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
  22 #ifndef AUTOCONF_INCLUDED
  23 #include <linux/config.h>
  24 #endif
  25 #include <linux/kernel.h>
  26 #include <linux/mm.h>
  27 #include <linux/string.h>
  28 #include <linux/stat.h>
  29 #include <linux/errno.h>
  30 #include <linux/smp_lock.h>
  31 #include <linux/unistd.h>
  32 #include <linux/version.h>
  33 #include <asm/system.h>
  34 #include <asm/uaccess.h>
  35
  36 #include <linux/fs.h>
  37 #include <linux/stat.h>
  38 #include <asm/uaccess.h>
  39 #include <asm/segment.h>
  40 #include <linux/mm.h>
  41 #include <linux/pagemap.h>
  42 #include <linux/smp_lock.h>
  43
  44 #define DEBUG_SUBSYSTEM S_LLITE
  45
  46 //#include <lustre_mdc.h>
  47 #include <lustre_lite.h>
  48 #include "llite_internal.h"
  49 #include <linux/lustre_compat25.h>
  50
  51 #define VMA_DEBUG(vma, fmt, arg...)                                     \
  52         CDEBUG(D_MMAP, "vma(%p) start(%ld) end(%ld) pgoff(%ld) inode(%p) "   \
  53                "ino(%lu) iname(%s): " fmt, vma, vma->vm_start, vma->vm_end,  \
  54                vma->vm_pgoff, vma->vm_file->f_dentry->d_inode,               \
  55                vma->vm_file->f_dentry->d_inode->i_ino,                       \
  56                vma->vm_file->f_dentry->d_iname, ## arg);                     \
  57
  58
  59 struct ll_lock_tree_node {
  60         rb_node_t               lt_node;
  61         struct list_head        lt_locked_item;
  62         __u64                   lt_oid;
  63         ldlm_policy_data_t      lt_policy;
  64         struct lustre_handle    lt_lockh;
  65         ldlm_mode_t             lt_mode;
  66         struct inode           *lt_inode;
  67 };
  68
  69 int lt_get_mmap_locks(struct ll_lock_tree *tree,
  70                       unsigned long addr, size_t count);
  71
  72 struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
  73                        int *type);
  74
  75 struct ll_lock_tree_node * ll_node_from_inode(struct inode *inode, __u64 start,
  76                                               __u64 end, ldlm_mode_t mode)
  77 {
  78         struct ll_lock_tree_node *node;
  79
  80         OBD_ALLOC(node, sizeof(*node));
  81         if (node == NULL)
  82                 RETURN(ERR_PTR(-ENOMEM));
  83
  84         node->lt_inode = inode;
  85         node->lt_oid = ll_i2info(inode)->lli_smd->lsm_object_id;
  86         node->lt_policy.l_extent.start = start;
  87         node->lt_policy.l_extent.end = end;
  88         memset(&node->lt_lockh, 0, sizeof(node->lt_lockh));
  89         INIT_LIST_HEAD(&node->lt_locked_item);
  90         node->lt_mode = mode;
  91
  92         return node;
  93 }
  94
  95 int lt_compare(struct ll_lock_tree_node *one, struct ll_lock_tree_node *two)
  96 {
  97         /* To avoid multiple fs deadlock */
  98         if (one->lt_inode->i_sb->s_dev < two->lt_inode->i_sb->s_dev)
  99                 return -1;
 100         if (one->lt_inode->i_sb->s_dev > two->lt_inode->i_sb->s_dev)
 101                 return 1;
 102
 103         if (one->lt_oid < two->lt_oid)
 104                 return -1;
 105         if (one->lt_oid > two->lt_oid)
 106                 return 1;
 107
 108         if (one->lt_policy.l_extent.end < two->lt_policy.l_extent.start)
 109                 return -1;
 110         if (one->lt_policy.l_extent.start > two->lt_policy.l_extent.end)
 111                 return 1;
 112
 113         return 0; /* they are the same object and overlap */
 114 }
 115
 116 static void lt_merge(struct ll_lock_tree_node *dst,
 117                      struct ll_lock_tree_node *src)
 118 {
 119         dst->lt_policy.l_extent.start = min(dst->lt_policy.l_extent.start,
 120                                             src->lt_policy.l_extent.start);
 121         dst->lt_policy.l_extent.end = max(dst->lt_policy.l_extent.end,
 122                                           src->lt_policy.l_extent.end);
 123
 124         /* XXX could be a real call to the dlm to find superset modes */
 125         if (src->lt_mode == LCK_PW && dst->lt_mode != LCK_PW)
