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 == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
 259                 ret = wait_on_sync_kiocb(iocb);
 260         if (ret > 0)
 261                 XFS_STATS_ADD(xs_read_bytes, ret);
 262
 263         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 264         return ret;
 265 }
 266
 267 ssize_t
 268 xfs_splice_read(
 269         xfs_inode_t             *ip,
 270         struct file             *infilp,
 271         loff_t                  *ppos,
 272         struct pipe_inode_info  *pipe,
 273         size_t                  count,
 274         int                     flags,
 275         int                     ioflags)
 276 {
 277         xfs_mount_t             *mp = ip->i_mount;
 278         ssize_t                 ret;
 279
 280         XFS_STATS_INC(xs_read_calls);
 281         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 282                 return -EIO;
 283
 284         xfs_ilock(ip, XFS_IOLOCK_SHARED);
 285
 286         if (DM_EVENT_ENABLED(ip, DM_EVENT_READ) && !(ioflags & IO_INVIS)) {
 287                 int iolock = XFS_IOLOCK_SHARED;
 288                 int error;
 289
 290                 error = XFS_SEND_DATA(mp, DM_EVENT_READ, ip, *ppos, count,
 291                                         FILP_DELAY_FLAG(infilp), &iolock);
 292                 if (error) {
 293                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 294                         return -error;
 295                 }
 296         }
 297         xfs_rw_enter_trace(XFS_SPLICE_READ_ENTER, ip,
 298                            pipe, count, *ppos, ioflags);
 299         ret = generic_file_splice_read(infilp, ppos, pipe, count, flags);
 300         if (ret > 0)
 301                 XFS_STATS_ADD(xs_read_bytes, ret);
 302
 303         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 304         return ret;
 305 }
 306
 307 ssize_t
 308 xfs_splice_write(
 309         xfs_inode_t             *ip,
 310         struct pipe_inode_info  *pipe,
 311         struct file             *outfilp,
 312         loff_t                  *ppos,
 313         size_t                  count,
 314         int                     flags,
 315         int                     ioflags)
 316 {
 317         xfs_mount_t             *mp = ip->i_mount;
 318         ssize_t                 ret;
 319         struct inode            *inode = outfilp->f_mapping->host;
 320         xfs_fsize_t             isize, new_size;
 321
 322         XFS_STATS_INC(xs_write_calls);
 323         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 324                 return -EIO;
 325
 326         xfs_ilock(ip, XFS_IOLOCK_EXCL);
 327
 328         if (DM_EVENT_ENABLED(ip, DM_EVENT_WRITE) && !(ioflags & IO_INVIS)) {
 329                 int iolock = XFS_IOLOCK_EXCL;
 330                 int error;
 331
 332                 error = XFS_SEND_DATA(mp, DM_EVENT_WRITE, ip, *ppos, count,
 333                                         FILP_DELAY_FLAG(outfilp), &iolock);
 334                 if (error) {
 335                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 336                         return -error;
 337                 }
 338         }
 339
 340         new_size = *ppos + count;
 341
 342         xfs_ilock(ip, XFS_ILOCK_EXCL);
 343         if (new_size > ip->i_size)
 344                 ip->i_new_size = new_size;
 345         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 346
 347         xfs_rw_enter_trace(XFS_SPLICE_WRITE_ENTER, ip,
 348                            pipe, count, *ppos, ioflags);
 349         ret = generic_file_splice_write(pipe, outfilp, ppos, count, flags);
 350         if (ret > 0)
 351                 XFS_STATS_ADD(xs_write_bytes, ret);
 352
 353         isize = i_size_read(inode);
 354         if (unlikely(ret < 0 && ret != -EFAULT && *ppos > isize))
 355                 *ppos = isize;
 356
 357         if (*ppos > ip->i_size) {
 358                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 359                 if (*ppos > ip->i_size)
 360                         ip->i_size = *ppos;
 361                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 362         }
 363
 364         if (ip->i_new_size) {
 365                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 366                 ip->i_new_size = 0;
 367                 if (ip->i_d.di_size > ip->i_size)
 368                         ip->i_d.di_size = ip->i_size;
 369                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 370         }
 371         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 372         return ret;
 373 }
 374
 375 /*
 376  * This routine is called to handle zeroing any space in the last
 377  * block of the file that is beyond the EOF.  We do this since the
 378  * size is being increased without writing anything to that block
 379  * and we don't want anyone to read the garbage on the disk.
