SAFE public projects git trees. - safe/jmp/linux-2.6/blob - fs/xfs/linux-2.6/xfs_lrw.c

   1 /*
   2  * Copyright (c) 2000-2003,2005 Silicon Graphics, Inc.
   3  * All Rights Reserved.
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU General Public License as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it would be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write the Free Software Foundation,
  16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17  */
  18 #include "xfs.h"
  19 #include "xfs_fs.h"
  20 #include "xfs_bit.h"
  21 #include "xfs_log.h"
  22 #include "xfs_inum.h"
  23 #include "xfs_trans.h"
  24 #include "xfs_sb.h"
  25 #include "xfs_ag.h"
  26 #include "xfs_dir2.h"
  27 #include "xfs_alloc.h"
  28 #include "xfs_dmapi.h"
  29 #include "xfs_quota.h"
  30 #include "xfs_mount.h"
  31 #include "xfs_bmap_btree.h"
  32 #include "xfs_alloc_btree.h"
  33 #include "xfs_ialloc_btree.h"
  34 #include "xfs_dir2_sf.h"
  35 #include "xfs_attr_sf.h"
  36 #include "xfs_dinode.h"
  37 #include "xfs_inode.h"
  38 #include "xfs_bmap.h"
  39 #include "xfs_btree.h"
  40 #include "xfs_ialloc.h"
  41 #include "xfs_rtalloc.h"
  42 #include "xfs_error.h"
  43 #include "xfs_itable.h"
  44 #include "xfs_rw.h"
  45 #include "xfs_attr.h"
  46 #include "xfs_inode_item.h"
  47 #include "xfs_buf_item.h"
  48 #include "xfs_utils.h"
  49 #include "xfs_iomap.h"
  50 #include "xfs_vnodeops.h"
  51
  52 #include <linux/capability.h>
  53 #include <linux/writeback.h>
  54
  55
  56 #if defined(XFS_RW_TRACE)
  57 void
  58 xfs_rw_enter_trace(
  59         int                     tag,
  60         xfs_inode_t             *ip,
  61         void                    *data,
  62         size_t                  segs,
  63         loff_t                  offset,
  64         int                     ioflags)
  65 {
  66         if (ip->i_rwtrace == NULL)
  67                 return;
  68         ktrace_enter(ip->i_rwtrace,
  69                 (void *)(unsigned long)tag,
  70                 (void *)ip,
  71                 (void *)((unsigned long)((ip->i_d.di_size >> 32) & 0xffffffff)),
  72                 (void *)((unsigned long)(ip->i_d.di_size & 0xffffffff)),
  73                 (void *)data,
  74                 (void *)((unsigned long)segs),
  75                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
  76                 (void *)((unsigned long)(offset & 0xffffffff)),
  77                 (void *)((unsigned long)ioflags),
  78                 (void *)((unsigned long)((ip->i_new_size >> 32) & 0xffffffff)),
  79                 (void *)((unsigned long)(ip->i_new_size & 0xffffffff)),
  80                 (void *)((unsigned long)current_pid()),
  81                 (void *)NULL,
  82                 (void *)NULL,
  83                 (void *)NULL,
  84                 (void *)NULL);
  85 }
  86
  87 void
  88 xfs_inval_cached_trace(
  89         xfs_inode_t     *ip,
  90         xfs_off_t       offset,
  91         xfs_off_t       len,
  92         xfs_off_t       first,
  93         xfs_off_t       last)
  94 {
  95
  96         if (ip->i_rwtrace == NULL)
  97                 return;
  98         ktrace_enter(ip->i_rwtrace,
  99                 (void *)(__psint_t)XFS_INVAL_CACHED,
 100                 (void *)ip,
 101                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
 102                 (void *)((unsigned long)(offset & 0xffffffff)),
 103                 (void *)((unsigned long)((len >> 32) & 0xffffffff)),
 104                 (void *)((unsigned long)(len & 0xffffffff)),
 105                 (void *)((unsigned long)((first >> 32) & 0xffffffff)),
 106                 (void *)((unsigned long)(first & 0xffffffff)),
 107                 (void *)((unsigned long)((last >> 32) & 0xffffffff)),
 108                 (void *)((unsigned long)(last & 0xffffffff)),
 109                 (void *)((unsigned long)current_pid()),
 110                 (void *)NULL,
 111                 (void *)NULL,
 112                 (void *)NULL,
 113                 (void *)NULL,
 114                 (void *)NULL);
 115 }
 116 #endif
 117
 118 /*
 119  *      xfs_iozero
 120  *
 121  *      xfs_iozero clears the specified range of buffer supplied,
 122  *      and marks all the affected blocks as valid and modified.  If
 123  *      an affected block is not allocated, it will be allocated.  If
 124  *      an affected block is not completely overwritten, and is not
 125  *      valid before the operation, it will be read from disk before
 126  *      being partially zeroed.
