fs/ocfs2/file.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * file.c
   5  *
   6  * File open, close, extend, truncate
   7  *
   8  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2 of the License, or (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public
  21  * License along with this program; if not, write to the
  22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  23  * Boston, MA 021110-1307, USA.
  24  */
  25
  26 #include <linux/capability.h>
  27 #include <linux/fs.h>
  28 #include <linux/types.h>
  29 #include <linux/slab.h>
  30 #include <linux/highmem.h>
  31 #include <linux/pagemap.h>
  32 #include <linux/uio.h>
  33 #include <linux/sched.h>
  34 #include <linux/splice.h>
  35 #include <linux/mount.h>
  36 #include <linux/writeback.h>
  37
  38 #define MLOG_MASK_PREFIX ML_INODE
  39 #include <cluster/masklog.h>
  40
  41 #include "ocfs2.h"
  42
  43 #include "alloc.h"
  44 #include "aops.h"
  45 #include "dir.h"
  46 #include "dlmglue.h"
  47 #include "extent_map.h"
  48 #include "file.h"
  49 #include "sysfile.h"
  50 #include "inode.h"
  51 #include "ioctl.h"
  52 #include "journal.h"
  53 #include "mmap.h"
  54 #include "suballoc.h"
  55 #include "super.h"
  56
  57 #include "buffer_head_io.h"
  58
  59 static int ocfs2_sync_inode(struct inode *inode)
  60 {
  61         filemap_fdatawrite(inode->i_mapping);
  62         return sync_mapping_buffers(inode->i_mapping);
  63 }
  64
  65 static int ocfs2_file_open(struct inode *inode, struct file *file)
  66 {
  67         int status;
  68         int mode = file->f_flags;
  69         struct ocfs2_inode_info *oi = OCFS2_I(inode);
  70
  71         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
  72                    file->f_path.dentry->d_name.len, file->f_path.dentry->d_name.name);
  73
  74         spin_lock(&oi->ip_lock);
  75
  76         /* Check that the inode hasn't been wiped from disk by another
  77          * node. If it hasn't then we're safe as long as we hold the
  78          * spin lock until our increment of open count. */
  79         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
  80                 spin_unlock(&oi->ip_lock);
  81
  82                 status = -ENOENT;
  83                 goto leave;
  84         }
  85
  86         if (mode & O_DIRECT)
  87                 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
  88
  89         oi->ip_open_count++;
  90         spin_unlock(&oi->ip_lock);
  91         status = 0;
  92 leave:
  93         mlog_exit(status);
  94         return status;
  95 }
  96
  97 static int ocfs2_file_release(struct inode *inode, struct file *file)
  98 {
  99         struct ocfs2_inode_info *oi = OCFS2_I(inode);
 100
 101         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
 102                        file->f_path.dentry->d_name.len,
 103                        file->f_path.dentry->d_name.name);
 104
 105         spin_lock(&oi->ip_lock);
 106         if (!--oi->ip_open_count)
 107                 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
 108         spin_unlock(&oi->ip_lock);
 109
 110         mlog_exit(0);
 111
 112         return 0;
 113 }
 114
 115 static int ocfs2_sync_file(struct file *file,
 116                            struct dentry *dentry,
 117                            int datasync)
 118 {
 119         int err = 0;
 120         journal_t *journal;
 121         struct inode *inode = dentry->d_inode;
 122         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 123
 124         mlog_entry("(0x%p, 0x%p, %d, '%.*s')\n", file, dentry, datasync,
 125                    dentry->d_name.len, dentry->d_name.name);
 126
 127         err = ocfs2_sync_inode(dentry->d_inode);
 128         if (err)
 129                 goto bail;
 130
 131         journal = osb->journal->j_journal;
 132         err = journal_force_commit(journal);
 133
 134 bail:
 135         mlog_exit(err);
 136
 137         return (err < 0) ? -EIO : 0;
 138 }
 139
 140 int ocfs2_should_update_atime(struct inode *inode,
 141                               struct vfsmount *vfsmnt)
 142 {
 143         struct timespec now;
 144         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 145
 146         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
 147                 return 0;
 148
 149         if ((inode->i_flags & S_NOATIME) ||
 150             ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
 151                 return 0;
 152
 153         /*
 154          * We can be called with no vfsmnt structure - NFSD will
 155          * sometimes do this.
 156          *
 157          * Note that our action here is different than touch_atime() -
 158          * if we can't tell whether this is a noatime mount, then we
 159          * don't know whether to trust the value of s_atime_quantum.
 160          */
 161         if (vfsmnt == NULL)
 162                 return 0;
 163
 164         if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
 165             ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
 166                 return 0;
 167
 168         if (vfsmnt->mnt_flags & MNT_RELATIME) {
 169                 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
 170                     (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
 171                         return 1;
 172
 173                 return 0;
 174         }
 175
 176         now = CURRENT_TIME;
 177         if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
 178                 return 0;
 179         else
 180                 return 1;
 181 }
 182
 183 int ocfs2_update_inode_atime(struct inode *inode,
 184                              struct buffer_head *bh)
 185 {
 186         int ret;
 187         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 188         handle_t *handle;
 189
 190         mlog_entry_void();
 191
 192         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 193         if (handle == NULL) {
 194                 ret = -ENOMEM;
 195                 mlog_errno(ret);
 196                 goto out;
 197         }
 198
 199         inode->i_atime = CURRENT_TIME;
 200         ret = ocfs2_mark_inode_dirty(handle, inode, bh);
 201         if (ret < 0)
 202                 mlog_errno(ret);
 203
 204         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
 205 out:
 206         mlog_exit(ret);
 207         return ret;
 208 }
 209
 210 static int ocfs2_set_inode_size(handle_t *handle,
 211                                 struct inode *inode,
 212                                 struct buffer_head *fe_bh,
 213                                 u64 new_i_size)
 214 {
 215         int status;
 216
 217         mlog_entry_void();
 218         i_size_write(inode, new_i_size);
 219         inode->i_blocks = ocfs2_inode_sector_count(inode);
 220         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 221
 222         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
 223         if (status < 0) {
 224                 mlog_errno(status);
 225                 goto bail;
 226         }
 227
 228 bail:
 229         mlog_exit(status);
 230         return status;
 231 }
 232
 233 static int ocfs2_simple_size_update(struct inode *inode,
 234                                     struct buffer_head *di_bh,
 235                                     u64 new_i_size)
 236 {
 237         int ret;
 238         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 239         handle_t *handle = NULL;
 240
 241         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 242         if (handle == NULL) {
 243                 ret = -ENOMEM;
 244                 mlog_errno(ret);
 245                 goto out;
 246         }
 247
 248         ret = ocfs2_set_inode_size(handle, inode, di_bh,
 249                                    new_i_size);
 250         if (ret < 0)
 251                 mlog_errno(ret);
 252
 253         ocfs2_commit_trans(osb, handle);
 254 out:
 255         return ret;
 256 }
 257
 258 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
 259                                      struct inode *inode,
 260                                      struct buffer_head *fe_bh,
 261                                      u64 new_i_size)
 262 {
 263         int status;
 264         handle_t *handle;
 265         struct ocfs2_dinode *di;
 266
 267         mlog_entry_void();
 268
 269         /* TODO: This needs to actually orphan the inode in this
 270          * transaction. */
 271
 272         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 273         if (IS_ERR(handle)) {
 274                 status = PTR_ERR(handle);
 275                 mlog_errno(status);
 276                 goto out;
 277         }
 278
 279         status = ocfs2_journal_access(handle, inode, fe_bh,
 280                                       OCFS2_JOURNAL_ACCESS_WRITE);
 281         if (status < 0) {
 282                 mlog_errno(status);
 283                 goto out_commit;
 284         }
 285
 286         /*
 287          * Do this before setting i_size.
