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 <linux/mempolicy.h>
  15 #include <linux/cpuset.h>
  16
  17 #include <asm/page.h>
  18 #include <asm/pgtable.h>
  19
  20 #include <linux/hugetlb.h>
  21 #include <linux/vs_memory.h>
  22
  23 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
  24 static unsigned long nr_huge_pages, free_huge_pages;
  25 unsigned long max_huge_pages;
  26 static struct list_head hugepage_freelists[MAX_NUMNODES];
  27 static unsigned int nr_huge_pages_node[MAX_NUMNODES];
  28 static unsigned int free_huge_pages_node[MAX_NUMNODES];
  29
  30 /*
  31  * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
  32  */
  33 static DEFINE_SPINLOCK(hugetlb_lock);
  34
  35 static void enqueue_huge_page(struct page *page)
  36 {
  37         int nid = page_to_nid(page);
  38         list_add(&page->lru, &hugepage_freelists[nid]);
  39         free_huge_pages++;
  40         free_huge_pages_node[nid]++;
  41 }
  42
  43 static struct page *dequeue_huge_page(struct vm_area_struct *vma,
  44                                 unsigned long address)
  45 {
  46         int nid = numa_node_id();
  47         struct page *page = NULL;
  48         struct zonelist *zonelist = huge_zonelist(vma, address);
  49         struct zone **z;
  50
  51         for (z = zonelist->zones; *z; z++) {
  52                 nid = (*z)->zone_pgdat->node_id;
  53                 if (cpuset_zone_allowed(*z, GFP_HIGHUSER) &&
  54                     !list_empty(&hugepage_freelists[nid]))
  55                         break;
  56         }
  57
  58         if (*z) {
  59                 page = list_entry(hugepage_freelists[nid].next,
  60                                   struct page, lru);
  61                 list_del(&page->lru);
  62                 free_huge_pages--;
  63                 free_huge_pages_node[nid]--;
  64         }
  65         return page;
  66 }
  67
  68 static struct page *alloc_fresh_huge_page(void)
  69 {
  70         static int nid = 0;
  71         struct page *page;
  72         page = alloc_pages_node(nid, GFP_HIGHUSER|__GFP_COMP|__GFP_NOWARN,
  73                                         HUGETLB_PAGE_ORDER);
  74         nid = (nid + 1) % num_online_nodes();
  75         if (page) {
  76                 spin_lock(&hugetlb_lock);
  77                 nr_huge_pages++;
  78                 nr_huge_pages_node[page_to_nid(page)]++;
  79                 spin_unlock(&hugetlb_lock);
  80         }
  81         return page;
  82 }
  83
  84 void free_huge_page(struct page *page)
  85 {
  86         BUG_ON(page_count(page));
  87
  88         INIT_LIST_HEAD(&page->lru);
  89         page[1].lru.next = NULL;                        /* reset dtor */
  90
  91         spin_lock(&hugetlb_lock);
  92         enqueue_huge_page(page);
  93         spin_unlock(&hugetlb_lock);
  94 }
  95
  96 struct page *alloc_huge_page(struct vm_area_struct *vma, unsigned long addr)
  97 {
  98         struct page *page;
  99         int i;
 100
 101         spin_lock(&hugetlb_lock);
 102         page = dequeue_huge_page(vma, addr);
 103         if (!page) {
 104                 spin_unlock(&hugetlb_lock);
 105                 return NULL;
 106         }
 107         spin_unlock(&hugetlb_lock);
 108         set_page_count(page, 1);
 109         page[1].lru.next = (void *)free_huge_page;      /* set dtor */
 110         for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); ++i)
 111                 clear_user_highpage(&page[i], addr);
 112         return page;
 113 }
 114
 115 static int __init hugetlb_init(void)
 116 {
 117         unsigned long i;
 118         struct page *page;
 119
 120         if (HPAGE_SHIFT == 0)
 121                 return 0;
 122
 123         for (i = 0; i < MAX_NUMNODES; ++i)