 127  */
 128 STATIC int
 129 xfs_iozero(
 130         struct xfs_inode        *ip,    /* inode                        */
 131         loff_t                  pos,    /* offset in file               */
 132         size_t                  count)  /* size of data to zero         */
 133 {
 134         struct page             *page;
 135         struct address_space    *mapping;
 136         int                     status;
 137
 138         mapping = VFS_I(ip)->i_mapping;
 139         do {
 140                 unsigned offset, bytes;
 141                 void *fsdata;
 142
 143                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
 144                 bytes = PAGE_CACHE_SIZE - offset;
 145                 if (bytes > count)
 146                         bytes = count;
 147
 148                 status = pagecache_write_begin(NULL, mapping, pos, bytes,
 149                                         AOP_FLAG_UNINTERRUPTIBLE,
 150                                         &page, &fsdata);
 151                 if (status)
 152                         break;
 153
 154                 zero_user(page, offset, bytes);
 155
 156                 status = pagecache_write_end(NULL, mapping, pos, bytes, bytes,
 157                                         page, fsdata);
 158                 WARN_ON(status <= 0); /* can't return less than zero! */
 159                 pos += bytes;
 160                 count -= bytes;
 161                 status = 0;
 162         } while (count);
 163
 164         return (-status);
 165 }
 166
 167 ssize_t                 /* bytes read, or (-)  error */
 168 xfs_read(
 169         xfs_inode_t             *ip,
 170         struct kiocb            *iocb,
 171         const struct iovec      *iovp,
 172         unsigned int            segs,
 173         loff_t                  *offset,
 174         int                     ioflags)
 175 {
 176         struct file             *file = iocb->ki_filp;
 177         struct inode            *inode = file->f_mapping->host;
 178         xfs_mount_t             *mp = ip->i_mount;
 179         size_t                  size = 0;
 180         ssize_t                 ret = 0;
 181         xfs_fsize_t             n;
 182         unsigned long           seg;
 183
 184
 185         XFS_STATS_INC(xs_read_calls);
 186
 187         /* START copy & waste from filemap.c */
 188         for (seg = 0; seg < segs; seg++) {
 189                 const struct iovec *iv = &iovp[seg];
 190
 191                 /*
 192                  * If any segment has a negative length, or the cumulative
 193                  * length ever wraps negative then return -EINVAL.
 194                  */
 195                 size += iv->iov_len;
 196                 if (unlikely((ssize_t)(size|iv->iov_len) < 0))
 197                         return XFS_ERROR(-EINVAL);