 126                 dst->lt_mode = LCK_PW;
 127 }
 128
 129 static void lt_insert(struct ll_lock_tree *tree,
 130                       struct ll_lock_tree_node *node)
 131 {
 132         struct ll_lock_tree_node *walk;
 133         rb_node_t **p, *parent;
 134         ENTRY;
 135
 136 restart:
 137         p = &tree->lt_root.rb_node;
 138         parent = NULL;
 139         while (*p) {
 140                 parent = *p;
 141                 walk = rb_entry(parent, struct ll_lock_tree_node, lt_node);
 142                 switch (lt_compare(node, walk)) {
 143                 case -1:
 144                         p = &(*p)->rb_left;
 145                         break;
 146                 case 1:
 147                         p = &(*p)->rb_right;
 148                         break;
 149                 case 0:
 150                         lt_merge(node, walk);
 151                         rb_erase(&walk->lt_node, &tree->lt_root);
 152                         OBD_FREE(walk, sizeof(*walk));
 153                         goto restart;
 154                         break;
 155                 default:
 156                         LBUG();
 157                         break;
 158                 }
 159         }
 160         rb_link_node(&node->lt_node, parent, p);
 161         rb_insert_color(&node->lt_node, &tree->lt_root);
 162         EXIT;
 163 }
 164
 165 static struct ll_lock_tree_node *lt_least_node(struct ll_lock_tree *tree)
 166 {
 167         rb_node_t *rbnode;
 168         struct ll_lock_tree_node *node = NULL;
 169
 170         for ( rbnode = tree->lt_root.rb_node; rbnode != NULL;
 171               rbnode = rbnode->rb_left) {
 172                 if (rbnode->rb_left == NULL) {
 173                         node = rb_entry(rbnode, struct ll_lock_tree_node,
 174                                         lt_node);
 175                         break;
 176                 }
 177         }
 178         RETURN(node);
 179 }
 180
 181 int ll_tree_unlock(struct ll_lock_tree *tree)
 182 {
 183         struct ll_lock_tree_node *node;
 184         struct list_head *pos, *n;
 185         struct inode *inode;
 186         int rc = 0;
 187         ENTRY;
 188
 189         list_for_each_safe(pos, n, &tree->lt_locked_list) {
 190                 node = list_entry(pos, struct ll_lock_tree_node,
 191                                   lt_locked_item);
 192
 193                 inode = node->lt_inode;
 194                 rc = ll_extent_unlock(tree->lt_fd, inode,
 195                                       ll_i2info(inode)->lli_smd, node->lt_mode,
 196                                       &node->lt_lockh);
 197                 if (rc != 0) {
 198                         /* XXX better message */
 199                         CERROR("couldn't unlock %d\n", rc);
 200                 }
 201                 list_del(&node->lt_locked_item);
 202                 OBD_FREE(node, sizeof(*node));
 203         }
 204
 205         while ((node = lt_least_node(tree))) {
 206                 rb_erase(&node->lt_node, &tree->lt_root);
 207                 OBD_FREE(node, sizeof(*node));
 208         }
 209
 210         RETURN(rc);
 211 }
 212
 213 int ll_tree_lock(struct ll_lock_tree *tree,
 214                  struct ll_lock_tree_node *first_node,
 215                  const char *buf, size_t count, int ast_flags)
 216 {
 217         struct ll_lock_tree_node *node;
 218         int rc = 0;
 219         ENTRY;
 220
 221         tree->lt_root.rb_node = NULL;
 222         INIT_LIST_HEAD(&tree->lt_locked_list);
 223         if (first_node != NULL)
 224                 lt_insert(tree, first_node);
 225
 226         /* To avoid such subtle deadlock case: client1 try to read file1 to
 227          * mmapped file2, on the same time, client2 try to read file2 to
 228          * mmapped file1.*/
 229         rc = lt_get_mmap_locks(tree, (unsigned long)buf, count);
 230         if (rc)
 231                 GOTO(out, rc);
 232
 233         while ((node = lt_least_node(tree))) {
 234                 struct inode *inode = node->lt_inode;