 288          */
 289         status = ocfs2_zero_tail_for_truncate(inode, handle, new_i_size);
 290         if (status) {
 291                 mlog_errno(status);
 292                 goto out_commit;
 293         }
 294
 295         i_size_write(inode, new_i_size);
 296         inode->i_blocks = ocfs2_align_bytes_to_sectors(new_i_size);
 297         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
 298
 299         di = (struct ocfs2_dinode *) fe_bh->b_data;
 300         di->i_size = cpu_to_le64(new_i_size);
 301         di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
 302         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
 303
 304         status = ocfs2_journal_dirty(handle, fe_bh);
 305         if (status < 0)
 306                 mlog_errno(status);
 307
 308 out_commit:
 309         ocfs2_commit_trans(osb, handle);
 310 out:
 311
 312         mlog_exit(status);
 313         return status;
 314 }
 315
 316 static int ocfs2_truncate_file(struct inode *inode,
 317                                struct buffer_head *di_bh,
 318                                u64 new_i_size)
 319 {
 320         int status = 0;
 321         struct ocfs2_dinode *fe = NULL;
 322         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 323         struct ocfs2_truncate_context *tc = NULL;
 324
 325         mlog_entry("(inode = %llu, new_i_size = %llu\n",
 326                    (unsigned long long)OCFS2_I(inode)->ip_blkno,
 327                    (unsigned long long)new_i_size);
 328
 329         fe = (struct ocfs2_dinode *) di_bh->b_data;
 330         if (!OCFS2_IS_VALID_DINODE(fe)) {
 331                 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
 332                 status = -EIO;
 333                 goto bail;
 334         }
 335
 336         mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
 337                         "Inode %llu, inode i_size = %lld != di "
 338                         "i_size = %llu, i_flags = 0x%x\n",
 339                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
 340                         i_size_read(inode),
 341                         (unsigned long long)le64_to_cpu(fe->i_size),
 342                         le32_to_cpu(fe->i_flags));
 343
 344         if (new_i_size > le64_to_cpu(fe->i_size)) {
 345                 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
 346                      (unsigned long long)le64_to_cpu(fe->i_size),
 347                      (unsigned long long)new_i_size);
 348                 status = -EINVAL;
 349                 mlog_errno(status);
 350                 goto bail;
 351         }
 352
 353         mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
 354              (unsigned long long)le64_to_cpu(fe->i_blkno),
 355              (unsigned long long)le64_to_cpu(fe->i_size),
 356              (unsigned long long)new_i_size);
 357
 358         /* lets handle the simple truncate cases before doing any more
 359          * cluster locking. */
 360         if (new_i_size == le64_to_cpu(fe->i_size))
 361                 goto bail;
 362
 363         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 364
 365         /* This forces other nodes to sync and drop their pages. Do
 366          * this even if we have a truncate without allocation change -
 367          * ocfs2 cluster sizes can be much greater than page size, so
 368          * we have to truncate them anyway.  */
 369         status = ocfs2_data_lock(inode, 1);
 370         if (status < 0) {
 371                 up_write(&OCFS2_I(inode)->ip_alloc_sem);
 372
 373                 mlog_errno(status);
 374                 goto bail;
 375         }
 376
 377         unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
 378         truncate_inode_pages(inode->i_mapping, new_i_size);
 379
 380         /* alright, we're going to need to do a full blown alloc size
 381          * change. Orphan the inode so that recovery can complete the
 382          * truncate if necessary. This does the task of marking
 383          * i_size. */
 384         status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
 385         if (status < 0) {
 386                 mlog_errno(status);
 387                 goto bail_unlock_data;
 388         }
 389
 390         status = ocfs2_prepare_truncate(osb, inode, di_bh, &tc);
 391         if (status < 0) {
 392                 mlog_errno(status);
 393                 goto bail_unlock_data;
 394         }
 395
 396         status = ocfs2_commit_truncate(osb, inode, di_bh, tc);
 397         if (status < 0) {
 398                 mlog_errno(status);
 399                 goto bail_unlock_data;
 400         }
 401
 402         /* TODO: orphan dir cleanup here. */
 403 bail_unlock_data:
 404         ocfs2_data_unlock(inode, 1);
 405
 406         up_write(&OCFS2_I(inode)->ip_alloc_sem);
 407
 408 bail:
 409
 410         mlog_exit(status);
 411         return status;
 412 }
 413
 414 /*
 415  * extend allocation only here.
 416  * we'll update all the disk stuff, and oip->alloc_size
 417  *
 418  * expect stuff to be locked, a transaction started and enough data /
 419  * metadata reservations in the contexts.
 420  *
 421  * Will return -EAGAIN, and a reason if a restart is needed.
 422  * If passed in, *reason will always be set, even in error.
 423  */
 424 int ocfs2_do_extend_allocation(struct ocfs2_super *osb,
 425                                struct inode *inode,
 426                                u32 *logical_offset,
 427                                u32 clusters_to_add,
 428                                struct buffer_head *fe_bh,
 429                                handle_t *handle,
 430                                struct ocfs2_alloc_context *data_ac,
 431                                struct ocfs2_alloc_context *meta_ac,
 432                                enum ocfs2_alloc_restarted *reason_ret)
 433 {
 434         int status = 0;
 435         int free_extents;
 436         struct ocfs2_dinode *fe = (struct ocfs2_dinode *) fe_bh->b_data;
 437         enum ocfs2_alloc_restarted reason = RESTART_NONE;
 438         u32 bit_off, num_bits;
 439         u64 block;
 440
 441         BUG_ON(!clusters_to_add);
 442
 443         free_extents = ocfs2_num_free_extents(osb, inode, fe);
 444         if (free_extents < 0) {
 445                 status = free_extents;
 446                 mlog_errno(status);
 447                 goto leave;
 448         }
 449
 450         /* there are two cases which could cause us to EAGAIN in the
 451          * we-need-more-metadata case:
 452          * 1) we haven't reserved *any*
 453          * 2) we are so fragmented, we've needed to add metadata too
 454          *    many times. */
 455         if (!free_extents && !meta_ac) {
 456                 mlog(0, "we haven't reserved any metadata!\n");
 457                 status = -EAGAIN;
 458                 reason = RESTART_META;
 459                 goto leave;
 460         } else if ((!free_extents)
 461                    && (ocfs2_alloc_context_bits_left(meta_ac)
 462                        < ocfs2_extend_meta_needed(fe))) {
 463                 mlog(0, "filesystem is really fragmented...\n");
 464                 status = -EAGAIN;
 465                 reason = RESTART_META;
 466                 goto leave;
 467         }
 468
 469         status = ocfs2_claim_clusters(osb, handle, data_ac, 1,
 470                                       &bit_off, &num_bits);
 471         if (status < 0) {
 472                 if (status != -ENOSPC)
 473                         mlog_errno(status);
 474                 goto leave;
 475         }
 476
 477         BUG_ON(num_bits > clusters_to_add);
 478
 479         /* reserve our write early -- insert_extent may update the inode */
 480         status = ocfs2_journal_access(handle, inode, fe_bh,
 481                                       OCFS2_JOURNAL_ACCESS_WRITE);
 482         if (status < 0) {
 483                 mlog_errno(status);
 484                 goto leave;
 485         }
 486
 487         block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
 488         mlog(0, "Allocating %u clusters at block %u for inode %llu\n",
 489              num_bits, bit_off, (unsigned long long)OCFS2_I(inode)->ip_blkno);
 490         status = ocfs2_insert_extent(osb, handle, inode, fe_bh,
 491                                      *logical_offset, block, num_bits,
 492                                      meta_ac);
 493         if (status < 0) {
 494                 mlog_errno(status);
 495                 goto leave;
 496         }
 497
 498         status = ocfs2_journal_dirty(handle, fe_bh);
 499         if (status < 0) {
 500                 mlog_errno(status);
 501                 goto leave;
 502         }
 503
 504         clusters_to_add -= num_bits;
 505         *logical_offset += num_bits;
 506
 507         if (clusters_to_add) {
 508                 mlog(0, "need to alloc once more, clusters = %u, wanted = "
 509                      "%u\n", fe->i_clusters, clusters_to_add);
 510                 status = -EAGAIN;
 511                 reason = RESTART_TRANS;
 512         }
 513
 514 leave:
 515         mlog_exit(status);
 516         if (reason_ret)
 517                 *reason_ret = reason;
 518         return status;
 519 }
 520
 521 /*
 522  * For a given allocation, determine which allocators will need to be
 523  * accessed, and lock them, reserving the appropriate number of bits.