 198         }
 199         /* END copy & waste from filemap.c */
 200
 201         if (unlikely(ioflags & IO_ISDIRECT)) {
 202                 xfs_buftarg_t   *target =
 203                         XFS_IS_REALTIME_INODE(ip) ?
 204                                 mp->m_rtdev_targp : mp->m_ddev_targp;
 205                 if ((*offset & target->bt_smask) ||
 206                     (size & target->bt_smask)) {
 207                         if (*offset == ip->i_size) {
 208                                 return (0);
 209                         }
 210                         return -XFS_ERROR(EINVAL);
 211                 }
 212         }
 213
 214         n = XFS_MAXIOFFSET(mp) - *offset;
 215         if ((n <= 0) || (size == 0))
 216                 return 0;
 217
 218         if (n < size)
 219                 size = n;
 220
 221         if (XFS_FORCED_SHUTDOWN(mp))
 222                 return -EIO;
 223
 224         if (unlikely(ioflags & IO_ISDIRECT))
 225                 mutex_lock(&inode->i_mutex);
 226         xfs_ilock(ip, XFS_IOLOCK_SHARED);
 227
 228         if (DM_EVENT_ENABLED(ip, DM_EVENT_READ) && !(ioflags & IO_INVIS)) {
 229                 int dmflags = FILP_DELAY_FLAG(file) | DM_SEM_FLAG_RD(ioflags);
 230                 int iolock = XFS_IOLOCK_SHARED;
 231
 232                 ret = -XFS_SEND_DATA(mp, DM_EVENT_READ, ip, *offset, size,
 233                                         dmflags, &iolock);
 234                 if (ret) {
 235                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 236                         if (unlikely(ioflags & IO_ISDIRECT))
 237                                 mutex_unlock(&inode->i_mutex);
 238                         return ret;
 239                 }
 240         }
 241
 242         if (unlikely(ioflags & IO_ISDIRECT)) {
 243                 if (inode->i_mapping->nrpages)
 244                         ret = -xfs_flushinval_pages(ip, (*offset & PAGE_CACHE_MASK),
 245                                                     -1, FI_REMAPF_LOCKED);
 246                 mutex_unlock(&inode->i_mutex);
 247                 if (ret) {
 248                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 249                         return ret;
 250                 }
 251         }
 252
 253         xfs_rw_enter_trace(XFS_READ_ENTER, ip,
 254                                 (void *)iovp, segs, *offset, ioflags);
 255
 256         iocb->ki_pos = *offset;
 257         ret = generic_file_aio_read(iocb, iovp, segs, *offset);
 258         if (ret > 0)
 259                 XFS_STATS_ADD(xs_read_bytes, ret);
 260
 261         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 262         return ret;
 263 }
 264
 265 ssize_t
 266 xfs_splice_read(
 267         xfs_inode_t             *ip,
 268         struct file             *infilp,
 269         loff_t                  *ppos,
 270         struct pipe_inode_info  *pipe,
 271         size_t                  count,
 272         int                     flags,
 273         int                     ioflags)
 274 {
 275         xfs_mount_t             *mp = ip->i_mount;
 276         ssize_t                 ret;
 277
 278         XFS_STATS_INC(xs_read_calls);
 279         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 280                 return -EIO;
 281
 282         xfs_ilock(ip, XFS_IOLOCK_SHARED);
 283
 284         if (DM_EVENT_ENABLED(ip, DM_EVENT_READ) && !(ioflags & IO_INVIS)) {
 285                 int iolock = XFS_IOLOCK_SHARED;
 286                 int error;
 287
 288                 error = XFS_SEND_DATA(mp, DM_EVENT_READ, ip, *ppos, count,
 289                                         FILP_DELAY_FLAG(infilp), &iolock);
 290                 if (error) {
 291                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 292                         return -error;
 293                 }
 294         }
 295         xfs_rw_enter_trace(XFS_SPLICE_READ_ENTER, ip,
 296                            pipe, count, *ppos, ioflags);
 297         ret = generic_file_splice_read(infilp, ppos, pipe, count, flags);
 298         if (ret > 0)
 299                 XFS_STATS_ADD(xs_read_bytes, ret);
 300
 301         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 302         return ret;
 303 }
 304
 305 ssize_t
 306 xfs_splice_write(
 307         xfs_inode_t             *ip,
 308         struct pipe_inode_info  *pipe,
 309         struct file             *outfilp,
 310         loff_t                  *ppos,
 311         size_t                  count,
 312         int                     flags,
 313         int                     ioflags)
 314 {
 315         xfs_mount_t             *mp = ip->i_mount;
 316         ssize_t                 ret;
 317         struct inode            *inode = outfilp->f_mapping->host;
 318         xfs_fsize_t             isize, new_size;
 319
 320         XFS_STATS_INC(xs_write_calls);
 321         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 322                 return -EIO;