 235                 rc = ll_extent_lock(tree->lt_fd, inode,
 236                                     ll_i2info(inode)->lli_smd, node->lt_mode,
 237                                     &node->lt_policy, &node->lt_lockh,
 238                                     ast_flags);
 239                 if (rc != 0)
 240                         GOTO(out, rc);
 241
 242                 rb_erase(&node->lt_node, &tree->lt_root);
 243                 list_add_tail(&node->lt_locked_item, &tree->lt_locked_list);
 244         }
 245         RETURN(rc);
 246 out:
 247         ll_tree_unlock(tree);
 248         RETURN(rc);
 249 }
 250
 251 static ldlm_mode_t mode_from_vma(struct vm_area_struct *vma)
 252 {
 253         /* we only want to hold PW locks if the mmap() can generate
 254          * writes back to the file and that only happens in shared
 255          * writable vmas */
 256         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
 257                 return LCK_PW;
 258         return LCK_PR;
 259 }
 260
 261 static void policy_from_vma(ldlm_policy_data_t *policy,
 262                             struct vm_area_struct *vma, unsigned long addr,
 263                             size_t count)
 264 {
 265         policy->l_extent.start = ((addr - vma->vm_start) & CFS_PAGE_MASK) +
 266                                  (vma->vm_pgoff << CFS_PAGE_SHIFT);
 267         policy->l_extent.end = (policy->l_extent.start + count - 1) |
 268                                ~CFS_PAGE_MASK;
 269 }
 270
 271 static struct vm_area_struct * our_vma(unsigned long addr, size_t count)
 272 {
 273         struct mm_struct *mm = current->mm;
 274         struct vm_area_struct *vma, *ret = NULL;
 275         ENTRY;
 276
 277         /* No MM (e.g. NFS)? No vmas too. */
 278         if (!mm)
 279                 RETURN(NULL);
 280
 281         spin_lock(&mm->page_table_lock);
 282         for(vma = find_vma(mm, addr);
 283             vma != NULL && vma->vm_start < (addr + count); vma = vma->vm_next) {
 284                 if (vma->vm_ops && vma->vm_ops->nopage == ll_nopage &&
 285                     vma->vm_flags & VM_SHARED) {
 286                         ret = vma;
 287                         break;
 288                 }
 289         }
 290         spin_unlock(&mm->page_table_lock);
 291         RETURN(ret);
 292 }
 293
 294 int lt_get_mmap_locks(struct ll_lock_tree *tree,
 295                       unsigned long addr, size_t count)
 296 {
 297         struct vm_area_struct *vma;
 298         struct ll_lock_tree_node *node;
 299         ldlm_policy_data_t policy;
 300         struct inode *inode;
 301         ENTRY;
 302
 303         if (count == 0)
 304                 RETURN(0);
 305
 306         /* we need to look up vmas on page aligned addresses */
 307         count += addr & (~CFS_PAGE_MASK);
 308         addr &= CFS_PAGE_MASK;
 309
 310         while ((vma = our_vma(addr, count)) != NULL) {
 311                 LASSERT(vma->vm_file);
 312
 313                 inode = vma->vm_file->f_dentry->d_inode;
 314                 policy_from_vma(&policy, vma, addr, count);
 315                 node = ll_node_from_inode(inode, policy.l_extent.start,
 316                                           policy.l_extent.end,
 317                                           mode_from_vma(vma));
 318                 if (IS_ERR(node)) {
 319                         CERROR("not enough mem for lock_tree_node!\n");
 320                         RETURN(-ENOMEM);
 321                 }
 322                 lt_insert(tree, node);
 323
 324                 if (vma->vm_end - addr >= count)
 325                         break;
 326                 count -= vma->vm_end - addr;
 327                 addr = vma->vm_end;
 328         }
 329         RETURN(0);
 330 }
 331
 332 /* FIXME: there is a pagefault race goes as follow (only 2.4):
 333  * 1. A user process on node A accesses a portion of a mapped file,
 334  *    resulting in a page fault.  The pagefault handler invokes the
 335  *    ll_nopage function, which reads the page into memory.