 524  *
 525  * Called from ocfs2_extend_allocation() for file systems which don't
 526  * support holes, and from ocfs2_write() for file systems which
 527  * understand sparse inodes.
 528  */
 529 int ocfs2_lock_allocators(struct inode *inode, struct ocfs2_dinode *di,
 530                           u32 clusters_to_add,
 531                           struct ocfs2_alloc_context **data_ac,
 532                           struct ocfs2_alloc_context **meta_ac)
 533 {
 534         int ret, num_free_extents;
 535         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 536
 537         *meta_ac = NULL;
 538         *data_ac = NULL;
 539
 540         mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
 541              "clusters_to_add = %u\n",
 542              (unsigned long long)OCFS2_I(inode)->ip_blkno, i_size_read(inode),
 543              le32_to_cpu(di->i_clusters), clusters_to_add);
 544
 545         num_free_extents = ocfs2_num_free_extents(osb, inode, di);
 546         if (num_free_extents < 0) {
 547                 ret = num_free_extents;
 548                 mlog_errno(ret);
 549                 goto out;
 550         }
 551
 552         /*
 553          * Sparse allocation file systems need to be more conservative
 554          * with reserving room for expansion - the actual allocation
 555          * happens while we've got a journal handle open so re-taking
 556          * a cluster lock (because we ran out of room for another
 557          * extent) will violate ordering rules.
 558          *
 559          * Most of the time we'll only be seeing this 1 cluster at a time
 560          * anyway.
 561          */
 562         if (!num_free_extents ||
 563             (ocfs2_sparse_alloc(osb) && num_free_extents < clusters_to_add)) {
 564                 ret = ocfs2_reserve_new_metadata(osb, di, meta_ac);
 565                 if (ret < 0) {
 566                         if (ret != -ENOSPC)
 567                                 mlog_errno(ret);
 568                         goto out;
 569                 }
 570         }
 571
 572         ret = ocfs2_reserve_clusters(osb, clusters_to_add, data_ac);
 573         if (ret < 0) {
 574                 if (ret != -ENOSPC)
 575                         mlog_errno(ret);
 576                 goto out;
 577         }
 578
 579 out:
 580         if (ret) {
 581                 if (*meta_ac) {
 582                         ocfs2_free_alloc_context(*meta_ac);
 583                         *meta_ac = NULL;
 584                 }
 585
 586                 /*
 587                  * We cannot have an error and a non null *data_ac.
 588                  */
 589         }
 590
 591         return ret;
 592 }
 593
 594 static int ocfs2_extend_allocation(struct inode *inode,
 595                                    u32 clusters_to_add)
 596 {
 597         int status = 0;
 598         int restart_func = 0;
 599         int drop_alloc_sem = 0;
 600         int credits;
 601         u32 prev_clusters, logical_start;
 602         struct buffer_head *bh = NULL;
 603         struct ocfs2_dinode *fe = NULL;
 604         handle_t *handle = NULL;
 605         struct ocfs2_alloc_context *data_ac = NULL;
 606         struct ocfs2_alloc_context *meta_ac = NULL;
 607         enum ocfs2_alloc_restarted why;
 608         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 609
 610         mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
 611
 612         /*
 613          * This function only exists for file systems which don't
 614          * support holes.
 615          */
 616         BUG_ON(ocfs2_sparse_alloc(osb));
 617
 618         status = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &bh,
 619                                   OCFS2_BH_CACHED, inode);
 620         if (status < 0) {
 621                 mlog_errno(status);
 622                 goto leave;
 623         }
 624
 625         fe = (struct ocfs2_dinode *) bh->b_data;
 626         if (!OCFS2_IS_VALID_DINODE(fe)) {
 627                 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
 628                 status = -EIO;
 629                 goto leave;
 630         }
 631
 632         logical_start = OCFS2_I(inode)->ip_clusters;
 633
 634 restart_all:
 635         BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
 636
 637         /* blocks peope in read/write from reading our allocation
 638          * until we're done changing it. We depend on i_mutex to block
 639          * other extend/truncate calls while we're here. Ordering wrt
 640          * start_trans is important here -- always do it before! */
 641         down_write(&OCFS2_I(inode)->ip_alloc_sem);
 642         drop_alloc_sem = 1;
 643
 644         status = ocfs2_lock_allocators(inode, fe, clusters_to_add, &data_ac,
 645                                        &meta_ac);
 646         if (status) {
 647                 mlog_errno(status);
 648                 goto leave;
 649         }
 650
 651         credits = ocfs2_calc_extend_credits(osb->sb, fe, clusters_to_add);
 652         handle = ocfs2_start_trans(osb, credits);
 653         if (IS_ERR(handle)) {
 654                 status = PTR_ERR(handle);
 655                 handle = NULL;
 656                 mlog_errno(status);
 657                 goto leave;
 658         }
 659
 660 restarted_transaction:
 661         /* reserve a write to the file entry early on - that we if we
 662          * run out of credits in the allocation path, we can still
 663          * update i_size. */
 664         status = ocfs2_journal_access(handle, inode, bh,
 665                                       OCFS2_JOURNAL_ACCESS_WRITE);
 666         if (status < 0) {
 667                 mlog_errno(status);
 668                 goto leave;
 669         }
 670
 671         prev_clusters = OCFS2_I(inode)->ip_clusters;
 672
 673         status = ocfs2_do_extend_allocation(osb,
 674                                             inode,
 675                                             &logical_start,
 676                                             clusters_to_add,
 677                                             bh,
 678                                             handle,
 679                                             data_ac,
 680                                             meta_ac,
 681                                             &why);
 682         if ((status < 0) && (status != -EAGAIN)) {
 683                 if (status != -ENOSPC)
 684                         mlog_errno(status);
 685                 goto leave;
 686         }
 687
 688         status = ocfs2_journal_dirty(handle, bh);
 689         if (status < 0) {
 690                 mlog_errno(status);
 691                 goto leave;
 692         }
 693
 694         spin_lock(&OCFS2_I(inode)->ip_lock);
 695         clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
 696         spin_unlock(&OCFS2_I(inode)->ip_lock);
 697
 698         if (why != RESTART_NONE && clusters_to_add) {
 699                 if (why == RESTART_META) {
 700                         mlog(0, "restarting function.\n");
 701                         restart_func = 1;
 702                 } else {
 703                         BUG_ON(why != RESTART_TRANS);
 704
 705                         mlog(0, "restarting transaction.\n");
 706                         /* TODO: This can be more intelligent. */
 707                         credits = ocfs2_calc_extend_credits(osb->sb,