 323
 324         xfs_ilock(ip, XFS_IOLOCK_EXCL);
 325
 326         if (DM_EVENT_ENABLED(ip, DM_EVENT_WRITE) && !(ioflags & IO_INVIS)) {
 327                 int iolock = XFS_IOLOCK_EXCL;
 328                 int error;
 329
 330                 error = XFS_SEND_DATA(mp, DM_EVENT_WRITE, ip, *ppos, count,
 331                                         FILP_DELAY_FLAG(outfilp), &iolock);
 332                 if (error) {
 333                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 334                         return -error;
 335                 }
 336         }
 337
 338         new_size = *ppos + count;
 339
 340         xfs_ilock(ip, XFS_ILOCK_EXCL);
 341         if (new_size > ip->i_size)
 342                 ip->i_new_size = new_size;
 343         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 344
 345         xfs_rw_enter_trace(XFS_SPLICE_WRITE_ENTER, ip,
 346                            pipe, count, *ppos, ioflags);
 347         ret = generic_file_splice_write(pipe, outfilp, ppos, count, flags);
 348         if (ret > 0)
 349                 XFS_STATS_ADD(xs_write_bytes, ret);
 350
 351         isize = i_size_read(inode);
 352         if (unlikely(ret < 0 && ret != -EFAULT && *ppos > isize))
 353                 *ppos = isize;
 354
 355         if (*ppos > ip->i_size) {
 356                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 357                 if (*ppos > ip->i_size)
 358                         ip->i_size = *ppos;
 359                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 360         }
 361
 362         if (ip->i_new_size) {
 363                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 364                 ip->i_new_size = 0;
 365                 if (ip->i_d.di_size > ip->i_size)
 366                         ip->i_d.di_size = ip->i_size;
 367                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 368         }
 369         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 370         return ret;
 371 }
 372
 373 /*
 374  * This routine is called to handle zeroing any space in the last
 375  * block of the file that is beyond the EOF.  We do this since the
 376  * size is being increased without writing anything to that block
 377  * and we don't want anyone to read the garbage on the disk.
 378  */
 379 STATIC int                              /* error (positive) */
 380 xfs_zero_last_block(
 381         xfs_inode_t     *ip,
 382         xfs_fsize_t     offset,
 383         xfs_fsize_t     isize)
 384 {
 385         xfs_fileoff_t   last_fsb;
 386         xfs_mount_t     *mp = ip->i_mount;
 387         int             nimaps;
 388         int             zero_offset;
 389         int             zero_len;
 390         int             error = 0;
 391         xfs_bmbt_irec_t imap;
 392
 393         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 394
 395         zero_offset = XFS_B_FSB_OFFSET(mp, isize);
 396         if (zero_offset == 0) {
 397                 /*
 398                  * There are no extra bytes in the last block on disk to
 399                  * zero, so return.
 400                  */
 401                 return 0;
 402         }
 403
 404         last_fsb = XFS_B_TO_FSBT(mp, isize);
 405         nimaps = 1;
 406         error = xfs_bmapi(NULL, ip, last_fsb, 1, 0, NULL, 0, &imap,
 407                           &nimaps, NULL, NULL);
 408         if (error) {
 409                 return error;
 410         }
 411         ASSERT(nimaps > 0);
 412         /*
 413          * If the block underlying isize is just a hole, then there
 414          * is nothing to zero.
 415          */
 416         if (imap.br_startblock == HOLESTARTBLOCK) {
 417                 return 0;
 418         }
 419         /*
 420          * Zero the part of the last block beyond the EOF, and write it
 421          * out sync.  We need to drop the ilock while we do this so we
 422          * don't deadlock when the buffer cache calls back to us.
 423          */
 424         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 425
 426         zero_len = mp->m_sb.sb_blocksize - zero_offset;
 427         if (isize + zero_len > offset)
 428                 zero_len = offset - isize;
 429         error = xfs_iozero(ip, isize, zero_len);
 430
 431         xfs_ilock(ip, XFS_ILOCK_EXCL);
 432         ASSERT(error >= 0);
 433         return error;
 434 }
 435
 436 /*
 437  * Zero any on disk space between the current EOF and the new,
 438  * larger EOF.  This handles the normal case of zeroing the remainder
 439  * of the last block in the file and the unusual case of zeroing blocks
 440  * out beyond the size of the file.  This second case only happens
 441  * with fixed size extents and when the system crashes before the inode
 442  * size was updated but after blocks were allocated.  If fill is set,
 443  * then any holes in the range are filled and zeroed.  If not, the holes
 444  * are left alone as holes.
