lustre/llite/llite_mmap.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of Lustre, http://www.lustre.org.
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  */
  21 #ifdef HAVE_KERNEL_CONFIG_H
  22 #include <linux/config.h>
  23 #endif
  24 #include <linux/kernel.h>
  25 #include <linux/mm.h>
  26 #include <linux/string.h>
  27 #include <linux/stat.h>
  28 #include <linux/errno.h>
  29 #include <linux/smp_lock.h>
  30 #include <linux/unistd.h>
  31 #include <linux/version.h>
  32 #include <asm/system.h>
  33 #include <asm/uaccess.h>
  34
  35 #include <linux/fs.h>
  36 #include <linux/stat.h>
  37 #include <asm/uaccess.h>
  38 #include <asm/segment.h>
  39 #include <linux/mm.h>
  40 #include <linux/pagemap.h>
  41 #include <linux/smp_lock.h>
  42 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
  43 #include <linux/iobuf.h>
  44 #endif
  45
  46 #define DEBUG_SUBSYSTEM S_LLITE
  47
  48 #include <lustre_lite.h>
  49 #include "llite_internal.h"
  50 #include <linux/lustre_compat25.h>
  51
  52 #define VMA_DEBUG(vma, fmt, arg...)                                     \
  53         CDEBUG(D_MMAP, "vma(%p) start(%ld) end(%ld) pgoff(%ld) inode(%p) "   \
  54                "ino(%lu) iname(%s): " fmt, vma, vma->vm_start, vma->vm_end,  \
  55                vma->vm_pgoff, vma->vm_file->f_dentry->d_inode,               \
  56                vma->vm_file->f_dentry->d_inode->i_ino,                       \
  57                vma->vm_file->f_dentry->d_iname, ## arg);                     \
  58
  59
  60 struct ll_lock_tree_node {
  61         rb_node_t               lt_node;
  62         struct list_head        lt_locked_item;
  63         __u64                   lt_oid;
  64         ldlm_policy_data_t      lt_policy;
  65         struct lustre_handle    lt_lockh;
  66         ldlm_mode_t             lt_mode;
  67         struct inode           *lt_inode;
  68 };
  69
  70 int lt_get_mmap_locks(struct ll_lock_tree *tree,
  71                       unsigned long addr, size_t count);
  72
  73 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  74 struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
  75                        int *type);
  76 #else
  77
  78 struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
  79                        int unused);
  80 #endif
  81
  82 struct ll_lock_tree_node * ll_node_from_inode(struct inode *inode, __u64 start,
  83                                               __u64 end, ldlm_mode_t mode)
  84 {
  85         struct ll_lock_tree_node *node;
  86
  87         OBD_ALLOC(node, sizeof(*node));
  88         if (node == NULL)
  89                 RETURN(ERR_PTR(-ENOMEM));
  90
  91         node->lt_inode = inode;
  92         node->lt_oid = ll_i2info(inode)->lli_smd->lsm_object_id;
  93         node->lt_policy.l_extent.start = start;
  94         node->lt_policy.l_extent.end = end;
  95         memset(&node->lt_lockh, 0, sizeof(node->lt_lockh));
  96         INIT_LIST_HEAD(&node->lt_locked_item);
  97         node->lt_mode = mode;
  98
  99         return node;
 100 }
 101
 102 int lt_compare(struct ll_lock_tree_node *one, struct ll_lock_tree_node *two)
 103 {
 104         /* To avoid multiple fs deadlock */
 105         if (one->lt_inode->i_sb->s_dev < two->lt_inode->i_sb->s_dev)
 106                 return -1;
 107         if (one->lt_inode->i_sb->s_dev > two->lt_inode->i_sb->s_dev)
 108                 return 1;
 109
 110         if (one->lt_oid < two->lt_oid)
 111                 return -1;
 112         if (one->lt_oid > two->lt_oid)
 113                 return 1;
 114
 115         if (one->lt_policy.l_extent.end < two->lt_policy.l_extent.start)
 116                 return -1;
 117         if (one->lt_policy.l_extent.start > two->lt_policy.l_extent.end)
 118                 return 1;
 119
 120         return 0; /* they are the same object and overlap */
 121 }
 122
 123 static void lt_merge(struct ll_lock_tree_node *dst,
 124                      struct ll_lock_tree_node *src)
 125 {
 126         dst->lt_policy.l_extent.start = min(dst->lt_policy.l_extent.start,
 127                                             src->lt_policy.l_extent.start);
 128         dst->lt_policy.l_extent.end = max(dst->lt_policy.l_extent.end,
 129                                           src->lt_policy.l_extent.end);