 336  * 2. A user process on node B writes to the same portion of the file
 337  *    (either via mmap or write()), that cause node A to cancel the
 338  *    lock and truncate the page.
 339  * 3. Node A then executes the rest of do_no_page(), entering the
 340  *    now-invalid page into the PTEs.
 341  *
 342  * Make the whole do_no_page as a hook to cover both the page cache
 343  * and page mapping installing with dlm lock would eliminate this race.
 344  *
 345  * In 2.6, the truncate_count of address_space can cover this race.
 346  */
 347 struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
 348                        int *type)
 349 {
 350         struct file *filp = vma->vm_file;
 351         struct ll_file_data *fd = LUSTRE_FPRIVATE(filp);
 352         struct inode *inode = filp->f_dentry->d_inode;
 353         struct lustre_handle lockh = { 0 };
 354         ldlm_policy_data_t policy;
 355         ldlm_mode_t mode;
 356         struct page *page = NULL;
 357         struct ll_inode_info *lli = ll_i2info(inode);
 358         struct lov_stripe_md *lsm;
 359         struct ost_lvb lvb;
 360         __u64 kms, old_mtime;
 361         unsigned long pgoff, size, rand_read, seq_read;
 362         int rc = 0;
 363         ENTRY;
 364
 365         if (lli->lli_smd == NULL) {
 366                 CERROR("No lsm on fault?\n");
 367                 RETURN(NULL);
 368         }
 369
 370         /* start and end the lock on the first and last bytes in the page */
 371         policy_from_vma(&policy, vma, address, CFS_PAGE_SIZE);
 372
 373         CDEBUG(D_MMAP, "nopage vma %p inode %lu, locking ["LPU64", "LPU64"]\n",
 374                vma, inode->i_ino, policy.l_extent.start, policy.l_extent.end);
 375
 376         mode = mode_from_vma(vma);
 377         old_mtime = LTIME_S(inode->i_mtime);
 378
 379         lsm = lli->lli_smd;
 380         rc = ll_extent_lock(fd, inode, lsm, mode, &policy,
 381                             &lockh, LDLM_FL_CBPENDING | LDLM_FL_NO_LRU);
 382         if (rc != 0)
 383                 RETURN(NULL);
 384
 385         if (vma->vm_flags & VM_EXEC && LTIME_S(inode->i_mtime) != old_mtime)
 386                 CWARN("binary changed. inode %lu\n", inode->i_ino);
 387
 388         lov_stripe_lock(lsm);
 389         inode_init_lvb(inode, &lvb);
 390         obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1);
 391         kms = lvb.lvb_size;
 392
 393         pgoff = ((address - vma->vm_start) >> CFS_PAGE_SHIFT) + vma->vm_pgoff;
 394         size = (kms + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
 395
 396         if (pgoff >= size) {
 397                 lov_stripe_unlock(lsm);
 398                 ll_glimpse_size(inode, LDLM_FL_BLOCK_GRANTED);
 399         } else {
 400                 /* XXX change inode size without ll_inode_size_lock() held!