 445  */
 446
 447 int                                     /* error (positive) */
 448 xfs_zero_eof(
 449         xfs_inode_t     *ip,
 450         xfs_off_t       offset,         /* starting I/O offset */
 451         xfs_fsize_t     isize)          /* current inode size */
 452 {
 453         xfs_mount_t     *mp = ip->i_mount;
 454         xfs_fileoff_t   start_zero_fsb;
 455         xfs_fileoff_t   end_zero_fsb;
 456         xfs_fileoff_t   zero_count_fsb;
 457         xfs_fileoff_t   last_fsb;
 458         xfs_fileoff_t   zero_off;
 459         xfs_fsize_t     zero_len;
 460         int             nimaps;
 461         int             error = 0;
 462         xfs_bmbt_irec_t imap;
 463
 464         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_IOLOCK_EXCL));
 465         ASSERT(offset > isize);
 466
 467         /*
 468          * First handle zeroing the block on which isize resides.
 469          * We only zero a part of that block so it is handled specially.
 470          */
 471         error = xfs_zero_last_block(ip, offset, isize);
 472         if (error) {
 473                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_IOLOCK_EXCL));
 474                 return error;
 475         }
 476
 477         /*
 478          * Calculate the range between the new size and the old
 479          * where blocks needing to be zeroed may exist.  To get the
 480          * block where the last byte in the file currently resides,
 481          * we need to subtract one from the size and truncate back
 482          * to a block boundary.  We subtract 1 in case the size is
 483          * exactly on a block boundary.
 484          */
 485         last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
 486         start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
 487         end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
 488         ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
 489         if (last_fsb == end_zero_fsb) {
 490                 /*
 491                  * The size was only incremented on its last block.
 492                  * We took care of that above, so just return.
 493                  */
 494                 return 0;
 495         }
 496
 497         ASSERT(start_zero_fsb <= end_zero_fsb);
 498         while (start_zero_fsb <= end_zero_fsb) {
 499                 nimaps = 1;
 500                 zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
 501                 error = xfs_bmapi(NULL, ip, start_zero_fsb, zero_count_fsb,
 502                                   0, NULL, 0, &imap, &nimaps, NULL, NULL);
 503                 if (error) {
 504                         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_IOLOCK_EXCL));
 505                         return error;
 506                 }
 507                 ASSERT(nimaps > 0);
 508
 509                 if (imap.br_state == XFS_EXT_UNWRITTEN ||
 510                     imap.br_startblock == HOLESTARTBLOCK) {
 511                         /*
 512                          * This loop handles initializing pages that were
 513                          * partially initialized by the code below this
 514                          * loop. It basically zeroes the part of the page
 515                          * that sits on a hole and sets the page as P_HOLE
 516                          * and calls remapf if it is a mapped file.
 517                          */
 518                         start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 519                         ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 520                         continue;
 521                 }
 522
 523                 /*
 524                  * There are blocks we need to zero.
 525                  * Drop the inode lock while we're doing the I/O.
 526                  * We'll still have the iolock to protect us.
 527                  */
 528                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 529
 530                 zero_off = XFS_FSB_TO_B(mp, start_zero_fsb);
 531                 zero_len = XFS_FSB_TO_B(mp, imap.br_blockcount);
 532
 533                 if ((zero_off + zero_len) > offset)
 534                         zero_len = offset - zero_off;
 535
 536                 error = xfs_iozero(ip, zero_off, zero_len);
 537                 if (error) {
 538                         goto out_lock;
 539                 }
 540
 541                 start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 542                 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 543
 544                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 545         }
 546
 547         return 0;
 548
 549 out_lock:
 550         xfs_ilock(ip, XFS_ILOCK_EXCL);
 551         ASSERT(error >= 0);
 552         return error;
 553 }
 554
 555 ssize_t                         /* bytes written, or (-) error */
 556 xfs_write(
 557         struct xfs_inode        *xip,
 558         struct kiocb            *iocb,
 559         const struct iovec      *iovp,
 560         unsigned int            nsegs,
 561         loff_t                  *offset,
 562         int                     ioflags)
 563 {
 564         struct file             *file = iocb->ki_filp;