 130
 131         /* XXX could be a real call to the dlm to find superset modes */
 132         if (src->lt_mode == LCK_PW && dst->lt_mode != LCK_PW)
 133                 dst->lt_mode = LCK_PW;
 134 }
 135
 136 static void lt_insert(struct ll_lock_tree *tree,
 137                       struct ll_lock_tree_node *node)
 138 {
 139         struct ll_lock_tree_node *walk;
 140         rb_node_t **p, *parent;
 141         ENTRY;
 142
 143 restart:
 144         p = &tree->lt_root.rb_node;
 145         parent = NULL;
 146         while (*p) {
 147                 parent = *p;
 148                 walk = rb_entry(parent, struct ll_lock_tree_node, lt_node);
 149                 switch (lt_compare(node, walk)) {
 150                 case -1:
 151                         p = &(*p)->rb_left;
 152                         break;
 153                 case 1:
 154                         p = &(*p)->rb_right;
 155                         break;
 156                 case 0:
 157                         lt_merge(node, walk);
 158                         rb_erase(&walk->lt_node, &tree->lt_root);
 159                         OBD_FREE(walk, sizeof(*walk));
 160                         goto restart;
 161                         break;
 162                 default:
 163                         LBUG();
 164                         break;
 165                 }
 166         }
 167         rb_link_node(&node->lt_node, parent, p);
 168         rb_insert_color(&node->lt_node, &tree->lt_root);
 169         EXIT;
 170 }
 171
 172 static struct ll_lock_tree_node *lt_least_node(struct ll_lock_tree *tree)
 173 {
 174         rb_node_t *rbnode;
 175         struct ll_lock_tree_node *node = NULL;
 176
 177         for ( rbnode = tree->lt_root.rb_node; rbnode != NULL;
 178               rbnode = rbnode->rb_left) {
 179                 if (rbnode->rb_left == NULL) {
 180                         node = rb_entry(rbnode, struct ll_lock_tree_node,
 181                                         lt_node);
 182                         break;
 183                 }
 184         }
 185         RETURN(node);
 186 }
 187
 188 int ll_tree_unlock(struct ll_lock_tree *tree)
 189 {
 190         struct ll_lock_tree_node *node;
 191         struct list_head *pos, *n;
 192         struct inode *inode;
 193         int rc = 0;
 194         ENTRY;
 195
 196         list_for_each_safe(pos, n, &tree->lt_locked_list) {
 197                 node = list_entry(pos, struct ll_lock_tree_node,
 198                                   lt_locked_item);
 199
 200                 inode = node->lt_inode;
 201                 rc = ll_extent_unlock(tree->lt_fd, inode,
 202                                       ll_i2info(inode)->lli_smd, node->lt_mode,
 203                                       &node->lt_lockh);
 204                 if (rc != 0) {
 205                         /* XXX better message */
 206                         CERROR("couldn't unlock %d\n", rc);
 207                 }
 208                 list_del(&node->lt_locked_item);
 209                 OBD_FREE(node, sizeof(*node));
 210         }
 211
 212         while ((node = lt_least_node(tree))) {
 213                 rb_erase(&node->lt_node, &tree->lt_root);
 214                 OBD_FREE(node, sizeof(*node));
 215         }
 216
 217         RETURN(rc);
 218 }
 219
 220 int ll_tree_lock(struct ll_lock_tree *tree,
 221                  struct ll_lock_tree_node *first_node,
 222                  const char *buf, size_t count, int ast_flags)
 223 {
 224         struct ll_lock_tree_node *node;
 225         int rc = 0;
 226         ENTRY;
 227
 228         tree->lt_root.rb_node = NULL;
 229         INIT_LIST_HEAD(&tree->lt_locked_list);
 230         if (first_node != NULL)
 231                 lt_insert(tree, first_node);
 232
 233         /* To avoid such subtle deadlock case: client1 try to read file1 to
 234          * mmapped file2, on the same time, client2 try to read file2 to
 235          * mmapped file1.*/
 236         rc = lt_get_mmap_locks(tree, (unsigned long)buf, count);
 237         if (rc)
 238                 GOTO(out, rc);
 239
 240         while ((node = lt_least_node(tree))) {
 241                 struct inode *inode = node->lt_inode;
 242                 rc = ll_extent_lock(tree->lt_fd, inode,