 401                  *     there is a race condition with truncate path. (see
 402                  *     ll_extent_lock) */
 403                 /* XXX i_size_write() is not used because it is not safe to
 404                  *     take the ll_inode_size_lock() due to a potential lock
 405                  *     inversion (bug 6077).  And since it's not safe to use
 406                  *     i_size_write() without a covering mutex we do the
 407                  *     assignment directly.  It is not critical that the
 408                  *     size be correct. */
 409                 /* region is within kms and, hence, within real file size (A).
 410                  * We need to increase i_size to cover the read region so that
 411                  * generic_file_read() will do its job, but that doesn't mean
 412                  * the kms size is _correct_, it is only the _minimum_ size.
 413                  * If someone does a stat they will get the correct size which
 414                  * will always be >= the kms value here.  b=11081 */
 415                 if (i_size_read(inode) < kms) {
 416                         inode->i_size = kms;
 417                         CDEBUG(D_INODE, "ino=%lu, updating i_size %llu\n",
 418                                inode->i_ino, i_size_read(inode));
 419                 }
 420                 lov_stripe_unlock(lsm);
 421         }
 422
 423         /* If mapping is writeable, adjust kms to cover this page,
 424          * but do not extend kms beyond actual file size.
 425          * policy.l_extent.end is set to the end of the page by policy_from_vma
 426          * bug 10919 */
 427         lov_stripe_lock(lsm);
 428         if (mode == LCK_PW)
 429                 obd_adjust_kms(ll_i2dtexp(inode), lsm,
 430                                min_t(loff_t, policy.l_extent.end + 1,
 431                                i_size_read(inode)), 0);
 432         lov_stripe_unlock(lsm);
 433
 434         /* disable VM_SEQ_READ and use VM_RAND_READ to make sure that
 435          * the kernel will not read other pages not covered by ldlm in
 436          * filemap_nopage. we do our readahead in ll_readpage.
 437          */
 438         rand_read = vma->vm_flags & VM_RAND_READ;
 439         seq_read = vma->vm_flags & VM_SEQ_READ;
 440         vma->vm_flags &= ~ VM_SEQ_READ;
 441         vma->vm_flags |= VM_RAND_READ;
 442
 443         page = filemap_nopage(vma, address, type);
 444         LL_CDEBUG_PAGE(D_PAGE, page, "got addr %lu type %lx\n", address,
 445                        (long)type);
 446         vma->vm_flags &= ~VM_RAND_READ;
 447         vma->vm_flags |= (rand_read | seq_read);
 448
 449         ll_extent_unlock(fd, inode, ll_i2info(inode)->lli_smd, mode, &lockh);
 450         RETURN(page);
 451 }
 452
 453 /* To avoid cancel the locks covering mmapped region for lock cache pressure,
 454  * we track the mapped vma count by lli_mmap_cnt.
 455  * ll_vm_open():  when first vma is linked, split locks from lru.
 456  * ll_vm_close(): when last vma is unlinked, join all this file's locks to lru.
 457  *
 458  * XXX we don't check the if the region of vma/lock for performance.