 565         struct address_space    *mapping = file->f_mapping;
 566         struct inode            *inode = mapping->host;
 567         unsigned long           segs = nsegs;
 568         xfs_mount_t             *mp;
 569         ssize_t                 ret = 0, error = 0;
 570         xfs_fsize_t             isize, new_size;
 571         int                     iolock;
 572         int                     eventsent = 0;
 573         size_t                  ocount = 0, count;
 574         loff_t                  pos;
 575         int                     need_i_mutex;
 576
 577         XFS_STATS_INC(xs_write_calls);
 578
 579         error = generic_segment_checks(iovp, &segs, &ocount, VERIFY_READ);
 580         if (error)
 581                 return error;
 582
 583         count = ocount;
 584         pos = *offset;
 585
 586         if (count == 0)
 587                 return 0;
 588
 589         mp = xip->i_mount;
 590
 591         xfs_wait_for_freeze(mp, SB_FREEZE_WRITE);
 592
 593         if (XFS_FORCED_SHUTDOWN(mp))
 594                 return -EIO;
 595
 596 relock:
 597         if (ioflags & IO_ISDIRECT) {
 598                 iolock = XFS_IOLOCK_SHARED;
 599                 need_i_mutex = 0;
 600         } else {
 601                 iolock = XFS_IOLOCK_EXCL;
 602                 need_i_mutex = 1;
 603                 mutex_lock(&inode->i_mutex);
 604         }
 605
 606         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
 607
 608 start:
 609         error = -generic_write_checks(file, &pos, &count,
 610                                         S_ISBLK(inode->i_mode));
 611         if (error) {
 612                 xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 613                 goto out_unlock_mutex;
 614         }
 615
 616         if ((DM_EVENT_ENABLED(xip, DM_EVENT_WRITE) &&
 617             !(ioflags & IO_INVIS) && !eventsent)) {
 618                 int             dmflags = FILP_DELAY_FLAG(file);
 619
 620                 if (need_i_mutex)
 621                         dmflags |= DM_FLAGS_IMUX;
 622
 623                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 624                 error = XFS_SEND_DATA(xip->i_mount, DM_EVENT_WRITE, xip,
 625                                       pos, count, dmflags, &iolock);
 626                 if (error) {
 627                         goto out_unlock_internal;
 628                 }
 629                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 630                 eventsent = 1;
 631
 632                 /*
 633                  * The iolock was dropped and reacquired in XFS_SEND_DATA
 634                  * so we have to recheck the size when appending.
 635                  * We will only "goto start;" once, since having sent the
 636                  * event prevents another call to XFS_SEND_DATA, which is
 637                  * what allows the size to change in the first place.
 638                  */
 639                 if ((file->f_flags & O_APPEND) && pos != xip->i_size)
 640                         goto start;
 641         }
 642
 643         if (ioflags & IO_ISDIRECT) {
 644                 xfs_buftarg_t   *target =
 645                         XFS_IS_REALTIME_INODE(xip) ?
 646                                 mp->m_rtdev_targp : mp->m_ddev_targp;
 647
 648                 if ((pos & target->bt_smask) || (count & target->bt_smask)) {
 649                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 650                         return XFS_ERROR(-EINVAL);
 651                 }
 652
 653                 if (!need_i_mutex && (mapping->nrpages || pos > xip->i_size)) {
 654                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 655                         iolock = XFS_IOLOCK_EXCL;
 656                         need_i_mutex = 1;
 657                         mutex_lock(&inode->i_mutex);
 658                         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
 659                         goto start;
 660                 }
 661         }
 662
 663         new_size = pos + count;
 664         if (new_size > xip->i_size)
 665                 xip->i_new_size = new_size;
 666
 667         if (likely(!(ioflags & IO_INVIS)))
 668                 file_update_time(file);
 669
 670         /*
 671          * If the offset is beyond the size of the file, we have a couple
 672          * of things to do. First, if there is already space allocated
 673          * we need to either create holes or zero the disk or ...
 674          *
 675          * If there is a page where the previous size lands, we need
 676          * to zero it out up to the new size.
 677          */
 678
 679         if (pos > xip->i_size) {
 680                 error = xfs_zero_eof(xip, pos, xip->i_size);
 681                 if (error) {
 682                         xfs_iunlock(xip, XFS_ILOCK_EXCL);
 683                         goto out_unlock_internal;
 684                 }
 685         }
 686         xfs_iunlock(xip, XFS_ILOCK_EXCL);
 687
 688         /*
 689          * If we're writing the file then make sure to clear the
 690          * setuid and setgid bits if the process is not being run
 691          * by root.  This keeps people from modifying setuid and
 692          * setgid binaries.