 380  */
 381 STATIC int                              /* error (positive) */
 382 xfs_zero_last_block(
 383         xfs_inode_t     *ip,
 384         xfs_fsize_t     offset,
 385         xfs_fsize_t     isize)
 386 {
 387         xfs_fileoff_t   last_fsb;
 388         xfs_mount_t     *mp = ip->i_mount;
 389         int             nimaps;
 390         int             zero_offset;
 391         int             zero_len;
 392         int             error = 0;
 393         xfs_bmbt_irec_t imap;
 394
 395         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 396
 397         zero_offset = XFS_B_FSB_OFFSET(mp, isize);
 398         if (zero_offset == 0) {
 399                 /*
 400                  * There are no extra bytes in the last block on disk to
 401                  * zero, so return.
 402                  */
 403                 return 0;
 404         }
 405
 406         last_fsb = XFS_B_TO_FSBT(mp, isize);
 407         nimaps = 1;
 408         error = xfs_bmapi(NULL, ip, last_fsb, 1, 0, NULL, 0, &imap,
 409                           &nimaps, NULL, NULL);
 410         if (error) {
 411                 return error;
 412         }
 413         ASSERT(nimaps > 0);
 414         /*
 415          * If the block underlying isize is just a hole, then there
 416          * is nothing to zero.
 417          */
 418         if (imap.br_startblock == HOLESTARTBLOCK) {
 419                 return 0;
 420         }
 421         /*
 422          * Zero the part of the last block beyond the EOF, and write it
 423          * out sync.  We need to drop the ilock while we do this so we
 424          * don't deadlock when the buffer cache calls back to us.
 425          */
 426         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 427
 428         zero_len = mp->m_sb.sb_blocksize - zero_offset;
 429         if (isize + zero_len > offset)
 430                 zero_len = offset - isize;
 431         error = xfs_iozero(ip, isize, zero_len);
 432
 433         xfs_ilock(ip, XFS_ILOCK_EXCL);
 434         ASSERT(error >= 0);
 435         return error;
 436 }
 437
 438 /*
 439  * Zero any on disk space between the current EOF and the new,
 440  * larger EOF.  This handles the normal case of zeroing the remainder
 441  * of the last block in the file and the unusual case of zeroing blocks
 442  * out beyond the size of the file.  This second case only happens
 443  * with fixed size extents and when the system crashes before the inode
 444  * size was updated but after blocks were allocated.  If fill is set,
 445  * then any holes in the range are filled and zeroed.  If not, the holes
 446  * are left alone as holes.
 447  */
 448
 449 int                                     /* error (positive) */
 450 xfs_zero_eof(
 451         xfs_inode_t     *ip,
 452         xfs_off_t       offset,         /* starting I/O offset */
 453         xfs_fsize_t     isize)          /* current inode size */
 454 {
 455         xfs_mount_t     *mp = ip->i_mount;
 456         xfs_fileoff_t   start_zero_fsb;
 457         xfs_fileoff_t   end_zero_fsb;
 458         xfs_fileoff_t   zero_count_fsb;
 459         xfs_fileoff_t   last_fsb;
 460         xfs_fileoff_t   zero_off;
 461         xfs_fsize_t     zero_len;
 462         int             nimaps;
 463         int             error = 0;
 464         xfs_bmbt_irec_t imap;
 465
 466         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_IOLOCK_EXCL));
 467         ASSERT(offset > isize);
 468
 469         /*
 470          * First handle zeroing the block on which isize resides.
 471          * We only zero a part of that block so it is handled specially.
 472          */
 473         error = xfs_zero_last_block(ip, offset, isize);
 474         if (error) {
 475                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_IOLOCK_EXCL));
 476                 return error;
 477         }
 478
 479         /*
 480          * Calculate the range between the new size and the old
 481          * where blocks needing to be zeroed may exist.  To get the
 482          * block where the last byte in the file currently resides,
 483          * we need to subtract one from the size and truncate back
 484          * to a block boundary.  We subtract 1 in case the size is
 485          * exactly on a block boundary.