 708                                                             fe,
 709                                                             clusters_to_add);
 710                         status = ocfs2_extend_trans(handle, credits);
 711                         if (status < 0) {
 712                                 /* handle still has to be committed at
 713                                  * this point. */
 714                                 status = -ENOMEM;
 715                                 mlog_errno(status);
 716                                 goto leave;
 717                         }
 718                         goto restarted_transaction;
 719                 }
 720         }
 721
 722         mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
 723              le32_to_cpu(fe->i_clusters),
 724              (unsigned long long)le64_to_cpu(fe->i_size));
 725         mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
 726              OCFS2_I(inode)->ip_clusters, i_size_read(inode));
 727
 728 leave:
 729         if (drop_alloc_sem) {
 730                 up_write(&OCFS2_I(inode)->ip_alloc_sem);
 731                 drop_alloc_sem = 0;
 732         }
 733         if (handle) {
 734                 ocfs2_commit_trans(osb, handle);
 735                 handle = NULL;
 736         }
 737         if (data_ac) {
 738                 ocfs2_free_alloc_context(data_ac);
 739                 data_ac = NULL;
 740         }
 741         if (meta_ac) {
 742                 ocfs2_free_alloc_context(meta_ac);
 743                 meta_ac = NULL;
 744         }
 745         if ((!status) && restart_func) {
 746                 restart_func = 0;
 747                 goto restart_all;
 748         }
 749         if (bh) {
 750                 brelse(bh);
 751                 bh = NULL;
 752         }
 753
 754         mlog_exit(status);
 755         return status;
 756 }
 757
 758 /* Some parts of this taken from generic_cont_expand, which turned out
 759  * to be too fragile to do exactly what we need without us having to
 760  * worry about recursive locking in ->prepare_write() and
 761  * ->commit_write(). */
 762 static int ocfs2_write_zero_page(struct inode *inode,
 763                                  u64 size)
 764 {
 765         struct address_space *mapping = inode->i_mapping;
 766         struct page *page;
 767         unsigned long index;
 768         unsigned int offset;
 769         handle_t *handle = NULL;
 770         int ret;
 771
 772         offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
 773         /* ugh.  in prepare/commit_write, if from==to==start of block, we
 774         ** skip the prepare.  make sure we never send an offset for the start
 775         ** of a block
 776         */
 777         if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
 778                 offset++;
 779         }
 780         index = size >> PAGE_CACHE_SHIFT;
 781
 782         page = grab_cache_page(mapping, index);
 783         if (!page) {
 784                 ret = -ENOMEM;
 785                 mlog_errno(ret);
 786                 goto out;
 787         }
 788
 789         ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
 790         if (ret < 0) {
 791                 mlog_errno(ret);
 792                 goto out_unlock;
 793         }
 794
 795         if (ocfs2_should_order_data(inode)) {
 796                 handle = ocfs2_start_walk_page_trans(inode, page, offset,
 797                                                      offset);
 798                 if (IS_ERR(handle)) {
 799                         ret = PTR_ERR(handle);
 800                         handle = NULL;
 801                         goto out_unlock;
 802                 }
 803         }
 804
 805         /* must not update i_size! */
 806         ret = block_commit_write(page, offset, offset);
 807         if (ret < 0)
 808                 mlog_errno(ret);
 809         else
 810                 ret = 0;
 811
 812         if (handle)
 813                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
 814 out_unlock:
 815         unlock_page(page);
 816         page_cache_release(page);
 817 out:
 818         return ret;
 819 }
 820
 821 static int ocfs2_zero_extend(struct inode *inode,
 822                              u64 zero_to_size)
 823 {
 824         int ret = 0;
 825         u64 start_off;
 826         struct super_block *sb = inode->i_sb;
 827
 828         start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
 829         while (start_off < zero_to_size) {
 830                 ret = ocfs2_write_zero_page(inode, start_off);
 831                 if (ret < 0) {
 832                         mlog_errno(ret);
 833                         goto out;
 834                 }
 835
 836                 start_off += sb->s_blocksize;
 837
 838                 /*
 839                  * Very large extends have the potential to lock up
 840                  * the cpu for extended periods of time.
 841                  */
 842                 cond_resched();
 843         }
 844
 845 out:
 846         return ret;
 847 }
 848
 849 /*
 850  * A tail_to_skip value > 0 indicates that we're being called from
 851  * ocfs2_file_aio_write(). This has the following implications:
 852  *
 853  * - we don't want to update i_size
 854  * - di_bh will be NULL, which is fine because it's only used in the
 855  *   case where we want to update i_size.
 856  * - ocfs2_zero_extend() will then only be filling the hole created
 857  *   between i_size and the start of the write.
 858  */
 859 static int ocfs2_extend_file(struct inode *inode,
 860                              struct buffer_head *di_bh,
 861                              u64 new_i_size,
 862                              size_t tail_to_skip)
 863 {
 864         int ret = 0;
 865         u32 clusters_to_add = 0;
 866
 867         BUG_ON(!tail_to_skip && !di_bh);
 868
 869         /* setattr sometimes calls us like this. */
 870         if (new_i_size == 0)
 871                 goto out;
 872
 873         if (i_size_read(inode) == new_i_size)
 874                 goto out;
 875         BUG_ON(new_i_size < i_size_read(inode));
 876
 877         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
 878                 BUG_ON(tail_to_skip != 0);
 879                 goto out_update_size;
 880         }
 881
 882         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size) -
 883                 OCFS2_I(inode)->ip_clusters;
 884
 885         /*
 886          * protect the pages that ocfs2_zero_extend is going to be
 887          * pulling into the page cache.. we do this before the
 888          * metadata extend so that we don't get into the situation
 889          * where we've extended the metadata but can't get the data
 890          * lock to zero.
 891          */
 892         ret = ocfs2_data_lock(inode, 1);
 893         if (ret < 0) {
 894                 mlog_errno(ret);
 895                 goto out;
 896         }
 897
 898         if (clusters_to_add) {
 899                 ret = ocfs2_extend_allocation(inode, clusters_to_add);
 900                 if (ret < 0) {
 901                         mlog_errno(ret);
 902                         goto out_unlock;
 903                 }
 904         }
 905
 906         /*
 907          * Call this even if we don't add any clusters to the tree. We
 908          * still need to zero the area between the old i_size and the
 909          * new i_size.