 693          */
 694
 695         if (((xip->i_d.di_mode & S_ISUID) ||
 696             ((xip->i_d.di_mode & (S_ISGID | S_IXGRP)) ==
 697                 (S_ISGID | S_IXGRP))) &&
 698              !capable(CAP_FSETID)) {
 699                 error = xfs_write_clear_setuid(xip);
 700                 if (likely(!error))
 701                         error = -file_remove_suid(file);
 702                 if (unlikely(error)) {
 703                         goto out_unlock_internal;
 704                 }
 705         }
 706
 707         /* We can write back this queue in page reclaim */
 708         current->backing_dev_info = mapping->backing_dev_info;
 709
 710         if ((ioflags & IO_ISDIRECT)) {
 711                 if (mapping->nrpages) {
 712                         WARN_ON(need_i_mutex == 0);
 713                         xfs_inval_cached_trace(xip, pos, -1,
 714                                         (pos & PAGE_CACHE_MASK), -1);
 715                         error = xfs_flushinval_pages(xip,
 716                                         (pos & PAGE_CACHE_MASK),
 717                                         -1, FI_REMAPF_LOCKED);
 718                         if (error)
 719                                 goto out_unlock_internal;
 720                 }
 721
 722                 if (need_i_mutex) {
 723                         /* demote the lock now the cached pages are gone */
 724                         xfs_ilock_demote(xip, XFS_IOLOCK_EXCL);
 725                         mutex_unlock(&inode->i_mutex);
 726
 727                         iolock = XFS_IOLOCK_SHARED;
 728                         need_i_mutex = 0;
 729                 }
 730
 731                 xfs_rw_enter_trace(XFS_DIOWR_ENTER, xip, (void *)iovp, segs,
 732                                 *offset, ioflags);
 733                 ret = generic_file_direct_write(iocb, iovp,
 734                                 &segs, pos, offset, count, ocount);
 735
 736                 /*
 737                  * direct-io write to a hole: fall through to buffered I/O
 738                  * for completing the rest of the request.
 739                  */
 740                 if (ret >= 0 && ret != count) {
 741                         XFS_STATS_ADD(xs_write_bytes, ret);
 742
 743                         pos += ret;
 744                         count -= ret;
 745
 746                         ioflags &= ~IO_ISDIRECT;
 747                         xfs_iunlock(xip, iolock);
 748                         goto relock;
 749                 }
 750         } else {
 751                 int enospc = 0;
 752                 ssize_t ret2 = 0;
 753
 754 write_retry:
 755                 xfs_rw_enter_trace(XFS_WRITE_ENTER, xip, (void *)iovp, segs,
 756                                 *offset, ioflags);
 757                 ret2 = generic_file_buffered_write(iocb, iovp, segs,
 758                                 pos, offset, count, ret);
 759                 /*
 760                  * if we just got an ENOSPC, flush the inode now we
 761                  * aren't holding any page locks and retry *once*
 762                  */
 763                 if (ret2 == -ENOSPC && !enospc) {
 764                         error = xfs_flush_pages(xip, 0, -1, 0, FI_NONE);
 765                         if (error)
 766                                 goto out_unlock_internal;
 767                         enospc = 1;
 768                         goto write_retry;
 769                 }
 770                 ret = ret2;
 771         }
 772
 773         current->backing_dev_info = NULL;
 774
 775         isize = i_size_read(inode);
 776         if (unlikely(ret < 0 && ret != -EFAULT && *offset > isize))
 777                 *offset = isize;
 778
 779         if (*offset > xip->i_size) {
 780                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 781                 if (*offset > xip->i_size)
 782                         xip->i_size = *offset;
 783                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 784         }
 785
 786         if (ret == -ENOSPC &&
 787             DM_EVENT_ENABLED(xip, DM_EVENT_NOSPACE) && !(ioflags & IO_INVIS)) {
 788                 xfs_iunlock(xip, iolock);
 789                 if (need_i_mutex)
 790                         mutex_unlock(&inode->i_mutex);
 791                 error = XFS_SEND_NAMESP(xip->i_mount, DM_EVENT_NOSPACE, xip,
 792                                 DM_RIGHT_NULL, xip, DM_RIGHT_NULL, NULL, NULL,
 793                                 0, 0, 0); /* Delay flag intentionally  unused */
 794                 if (need_i_mutex)
 795                         mutex_lock(&inode->i_mutex);