 243                                     ll_i2info(inode)->lli_smd, node->lt_mode,
 244                                     &node->lt_policy, &node->lt_lockh,
 245                                     ast_flags);
 246                 if (rc != 0)
 247                         GOTO(out, rc);
 248
 249                 rb_erase(&node->lt_node, &tree->lt_root);
 250                 list_add_tail(&node->lt_locked_item, &tree->lt_locked_list);
 251         }
 252         RETURN(rc);
 253 out:
 254         ll_tree_unlock(tree);
 255         RETURN(rc);
 256 }
 257
 258 static ldlm_mode_t mode_from_vma(struct vm_area_struct *vma)
 259 {
 260         /* we only want to hold PW locks if the mmap() can generate
 261          * writes back to the file and that only happens in shared
 262          * writable vmas */
 263         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
 264                 return LCK_PW;
 265         return LCK_PR;
 266 }
 267
 268 static void policy_from_vma(ldlm_policy_data_t *policy,
 269                             struct vm_area_struct *vma, unsigned long addr,
 270                             size_t count)
 271 {
 272         policy->l_extent.start = ((addr - vma->vm_start) & CFS_PAGE_MASK) +
 273                                  ((__u64)vma->vm_pgoff << CFS_PAGE_SHIFT);
 274         policy->l_extent.end = (policy->l_extent.start + count - 1) |
 275                                ~CFS_PAGE_MASK;
 276 }
 277
 278 static struct vm_area_struct * our_vma(unsigned long addr, size_t count)
 279 {
 280         struct mm_struct *mm = current->mm;
 281         struct vm_area_struct *vma, *ret = NULL;
 282         ENTRY;
 283
 284         /* No MM (e.g. NFS)? No vmas too. */
 285         if (!mm)
 286                 RETURN(NULL);
 287
 288         spin_lock(&mm->page_table_lock);
 289         for(vma = find_vma(mm, addr);
 290             vma != NULL && vma->vm_start < (addr + count); vma = vma->vm_next) {
 291                 if (vma->vm_ops && vma->vm_ops->nopage == ll_nopage &&
 292                     vma->vm_flags & VM_SHARED) {
 293                         ret = vma;
 294                         break;
 295                 }
 296         }
 297         spin_unlock(&mm->page_table_lock);
 298         RETURN(ret);
 299 }
 300
 301 int lt_get_mmap_locks(struct ll_lock_tree *tree,
 302                       unsigned long addr, size_t count)
 303 {
 304         struct vm_area_struct *vma;
 305         struct ll_lock_tree_node *node;
 306         ldlm_policy_data_t policy;
 307         struct inode *inode;
 308         ENTRY;
 309
 310         if (count == 0)
 311                 RETURN(0);
 312
 313         /* we need to look up vmas on page aligned addresses */
 314         count += addr & (~CFS_PAGE_MASK);
 315         addr &= CFS_PAGE_MASK;
 316
 317         while ((vma = our_vma(addr, count)) != NULL) {
 318                 LASSERT(vma->vm_file);
 319
 320                 inode = vma->vm_file->f_dentry->d_inode;
 321                 policy_from_vma(&policy, vma, addr, count);
 322                 node = ll_node_from_inode(inode, policy.l_extent.start,
 323                                           policy.l_extent.end,
 324                                           mode_from_vma(vma));
 325                 if (IS_ERR(node)) {
 326                         CERROR("not enough mem for lock_tree_node!\n");
 327                         RETURN(-ENOMEM);
 328                 }
 329                 lt_insert(tree, node);
 330
 331                 if (vma->vm_end - addr >= count)
 332                         break;
 333                 count -= vma->vm_end - addr;
 334                 addr = vma->vm_end;
 335         }
 336         RETURN(0);
 337 }
 338
 339 /* FIXME: there is a pagefault race goes as follow (only 2.4):
 340  * 1. A user process on node A accesses a portion of a mapped file,
 341  *    resulting in a page fault.  The pagefault handler invokes the
 342  *    ll_nopage function, which reads the page into memory.
 343  * 2. A user process on node B writes to the same portion of the file
 344  *    (either via mmap or write()), that cause node A to cancel the
 345  *    lock and truncate the page.
 346  * 3. Node A then executes the rest of do_no_page(), entering the
 347  *    now-invalid page into the PTEs.