 910          */
 911         ret = ocfs2_zero_extend(inode, (u64)new_i_size - tail_to_skip);
 912         if (ret < 0) {
 913                 mlog_errno(ret);
 914                 goto out_unlock;
 915         }
 916
 917 out_update_size:
 918         if (!tail_to_skip) {
 919                 /* We're being called from ocfs2_setattr() which wants
 920                  * us to update i_size */
 921                 ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
 922                 if (ret < 0)
 923                         mlog_errno(ret);
 924         }
 925
 926 out_unlock:
 927         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
 928                 ocfs2_data_unlock(inode, 1);
 929
 930 out:
 931         return ret;
 932 }
 933
 934 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
 935 {
 936         int status = 0, size_change;
 937         struct inode *inode = dentry->d_inode;
 938         struct super_block *sb = inode->i_sb;
 939         struct ocfs2_super *osb = OCFS2_SB(sb);
 940         struct buffer_head *bh = NULL;
 941         handle_t *handle = NULL;
 942
 943         mlog_entry("(0x%p, '%.*s')\n", dentry,
 944                    dentry->d_name.len, dentry->d_name.name);
 945
 946         if (attr->ia_valid & ATTR_MODE)
 947                 mlog(0, "mode change: %d\n", attr->ia_mode);
 948         if (attr->ia_valid & ATTR_UID)
 949                 mlog(0, "uid change: %d\n", attr->ia_uid);
 950         if (attr->ia_valid & ATTR_GID)
 951                 mlog(0, "gid change: %d\n", attr->ia_gid);
 952         if (attr->ia_valid & ATTR_SIZE)
 953                 mlog(0, "size change...\n");
 954         if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
 955                 mlog(0, "time change...\n");
 956
 957 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
 958                            | ATTR_GID | ATTR_UID | ATTR_MODE)
 959         if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
 960                 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
 961                 return 0;
 962         }
 963
 964         status = inode_change_ok(inode, attr);
 965         if (status)
 966                 return status;
 967
 968         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
 969         if (size_change) {
 970                 status = ocfs2_rw_lock(inode, 1);
 971                 if (status < 0) {
 972                         mlog_errno(status);
 973                         goto bail;
 974                 }
 975         }
 976
 977         status = ocfs2_meta_lock(inode, &bh, 1);
 978         if (status < 0) {
 979                 if (status != -ENOENT)
 980                         mlog_errno(status);
 981                 goto bail_unlock_rw;
 982         }
 983
 984         if (size_change && attr->ia_size != i_size_read(inode)) {
 985                 if (i_size_read(inode) > attr->ia_size)
 986                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
 987                 else
 988                         status = ocfs2_extend_file(inode, bh, attr->ia_size, 0);
 989                 if (status < 0) {
 990                         if (status != -ENOSPC)
 991                                 mlog_errno(status);
 992                         status = -ENOSPC;
 993                         goto bail_unlock;
 994                 }
 995         }
 996
 997         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
 998         if (IS_ERR(handle)) {
 999                 status = PTR_ERR(handle);
1000                 mlog_errno(status);
1001                 goto bail_unlock;
1002         }
1003
1004         status = inode_setattr(inode, attr);
1005         if (status < 0) {
1006                 mlog_errno(status);
1007                 goto bail_commit;
1008         }
1009
1010         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1011         if (status < 0)
1012                 mlog_errno(status);
1013
1014 bail_commit:
1015         ocfs2_commit_trans(osb, handle);
1016 bail_unlock:
1017         ocfs2_meta_unlock(inode, 1);
1018 bail_unlock_rw:
1019         if (size_change)
1020                 ocfs2_rw_unlock(inode, 1);
1021 bail:
1022         if (bh)
1023                 brelse(bh);
1024
1025         mlog_exit(status);
1026         return status;
1027 }
1028
1029 int ocfs2_getattr(struct vfsmount *mnt,
1030                   struct dentry *dentry,
1031                   struct kstat *stat)
1032 {
1033         struct inode *inode = dentry->d_inode;
1034         struct super_block *sb = dentry->d_inode->i_sb;
1035         struct ocfs2_super *osb = sb->s_fs_info;
1036         int err;
1037
1038         mlog_entry_void();
1039
1040         err = ocfs2_inode_revalidate(dentry);
1041         if (err) {
1042                 if (err != -ENOENT)
1043                         mlog_errno(err);
1044                 goto bail;
1045         }
1046
1047         generic_fillattr(inode, stat);
1048
1049         /* We set the blksize from the cluster size for performance */
1050         stat->blksize = osb->s_clustersize;
1051
1052 bail:
1053         mlog_exit(err);
1054
1055         return err;
1056 }
1057
1058 int ocfs2_permission(struct inode *inode, int mask, struct nameidata *nd)
1059 {
1060         int ret;
1061
1062         mlog_entry_void();
1063
1064         ret = ocfs2_meta_lock(inode, NULL, 0);
1065         if (ret) {
1066                 if (ret != -ENOENT)
1067                         mlog_errno(ret);
1068                 goto out;
1069         }
1070
1071         ret = generic_permission(inode, mask, NULL);
1072
1073         ocfs2_meta_unlock(inode, 0);
1074 out:
1075         mlog_exit(ret);
1076         return ret;
1077 }
1078
1079 static int ocfs2_write_remove_suid(struct inode *inode)
1080 {
1081         int ret;
1082         struct buffer_head *bh = NULL;
1083         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1084         handle_t *handle;
1085         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1086         struct ocfs2_dinode *di;
1087
1088         mlog_entry("(Inode %llu, mode 0%o)\n",
1089                    (unsigned long long)oi->ip_blkno, inode->i_mode);
1090
1091         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1092         if (handle == NULL) {
1093                 ret = -ENOMEM;
1094                 mlog_errno(ret);
1095                 goto out;
1096         }
1097
1098         ret = ocfs2_read_block(osb, oi->ip_blkno, &bh, OCFS2_BH_CACHED, inode);
1099         if (ret < 0) {
1100                 mlog_errno(ret);
1101                 goto out_trans;
1102         }
1103
1104         ret = ocfs2_journal_access(handle, inode, bh,
1105                                    OCFS2_JOURNAL_ACCESS_WRITE);
1106         if (ret < 0) {
1107                 mlog_errno(ret);
1108                 goto out_bh;
1109         }
1110
1111         inode->i_mode &= ~S_ISUID;
1112         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1113                 inode->i_mode &= ~S_ISGID;
1114
1115         di = (struct ocfs2_dinode *) bh->b_data;
1116         di->i_mode = cpu_to_le16(inode->i_mode);
1117
1118         ret = ocfs2_journal_dirty(handle, bh);
1119         if (ret < 0)
1120                 mlog_errno(ret);
1121 out_bh:
1122         brelse(bh);
1123 out_trans:
1124         ocfs2_commit_trans(osb, handle);
1125 out:
1126         mlog_exit(ret);
1127         return ret;
1128 }
1129
1130 /*
1131  * Will look for holes and unwritten extents in the range starting at
1132  * pos for count bytes (inclusive).
1133  */
1134 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1135                                        size_t count)
1136 {
1137         int ret = 0;
1138         unsigned int extent_flags;
1139         u32 cpos, clusters, extent_len, phys_cpos;
1140         struct super_block *sb = inode->i_sb;
1141
1142         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1143         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1144
1145         while (clusters) {
1146                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1147                                          &extent_flags);
1148                 if (ret < 0) {
1149                         mlog_errno(ret);
1150                         goto out;
1151                 }
1152
1153                 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1154                         ret = 1;
1155                         break;
1156                 }
1157
1158                 if (extent_len > clusters)
1159                         extent_len = clusters;
1160
1161                 clusters -= extent_len;
1162                 cpos += extent_len;
1163         }
1164 out:
1165         return ret;
1166 }
1167
1168 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1169                                          loff_t *ppos,
1170                                          size_t count,
1171                                          int appending,
1172                                          int *direct_io)
1173 {
1174         int ret = 0, meta_level = appending;
1175         struct inode *inode = dentry->d_inode;
1176         u32 clusters;
1177         loff_t newsize, saved_pos;
1178
1179         /*
1180          * We sample i_size under a read level meta lock to see if our write
1181          * is extending the file, if it is we back off and get a write level
1182          * meta lock.