 486          */
 487         last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
 488         start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
 489         end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
 490         ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
 491         if (last_fsb == end_zero_fsb) {
 492                 /*
 493                  * The size was only incremented on its last block.
 494                  * We took care of that above, so just return.
 495                  */
 496                 return 0;
 497         }
 498
 499         ASSERT(start_zero_fsb <= end_zero_fsb);
 500         while (start_zero_fsb <= end_zero_fsb) {
 501                 nimaps = 1;
 502                 zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
 503                 error = xfs_bmapi(NULL, ip, start_zero_fsb, zero_count_fsb,
 504                                   0, NULL, 0, &imap, &nimaps, NULL, NULL);
 505                 if (error) {
 506                         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_IOLOCK_EXCL));
 507                         return error;
 508                 }
 509                 ASSERT(nimaps > 0);
 510
 511                 if (imap.br_state == XFS_EXT_UNWRITTEN ||
 512                     imap.br_startblock == HOLESTARTBLOCK) {
 513                         /*
 514                          * This loop handles initializing pages that were
 515                          * partially initialized by the code below this
 516                          * loop. It basically zeroes the part of the page
 517                          * that sits on a hole and sets the page as P_HOLE
 518                          * and calls remapf if it is a mapped file.
 519                          */
 520                         start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 521                         ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 522                         continue;
 523                 }
 524
 525                 /*
 526                  * There are blocks we need to zero.
 527                  * Drop the inode lock while we're doing the I/O.
 528                  * We'll still have the iolock to protect us.
 529                  */
 530                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 531
 532                 zero_off = XFS_FSB_TO_B(mp, start_zero_fsb);
 533                 zero_len = XFS_FSB_TO_B(mp, imap.br_blockcount);
 534
 535                 if ((zero_off + zero_len) > offset)
 536                         zero_len = offset - zero_off;
 537
 538                 error = xfs_iozero(ip, zero_off, zero_len);
 539                 if (error) {
 540                         goto out_lock;
 541                 }
 542
 543                 start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 544                 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 545
 546                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 547         }
 548
 549         return 0;
 550
 551 out_lock:
 552         xfs_ilock(ip, XFS_ILOCK_EXCL);
 553         ASSERT(error >= 0);
 554         return error;
 555 }
 556
 557 ssize_t                         /* bytes written, or (-) error */
 558 xfs_write(
 559         struct xfs_inode        *xip,
 560         struct kiocb            *iocb,
 561         const struct iovec      *iovp,
 562         unsigned int            nsegs,
 563         loff_t                  *offset,
 564         int                     ioflags)
 565 {
 566         struct file             *file = iocb->ki_filp;
 567         struct address_space    *mapping = file->f_mapping;
 568         struct inode            *inode = mapping->host;
 569         unsigned long           segs = nsegs;
 570         xfs_mount_t             *mp;
 571         ssize_t                 ret = 0, error = 0;
 572         xfs_fsize_t             isize, new_size;
 573         int                     iolock;
 574         int                     eventsent = 0;
 575         size_t                  ocount = 0, count;
 576         loff_t                  pos;
 577         int                     need_i_mutex;
 578
 579         XFS_STATS_INC(xs_write_calls);
 580
 581         error = generic_segment_checks(iovp, &segs, &ocount, VERIFY_READ);
 582         if (error)
 583                 return error;
 584
 585         count = ocount;
 586         pos = *offset;
 587
 588         if (count == 0)
 589                 return 0;
 590
 591         mp = xip->i_mount;
 592
 593         xfs_wait_for_freeze(mp, SB_FREEZE_WRITE);
 594
 595         if (XFS_FORCED_SHUTDOWN(mp))
 596                 return -EIO;
 597
 598 relock:
 599         if (ioflags & IO_ISDIRECT) {
 600                 iolock = XFS_IOLOCK_SHARED;
 601                 need_i_mutex = 0;
 602         } else {
 603                 iolock = XFS_IOLOCK_EXCL;
 604                 need_i_mutex = 1;
 605                 mutex_lock(&inode->i_mutex);
 606         }
 607
 608         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
 609
 610 start:
 611         error = -generic_write_checks(file, &pos, &count,
 612                                         S_ISBLK(inode->i_mode));
 613         if (error) {
 614                 xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 615                 goto out_unlock_mutex;
 616         }
 617
 618         if ((DM_EVENT_ENABLED(xip, DM_EVENT_WRITE) &&
 619             !(ioflags & IO_INVIS) && !eventsent)) {
 620                 int             dmflags = FILP_DELAY_FLAG(file);
 621
 622                 if (need_i_mutex)
 623                         dmflags |= DM_FLAGS_IMUX;
 624
 625                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 626                 error = XFS_SEND_DATA(xip->i_mount, DM_EVENT_WRITE, xip,
 627                                       pos, count, dmflags, &iolock);
 628                 if (error) {
 629                         goto out_unlock_internal;
 630                 }
 631                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 632                 eventsent = 1;
 633
 634                 /*
 635                  * The iolock was dropped and reacquired in XFS_SEND_DATA
 636                  * so we have to recheck the size when appending.