 348  *
 349  * Make the whole do_no_page as a hook to cover both the page cache
 350  * and page mapping installing with dlm lock would eliminate this race.
 351  *
 352  * In 2.6, the truncate_count of address_space can cover this race.
 353  */
 354 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
 355 struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
 356                        int *type)
 357 #else
 358 struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
 359                        int type /* unused */)
 360 #endif
 361 {
 362         struct file *filp = vma->vm_file;
 363         struct ll_file_data *fd = LUSTRE_FPRIVATE(filp);
 364         struct inode *inode = filp->f_dentry->d_inode;
 365         struct lustre_handle lockh = { 0 };
 366         ldlm_policy_data_t policy;
 367         ldlm_mode_t mode;
 368         struct page *page = NULL;
 369         struct ll_inode_info *lli = ll_i2info(inode);
 370         struct lov_stripe_md *lsm;
 371         struct ost_lvb lvb;
 372         __u64 kms, old_mtime;
 373         unsigned long pgoff, size, rand_read, seq_read;
 374         int rc = 0;
 375         ENTRY;
 376
 377         if (lli->lli_smd == NULL) {
 378                 CERROR("No lsm on fault?\n");
 379                 RETURN(NULL);
 380         }
 381
 382         /* start and end the lock on the first and last bytes in the page */
 383         policy_from_vma(&policy, vma, address, CFS_PAGE_SIZE);
 384
 385         CDEBUG(D_MMAP, "nopage vma %p inode %lu, locking ["LPU64", "LPU64"]\n",
 386                vma, inode->i_ino, policy.l_extent.start, policy.l_extent.end);
 387
 388         mode = mode_from_vma(vma);
 389         old_mtime = LTIME_S(inode->i_mtime);
 390
 391         lsm = lli->lli_smd;
 392         rc = ll_extent_lock(fd, inode, lsm, mode, &policy,
 393                             &lockh, LDLM_FL_CBPENDING | LDLM_FL_NO_LRU);
 394         if (rc != 0)
 395                 RETURN(NULL);
 396
 397         if (vma->vm_flags & VM_EXEC && LTIME_S(inode->i_mtime) != old_mtime)
 398                 CWARN("binary changed. inode %lu\n", inode->i_ino);
 399
 400         lov_stripe_lock(lsm);
 401         inode_init_lvb(inode, &lvb);
 402         obd_merge_lvb(ll_i2obdexp(inode), lsm, &lvb, 1);
 403         kms = lvb.lvb_size;
 404
 405         pgoff = ((address - vma->vm_start) >> CFS_PAGE_SHIFT) + vma->vm_pgoff;
 406         size = (kms + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
 407
 408         if (pgoff >= size) {
 409                 lov_stripe_unlock(lsm);
 410                 ll_glimpse_size(inode, LDLM_FL_BLOCK_GRANTED);
 411         } else {
 412                 /* XXX change inode size without ll_inode_size_lock() held!
 413                  *     there is a race condition with truncate path. (see
 414                  *     ll_extent_lock) */
 415                /* region is within kms and, hence, within real file size (A).
 416                  * We need to increase i_size to cover the read region so that
 417                  * generic_file_read() will do its job, but that doesn't mean
 418                  * the kms size is _correct_, it is only the _minimum_ size.
 419                  * If someone does a stat they will get the correct size which
 420                  * will always be >= the kms value here.  b=11081 */
 421                 if (inode->i_size < kms) {
 422                         inode->i_size = kms;
 423                          CDEBUG(D_INODE, "ino=%lu, updating i_size %llu\n",
 424                                inode->i_ino, inode->i_size);
 425                 }
 426                 lov_stripe_unlock(lsm);
 427         }
 428
 429         /* If mapping is writeable, adjust kms to cover this page,
 430          * but do not extend kms beyond actual file size.
 431          * policy.l_extent.end is set to the end of the page by policy_from_vma
 432          * bug 10919 */
 433         lov_stripe_lock(lsm);
 434         if (mode == LCK_PW)
 435                 obd_adjust_kms(ll_i2obdexp(inode), lsm,
 436                                min_t(loff_t, policy.l_extent.end,inode->i_size),
 437                                0);
 438         lov_stripe_unlock(lsm);
 439
 440         /* disable VM_SEQ_READ and use VM_RAND_READ to make sure that
 441          * the kernel will not read other pages not covered by ldlm in
 442          * filemap_nopage. we do our readahead in ll_readpage.
