err.no Git - 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_acl.h"
  46 #include "xfs_attr.h"
  47 #include "xfs_inode_item.h"
  48 #include "xfs_buf_item.h"
  49 #include "xfs_utils.h"
  50 #include "xfs_iomap.h"
  51 #include "xfs_vnodeops.h"
  52
  53 #include <linux/capability.h>
  54 #include <linux/writeback.h>
  55
  56
  57 #if defined(XFS_RW_TRACE)
  58 void
  59 xfs_rw_enter_trace(
  60         int                     tag,
  61         xfs_inode_t             *ip,
  62         void                    *data,
  63         size_t                  segs,
  64         loff_t                  offset,
  65         int                     ioflags)
  66 {
  67         if (ip->i_rwtrace == NULL)
  68                 return;
  69         ktrace_enter(ip->i_rwtrace,
  70                 (void *)(unsigned long)tag,
  71                 (void *)ip,
  72                 (void *)((unsigned long)((ip->i_d.di_size >> 32) & 0xffffffff)),
  73                 (void *)((unsigned long)(ip->i_d.di_size & 0xffffffff)),
  74                 (void *)data,
  75                 (void *)((unsigned long)segs),
  76                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
  77                 (void *)((unsigned long)(offset & 0xffffffff)),
  78                 (void *)((unsigned long)ioflags),
  79                 (void *)((unsigned long)((ip->i_new_size >> 32) & 0xffffffff)),
  80                 (void *)((unsigned long)(ip->i_new_size & 0xffffffff)),
  81                 (void *)((unsigned long)current_pid()),
  82                 (void *)NULL,
  83                 (void *)NULL,
  84                 (void *)NULL,
  85                 (void *)NULL);
  86 }
  87
  88 void
  89 xfs_inval_cached_trace(
  90         xfs_inode_t     *ip,
  91         xfs_off_t       offset,
  92         xfs_off_t       len,
  93         xfs_off_t       first,
  94         xfs_off_t       last)
  95 {
  96
  97         if (ip->i_rwtrace == NULL)
  98                 return;
  99         ktrace_enter(ip->i_rwtrace,
 100                 (void *)(__psint_t)XFS_INVAL_CACHED,
 101                 (void *)ip,
 102                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
 103                 (void *)((unsigned long)(offset & 0xffffffff)),
 104                 (void *)((unsigned long)((len >> 32) & 0xffffffff)),
 105                 (void *)((unsigned long)(len & 0xffffffff)),
 106                 (void *)((unsigned long)((first >> 32) & 0xffffffff)),
 107                 (void *)((unsigned long)(first & 0xffffffff)),
 108                 (void *)((unsigned long)((last >> 32) & 0xffffffff)),
 109                 (void *)((unsigned long)(last & 0xffffffff)),
 110                 (void *)((unsigned long)current_pid()),
 111                 (void *)NULL,
 112                 (void *)NULL,
 113                 (void *)NULL,
 114                 (void *)NULL,
 115                 (void *)NULL);
 116 }
 117 #endif
 118
 119 /*
 120  *      xfs_iozero
 121  *
 122  *      xfs_iozero clears the specified range of buffer supplied,
 123  *      and marks all the affected blocks as valid and modified.  If
 124  *      an affected block is not allocated, it will be allocated.  If
 125  *      an affected block is not completely overwritten, and is not
 126  *      valid before the operation, it will be read from disk before
 127  *      being partially zeroed.