 637                  * We will only "goto start;" once, since having sent the
 638                  * event prevents another call to XFS_SEND_DATA, which is
 639                  * what allows the size to change in the first place.
 640                  */
 641                 if ((file->f_flags & O_APPEND) && pos != xip->i_size)
 642                         goto start;
 643         }
 644
 645         if (ioflags & IO_ISDIRECT) {
 646                 xfs_buftarg_t   *target =
 647                         XFS_IS_REALTIME_INODE(xip) ?
 648                                 mp->m_rtdev_targp : mp->m_ddev_targp;
 649
 650                 if ((pos & target->bt_smask) || (count & target->bt_smask)) {
 651                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 652                         return XFS_ERROR(-EINVAL);
 653                 }
 654
 655                 if (!need_i_mutex && (mapping->nrpages || pos > xip->i_size)) {
 656                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 657                         iolock = XFS_IOLOCK_EXCL;
 658                         need_i_mutex = 1;
 659                         mutex_lock(&inode->i_mutex);
 660                         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
 661                         goto start;
 662                 }
 663         }
 664
 665         new_size = pos + count;
 666         if (new_size > xip->i_size)
 667                 xip->i_new_size = new_size;
 668
 669         if (likely(!(ioflags & IO_INVIS)))
 670                 file_update_time(file);
 671
 672         /*
 673          * If the offset is beyond the size of the file, we have a couple
 674          * of things to do. First, if there is already space allocated
 675          * we need to either create holes or zero the disk or ...
 676          *
 677          * If there is a page where the previous size lands, we need
 678          * to zero it out up to the new size.
 679          */
 680
 681         if (pos > xip->i_size) {
 682                 error = xfs_zero_eof(xip, pos, xip->i_size);
 683                 if (error) {
 684                         xfs_iunlock(xip, XFS_ILOCK_EXCL);
 685                         goto out_unlock_internal;
 686                 }
 687         }
 688         xfs_iunlock(xip, XFS_ILOCK_EXCL);
 689
 690         /*
 691          * If we're writing the file then make sure to clear the
 692          * setuid and setgid bits if the process is not being run
 693          * by root.  This keeps people from modifying setuid and
 694          * setgid binaries.