 443          */
 444         rand_read = vma->vm_flags & VM_RAND_READ;
 445         seq_read = vma->vm_flags & VM_SEQ_READ;
 446         vma->vm_flags &= ~ VM_SEQ_READ;
 447         vma->vm_flags |= VM_RAND_READ;
 448
 449         page = filemap_nopage(vma, address, type);
 450         LL_CDEBUG_PAGE(D_PAGE, page, "got addr %lu type %lx\n", address,
 451                        (long)type);
 452         vma->vm_flags &= ~VM_RAND_READ;
 453         vma->vm_flags |= (rand_read | seq_read);
 454
 455         ll_extent_unlock(fd, inode, ll_i2info(inode)->lli_smd, mode, &lockh);
 456         RETURN(page);
 457 }
 458
 459 /* To avoid cancel the locks covering mmapped region for lock cache pressure,
 460  * we track the mapped vma count by lli_mmap_cnt.
 461  * ll_vm_open():  when first vma is linked, split locks from lru.
 462  * ll_vm_close(): when last vma is unlinked, join all this file's locks to lru.
 463  *
 464  * XXX we don't check the if the region of vma/lock for performance.
 465  */
 466 static void ll_vm_open(struct vm_area_struct * vma)
 467 {
 468         struct inode *inode = vma->vm_file->f_dentry->d_inode;
 469         struct ll_inode_info *lli = ll_i2info(inode);
 470         ENTRY;
 471
 472         LASSERT(vma->vm_file);
 473
 474         spin_lock(&lli->lli_lock);
 475         LASSERT(atomic_read(&lli->lli_mmap_cnt) >= 0);
 476
 477         atomic_inc(&lli->lli_mmap_cnt);
 478         if (atomic_read(&lli->lli_mmap_cnt) == 1) {
 479                 struct lov_stripe_md *lsm = lli->lli_smd;
 480                 struct ll_sb_info *sbi = ll_i2sbi(inode);
 481                 int count;
 482
 483                 spin_unlock(&lli->lli_lock);
 484
 485                 if (!lsm)
 486                         return;
 487                 count = obd_join_lru(sbi->ll_osc_exp, lsm, 0);
 488                 VMA_DEBUG(vma, "split %d unused locks from lru\n", count);
 489         } else {
 490                 spin_unlock(&lli->lli_lock);
 491         }
 492
 493 }
 494
 495 static void ll_vm_close(struct vm_area_struct *vma)
 496 {
 497         struct inode *inode = vma->vm_file->f_dentry->d_inode;
 498         struct ll_inode_info *lli = ll_i2info(inode);
 499         ENTRY;
 500
 501         LASSERT(vma->vm_file);
 502
 503         spin_lock(&lli->lli_lock);
 504         LASSERT(atomic_read(&lli->lli_mmap_cnt) > 0);
 505
 506         atomic_dec(&lli->lli_mmap_cnt);
 507         if (atomic_read(&lli->lli_mmap_cnt) == 0) {
 508                 struct lov_stripe_md *lsm = lli->lli_smd;
 509                 struct ll_sb_info *sbi = ll_i2sbi(inode);
 510                 int count;
 511
 512                 spin_unlock(&lli->lli_lock);
 513
 514                 if (!lsm)
 515                         return;
 516                 count = obd_join_lru(sbi->ll_osc_exp, lsm, 1);
 517                 VMA_DEBUG(vma, "join %d unused locks to lru\n", count);
 518         } else {
 519                 spin_unlock(&lli->lli_lock);
 520         }
 521 }
 522
 523 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
 524 #ifndef HAVE_FILEMAP_POPULATE
 525 static int (*filemap_populate)(struct vm_area_struct * area, unsigned long address, unsigned long len, pgprot_t prot, unsigned long pgoff, int nonblock);
 526 #endif
 527 static int ll_populate(struct vm_area_struct *area, unsigned long address,
 528                        unsigned long len, pgprot_t prot, unsigned long pgoff,
 529                        int nonblock)
 530 {
 531         int rc = 0;
 532         ENTRY;
 533
 534         /* always set nonblock as true to avoid page read ahead */
 535         rc = filemap_populate(area, address, len, prot, pgoff, 1);
 536         RETURN(rc);
 537 }
 538 #endif
 539
 540 /* return the user space pointer that maps to a file offset via a vma */
 541 static inline unsigned long file_to_user(struct vm_area_struct *vma, __u64 byte)
 542 {
 543         return vma->vm_start + (byte - ((__u64)vma->vm_pgoff << CFS_PAGE_SHIFT));
 544
 545 }
 546
 547 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 548 /* [first, last] are the byte offsets affected.