 128  */
 129 STATIC int
 130 xfs_iozero(
 131         struct xfs_inode        *ip,    /* inode                        */
 132         loff_t                  pos,    /* offset in file               */
 133         size_t                  count)  /* size of data to zero         */
 134 {
 135         struct page             *page;
 136         struct address_space    *mapping;
 137         int                     status;
 138
 139         mapping = ip->i_vnode->i_mapping;
 140         do {
 141                 unsigned offset, bytes;
 142                 void *fsdata;
 143
 144                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
 145                 bytes = PAGE_CACHE_SIZE - offset;
 146                 if (bytes > count)
 147                         bytes = count;
 148
 149                 status = pagecache_write_begin(NULL, mapping, pos, bytes,
 150                                         AOP_FLAG_UNINTERRUPTIBLE,
 151                                         &page, &fsdata);
 152                 if (status)
 153                         break;
 154
 155                 zero_user(page, offset, bytes);
 156
 157                 status = pagecache_write_end(NULL, mapping, pos, bytes, bytes,
 158                                         page, fsdata);
 159                 WARN_ON(status <= 0); /* can't return less than zero! */
 160                 pos += bytes;
 161                 count -= bytes;
 162                 status = 0;
 163         } while (count);
 164
 165         return (-status);
 166 }
 167
 168 ssize_t                 /* bytes read, or (-)  error */
 169 xfs_read(
 170         xfs_inode_t             *ip,
 171         struct kiocb            *iocb,
 172         const struct iovec      *iovp,
 173         unsigned int            segs,
 174         loff_t                  *offset,
 175         int                     ioflags)
 176 {
 177         struct file             *file = iocb->ki_filp;
 178         struct inode            *inode = file->f_mapping->host;
 179         xfs_mount_t             *mp = ip->i_mount;
 180         size_t                  size = 0;
 181         ssize_t                 ret = 0;
 182         xfs_fsize_t             n;
 183         unsigned long           seg;
 184
 185
 186         XFS_STATS_INC(xs_read_calls);
 187
 188         /* START copy & waste from filemap.c */
 189         for (seg = 0; seg < segs; seg++) {
 190                 const struct iovec *iv = &iovp[seg];
 191
 192                 /*
 193                  * If any segment has a negative length, or the cumulative
 194                  * length ever wraps negative then return -EINVAL.
 195                  */
 196                 size += iv->iov_len;
 197                 if (unlikely((ssize_t)(size|iv->iov_len) < 0))
 198                         return XFS_ERROR(-EINVAL);
 199         }
 200         /* END copy & waste from filemap.c */
 201
 202         if (unlikely(ioflags & IO_ISDIRECT)) {
 203                 xfs_buftarg_t   *target =
 204                         XFS_IS_REALTIME_INODE(ip) ?
 205                                 mp->m_rtdev_targp : mp->m_ddev_targp;
 206                 if ((*offset & target->bt_smask) ||
 207                     (size & target->bt_smask)) {
 208                         if (*offset == ip->i_size) {
 209                                 return (0);
 210                         }
 211                         return -XFS_ERROR(EINVAL);
 212                 }
 213         }
 214
 215         n = XFS_MAXIOFFSET(mp) - *offset;
 216         if ((n <= 0) || (size == 0))
 217                 return 0;
 218
 219         if (n < size)
 220                 size = n;
 221
 222         if (XFS_FORCED_SHUTDOWN(mp))
 223                 return -EIO;
 224
 225         if (unlikely(ioflags & IO_ISDIRECT))
 226                 mutex_lock(&inode->i_mutex);
 227         xfs_ilock(ip, XFS_IOLOCK_SHARED);
 228
 229         if (DM_EVENT_ENABLED(ip, DM_EVENT_READ) && !(ioflags & IO_INVIS)) {
 230                 int dmflags = FILP_DELAY_FLAG(file) | DM_SEM_FLAG_RD(ioflags);
 231                 int iolock = XFS_IOLOCK_SHARED;
 232
 233                 ret = -XFS_SEND_DATA(mp, DM_EVENT_READ, ip, *offset, size,
 234                                         dmflags, &iolock);
 235                 if (ret) {
 236                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 237                         if (unlikely(ioflags & IO_ISDIRECT))
 238                                 mutex_unlock(&inode->i_mutex);
 239                         return ret;
 240                 }
 241         }
 242
 243         if (unlikely(ioflags & IO_ISDIRECT)) {
 244                 if (inode->i_mapping->nrpages)
 245                         ret = xfs_flushinval_pages(ip, (*offset & PAGE_CACHE_MASK),
 246                                                     -1, FI_REMAPF_LOCKED);
 247                 mutex_unlock(&inode->i_mutex);
 248                 if (ret) {
 249                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 250                         return ret;
 251                 }
 252         }
 253
