGE_SHIFT == 0)
 107                 return 0;
 108
 109         for (i = 0; i < MAX_NUMNODES; ++i)
 110                 INIT_LIST_HEAD(&hugepage_freelists[i]);
 111
 112         for (i = 0; i < max_huge_pages; ++i) {
 113                 page = alloc_fresh_huge_page();
 114                 if (!page)
 115                         break;
 116                 spin_lock(&hugetlb_lock);
 117                 enqueue_huge_page(page);
 118                 spin_unlock(&hugetlb_lock);
 119         }
 120         max_huge_pages = free_huge_pages = nr_huge_pages = i;
 121         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
 122         return 0;
 123 }
 124 module_init(hugetlb_init);
 125
 126 static int __init hugetlb_setup(char *s)
 127 {
 128         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
 129                 max_huge_pages = 0;
 130         return 1;
 131 }
 132 __setup("hugepages=", hugetlb_setup);
 133
 134 #ifdef CONFIG_SYSCTL
 135 static void update_and_free_page(struct page *page)
 136 {
 137         int i;
 138         nr_huge_pages--;
 139         nr_huge_pages_node[page_zone(page)->zone_pgdat->node_id]--;
 140         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
 141                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
 142                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
 143                                 1 << PG_private | 1<< PG_writeback);
 144                 set_page_count(&page[i], 0);
 145         }
 146         set_page_count(page, 1);
 147         __free_pages(page, HUGETLB_PAGE_ORDER);
 148 }
 149
 150 #ifdef CONFIG_HIGHMEM
 151 static void try_to_free_low(unsigned long count)
 152 {
 153         int i, nid;
 154         for (i = 0; i < MAX_NUMNODES; ++i) {
 155                 struct page *page, *next;
 156                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
 157                         if (PageHighMem(page))
 158                                 continue;
 159                         list_del(&page->lru);
 160                         update_and_free_page(page);
 161                         nid = page_zone(page)->zone_pgdat->node_id;
 162                         free_huge_pages--;
 163                         free_huge_pages_node[nid]--;
 164                         if (count >= nr_huge_pages)
 165                                 return;
 166                 }
 167         }
 168 }
 169 #else
 170 static inline void try_to_free_low(unsigned long count)
 171 {
 172 }
 173 #endif
 174
 175 static unsigned long set_max_huge_pages(unsigned long count)
 176 {
 177         while (count > nr_huge_pages) {
 178                 struct page *page = alloc_fresh_huge_page();
 179                 if (!page)
 180                         return nr_huge_pages;
 181                 spin_lock(&hugetlb_lock);
 182                 enqueue_huge_page(page);
 183                 spin_unlock(&hugetlb_lock);
 184         }
 185         if (count >= nr_huge_pages)
 186                 return nr_huge_pages;
 187
 188         spin_lock(&hugetlb_lock);
 189         try_to_free_low(count);
 190         while (count < nr_huge_pages) {
 191                 struct page *page = dequeue_huge_page();
 192                 if (!page)
 193                         break;
 194                 update_and_free_page(page);
 195         }
 196         spin_unlock(&hugetlb_lock);
 197         return nr_huge_pages;
 198 }
 199
 200 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
 201                            struct file *file, void __user *buffer,
 202                            size_t *length, loff_t *ppos)
 203 {
 204         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
 205         max_huge_pages = set_max_huge_pages(max_huge_pages);
 206         return 0;
 207 }
 208 #endif /* CONFIG_SYSCTL */
 209
 210 int hugetlb_report_meminfo(char *buf)
 211 {
 212         return sprintf(buf,
 213                         "HugePages_Total: %5lu\n"
 214                         "HugePages_Free:  %5lu\n"
 215                         "Hugepagesize:    %5lu kB\n",
 216                         nr_huge_pages,
 217                         free_huge_pages,
 218                         HPAGE_SIZE/1024);
 219 }
 220
 221 int hugetlb_report_node_meminfo(int nid, char *buf)
 222 {
 223         return sprintf(buf,
 224                 "Node %d HugePages_Total: %5u\n"
 225                 "Node %d HugePages_Free:  %5u\n",
 226                 nid, nr_huge_pages_node[nid],
 227                 nid, free_huge_pages_node[nid]);
 228 }
 229
 230 int is_hugepage_mem_enough(size_t size)
 231 {
 232         return (size + ~HPAGE_MASK)/HPAGE_SIZE <= free_huge_pages;
 233 }
 234
 235 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
 236 unsigned long hugetlb_total_pages(void)
 237 {
 238         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
 239 }
 240 EXPORT_SYMBOL(hugetlb_total_pages);
 241
 242 /*
 243  * We cannot handle pagefaults against hugetlb pages at all.  They cause
 244  * handle_mm_fault() to try to instantiate regular-sized pages in the
 245  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
 246  * this far.