 695          */
 696
 697         if (((xip->i_d.di_mode & S_ISUID) ||
 698             ((xip->i_d.di_mode & (S_ISGID | S_IXGRP)) ==
 699                 (S_ISGID | S_IXGRP))) &&
 700              !capable(CAP_FSETID)) {
 701                 error = xfs_write_clear_setuid(xip);
 702                 if (likely(!error))
 703                         error = -file_remove_suid(file);
 704                 if (unlikely(error)) {
 705                         goto out_unlock_internal;
 706                 }
 707         }
 708
 709         /* We can write back this queue in page reclaim */
 710         current->backing_dev_info = mapping->backing_dev_info;
 711
 712         if ((ioflags & IO_ISDIRECT)) {
 713                 if (mapping->nrpages) {
 714                         WARN_ON(need_i_mutex == 0);
 715                         xfs_inval_cached_trace(xip, pos, -1,
 716                                         (pos & PAGE_CACHE_MASK), -1);
 717                         error = xfs_flushinval_pages(xip,
 718                                         (pos & PAGE_CACHE_MASK),
 719                                         -1, FI_REMAPF_LOCKED);
 720                         if (error)
 721                                 goto out_unlock_internal;
 722                 }
 723
 724                 if (need_i_mutex) {
 725                         /* demote the lock now the cached pages are gone */
 726                         xfs_ilock_demote(xip, XFS_IOLOCK_EXCL);
 727                         mutex_unlock(&inode->i_mutex);
 728
 729                         iolock = XFS_IOLOCK_SHARED;
 730                         need_i_mutex = 0;
 731                 }
 732
 733                 xfs_rw_enter_trace(XFS_DIOWR_ENTER, xip, (void *)iovp, segs,
 734                                 *offset, ioflags);
 735                 ret = generic_file_direct_write(iocb, iovp,
 736                                 &segs, pos, offset, count, ocount);
 737
 738                 /*
 739                  * direct-io write to a hole: fall through to buffered I/O
 740                  * for completing the rest of the request.
 741                  */
 742                 if (ret >= 0 && ret != count) {
 743                         XFS_STATS_ADD(xs_write_bytes, ret);
 744
 745                         pos += ret;
 746                         count -= ret;
 747
 748                         ioflags &= ~IO_ISDIRECT;
 749                         xfs_iunlock(xip, iolock);
 750                         goto relock;
 751                 }
 752         } else {
 753                 int enospc = 0;
 754                 ssize_t ret2 = 0;
 755
 756 write_retry:
 757                 xfs_rw_enter_trace(XFS_WRITE_ENTER, xip, (void *)iovp, segs,
 758                                 *offset, ioflags);
 759                 ret2 = generic_file_buffered_write(iocb, iovp, segs,
 760                                 pos, offset, count, ret);
 761                 /*
 762                  * if we just got an ENOSPC, flush the inode now we
 763                  * aren't holding any page locks and retry *once*
 764                  */
 765                 if (ret2 == -ENOSPC && !enospc) {
 766                         error = xfs_flush_pages(xip, 0, -1, 0, FI_NONE);
 767                         if (error)
 768                                 goto out_unlock_internal;
 769                         enospc = 1;
 770                         goto write_retry;
 771                 }
 772                 ret = ret2;
 773         }
 774
 775         current->backing_dev_info = NULL;
 776
 777         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
 778                 ret = wait_on_sync_kiocb(iocb);
 779
 780         isize = i_size_read(inode);
 781         if (unlikely(ret < 0 && ret != -EFAULT && *offset > isize))
 782                 *offset = isize;
 783
 784         if (*offset > xip->i_size) {
 785                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 786                 if (*offset > xip->i_size)
 787                         xip->i_size = *offset;
 788                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 789         }
 790
 791         if (ret == -ENOSPC &&
 792             DM_EVENT_ENABLED(xip, DM_EVENT_NOSPACE) && !(ioflags & IO_INVIS)) {
 793                 xfs_iunlock(xip, iolock);
 794                 if (need_i_mutex)
 795                         mutex_unlock(&inode->i_mutex);
 796                 error = XFS_SEND_NAMESP(xip->i_mount, DM_EVENT_NOSPACE, xip,
 797                                 DM_RIGHT_NULL, xip, DM_RIGHT_NULL, NULL, NULL,
 798                                 0, 0, 0); /* Delay flag intentionally  unused */
 799                 if (need_i_mutex)
 800                         mutex_lock(&inode->i_mutex);
 801                 xfs_ilock(xip, iolock);
 802                 if (error)
 803                         goto out_unlock_internal;
 804                 goto start;
 805         }
 806
 807         error = -ret;
 808         if (ret <= 0)
 809                 goto out_unlock_internal;
 810
 811         XFS_STATS_ADD(xs_write_bytes, ret);
 812
 813         /* Handle various SYNC-type writes */
 814         if ((file->f_flags & O_SYNC) || IS_SYNC(inode)) {
 815                 loff_t end = pos + ret - 1;
 816                 int error2;
 817
 818                 xfs_iunlock(xip, iolock);
 819                 if (need_i_mutex)
 820                         mutex_unlock(&inode->i_mutex);
 821
 822                 error2 = filemap_write_and_wait_range(mapping, pos, end);
 823                 if (!error)
 824                         error = error2;
 825                 if (need_i_mutex)
 826                         mutex_lock(&inode->i_mutex);
 827                 xfs_ilock(xip, iolock);
 828
 829                 error2 = xfs_fsync(xip);
 830                 if (!error)
 831                         error = error2;
 832         }
 833
 834  out_unlock_internal:
 835         if (xip->i_new_size) {
 836                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 837                 xip->i_new_size = 0;
 838                 /*
 839                  * If this was a direct or synchronous I/O that failed (such
 840                  * as ENOSPC) then part of the I/O may have been written to
 841                  * disk before the error occured.  In this case the on-disk
 842                  * file size may have been adjusted beyond the in-memory file
 843                  * size and now needs to be truncated back.