 254         xfs_rw_enter_trace(XFS_READ_ENTER, ip,
 255                                 (void *)iovp, segs, *offset, ioflags);
 256
 257         iocb->ki_pos = *offset;
 258         ret = generic_file_aio_read(iocb, iovp, segs, *offset);
 259         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
 260                 ret = wait_on_sync_kiocb(iocb);
 261         if (ret > 0)
 262                 XFS_STATS_ADD(xs_read_bytes, ret);
 263
 264         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 265         return ret;
 266 }
 267
 268 ssize_t
 269 xfs_splice_read(
 270         xfs_inode_t             *ip,
 271         struct file             *infilp,
 272         loff_t                  *ppos,
 273         struct pipe_inode_info  *pipe,
 274         size_t                  count,
 275         int                     flags,
 276         int                     ioflags)
 277 {
 278         xfs_mount_t             *mp = ip->i_mount;
 279         ssize_t                 ret;
 280
 281         XFS_STATS_INC(xs_read_calls);
 282         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 283                 return -EIO;
 284
 285         xfs_ilock(ip, XFS_IOLOCK_SHARED);
 286
 287         if (DM_EVENT_ENABLED(ip, DM_EVENT_READ) && !(ioflags & IO_INVIS)) {
 288                 int iolock = XFS_IOLOCK_SHARED;
 289                 int error;
 290
 291                 error = XFS_SEND_DATA(mp, DM_EVENT_READ, ip, *ppos, count,
 292                                         FILP_DELAY_FLAG(infilp), &iolock);
 293                 if (error) {
 294                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 295                         return -error;
 296                 }
 297         }
 298         xfs_rw_enter_trace(XFS_SPLICE_READ_ENTER, ip,
 299                            pipe, count, *ppos, ioflags);
 300         ret = generic_file_splice_read(infilp, ppos, pipe, count, flags);
 301         if (ret > 0)
 302                 XFS_STATS_ADD(xs_read_bytes, ret);
 303
 304         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 305         return ret;
 306 }
 307
 308 ssize_t
 309 xfs_splice_write(
 310         xfs_inode_t             *ip,
 311         struct pipe_inode_info  *pipe,
 312         struct file             *outfilp,
 313         loff_t                  *ppos,
 314         size_t                  count,
 315         int                     flags,
 316         int                     ioflags)
 317 {
 318         xfs_mount_t             *mp = ip->i_mount;
 319         ssize_t                 ret;
 320         struct inode            *inode = outfilp->f_mapping->host;
 321         xfs_fsize_t             isize, new_size;
 322
 323         XFS_STATS_INC(xs_write_calls);
 324         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 325                 return -EIO;
 326
 327         xfs_ilock(ip, XFS_IOLOCK_EXCL);
 328
 329         if (DM_EVENT_ENABLED(ip, DM_EVENT_WRITE) && !(ioflags & IO_INVIS)) {
 330                 int iolock = XFS_IOLOCK_EXCL;
 331                 int error;
 332
 333                 error = XFS_SEND_DATA(mp, DM_EVENT_WRITE, ip, *ppos, count,
 334                                         FILP_DELAY_FLAG(outfilp), &iolock);
 335                 if (error) {
 336                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 337                         return -error;
 338                 }
 339         }
 340
 341         new_size = *ppos + count;
 342
 343         xfs_ilock(ip, XFS_ILOCK_EXCL);
 344         if (new_size > ip->i_size)
 345                 ip->i_new_size = new_size;
 346         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 347
 348         xfs_rw_enter_trace(XFS_SPLICE_WRITE_ENTER, ip,
 349                            pipe, count, *ppos, ioflags);
 350         ret = generic_file_splice_write(pipe, outfilp, ppos, count, flags);
 351         if (ret > 0)
 352                 XFS_STATS_ADD(xs_write_bytes, ret);
 353
 354         isize = i_size_read(inode);
 355         if (unlikely(ret < 0 && ret != -EFAULT && *ppos > isize))
 356                 *ppos = isize;
 357
 358         if (*ppos > ip->i_size) {
 359                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 360                 if (*ppos > ip->i_size)
 361                         ip->i_size = *ppos;
 362                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 363         }
 364
 365         if (ip->i_new_size) {
 366                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 367                 ip->i_new_size = 0;
 368                 if (ip->i_d.di_size > ip->i_size)
 369                         ip->i_d.di_size = ip->i_size;
 370                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 371         }
 372         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 373         return ret;
 374 }
 375
 376 /*
 377  * This routine is called to handle zeroing any space in the last
 378  * block of the file that is beyond the EOF.  We do this since the
 379  * size is being increased without writing anything to that block
 380  * and we don't want anyone to read the garbage on the disk.