1183          */
1184         for(;;) {
1185                 ret = ocfs2_meta_lock(inode, NULL, meta_level);
1186                 if (ret < 0) {
1187                         meta_level = -1;
1188                         mlog_errno(ret);
1189                         goto out;
1190                 }
1191
1192                 /* Clear suid / sgid if necessary. We do this here
1193                  * instead of later in the write path because
1194                  * remove_suid() calls ->setattr without any hint that
1195                  * we may have already done our cluster locking. Since
1196                  * ocfs2_setattr() *must* take cluster locks to
1197                  * proceeed, this will lead us to recursively lock the
1198                  * inode. There's also the dinode i_size state which
1199                  * can be lost via setattr during extending writes (we
1200                  * set inode->i_size at the end of a write. */
1201                 if (should_remove_suid(dentry)) {
1202                         if (meta_level == 0) {
1203                                 ocfs2_meta_unlock(inode, meta_level);
1204                                 meta_level = 1;
1205                                 continue;
1206                         }
1207
1208                         ret = ocfs2_write_remove_suid(inode);
1209                         if (ret < 0) {
1210                                 mlog_errno(ret);
1211                                 goto out_unlock;
1212                         }
1213                 }
1214
1215                 /* work on a copy of ppos until we're sure that we won't have
1216                  * to recalculate it due to relocking. */
1217                 if (appending) {
1218                         saved_pos = i_size_read(inode);
1219                         mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1220                 } else {
1221                         saved_pos = *ppos;
1222                 }
1223
1224                 if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
1225                         loff_t end = saved_pos + count;
1226
1227                         /*
1228                          * Skip the O_DIRECT checks if we don't need
1229                          * them.
1230                          */
1231                         if (!direct_io || !(*direct_io))
1232                                 break;
1233
1234                         /*
1235                          * Allowing concurrent direct writes means
1236                          * i_size changes wouldn't be synchronized, so
1237                          * one node could wind up truncating another
1238                          * nodes writes.
1239                          */
1240                         if (end > i_size_read(inode)) {
1241                                 *direct_io = 0;
1242                                 break;
1243                         }
1244
1245                         /*
1246                          * We don't fill holes during direct io, so
1247                          * check for them here. If any are found, the
1248                          * caller will have to retake some cluster
1249                          * locks and initiate the io as buffered.
1250                          */
1251                         ret = ocfs2_check_range_for_holes(inode, saved_pos,
1252                                                           count);
1253                         if (ret == 1) {
1254                                 *direct_io = 0;
1255                                 ret = 0;
1256                         } else if (ret < 0)
1257                                 mlog_errno(ret);
1258                         break;
1259                 }
1260
1261                 /*
1262                  * The rest of this loop is concerned with legacy file
1263                  * systems which don't support sparse files.
1264                  */
1265
1266                 newsize = count + saved_pos;
1267
1268                 mlog(0, "pos=%lld newsize=%lld cursize=%lld\n",
1269                      (long long) saved_pos, (long long) newsize,
1270                      (long long) i_size_read(inode));
1271
1272                 /* No need for a higher level metadata lock if we're
1273                  * never going past i_size. */
1274                 if (newsize <= i_size_read(inode))
1275                         break;
1276
1277                 if (meta_level == 0) {
1278                         ocfs2_meta_unlock(inode, meta_level);
1279                         meta_level = 1;
1280                         continue;
1281                 }
1282
1283                 spin_lock(&OCFS2_I(inode)->ip_lock);
1284                 clusters = ocfs2_clusters_for_bytes(inode->i_sb, newsize) -
1285                         OCFS2_I(inode)->ip_clusters;
1286                 spin_unlock(&OCFS2_I(inode)->ip_lock);
1287
1288                 mlog(0, "Writing at EOF, may need more allocation: "
1289                      "i_size = %lld, newsize = %lld, need %u clusters\n",
1290                      (long long) i_size_read(inode), (long long) newsize,
1291                      clusters);
1292
1293                 /* We only want to continue the rest of this loop if
1294                  * our extend will actually require more
1295                  * allocation. */
1296                 if (!clusters)
1297                         break;
1298
1299                 ret = ocfs2_extend_file(inode, NULL, newsize, count);
1300                 if (ret < 0) {
1301                         if (ret != -ENOSPC)
1302                                 mlog_errno(ret);
1303                         goto out_unlock;
1304                 }
1305                 break;
1306         }
1307
1308         if (appending)
1309                 *ppos = saved_pos;
1310
1311 out_unlock:
1312         ocfs2_meta_unlock(inode, meta_level);
1313
1314 out:
1315         return ret;
1316 }
1317
1318 static inline void
1319 ocfs2_set_next_iovec(const struct iovec **iovp, size_t *basep, size_t bytes)
1320 {
1321         const struct iovec *iov = *iovp;
1322         size_t base = *basep;
1323
1324         do {
1325                 int copy = min(bytes, iov->iov_len - base);
1326
1327                 bytes -= copy;
1328                 base += copy;
1329                 if (iov->iov_len == base) {
1330                         iov++;
1331                         base = 0;
1332                 }
1333         } while (bytes);
1334         *iovp = iov;
1335         *basep = base;
1336 }
1337
1338 static struct page * ocfs2_get_write_source(char **ret_src_buf,
1339                                             const struct iovec *cur_iov,
1340                                             size_t iov_offset)
1341 {
1342         int ret;
1343         char *buf = cur_iov->iov_base + iov_offset;
1344         struct page *src_page = NULL;
1345         unsigned long off;
1346
1347         off = (unsigned long)(buf) & ~PAGE_CACHE_MASK;
1348
1349         if (!segment_eq(get_fs(), KERNEL_DS)) {
1350                 /*
1351                  * Pull in the user page. We want to do this outside
1352                  * of the meta data locks in order to preserve locking
1353                  * order in case of page fault.
1354                  */
1355                 ret = get_user_pages(current, current->mm,
1356                                      (unsigned long)buf & PAGE_CACHE_MASK, 1,
1357                                      0, 0, &src_page, NULL);
1358                 if (ret == 1)
1359                         *ret_src_buf = kmap(src_page) + off;
1360                 else
1361                         src_page = ERR_PTR(-EFAULT);
1362         } else {
1363                 *ret_src_buf = buf;
1364         }
1365
1366         return src_page;
1367 }
1368
1369 static void ocfs2_put_write_source(struct page *page)
1370 {
1371         if (page) {
1372                 kunmap(page);
1373                 page_cache_release(page);
1374         }
1375 }
1376
1377 static ssize_t ocfs2_file_buffered_write(struct file *file, loff_t *ppos,
1378                                          const struct iovec *iov,
1379                                          unsigned long nr_segs,
1380                                          size_t count,
1381                                          ssize_t o_direct_written)
1382 {
1383         int ret = 0;
1384         ssize_t copied, total = 0;
1385         size_t iov_offset = 0, bytes;
1386         loff_t pos;
1387         const struct iovec *cur_iov = iov;
1388         struct page *user_page, *page;
1389         char *buf, *dst;
1390         void *fsdata;
1391
1392         /*
1393          * handle partial DIO write.  Adjust cur_iov if needed.