 459  */
 460 static void ll_vm_open(struct vm_area_struct * vma)
 461 {
 462         struct inode *inode = vma->vm_file->f_dentry->d_inode;
 463         struct ll_inode_info *lli = ll_i2info(inode);
 464         ENTRY;
 465
 466         LASSERT(vma->vm_file);
 467
 468         spin_lock(&lli->lli_lock);
 469         LASSERT(atomic_read(&lli->lli_mmap_cnt) >= 0);
 470
 471         atomic_inc(&lli->lli_mmap_cnt);
 472         if (atomic_read(&lli->lli_mmap_cnt) == 1) {
 473                 struct lov_stripe_md *lsm = lli->lli_smd;
 474                 struct ll_sb_info *sbi = ll_i2sbi(inode);
 475                 int count;
 476
 477                 spin_unlock(&lli->lli_lock);
 478
 479                 if (!lsm)
 480                         return;
 481                 count = obd_join_lru(sbi->ll_dt_exp, lsm, 0);
 482                 VMA_DEBUG(vma, "split %d unused locks from lru\n", count);
 483         } else {
 484                 spin_unlock(&lli->lli_lock);
 485         }
 486
 487 }
 488
 489 static void ll_vm_close(struct vm_area_struct *vma)
 490 {
 491         struct inode *inode = vma->vm_file->f_dentry->d_inode;
 492         struct ll_inode_info *lli = ll_i2info(inode);
 493         ENTRY;
 494
 495         LASSERT(vma->vm_file);
 496
 497         spin_lock(&lli->lli_lock);
 498         LASSERT(atomic_read(&lli->lli_mmap_cnt) > 0);
 499
 500         atomic_dec(&lli->lli_mmap_cnt);
 501         if (atomic_read(&lli->lli_mmap_cnt) == 0) {
 502                 struct lov_stripe_md *lsm = lli->lli_smd;
 503                 struct ll_sb_info *sbi = ll_i2sbi(inode);
 504                 int count;
 505
 506                 spin_unlock(&lli->lli_lock);
 507
 508                 if (!lsm)
 509                         return;
 510                 count = obd_join_lru(sbi->ll_dt_exp, lsm, 1);
 511                 VMA_DEBUG(vma, "join %d unused locks to lru\n", count);
 512         } else {
 513                 spin_unlock(&lli->lli_lock);
 514         }
 515 }
 516
 517 #ifndef HAVE_FILEMAP_POPULATE
 518 static int (*filemap_populate)(struct vm_area_struct * area, unsigned long address, unsigned long len, pgprot_t prot, unsigned long pgoff, int nonblock);
 519 #endif
 520 static int ll_populate(struct vm_area_struct *area, unsigned long address,
 521                        unsigned long len, pgprot_t prot, unsigned long pgoff,
 522                        int nonblock)
 523 {
 524         int rc = 0;
 525         ENTRY;
 526
 527         /* always set nonblock as true to avoid page read ahead */
 528         rc = filemap_populate(area, address, len, prot, pgoff, 1);
 529         RETURN(rc);
 530 }
 531
 532 /* return the user space pointer that maps to a file offset via a vma */
 533 static inline unsigned long file_to_user(struct vm_area_struct *vma, __u64 byte)
 534 {
 535         return vma->vm_start + (byte - ((__u64)vma->vm_pgoff << CFS_PAGE_SHIFT));
 536
 537 }
 538
 539 /* XXX put nice comment here.  talk about __free_pte -> dirty pages and
 540  * nopage's reference passing to the pte */
 541 int ll_teardown_mmaps(struct address_space *mapping, __u64 first, __u64 last)
 542 {
 543         int rc = -ENOENT;
 544         ENTRY;
 545
 546         LASSERTF(last > first, "last "LPU64" first "LPU64"\n", last, first);
 547         if (mapping_mapped(mapping)) {
 548                 rc = 0;
 549                 unmap_mapping_range(mapping, first + CFS_PAGE_SIZE - 1,
 550                                     last - first + 1, 0);
 551         }
 552
 553         RETURN(rc);
 554 }
 555
 556 static struct vm_operations_struct ll_file_vm_ops = {
 557         .nopage         = ll_nopage,
 558         .open           = ll_vm_open,
 559         .close          = ll_vm_close,
 560         .populate       = ll_populate,
 561 };
 562
 563 int ll_file_mmap(struct file * file, struct vm_area_struct * vma)
 564 {
 565         int rc;
 566         ENTRY;
 567
 568         ll_stats_ops_tally(ll_i2sbi(file->f_dentry->d_inode), LPROC_LL_MAP, 1);
 569         rc = generic_file_mmap(file, vma);
 570         if (rc == 0) {
 571 #if !defined(HAVE_FILEMAP_POPULATE)
 572                 if (!filemap_populate)
 573                         filemap_populate = vma->vm_ops->populate;
 574 #endif
 575                 vma->vm_ops = &ll_file_vm_ops;
 576                 vma->vm_ops->open(vma);
 577                 /* update the inode's size and mtime */
 578                 rc = ll_glimpse_size(file->f_dentry->d_inode, 0);
 579         }
 580
 581         RETURN(rc);
 582 }