 381  */
 382 STATIC int                              /* error (positive) */
 383 xfs_zero_last_block(
 384         xfs_inode_t     *ip,
 385         xfs_fsize_t     offset,
 386         xfs_fsize_t     isize)
 387 {
 388         xfs_fileoff_t   last_fsb;
 389         xfs_mount_t     *mp = ip->i_mount;
 390         int             nimaps;
 391         int             zero_offset;
 392         int             zero_len;
 393         int             error = 0;
 394         xfs_bmbt_irec_t imap;
 395
 396         ASSERT(ismrlocked(&ip->i_lock, MR_UPDATE) != 0);
 397
 398         zero_offset = XFS_B_FSB_OFFSET(mp, isize);
 399         if (zero_offset == 0) {
 400                 /*
 401                  * There are no extra bytes in the last block on disk to
 402                  * zero, so return.
 403                  */
 404                 return 0;
 405         }
 406
 407         last_fsb = XFS_B_TO_FSBT(mp, isize);
 408         nimaps = 1;
 409         error = xfs_bmapi(NULL, ip, last_fsb, 1, 0, NULL, 0, &imap,
 410                           &nimaps, NULL, NULL);
 411         if (error) {
 412                 return error;
 413         }
 414         ASSERT(nimaps > 0);
 415         /*
 416          * If the block underlying isize is just a hole, then there
 417          * is nothing to zero.
 418          */
 419         if (imap.br_startblock == HOLESTARTBLOCK) {
 420                 return 0;
 421         }
 422         /*
 423          * Zero the part of the last block beyond the EOF, and write it
 424          * out sync.  We need to drop the ilock while we do this so we
 425          * don't deadlock when the buffer cache calls back to us.
 426          */
 427         xfs_iunlock(ip, XFS_ILOCK_EXCL| XFS_EXTSIZE_RD);
 428
 429         zero_len = mp->m_sb.sb_blocksize - zero_offset;
 430         if (isize + zero_len > offset)
 431                 zero_len = offset - isize;
 432         error = xfs_iozero(ip, isize, zero_len);
 433
 434         xfs_ilock(ip, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 435         ASSERT(error >= 0);
 436         return error;
 437 }
 438
 439 /*
 440  * Zero any on disk space between the current EOF and the new,
 441  * larger EOF.  This handles the normal case of zeroing the remainder
 442  * of the last block in the file and the unusual case of zeroing blocks
 443  * out beyond the size of the file.  This second case only happens
 444  * with fixed size extents and when the system crashes before the inode
 445  * size was updated but after blocks were allocated.  If fill is set,
 446  * then any holes in the range are filled and zeroed.  If not, the holes
 447  * are left alone as holes.
 448  */
 449
 450 int                                     /* error (positive) */
 451 xfs_zero_eof(
 452         xfs_inode_t     *ip,
 453         xfs_off_t       offset,         /* starting I/O offset */
 454         xfs_fsize_t     isize)          /* current inode size */
 455 {
 456         xfs_mount_t     *mp = ip->i_mount;
 457         xfs_fileoff_t   start_zero_fsb;
 458         xfs_fileoff_t   end_zero_fsb;
 459         xfs_fileoff_t   zero_count_fsb;
 460         xfs_fileoff_t   last_fsb;
 461         xfs_fileoff_t   zero_off;
 462         xfs_fsize_t     zero_len;
 463         int             nimaps;
 464         int             error = 0;
 465         xfs_bmbt_irec_t imap;
 466
 467         ASSERT(ismrlocked(&ip->i_lock, MR_UPDATE));
 468         ASSERT(ismrlocked(&ip->i_iolock, MR_UPDATE));
 469         ASSERT(offset > isize);
 470
 471         /*
 472          * First handle zeroing the block on which isize resides.
 473          * We only zero a part of that block so it is handled specially.