 247  */
 248 static struct page *hugetlb_nopage(struct vm_area_struct *vma,
 249                                 unsigned long address, int *unused)
 250 {
 251         BUG();
 252         return NULL;
 253 }
 254
 255 struct vm_operations_struct hugetlb_vm_ops = {
 256         .nopage = hugetlb_nopage,
 257 };
 258
 259 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page)
 260 {
 261         pte_t entry;
 262
 263         if (vma->vm_flags & VM_WRITE) {
 264                 entry =
 265                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
 266         } else {
 267                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
 268         }
 269         entry = pte_mkyoung(entry);
 270         entry = pte_mkhuge(entry);
 271
 272         return entry;
 273 }
 274
 275 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 276                             struct vm_area_struct *vma)
 277 {
 278         pte_t *src_pte, *dst_pte, entry;
 279         struct page *ptepage;
 280         unsigned long addr;
 281
 282         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 283                 src_pte = huge_pte_offset(src, addr);
 284                 if (!src_pte)
 285                         continue;
 286                 dst_pte = huge_pte_alloc(dst, addr);
 287                 if (!dst_pte)
 288                         goto nomem;
 289                 spin_lock(&dst->page_table_lock);
 290                 spin_lock(&src->page_table_lock);
 291                 if (!pte_none(*src_pte)) {
 292                         entry = *src_pte;
 293                         ptepage = pte_page(entry);
 294                         get_page(ptepage);
 295                         add_mm_counter(dst, file_rss, HPAGE_SIZE / PAGE_SIZE);
 296                         set_huge_pte_at(dst, addr, dst_pte, entry);
 297                 }
 298                 spin_unlock(&src->page_table_lock);
 299                 spin_unlock(&dst->page_table_lock);
 300         }
 301         return 0;
 302
 303 nomem:
 304         return -ENOMEM;
 305 }
 306
 307 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 308                           unsigned long end)
 309 {
 310         struct mm_struct *mm = vma->vm_mm;
 311         unsigned long address;
 312         pte_t *ptep;
 313         pte_t pte;
 314         struct page *page;
 315
 316         WARN_ON(!is_vm_hugetlb_page(vma));
 317         BUG_ON(start & ~HPAGE_MASK);
 318         BUG_ON(end & ~HPAGE_MASK);
 319
 320         spin_lock(&mm->page_table_lock);
 321
 322         /* Update high watermark before we lower rss */
 323         update_hiwater_rss(mm);
 324
 325         for (address = start; address < end; address += HPAGE_SIZE) {
 326                 ptep = huge_pte_offset(mm, address);
 327                 if (!ptep)
 328                         continue;
 329
 330                 pte = huge_ptep_get_and_clear(mm, address, ptep);
 331                 if (pte_none(pte))
 332                         continue;
 333
 334                 page = pte_page(pte);
 335                 put_page(page);
 336                 add_mm_counter(mm, file_rss, (int) -(HPAGE_SIZE / PAGE_SIZE));
 337         }
 338
 339         spin_unlock(&mm->page_table_lock);
 340         flush_tlb_range(vma, start, end);
 341 }
 342
 343 static struct page *find_lock_huge_page(struct address_space *mapping,
 344                         unsigned long idx)
 345 {
 346         struct page *page;
 347         int err;
 348         struct inode *inode = mapping->host;
 349         unsigned long size;
 350
 351 retry:
 352         page = find_lock_page(mapping, idx);
 353         if (page)
 354                 goto out;
 355
 356         /* Check to make sure the mapping hasn't been truncated */
 357         size = i_size_read(inode) >> HPAGE_SHIFT;
 358         if (idx >= size)
 359                 goto out;
 360
 361         if (hugetlb_get_quota(mapping))
 362                 goto out;
 363         page = alloc_huge_page();
 364         if (!page) {
 365                 hugetlb_put_quota(mapping);
 366                 goto out;