 124                 INIT_LIST_HEAD(&hugepage_freelists[i]);
 125
 126         for (i = 0; i < max_huge_pages; ++i) {
 127                 page = alloc_fresh_huge_page();
 128                 if (!page)
 129                         break;
 130                 spin_lock(&hugetlb_lock);
 131                 enqueue_huge_page(page);
 132                 spin_unlock(&hugetlb_lock);
 133         }
 134         max_huge_pages = free_huge_pages = nr_huge_pages = i;
 135         printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
 136         return 0;
 137 }
 138 module_init(hugetlb_init);
 139
 140 static int __init hugetlb_setup(char *s)
 141 {
 142         if (sscanf(s, "%lu", &max_huge_pages) <= 0)
 143                 max_huge_pages = 0;
 144         return 1;
 145 }
 146 __setup("hugepages=", hugetlb_setup);
 147
 148 #ifdef CONFIG_SYSCTL
 149 static void update_and_free_page(struct page *page)
 150 {
 151         int i;
 152         nr_huge_pages--;
 153         nr_huge_pages_node[page_zone(page)->zone_pgdat->node_id]--;
 154         for (i = 0; i < (HPAGE_SIZE / PAGE_SIZE); i++) {
 155                 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
 156                                 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
 157                                 1 << PG_private | 1<< PG_writeback);
 158                 set_page_count(&page[i], 0);
 159         }
 160         set_page_count(page, 1);
 161         __free_pages(page, HUGETLB_PAGE_ORDER);
 162 }
 163
 164 #ifdef CONFIG_HIGHMEM
 165 static void try_to_free_low(unsigned long count)
 166 {
 167         int i, nid;
 168         for (i = 0; i < MAX_NUMNODES; ++i) {
 169                 struct page *page, *next;
 170                 list_for_each_entry_safe(page, next, &hugepage_freelists[i], lru) {
 171                         if (PageHighMem(page))
 172                                 continue;
 173                         list_del(&page->lru);
 174                         update_and_free_page(page);
 175                         nid = page_zone(page)->zone_pgdat->node_id;
 176                         free_huge_pages--;
 177                         free_huge_pages_node[nid]--;
 178                         if (count >= nr_huge_pages)
 179                                 return;
 180                 }
 181         }
 182 }
 183 #else
 184 static inline void try_to_free_low(unsigned long count)
 185 {
 186 }
 187 #endif
 188
 189 static unsigned long set_max_huge_pages(unsigned long count)
 190 {
 191         while (count > nr_huge_pages) {
 192                 struct page *page = alloc_fresh_huge_page();
 193                 if (!page)
 194                         return nr_huge_pages;
 195                 spin_lock(&hugetlb_lock);
 196                 enqueue_huge_page(page);
 197                 spin_unlock(&hugetlb_lock);
 198         }
 199         if (count >= nr_huge_pages)
 200                 return nr_huge_pages;
 201
 202         spin_lock(&hugetlb_lock);
 203         try_to_free_low(count);
 204         while (count < nr_huge_pages) {
 205                 struct page *page = dequeue_huge_page(NULL, 0);
 206                 if (!page)
 207                         break;
 208                 update_and_free_page(page);
 209         }
 210         spin_unlock(&hugetlb_lock);
 211         return nr_huge_pages;
 212 }
 213
 214 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
 215                            struct file *file, void __user *buffer,
 216                            size_t *length, loff_t *ppos)
 217 {
 218         proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
 219         max_huge_pages = set_max_huge_pages(max_huge_pages);
 220         return 0;
 221 }
 222 #endif /* CONFIG_SYSCTL */
 223
 224 int hugetlb_report_meminfo(char *buf)