 474          */
 475         error = xfs_zero_last_block(ip, offset, isize);
 476         if (error) {
 477                 ASSERT(ismrlocked(&ip->i_lock, MR_UPDATE));
 478                 ASSERT(ismrlocked(&ip->i_iolock, MR_UPDATE));
 479                 return error;
 480         }
 481
 482         /*
 483          * Calculate the range between the new size and the old
 484          * where blocks needing to be zeroed may exist.  To get the
 485          * block where the last byte in the file currently resides,
 486          * we need to subtract one from the size and truncate back
 487          * to a block boundary.  We subtract 1 in case the size is
 488          * exactly on a block boundary.
 489          */
 490         last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
 491         start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
 492         end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
 493         ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
 494         if (last_fsb == end_zero_fsb) {
 495                 /*
 496                  * The size was only incremented on its last block.
 497                  * We took care of that above, so just return.
 498                  */
 499                 return 0;
 500         }
 501
 502         ASSERT(start_zero_fsb <= end_zero_fsb);
 503         while (start_zero_fsb <= end_zero_fsb) {
 504                 nimaps = 1;
 505                 zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
 506                 error = xfs_bmapi(NULL, ip, start_zero_fsb, zero_count_fsb,
 507                                   0, NULL, 0, &imap, &nimaps, NULL, NULL);
 508                 if (error) {
 509                         ASSERT(ismrlocked(&ip->i_lock, MR_UPDATE));
 510                         ASSERT(ismrlocked(&ip->i_iolock, MR_UPDATE));
 511                         return error;
 512                 }
 513                 ASSERT(nimaps > 0);
 514
 515                 if (imap.br_state == XFS_EXT_UNWRITTEN ||
 516                     imap.br_startblock == HOLESTARTBLOCK) {
 517                         /*
 518                          * This loop handles initializing pages that were
 519                          * partially initialized by the code below this
 520                          * loop. It basically zeroes the part of the page
 521                          * that sits on a hole and sets the page as P_HOLE
 522                          * and calls remapf if it is a mapped file.
 523                          */
 524                         start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 525                         ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 526                         continue;
 527                 }
 528
 529                 /*
 530                  * There are blocks we need to zero.
 531                  * Drop the inode lock while we're doing the I/O.
 532                  * We'll still have the iolock to protect us.
 533                  */
 534                 xfs_iunlock(ip, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 535
 536                 zero_off = XFS_FSB_TO_B(mp, start_zero_fsb);
 537                 zero_len = XFS_FSB_TO_B(mp, imap.br_blockcount);
 538
 539                 if ((zero_off + zero_len) > offset)
 540                         zero_len = offset - zero_off;
 541
 542                 error = xfs_iozero(ip, zero_off, zero_len);
 543                 if (error) {
 544                         goto out_lock;
 545                 }
 546
 547                 start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 548                 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 549
 550                 xfs_ilock(ip, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 551         }
 552
 553         return 0;
 554
 555 out_lock:
 556         xfs_ilock(ip, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 557         ASSERT(error >= 0);
 558         return error;
 559 }
 560
 561 ssize_t                         /* bytes written, or (-) error */
 562 xfs_write(
 563         struct xfs_inode        *xip,
 564         struct kiocb            *iocb,
 565         const struct iovec      *iovp,
 566         unsigned int            nsegs,
 567         loff_t                  *offset,
 568         int                     ioflags)
 569 {
 570         struct file             *file = iocb->ki_filp;
 571         struct address_space    *mapping = file->f_mapping;
 572         struct inode            *inode = mapping->host;
 573         unsigned long           segs = nsegs;
 574         xfs_mount_t             *mp;
 575         ssize_t                 ret = 0, error = 0;
 576         xfs_fsize_t             isize, new_size;
 577         int                     iolock;
 578         int                     eventsent = 0;
 579         size_t                  ocount = 0, count;
 580         loff_t                  pos;