1394          */
1395         ocfs2_set_next_iovec(&cur_iov, &iov_offset, o_direct_written);
1396
1397         do {
1398                 pos = *ppos;
1399
1400                 user_page = ocfs2_get_write_source(&buf, cur_iov, iov_offset);
1401                 if (IS_ERR(user_page)) {
1402                         ret = PTR_ERR(user_page);
1403                         goto out;
1404                 }
1405
1406                 /* Stay within our page boundaries */
1407                 bytes = min((PAGE_CACHE_SIZE - ((unsigned long)pos & ~PAGE_CACHE_MASK)),
1408                             (PAGE_CACHE_SIZE - ((unsigned long)buf & ~PAGE_CACHE_MASK)));
1409                 /* Stay within the vector boundary */
1410                 bytes = min_t(size_t, bytes, cur_iov->iov_len - iov_offset);
1411                 /* Stay within count */
1412                 bytes = min(bytes, count);
1413
1414                 page = NULL;
1415                 ret = ocfs2_write_begin(file, file->f_mapping, pos, bytes, 0,
1416                                         &page, &fsdata);
1417                 if (ret) {
1418                         mlog_errno(ret);
1419                         goto out;
1420                 }
1421
1422                 dst = kmap_atomic(page, KM_USER0);
1423                 memcpy(dst + (pos & (PAGE_CACHE_SIZE - 1)), buf, bytes);
1424                 kunmap_atomic(dst, KM_USER0);
1425                 flush_dcache_page(page);
1426                 ocfs2_put_write_source(user_page);
1427
1428                 copied = ocfs2_write_end(file, file->f_mapping, pos, bytes,
1429                                          bytes, page, fsdata);
1430                 if (copied < 0) {
1431                         mlog_errno(copied);
1432                         ret = copied;
1433                         goto out;
1434                 }
1435
1436                 total += copied;
1437                 *ppos = pos + copied;
1438                 count -= copied;
1439
1440                 ocfs2_set_next_iovec(&cur_iov, &iov_offset, copied);
1441         } while(count);
1442
1443 out:
1444         return total ? total : ret;
1445 }
1446
1447 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
1448                                     const struct iovec *iov,
1449                                     unsigned long nr_segs,
1450                                     loff_t pos)
1451 {
1452         int ret, direct_io, appending, rw_level, have_alloc_sem  = 0;
1453         int can_do_direct, sync = 0;
1454         ssize_t written = 0;
1455         size_t ocount;          /* original count */
1456         size_t count;           /* after file limit checks */
1457         loff_t *ppos = &iocb->ki_pos;
1458         struct file *file = iocb->ki_filp;
1459         struct inode *inode = file->f_path.dentry->d_inode;
1460
1461         mlog_entry("(0x%p, %u, '%.*s')\n", file,
1462                    (unsigned int)nr_segs,
1463                    file->f_path.dentry->d_name.len,
1464                    file->f_path.dentry->d_name.name);
1465
1466         if (iocb->ki_left == 0)
1467                 return 0;
1468
1469         ret = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1470         if (ret)
1471                 return ret;
1472
1473         count = ocount;
1474
1475         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1476
1477         appending = file->f_flags & O_APPEND ? 1 : 0;
1478         direct_io = file->f_flags & O_DIRECT ? 1 : 0;
1479
1480         mutex_lock(&inode->i_mutex);
1481
1482 relock:
1483         /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
1484         if (direct_io) {
1485                 down_read(&inode->i_alloc_sem);
1486                 have_alloc_sem = 1;
1487         }
1488
1489         /* concurrent O_DIRECT writes are allowed */
1490         rw_level = !direct_io;
1491         ret = ocfs2_rw_lock(inode, rw_level);
1492         if (ret < 0) {
1493                 mlog_errno(ret);
1494                 goto out_sems;
1495         }
1496
1497         can_do_direct = direct_io;
1498         ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
1499                                             iocb->ki_left, appending,
1500                                             &can_do_direct);
1501         if (ret < 0) {
1502                 mlog_errno(ret);
1503                 goto out;
1504         }
1505
1506         /*
1507          * We can't complete the direct I/O as requested, fall back to
1508          * buffered I/O.
1509          */
1510         if (direct_io && !can_do_direct) {
1511                 ocfs2_rw_unlock(inode, rw_level);
1512                 up_read(&inode->i_alloc_sem);
1513
1514                 have_alloc_sem = 0;
1515                 rw_level = -1;
1516
1517                 direct_io = 0;
1518                 sync = 1;
1519                 goto relock;
1520         }
1521
1522         if (!sync && ((file->f_flags & O_SYNC) || IS_SYNC(inode)))
1523                 sync = 1;
1524
1525         /*
1526          * XXX: Is it ok to execute these checks a second time?
1527          */
1528         ret = generic_write_checks(file, ppos, &count, S_ISBLK(inode->i_mode));
1529         if (ret)
1530                 goto out;
1531
1532         /*
1533          * Set pos so that sync_page_range_nolock() below understands
1534          * where to start from. We might've moved it around via the
1535          * calls above. The range we want to actually sync starts from
1536          * *ppos here.
1537          *
1538          */
1539         pos = *ppos;
1540
1541         /* communicate with ocfs2_dio_end_io */
1542         ocfs2_iocb_set_rw_locked(iocb, rw_level);
1543
1544         if (direct_io) {
1545                 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
1546                                                     ppos, count, ocount);
1547                 if (written < 0) {
1548                         ret = written;
1549                         goto out_dio;
1550                 }
1551         } else {
1552                 written = ocfs2_file_buffered_write(file, ppos, iov, nr_segs,
1553                                                     count, written);
1554                 if (written < 0) {
1555                         ret = written;
1556                         if (ret != -EFAULT || ret != -ENOSPC)
1557                                 mlog_errno(ret);
1558                         goto out;
1559                 }
1560         }
1561
1562 out_dio:
1563         /* buffered aio wouldn't have proper lock coverage today */
1564         BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
1565
1566         /*
1567          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
1568          * function pointer which is called when o_direct io completes so that
1569          * it can unlock our rw lock.  (it's the clustered equivalent of
1570          * i_alloc_sem; protects truncate from racing with pending ios).
1571          * Unfortunately there are error cases which call end_io and others
1572          * that don't.  so we don't have to unlock the rw_lock if either an
1573          * async dio is going to do it in the future or an end_io after an
1574          * error has already done it.