 796                 xfs_ilock(xip, iolock);
 797                 if (error)
 798                         goto out_unlock_internal;
 799                 goto start;
 800         }
 801
 802         error = -ret;
 803         if (ret <= 0)
 804                 goto out_unlock_internal;
 805
 806         XFS_STATS_ADD(xs_write_bytes, ret);
 807
 808         /* Handle various SYNC-type writes */
 809         if ((file->f_flags & O_SYNC) || IS_SYNC(inode)) {
 810                 loff_t end = pos + ret - 1;
 811                 int error2;
 812
 813                 xfs_iunlock(xip, iolock);
 814                 if (need_i_mutex)
 815                         mutex_unlock(&inode->i_mutex);
 816
 817                 error2 = filemap_write_and_wait_range(mapping, pos, end);
 818                 if (!error)
 819                         error = error2;
 820                 if (need_i_mutex)
 821                         mutex_lock(&inode->i_mutex);
 822                 xfs_ilock(xip, iolock);
 823
 824                 error2 = xfs_fsync(xip);
 825                 if (!error)
 826                         error = error2;
 827         }
 828
 829  out_unlock_internal:
 830         if (xip->i_new_size) {
 831                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 832                 xip->i_new_size = 0;
 833                 /*
 834                  * If this was a direct or synchronous I/O that failed (such
 835                  * as ENOSPC) then part of the I/O may have been written to
 836                  * disk before the error occured.  In this case the on-disk
 837                  * file size may have been adjusted beyond the in-memory file
 838                  * size and now needs to be truncated back.
 839                  */
 840                 if (xip->i_d.di_size > xip->i_size)
 841                         xip->i_d.di_size = xip->i_size;
 842                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 843         }
 844         xfs_iunlock(xip, iolock);
 845  out_unlock_mutex:
 846         if (need_i_mutex)
 847                 mutex_unlock(&inode->i_mutex);
 848         return -error;
 849 }
 850
 851 /*
 852  * All xfs metadata buffers except log state machine buffers
 853  * get this attached as their b_bdstrat callback function.
 854  * This is so that we can catch a buffer
 855  * after prematurely unpinning it to forcibly shutdown the filesystem.
 856  */
 857 int
 858 xfs_bdstrat_cb(struct xfs_buf *bp)
 859 {
 860         if (XFS_FORCED_SHUTDOWN(bp->b_mount)) {
 861                 xfs_buftrace("XFS__BDSTRAT IOERROR", bp);
 862                 /*
 863                  * Metadata write that didn't get logged but
 864                  * written delayed anyway. These aren't associated
 865                  * with a transaction, and can be ignored.
 866                  */
 867                 if (XFS_BUF_IODONE_FUNC(bp) == NULL &&
 868                     (XFS_BUF_ISREAD(bp)) == 0)
 869                         return (xfs_bioerror_relse(bp));
 870                 else
 871                         return (xfs_bioerror(bp));
 872         }
 873
 874         xfs_buf_iorequest(bp);
 875         return 0;
 876 }
 877
 878 /*
 879  * Wrapper around bdstrat so that we can stop data from going to disk in case
 880  * we are shutting down the filesystem.  Typically user data goes thru this
 881  * path; one of the exceptions is the superblock.
 882  */
 883 void
 884 xfsbdstrat(
 885         struct xfs_mount        *mp,
 886         struct xfs_buf          *bp)
 887 {
 888         ASSERT(mp);
 889         if (!XFS_FORCED_SHUTDOWN(mp)) {
 890                 xfs_buf_iorequest(bp);
 891                 return;
 892         }
 893
 894         xfs_buftrace("XFSBDSTRAT IOERROR", bp);
 895         xfs_bioerror_relse(bp);
 896 }
 897
 898 /*
 899  * If the underlying (data/log/rt) device is readonly, there are some
 900  * operations that cannot proceed.
 901  */
 902 int
 903 xfs_dev_is_read_only(
 904         xfs_mount_t             *mp,
 905         char                    *message)
 906 {
 907         if (xfs_readonly_buftarg(mp->m_ddev_targp) ||
 908             xfs_readonly_buftarg(mp->m_logdev_targp) ||
 909             (mp->m_rtdev_targp && xfs_readonly_buftarg(mp->m_rtdev_targp))) {
 910                 cmn_err(CE_NOTE,
 911                         "XFS: %s required on read-only device.", message);
 912                 cmn_err(CE_NOTE,
 913                         "XFS: write access unavailable, cannot proceed.");
 914                 return EROFS;
 915         }
 916         return 0;
 917 }