 844                  */
 845                 if (xip->i_d.di_size > xip->i_size)
 846                         xip->i_d.di_size = xip->i_size;
 847                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 848         }
 849         xfs_iunlock(xip, iolock);
 850  out_unlock_mutex:
 851         if (need_i_mutex)
 852                 mutex_unlock(&inode->i_mutex);
 853         return -error;
 854 }
 855
 856 /*
 857  * All xfs metadata buffers except log state machine buffers
 858  * get this attached as their b_bdstrat callback function.
 859  * This is so that we can catch a buffer
 860  * after prematurely unpinning it to forcibly shutdown the filesystem.
 861  */
 862 int
 863 xfs_bdstrat_cb(struct xfs_buf *bp)
 864 {
 865         if (XFS_FORCED_SHUTDOWN(bp->b_mount)) {
 866                 xfs_buftrace("XFS__BDSTRAT IOERROR", bp);
 867                 /*
 868                  * Metadata write that didn't get logged but
 869                  * written delayed anyway. These aren't associated
 870                  * with a transaction, and can be ignored.
 871                  */
 872                 if (XFS_BUF_IODONE_FUNC(bp) == NULL &&
 873                     (XFS_BUF_ISREAD(bp)) == 0)
 874                         return (xfs_bioerror_relse(bp));
 875                 else
 876                         return (xfs_bioerror(bp));
 877         }
 878
 879         xfs_buf_iorequest(bp);
 880         return 0;
 881 }
 882
 883 /*
 884  * Wrapper around bdstrat so that we can stop data from going to disk in case
 885  * we are shutting down the filesystem.  Typically user data goes thru this
 886  * path; one of the exceptions is the superblock.
 887  */
 888 void
 889 xfsbdstrat(
 890         struct xfs_mount        *mp,
 891         struct xfs_buf          *bp)
 892 {
 893         ASSERT(mp);
 894         if (!XFS_FORCED_SHUTDOWN(mp)) {
 895                 xfs_buf_iorequest(bp);
 896                 return;
 897         }
 898
 899         xfs_buftrace("XFSBDSTRAT IOERROR", bp);
 900         xfs_bioerror_relse(bp);
 901 }
 902
 903 /*
 904  * If the underlying (data/log/rt) device is readonly, there are some
 905  * operations that cannot proceed.
 906  */
 907 int
 908 xfs_dev_is_read_only(
 909         xfs_mount_t             *mp,
 910         char                    *message)
 911 {
 912         if (xfs_readonly_buftarg(mp->m_ddev_targp) ||
 913             xfs_readonly_buftarg(mp->m_logdev_targp) ||
 914             (mp->m_rtdev_targp && xfs_readonly_buftarg(mp->m_rtdev_targp))) {
 915                 cmn_err(CE_NOTE,
 916                         "XFS: %s required on read-only device.", message);
 917                 cmn_err(CE_NOTE,
 918                         "XFS: write access unavailable, cannot proceed.");
 919                 return EROFS;
 920         }
 921         return 0;
 922 }