 367         }
 368
 369         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
 370         if (err) {
 371                 put_page(page);
 372                 hugetlb_put_quota(mapping);
 373                 if (err == -EEXIST)
 374                         goto retry;
 375                 page = NULL;
 376         }
 377 out:
 378         return page;
 379 }
 380
 381 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 382                         unsigned long address, int write_access)
 383 {
 384         int ret = VM_FAULT_SIGBUS;
 385         unsigned long idx;
 386         unsigned long size;
 387         pte_t *pte;
 388         struct page *page;
 389         struct address_space *mapping;
 390
 391         pte = huge_pte_alloc(mm, address);
 392         if (!pte)
 393                 goto out;
 394
 395         mapping = vma->vm_file->f_mapping;
 396         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
 397                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 398
 399         /*
 400          * Use page lock to guard against racing truncation
 401          * before we get page_table_lock.
 402          */
 403         page = find_lock_huge_page(mapping, idx);
 404         if (!page)
 405                 goto out;
 406
 407         spin_lock(&mm->page_table_lock);
 408         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 409         if (idx >= size)
 410                 goto backout;
 411
 412         ret = VM_FAULT_MINOR;
 413         if (!pte_none(*pte))
 414                 goto backout;
 415
 416         add_mm_counter(mm, file_rss, HPAGE_SIZE / PAGE_SIZE);
 417         set_huge_pte_at(mm, address, pte, make_huge_pte(vma, page));
 418         spin_unlock(&mm->page_table_lock);
 419         unlock_page(page);
 420 out:
 421         return ret;
 422
 423 backout:
 424         spin_unlock(&mm->page_table_lock);
 425         hugetlb_put_quota(mapping);
 426         unlock_page(page);
 427         put_page(page);
 428         goto out;
 429 }
 430
 431 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
 432                         struct page **pages, struct vm_area_struct **vmas,
 433                         unsigned long *position, int *length, int i)
 434 {
 435         unsigned long vpfn, vaddr = *position;
 436         int remainder = *length;
 437
 438         vpfn = vaddr/PAGE_SIZE;
 439         spin_lock(&mm->page_table_lock);
 440         while (vaddr < vma->vm_end && remainder) {
 441                 pte_t *pte;
 442                 struct page *page;
 443
 444                 /*
 445                  * Some archs (sparc64, sh*) have multiple pte_ts to
 446                  * each hugepage.  We have to make * sure we get the
 447                  * first, for the page indexing below to work.
 448                  */
 449                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
 450
 451                 if (!pte || pte_none(*pte)) {
 452                         int ret;
 453
 454                         spin_unlock(&mm->page_table_lock);
 455                         ret = hugetlb_fault(mm, vma, vaddr, 0);
 456                         spin_lock(&mm->page_table_lock);
 457                         if (ret == VM_FAULT_MINOR)
 458                                 continue;
 459
 460                         remainder = 0;
 461                         if (!i)
 462                                 i = -EFAULT;
 463                         break;
 464                 }
 465
 466                 if (pages) {
 467                         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
 468                         get_page(page);
 469                         pages[i] = page;
 470                 }
 471
 472                 if (vmas)
 473                         vmas[i] = vma;
 474
 475                 vaddr += PAGE_SIZE;
 476                 ++vpfn;
 477                 --remainder;
 478                 ++i;
 479         }
 480         spin_unlock(&mm->page_table_lock);
 481         *length = remainder;
 482         *position = vaddr;
 483
 484         return i;
 485 }