 225 {
 226         return sprintf(buf,
 227                         "HugePages_Total: %5lu\n"
 228                         "HugePages_Free:  %5lu\n"
 229                         "Hugepagesize:    %5lu kB\n",
 230                         nr_huge_pages,
 231                         free_huge_pages,
 232                         HPAGE_SIZE/1024);
 233 }
 234
 235 int hugetlb_report_node_meminfo(int nid, char *buf)
 236 {
 237         return sprintf(buf,
 238                 "Node %d HugePages_Total: %5u\n"
 239                 "Node %d HugePages_Free:  %5u\n",
 240                 nid, nr_huge_pages_node[nid],
 241                 nid, free_huge_pages_node[nid]);
 242 }
 243
 244 int is_hugepage_mem_enough(size_t size)
 245 {
 246         return (size + ~HPAGE_MASK)/HPAGE_SIZE <= free_huge_pages;
 247 }
 248
 249 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
 250 unsigned long hugetlb_total_pages(void)
 251 {
 252         return nr_huge_pages * (HPAGE_SIZE / PAGE_SIZE);
 253 }
 254
 255 /*
 256  * We cannot handle pagefaults against hugetlb pages at all.  They cause
 257  * handle_mm_fault() to try to instantiate regular-sized pages in the
 258  * hugegpage VMA.  do_page_fault() is supposed to trap this, so BUG is we get
 259  * this far.
 260  */
 261 static struct page *hugetlb_nopage(struct vm_area_struct *vma,
 262                                 unsigned long address, int *unused)
 263 {
 264         BUG();
 265         return NULL;
 266 }
 267
 268 struct vm_operations_struct hugetlb_vm_ops = {
 269         .nopage = hugetlb_nopage,
 270 };
 271
 272 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
 273                                 int writable)
 274 {
 275         pte_t entry;
 276
 277         if (writable) {
 278                 entry =
 279                     pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
 280         } else {
 281                 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
 282         }
 283         entry = pte_mkyoung(entry);
 284         entry = pte_mkhuge(entry);
 285
 286         return entry;
 287 }
 288
 289 static void set_huge_ptep_writable(struct vm_area_struct *vma,
 290                                    unsigned long address, pte_t *ptep)
 291 {
 292         pte_t entry;
 293
 294         entry = pte_mkwrite(pte_mkdirty(*ptep));
 295         ptep_set_access_flags(vma, address, ptep, entry, 1);
 296         update_mmu_cache(vma, address, entry);
 297         lazy_mmu_prot_update(entry);
 298 }
 299
 300
 301 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
 302                             struct vm_area_struct *vma)
 303 {
 304         pte_t *src_pte, *dst_pte, entry;
 305         struct page *ptepage;
 306         unsigned long addr;
 307         int cow;
 308
 309         cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
 310
 311         for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
 312                 src_pte = huge_pte_offset(src, addr);
 313                 if (!src_pte)
 314                         continue;
 315                 dst_pte = huge_pte_alloc(dst, addr);
 316                 if (!dst_pte)
 317                         goto nomem;
 318                 spin_lock(&dst->page_table_lock);
 319                 spin_lock(&src->page_table_lock);
 320                 if (!pte_none(*src_pte)) {
 321                         if (cow)
 322                                 ptep_set_wrprotect(src, addr, src_pte);
 323                         entry = *src_pte;
 324                         ptepage = pte_page(entry);
 325                         get_page(ptepage);
 326                         add_mm_counter(dst, file_rss, HPAGE_SIZE / PAGE_SIZE);
 327                         set_huge_pte_at(dst, addr, dst_pte, entry);
 328                 }