 549  * vm_{start, end} are user addresses of the first byte of the mapping and
 550  *      the next byte beyond it
 551  * vm_pgoff is the page index of the first byte in the mapping */
 552 static void teardown_vmas(struct vm_area_struct *vma, __u64 first,
 553                           __u64 last)
 554 {
 555         unsigned long address, len;
 556         for (; vma ; vma = vma->vm_next_share) {
 557                 if (last >> CFS_PAGE_SHIFT < vma->vm_pgoff)
 558                         continue;
 559                 if (first >> CFS_PAGE_SHIFT >= (vma->vm_pgoff +
 560                     ((vma->vm_end - vma->vm_start) >> CFS_PAGE_SHIFT)))
 561                         continue;
 562
 563                 /* XXX in case of unmap the cow pages of a running file,
 564                  * don't unmap these private writeable mapping here!
 565                  * though that will break private mappping a little.
 566                  *
 567                  * the clean way is to check the mapping of every page
 568                  * and just unmap the non-cow pages, just like
 569                  * unmap_mapping_range() with even_cow=0 in kernel 2.6.
 570                  */
 571                 if (!(vma->vm_flags & VM_SHARED) &&
 572                     (vma->vm_flags & VM_WRITE))
 573                         continue;
 574
 575                 address = max((unsigned long)vma->vm_start,
 576                               file_to_user(vma, first));
 577                 len = min((unsigned long)vma->vm_end,
 578                           file_to_user(vma, last) + 1) - address;
 579
 580                 VMA_DEBUG(vma, "zapping vma [first="LPU64" last="LPU64" "
 581                           "address=%ld len=%ld]\n", first, last, address, len);
 582                 LASSERT(len > 0);
 583                 ll_zap_page_range(vma, address, len);
 584         }
 585 }
 586 #endif
 587
 588 /* XXX put nice comment here.  talk about __free_pte -> dirty pages and
 589  * nopage's reference passing to the pte */
 590 int ll_teardown_mmaps(struct address_space *mapping, __u64 first, __u64 last)
 591 {
 592         int rc = -ENOENT;
 593         ENTRY;
 594
 595         LASSERTF(last > first, "last "LPU64" first "LPU64"\n", last, first);
 596 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
 597         if (mapping_mapped(mapping)) {
 598                 rc = 0;
 599                 unmap_mapping_range(mapping, first + CFS_PAGE_SIZE - 1,
 600                                     last - first + 1, 0);
 601         }
 602 #else
 603         spin_lock(&mapping->i_shared_lock);
 604         if (mapping->i_mmap != NULL) {
 605                 rc = 0;
 606                 teardown_vmas(mapping->i_mmap, first, last);
 607         }
 608         if (mapping->i_mmap_shared != NULL) {
 609                 rc = 0;
 610                 teardown_vmas(mapping->i_mmap_shared, first, last);
 611         }
 612         spin_unlock(&mapping->i_shared_lock);
 613 #endif
 614         RETURN(rc);
 615 }
 616
 617 static struct vm_operations_struct ll_file_vm_ops = {
 618         .nopage         = ll_nopage,
 619         .open           = ll_vm_open,
 620         .close          = ll_vm_close,
 621 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
 622         .populate       = ll_populate,
 623 #endif
 624 };
 625
 626 int ll_file_mmap(struct file * file, struct vm_area_struct * vma)
 627 {
 628         int rc;
 629         ENTRY;
 630
 631         ll_vfs_ops_tally(ll_i2sbi(file->f_dentry->d_inode), VFS_OPS_MMAP);
 632         rc = generic_file_mmap(file, vma);
 633         if (rc == 0) {
 634 #if !defined(HAVE_FILEMAP_POPULATE) && \
 635     (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
 636                 if (!filemap_populate)
 637                         filemap_populate = vma->vm_ops->populate;
 638 #endif
 639                 vma->vm_ops = &ll_file_vm_ops;
 640                 vma->vm_ops->open(vma);
 641                 /* update the inode's size and mtime */
 642                 rc = ll_glimpse_size(file->f_dentry->d_inode, 0);
 643         }
 644
 645         RETURN(rc);
 646 }