 581         int                     need_i_mutex;
 582
 583         XFS_STATS_INC(xs_write_calls);
 584
 585         error = generic_segment_checks(iovp, &segs, &ocount, VERIFY_READ);
 586         if (error)
 587                 return error;
 588
 589         count = ocount;
 590         pos = *offset;
 591
 592         if (count == 0)
 593                 return 0;
 594
 595         mp = xip->i_mount;
 596
 597         xfs_wait_for_freeze(mp, SB_FREEZE_WRITE);
 598
 599         if (XFS_FORCED_SHUTDOWN(mp))
 600                 return -EIO;
 601
 602 relock:
 603         if (ioflags & IO_ISDIRECT) {
 604                 iolock = XFS_IOLOCK_SHARED;
 605                 need_i_mutex = 0;
 606         } else {
 607                 iolock = XFS_IOLOCK_EXCL;
 608                 need_i_mutex = 1;
 609                 mutex_lock(&inode->i_mutex);
 610         }
 611
 612         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
 613
 614 start:
 615         error = -generic_write_checks(file, &pos, &count,
 616                                         S_ISBLK(inode->i_mode));
 617         if (error) {
 618                 xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 619                 goto out_unlock_mutex;
 620         }
 621
 622         if ((DM_EVENT_ENABLED(xip, DM_EVENT_WRITE) &&
 623             !(ioflags & IO_INVIS) && !eventsent)) {
 624                 int             dmflags = FILP_DELAY_FLAG(file);
 625
 626                 if (need_i_mutex)
 627                         dmflags |= DM_FLAGS_IMUX;
 628
 629                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 630                 error = XFS_SEND_DATA(xip->i_mount, DM_EVENT_WRITE, xip,
 631                                       pos, count, dmflags, &iolock);
 632                 if (error) {
 633                         goto out_unlock_internal;
 634                 }
 635                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 636                 eventsent = 1;
 637
 638                 /*
 639                  * The iolock was dropped and reacquired in XFS_SEND_DATA
 640                  * so we have to recheck the size when appending.
 641                  * We will only "goto start;" once, since having sent the
 642                  * event prevents another call to XFS_SEND_DATA, which is
 643                  * what allows the size to change in the first place.
 644                  */
 645                 if ((file->f_flags & O_APPEND) && pos != xip->i_size)
 646                         goto start;
 647         }
 648
 649         if (ioflags & IO_ISDIRECT) {
 650                 xfs_buftarg_t   *target =
 651                         XFS_IS_REALTIME_INODE(xip) ?
 652                                 mp->m_rtdev_targp : mp->m_ddev_targp;
 653
 654                 if ((pos & target->bt_smask) || (count & target->bt_smask)) {
 655                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 656                         return XFS_ERROR(-EINVAL);
 657                 }
 658
 659                 if (!need_i_mutex && (mapping->nrpages || pos > xip->i_size)) {
 660                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 661                         iolock = XFS_IOLOCK_EXCL;
 662                         need_i_mutex = 1;
 663                         mutex_lock(&inode->i_mutex);
 664                         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
 665                         goto start;
 666                 }
 667         }
 668
 669         new_size = pos + count;
 670         if (new_size > xip->i_size)
 671                 xip->i_new_size = new_size;
 672
 673         if (likely(!(ioflags & IO_INVIS))) {
 674                 file_update_time(file);
 675                 xfs_ichgtime_fast(xip, inode,
 676                                   XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
 677         }
 678
 679         /*
 680          * If the offset is beyond the size of the file, we have a couple
 681          * of things to do. First, if there is already space allocated
 682          * we need to either create holes or zero the disk or ...
 683          *
 684          * If there is a page where the previous size lands, we need
 685          * to zero it out up to the new size.
 686          */
 687
 688         if (pos > xip->i_size) {
 689                 error = xfs_zero_eof(xip, pos, xip->i_size);
 690                 if (error) {
 691                         xfs_iunlock(xip, XFS_ILOCK_EXCL);
 692                         goto out_unlock_internal;
 693                 }
 694         }
 695         xfs_iunlock(xip, XFS_ILOCK_EXCL);
 696
 697         /*
 698          * If we're writing the file then make sure to clear the
 699          * setuid and setgid bits if the process is not being run
 700          * by root.  This keeps people from modifying setuid and
 701          * setgid binaries.