 329                 spin_unlock(&src->page_table_lock);
 330                 spin_unlock(&dst->page_table_lock);
 331         }
 332         return 0;
 333
 334 nomem:
 335         return -ENOMEM;
 336 }
 337
 338 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 339                           unsigned long end)
 340 {
 341         struct mm_struct *mm = vma->vm_mm;
 342         unsigned long address;
 343         pte_t *ptep;
 344         pte_t pte;
 345         struct page *page;
 346
 347         WARN_ON(!is_vm_hugetlb_page(vma));
 348         BUG_ON(start & ~HPAGE_MASK);
 349         BUG_ON(end & ~HPAGE_MASK);
 350
 351         spin_lock(&mm->page_table_lock);
 352
 353         /* Update high watermark before we lower rss */
 354         update_hiwater_rss(mm);
 355
 356         for (address = start; address < end; address += HPAGE_SIZE) {
 357                 ptep = huge_pte_offset(mm, address);
 358                 if (!ptep)
 359                         continue;
 360
 361                 pte = huge_ptep_get_and_clear(mm, address, ptep);
 362                 if (pte_none(pte))
 363                         continue;
 364
 365                 page = pte_page(pte);
 366                 put_page(page);
 367                 add_mm_counter(mm, file_rss, (int) -(HPAGE_SIZE / PAGE_SIZE));
 368         }
 369
 370         spin_unlock(&mm->page_table_lock);
 371         flush_tlb_range(vma, start, end);
 372 }
 373
 374 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
 375                         unsigned long address, pte_t *ptep, pte_t pte)
 376 {
 377         struct page *old_page, *new_page;
 378         int i, avoidcopy;
 379
 380         old_page = pte_page(pte);
 381
 382         /* If no-one else is actually using this page, avoid the copy
 383          * and just make the page writable */
 384         avoidcopy = (page_count(old_page) == 1);
 385         if (avoidcopy) {
 386                 set_huge_ptep_writable(vma, address, ptep);
 387                 return VM_FAULT_MINOR;
 388         }
 389
 390         page_cache_get(old_page);
 391         new_page = alloc_huge_page(vma, address);
 392
 393         if (!new_page) {
 394                 page_cache_release(old_page);
 395                 return VM_FAULT_OOM;
 396         }
 397
 398         spin_unlock(&mm->page_table_lock);
 399         for (i = 0; i < HPAGE_SIZE/PAGE_SIZE; i++)
 400                 copy_user_highpage(new_page + i, old_page + i,
 401                                    address + i*PAGE_SIZE);
 402         spin_lock(&mm->page_table_lock);
 403
 404         ptep = huge_pte_offset(mm, address & HPAGE_MASK);
 405         if (likely(pte_same(*ptep, pte))) {
 406                 /* Break COW */
 407                 set_huge_pte_at(mm, address, ptep,
 408                                 make_huge_pte(vma, new_page, 1));
 409                 /* Make the old page be freed below */
 410                 new_page = old_page;
 411         }
 412         page_cache_release(new_page);
 413         page_cache_release(old_page);
 414         return VM_FAULT_MINOR;
 415 }
 416
 417 int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
 418                         unsigned long address, pte_t *ptep, int write_access)
 419 {
 420         int ret = VM_FAULT_SIGBUS;
 421         unsigned long idx;
 422         unsigned long size;
 423         struct page *page;
 424         struct address_space *mapping;
 425         pte_t new_pte;
 426
 427         mapping = vma->vm_file->f_mapping;
 428         idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
 429                 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
 430
 431         /*
 432          * Use page lock to guard against racing truncation
 433          * before we get page_table_lock.