1575          */
1576         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
1577                 rw_level = -1;
1578                 have_alloc_sem = 0;
1579         }
1580
1581 out:
1582         if (rw_level != -1)
1583                 ocfs2_rw_unlock(inode, rw_level);
1584
1585 out_sems:
1586         if (have_alloc_sem)
1587                 up_read(&inode->i_alloc_sem);
1588
1589         if (written > 0 && sync) {
1590                 ssize_t err;
1591
1592                 err = sync_page_range_nolock(inode, file->f_mapping, pos, count);
1593                 if (err < 0)
1594                         written = err;
1595         }
1596
1597         mutex_unlock(&inode->i_mutex);
1598
1599         mlog_exit(ret);
1600         return written ? written : ret;
1601 }
1602
1603 static int ocfs2_splice_write_actor(struct pipe_inode_info *pipe,
1604                                     struct pipe_buffer *buf,
1605                                     struct splice_desc *sd)
1606 {
1607         int ret, count;
1608         ssize_t copied = 0;
1609         struct file *file = sd->u.file;
1610         unsigned int offset;
1611         struct page *page = NULL;
1612         void *fsdata;
1613         char *src, *dst;
1614
1615         ret = buf->ops->confirm(pipe, buf);
1616         if (ret)
1617                 goto out;
1618
1619         offset = sd->pos & ~PAGE_CACHE_MASK;
1620         count = sd->len;
1621         if (count + offset > PAGE_CACHE_SIZE)
1622                 count = PAGE_CACHE_SIZE - offset;
1623
1624         ret = ocfs2_write_begin(file, file->f_mapping, sd->pos, count, 0,
1625                                 &page, &fsdata);
1626         if (ret) {
1627                 mlog_errno(ret);
1628                 goto out;
1629         }
1630
1631         src = buf->ops->map(pipe, buf, 1);
1632         dst = kmap_atomic(page, KM_USER1);
1633         memcpy(dst + offset, src + buf->offset, count);
1634         kunmap_atomic(page, KM_USER1);
1635         buf->ops->unmap(pipe, buf, src);
1636
1637         copied = ocfs2_write_end(file, file->f_mapping, sd->pos, count, count,
1638                                  page, fsdata);
1639         if (copied < 0) {
1640                 mlog_errno(copied);
1641                 ret = copied;
1642                 goto out;
1643         }
1644 out:
1645
1646         return copied ? copied : ret;
1647 }
1648
1649 static ssize_t __ocfs2_file_splice_write(struct pipe_inode_info *pipe,
1650                                          struct file *out,
1651                                          loff_t *ppos,
1652                                          size_t len,
1653                                          unsigned int flags)
1654 {
1655         int ret, err;
1656         struct address_space *mapping = out->f_mapping;
1657         struct inode *inode = mapping->host;
1658         struct splice_desc sd = {
1659                 .total_len = len,
1660                 .flags = flags,
1661                 .pos = *ppos,
1662                 .u.file = out,
1663         };
1664
1665         ret = __splice_from_pipe(pipe, &sd, ocfs2_splice_write_actor);
1666         if (ret > 0) {
1667                 *ppos += ret;
1668
1669                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
1670                         err = generic_osync_inode(inode, mapping,
1671                                                   OSYNC_METADATA|OSYNC_DATA);
1672                         if (err)
1673                                 ret = err;
1674                 }
1675         }
1676
1677         return ret;
1678 }
1679
1680 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
1681                                        struct file *out,
1682                                        loff_t *ppos,
1683                                        size_t len,
1684                                        unsigned int flags)
1685 {
1686         int ret;
1687         struct inode *inode = out->f_path.dentry->d_inode;
1688
1689         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
1690                    (unsigned int)len,
1691                    out->f_path.dentry->d_name.len,
1692                    out->f_path.dentry->d_name.name);
1693
1694         inode_double_lock(inode, pipe->inode);
1695
1696         ret = ocfs2_rw_lock(inode, 1);
1697         if (ret < 0) {
1698                 mlog_errno(ret);
1699                 goto out;
1700         }
1701
1702         ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, ppos, len, 0,
1703                                             NULL);
1704         if (ret < 0) {
1705                 mlog_errno(ret);
1706                 goto out_unlock;
1707         }
1708
1709         /* ok, we're done with i_size and alloc work */
1710         ret = __ocfs2_file_splice_write(pipe, out, ppos, len, flags);
1711
1712 out_unlock:
1713         ocfs2_rw_unlock(inode, 1);
1714 out:
1715         inode_double_unlock(inode, pipe->inode);
1716
1717         mlog_exit(ret);
1718         return ret;
1719 }
1720
1721 static ssize_t ocfs2_file_splice_read(struct file *in,
1722                                       loff_t *ppos,
1723                                       struct pipe_inode_info *pipe,
1724                                       size_t len,
1725                                       unsigned int flags)
1726 {
1727         int ret = 0;
1728         struct inode *inode = in->f_path.dentry->d_inode;
1729
1730         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
1731                    (unsigned int)len,
1732                    in->f_path.dentry->d_name.len,
1733                    in->f_path.dentry->d_name.name);
1734
1735         /*
1736          * See the comment in ocfs2_file_aio_read()
1737          */
1738         ret = ocfs2_meta_lock(inode, NULL, 0);
1739         if (ret < 0) {
1740                 mlog_errno(ret);
1741                 goto bail;
1742         }
1743         ocfs2_meta_unlock(inode, 0);
1744
1745         ret = generic_file_splice_read(in, ppos, pipe, len, flags);
1746
1747 bail:
1748         mlog_exit(ret);
1749         return ret;
1750 }
1751
1752 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
1753                                    const struct iovec *iov,
1754                                    unsigned long nr_segs,
1755                                    loff_t pos)
1756 {
1757         int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
1758         struct file *filp = iocb->ki_filp;
1759         struct inode *inode = filp->f_path.dentry->d_inode;
1760
1761         mlog_entry("(0x%p, %u, '%.*s')\n", filp,
1762                    (unsigned int)nr_segs,
1763                    filp->f_path.dentry->d_name.len,
1764                    filp->f_path.dentry->d_name.name);
1765
1766         if (!inode) {
1767                 ret = -EINVAL;
1768                 mlog_errno(ret);
1769                 goto bail;
1770         }
1771
1772         /*
1773          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
1774          * need locks to protect pending reads from racing with truncate.
1775          */
1776         if (filp->f_flags & O_DIRECT) {
1777                 down_read(&inode->i_alloc_sem);
1778                 have_alloc_sem = 1;
1779
1780                 ret = ocfs2_rw_lock(inode, 0);
1781                 if (ret < 0) {
1782                         mlog_errno(ret);
1783                         goto bail;
1784                 }
1785                 rw_level = 0;
1786                 /* communicate with ocfs2_dio_end_io */
1787                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
1788         }
1789
1790         /*
1791          * We're fine letting folks race truncates and extending
1792          * writes with read across the cluster, just like they can
1793          * locally. Hence no rw_lock during read.
1794          *
1795          * Take and drop the meta data lock to update inode fields
1796          * like i_size. This allows the checks down below
1797          * generic_file_aio_read() a chance of actually working.
1798          */
1799         ret = ocfs2_meta_lock_atime(inode, filp->f_vfsmnt, &lock_level);
1800         if (ret < 0) {
1801                 mlog_errno(ret);
1802                 goto bail;
1803         }
1804         ocfs2_meta_unlock(inode, lock_level);
1805
1806         ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
1807         if (ret == -EINVAL)
1808                 mlog(ML_ERROR, "generic_file_aio_read returned -EINVAL\n");
1809
1810         /* buffered aio wouldn't have proper lock coverage today */
1811         BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
1812
1813         /* see ocfs2_file_aio_write */
1814         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
1815                 rw_level = -1;
1816                 have_alloc_sem = 0;
1817         }
1818
1819 bail:
1820         if (have_alloc_sem)
1821                 up_read(&inode->i_alloc_sem);
1822         if (rw_level != -1)
1823                 ocfs2_rw_unlock(inode, rw_level);
1824         mlog_exit(ret);
1825
1826         return ret;
1827 }
1828
1829 const struct inode_operations ocfs2_file_iops = {
1830         .setattr        = ocfs2_setattr,
1831         .getattr        = ocfs2_getattr,
1832         .permission     = ocfs2_permission,
1833 };
1834
1835 const struct inode_operations ocfs2_special_file_iops = {
1836         .setattr        = ocfs2_setattr,
1837         .getattr        = ocfs2_getattr,
1838         .permission     = ocfs2_permission,
1839 };
1840
1841 const struct file_operations ocfs2_fops = {
1842         .read           = do_sync_read,
1843         .write          = do_sync_write,
1844         .mmap           = ocfs2_mmap,
1845         .fsync          = ocfs2_sync_file,
1846         .release        = ocfs2_file_release,
1847         .open           = ocfs2_file_open,
1848         .aio_read       = ocfs2_file_aio_read,
1849         .aio_write      = ocfs2_file_aio_write,
1850         .ioctl          = ocfs2_ioctl,
1851 #ifdef CONFIG_COMPAT
1852         .compat_ioctl   = ocfs2_compat_ioctl,
1853 #endif
1854         .splice_read    = ocfs2_file_splice_read,
1855         .splice_write   = ocfs2_file_splice_write,
1856 };
1857
1858 const struct file_operations ocfs2_dops = {
1859         .read           = generic_read_dir,
1860         .readdir        = ocfs2_readdir,
1861         .fsync          = ocfs2_sync_file,
1862         .ioctl          = ocfs2_ioctl,
1863 #ifdef CONFIG_COMPAT
1864         .compat_ioctl   = ocfs2_compat_ioctl,
1865 #endif
1866 };