 702          */
 703
 704         if (((xip->i_d.di_mode & S_ISUID) ||
 705             ((xip->i_d.di_mode & (S_ISGID | S_IXGRP)) ==
 706                 (S_ISGID | S_IXGRP))) &&
 707              !capable(CAP_FSETID)) {
 708                 error = xfs_write_clear_setuid(xip);
 709                 if (likely(!error))
 710                         error = -remove_suid(file->f_path.dentry);
 711                 if (unlikely(error)) {
 712                         goto out_unlock_internal;
 713                 }
 714         }
 715
 716 retry:
 717         /* We can write back this queue in page reclaim */
 718         current->backing_dev_info = mapping->backing_dev_info;
 719
 720         if ((ioflags & IO_ISDIRECT)) {
 721                 if (mapping->nrpages) {
 722                         WARN_ON(need_i_mutex == 0);
 723                         xfs_inval_cached_trace(xip, pos, -1,
 724                                         (pos & PAGE_CACHE_MASK), -1);
 725                         error = xfs_flushinval_pages(xip,
 726                                         (pos & PAGE_CACHE_MASK),
 727                                         -1, FI_REMAPF_LOCKED);
 728                         if (error)
 729                                 goto out_unlock_internal;
 730                 }
 731
 732                 if (need_i_mutex) {
 733                         /* demote the lock now the cached pages are gone */
 734                         xfs_ilock_demote(xip, XFS_IOLOCK_EXCL);
 735                         mutex_unlock(&inode->i_mutex);
 736
 737                         iolock = XFS_IOLOCK_SHARED;
 738                         need_i_mutex = 0;
 739                 }
 740
 741                 xfs_rw_enter_trace(XFS_DIOWR_ENTER, xip, (void *)iovp, segs,
 742                                 *offset, ioflags);
 743                 ret = generic_file_direct_write(iocb, iovp,
 744                                 &segs, pos, offset, count, ocount);
 745
 746                 /*
 747                  * direct-io write to a hole: fall through to buffered I/O
 748                  * for completing the rest of the request.
 749                  */
 750                 if (ret >= 0 && ret != count) {
 751                         XFS_STATS_ADD(xs_write_bytes, ret);
 752
 753                         pos += ret;
 754                         count -= ret;
 755
 756                         ioflags &= ~IO_ISDIRECT;
 757                         xfs_iunlock(xip, iolock);
 758                         goto relock;
 759                 }
 760         } else {
 761                 xfs_rw_enter_trace(XFS_WRITE_ENTER, xip, (void *)iovp, segs,
 762                                 *offset, ioflags);
 763                 ret = generic_file_buffered_write(iocb, iovp, segs,
 764                                 pos, offset, count, ret);
 765         }
 766
 767         current->backing_dev_info = NULL;
 768
 769         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
 770                 ret = wait_on_sync_kiocb(iocb);
 771
 772         if (ret == -ENOSPC &&
 773             DM_EVENT_ENABLED(xip, DM_EVENT_NOSPACE) && !(ioflags & IO_INVIS)) {
 774                 xfs_iunlock(xip, iolock);
 775                 if (need_i_mutex)
 776                         mutex_unlock(&inode->i_mutex);
 777                 error = XFS_SEND_NAMESP(xip->i_mount, DM_EVENT_NOSPACE, xip,
 778                                 DM_RIGHT_NULL, xip, DM_RIGHT_NULL, NULL, NULL,
 779                                 0, 0, 0); /* Delay flag intentionally  unused */
 780                 if (need_i_mutex)
 781                         mutex_lock(&inode->i_mutex);
 782                 xfs_ilock(xip, iolock);
 783                 if (error)
 784                         goto out_unlock_internal;
 785                 pos = xip->i_size;
 786                 ret = 0;
 787                 goto retry;
 788         }
 789
 790         isize = i_size_read(inode);
 791         if (unlikely(ret < 0 && ret != -EFAULT && *offset > isize))
 792                 *offset = isize;
 793
 794         if (*offset > xip->i_size) {
 795                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 796                 if (*offset > xip->i_size)
 797                         xip->i_size = *offset;
 798                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 799         }
 800
 801         error = -ret;
 802         if (ret <= 0)
 803                 goto out_unlock_internal;
 804
 805         XFS_STATS_ADD(xs_write_bytes, ret);
 806
 807         /* Handle various SYNC-type writes */
 808         if ((file->f_flags & O_SYNC) || IS_SYNC(inode)) {
 809                 int error2;
 810
 811                 xfs_iunlock(xip, iolock);
 812                 if (need_i_mutex)
 813                         mutex_unlock(&inode->i_mutex);
 814                 error2 = sync_page_range(inode, mapping, pos, ret);
 815                 if (!error)
 816                         error = error2;
 817                 if (need_i_mutex)
 818                         mutex_lock(&inode->i_mutex);
 819                 xfs_ilock(xip, iolock);
 820                 error2 = xfs_write_sync_logforce(mp, xip);
 821                 if (!error)
 822                         error = error2;
 823         }
 824
 825  out_unlock_internal:
 826         if (xip->i_new_size) {
 827                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 828                 xip->i_new_size = 0;
 829                 /*
 830                  * If this was a direct or synchronous I/O that failed (such
 831                  * as ENOSPC) then part of the I/O may have been written to
 832                  * disk before the error occured.  In this case the on-disk
 833                  * file size may have been adjusted beyond the in-memory file
 834                  * size and now needs to be truncated back.