 434          */
 435 retry:
 436         page = find_lock_page(mapping, idx);
 437         if (!page) {
 438                 if (hugetlb_get_quota(mapping))
 439                         goto out;
 440                 page = alloc_huge_page(vma, address);
 441                 if (!page) {
 442                         hugetlb_put_quota(mapping);
 443                         ret = VM_FAULT_OOM;
 444                         goto out;
 445                 }
 446
 447                 if (vma->vm_flags & VM_SHARED) {
 448                         int err;
 449
 450                         err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
 451                         if (err) {
 452                                 put_page(page);
 453                                 hugetlb_put_quota(mapping);
 454                                 if (err == -EEXIST)
 455                                         goto retry;
 456                                 goto out;
 457                         }
 458                 } else
 459                         lock_page(page);
 460         }
 461
 462         spin_lock(&mm->page_table_lock);
 463         size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 464         if (idx >= size)
 465                 goto backout;
 466
 467         ret = VM_FAULT_MINOR;
 468         if (!pte_none(*ptep))
 469                 goto backout;
 470
 471         add_mm_counter(mm, file_rss, HPAGE_SIZE / PAGE_SIZE);
 472         new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
 473                                 && (vma->vm_flags & VM_SHARED)));
 474         set_huge_pte_at(mm, address, ptep, new_pte);
 475
 476         if (write_access && !(vma->vm_flags & VM_SHARED)) {
 477                 /* Optimization, do the COW without a second fault */
 478                 ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
 479         }
 480
 481         spin_unlock(&mm->page_table_lock);
 482         unlock_page(page);
 483 out:
 484         return ret;
 485
 486 backout:
 487         spin_unlock(&mm->page_table_lock);
 488         hugetlb_put_quota(mapping);
 489         unlock_page(page);
 490         put_page(page);
 491         goto out;
 492 }
 493
 494 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 495                         unsigned long address, int write_access)
 496 {
 497         pte_t *ptep;
 498         pte_t entry;
 499         int ret;
 500
 501         ptep = huge_pte_alloc(mm, address);
 502         if (!ptep)
 503                 return VM_FAULT_OOM;
 504
 505         entry = *ptep;
 506         if (pte_none(entry))
 507                 return hugetlb_no_page(mm, vma, address, ptep, write_access);
 508
 509         ret = VM_FAULT_MINOR;
 510
 511         spin_lock(&mm->page_table_lock);
 512         /* Check for a racing update before calling hugetlb_cow */
 513         if (likely(pte_same(entry, *ptep)))
 514                 if (write_access && !pte_write(entry))
 515                         ret = hugetlb_cow(mm, vma, address, ptep, entry);
 516         spin_unlock(&mm->page_table_lock);
 517
 518         return ret;
 519 }
 520
 521 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
 522                         struct page **pages, struct vm_area_struct **vmas,
 523                         unsigned long *position, int *length, int i)
 524 {
 525         unsigned long vpfn, vaddr = *position;
 526         int remainder = *length;
 527
 528         vpfn = vaddr/PAGE_SIZE;
 529         spin_lock(&mm->page_table_lock);
 530         while (vaddr < vma->vm_end && remainder) {
 531                 pte_t *pte;
 532                 struct page *page;
 533
 534                 /*
 535                  * Some archs (sparc64, sh*) have multiple pte_ts to
 536                  * each hugepage.  We have to make * sure we get the
 537                  * first, for the page indexing below to work.
 538                  */
 539                 pte = huge_pte_offset(mm, vaddr & HPAGE_MASK);
 540
 541                 if (!pte || pte_none(*pte)) {
 542                         int ret;
 543
 544                         spin_unlock(&mm->page_table_lock);
 545                         ret = hugetlb_fault(mm, vma, vaddr, 0);
 546                         spin_lock(&mm->page_table_lock);
 547                         if (ret == VM_FAULT_MINOR)
 548                                 continue;
 549
 550                         remainder = 0;
 551                         if (!i)
 552                                 i = -EFAULT;
 553                         break;
 554                 }
 555
 556                 if (pages) {
 557                         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
 558                         get_page(page);
 559                         pages[i] = page;
 560                 }
 561
 562                 if (vmas)
 563                         vmas[i] = vma;
 564
 565                 vaddr += PAGE_SIZE;
 566                 ++vpfn;
 567                 --remainder;
 568                 ++i;
 569         }
 570         spin_unlock(&mm->page_table_lock);
 571         *length = remainder;
 572         *position = vaddr;
 573
 574         return i;
 575 }