 835                  */
 836                 if (xip->i_d.di_size > xip->i_size)
 837                         xip->i_d.di_size = xip->i_size;
 838                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 839         }
 840         xfs_iunlock(xip, iolock);
 841  out_unlock_mutex:
 842         if (need_i_mutex)
 843                 mutex_unlock(&inode->i_mutex);
 844         return -error;
 845 }
 846
 847 /*
 848  * All xfs metadata buffers except log state machine buffers
 849  * get this attached as their b_bdstrat callback function.
 850  * This is so that we can catch a buffer
 851  * after prematurely unpinning it to forcibly shutdown the filesystem.
 852  */
 853 int
 854 xfs_bdstrat_cb(struct xfs_buf *bp)
 855 {
 856         xfs_mount_t     *mp;
 857
 858         mp = XFS_BUF_FSPRIVATE3(bp, xfs_mount_t *);
 859         if (!XFS_FORCED_SHUTDOWN(mp)) {
 860                 xfs_buf_iorequest(bp);
 861                 return 0;
 862         } else {
 863                 xfs_buftrace("XFS__BDSTRAT IOERROR", bp);
 864                 /*
 865                  * Metadata write that didn't get logged but
 866                  * written delayed anyway. These aren't associated
 867                  * with a transaction, and can be ignored.
 868                  */
 869                 if (XFS_BUF_IODONE_FUNC(bp) == NULL &&
 870                     (XFS_BUF_ISREAD(bp)) == 0)
 871                         return (xfs_bioerror_relse(bp));
 872                 else
 873                         return (xfs_bioerror(bp));
 874         }
 875 }
 876
 877 /*
 878  * Wrapper around bdstrat so that we can stop data from going to disk in case
 879  * we are shutting down the filesystem.  Typically user data goes thru this
 880  * path; one of the exceptions is the superblock.
 881  */
 882 void
 883 xfsbdstrat(
 884         struct xfs_mount        *mp,
 885         struct xfs_buf          *bp)
 886 {
 887         ASSERT(mp);
 888         if (!XFS_FORCED_SHUTDOWN(mp)) {
 889                 xfs_buf_iorequest(bp);
 890                 return;
 891         }
 892
 893         xfs_buftrace("XFSBDSTRAT IOERROR", bp);
 894         xfs_bioerror_relse(bp);
 895 }
 896
 897 /*
 898  * If the underlying (data/log/rt) device is readonly, there are some
 899  * operations that cannot proceed.
 900  */
 901 int
 902 xfs_dev_is_read_only(
 903         xfs_mount_t             *mp,
 904         char                    *message)
 905 {
 906         if (xfs_readonly_buftarg(mp->m_ddev_targp) ||
 907             xfs_readonly_buftarg(mp->m_logdev_targp) ||
 908             (mp->m_rtdev_targp && xfs_readonly_buftarg(mp->m_rtdev_targp))) {
 909                 cmn_err(CE_NOTE,
 910                         "XFS: %s required on read-only device.", message);
 911                 cmn_err(CE_NOTE,
 912                         "XFS: write access unavailable, cannot proceed.");
 913                 return EROFS;
 914         }
 915         return 0;
 916 }