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 = VFS_I(ip)->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(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 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);
 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);
 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(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_IOLOCK_EXCL));
 468         ASSERT(offset > isize);
 469
 470         /*
 471          * First handle zeroing the block on which isize resides.
 472          * We only zero a part of that block so it is handled specially.
 473          */
 474         error = xfs_zero_last_block(ip, offset, isize);
 475         if (error) {
 476                 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_IOLOCK_EXCL));
 477                 return error;
 478         }
 479
 480         /*
 481          * Calculate the range between the new size and the old
 482          * where blocks needing to be zeroed may exist.  To get the
 483          * block where the last byte in the file currently resides,
 484          * we need to subtract one from the size and truncate back
 485          * to a block boundary.  We subtract 1 in case the size is
 486          * exactly on a block boundary.
 487          */
 488         last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
 489         start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
 490         end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
 491         ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
 492         if (last_fsb == end_zero_fsb) {
 493                 /*
 494                  * The size was only incremented on its last block.
 495                  * We took care of that above, so just return.
 496                  */
 497                 return 0;
 498         }
 499
 500         ASSERT(start_zero_fsb <= end_zero_fsb);
 501         while (start_zero_fsb <= end_zero_fsb) {
 502                 nimaps = 1;
 503                 zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
 504                 error = xfs_bmapi(NULL, ip, start_zero_fsb, zero_count_fsb,
 505                                   0, NULL, 0, &imap, &nimaps, NULL, NULL);
 506                 if (error) {
 507                         ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_IOLOCK_EXCL));
 508                         return error;
 509                 }
 510                 ASSERT(nimaps > 0);
 511
 512                 if (imap.br_state == XFS_EXT_UNWRITTEN ||
 513                     imap.br_startblock == HOLESTARTBLOCK) {
 514                         /*
 515                          * This loop handles initializing pages that were
 516                          * partially initialized by the code below this
 517                          * loop. It basically zeroes the part of the page
 518                          * that sits on a hole and sets the page as P_HOLE
 519                          * and calls remapf if it is a mapped file.
 520                          */
 521                         start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 522                         ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 523                         continue;
 524                 }
 525
 526                 /*
 527                  * There are blocks we need to zero.
 528                  * Drop the inode lock while we're doing the I/O.
 529                  * We'll still have the iolock to protect us.
 530                  */
 531                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 532
 533                 zero_off = XFS_FSB_TO_B(mp, start_zero_fsb);
 534                 zero_len = XFS_FSB_TO_B(mp, imap.br_blockcount);
 535
 536                 if ((zero_off + zero_len) > offset)
 537                         zero_len = offset - zero_off;
 538
 539                 error = xfs_iozero(ip, zero_off, zero_len);
 540                 if (error) {
 541                         goto out_lock;
 542                 }
 543
 544                 start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 545                 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 546
 547                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 548         }
 549
 550         return 0;
 551
 552 out_lock:
 553         xfs_ilock(ip, XFS_ILOCK_EXCL);
 554         ASSERT(error >= 0);
 555         return error;
 556 }
 557
 558 ssize_t                         /* bytes written, or (-) error */
 559 xfs_write(
 560         struct xfs_inode        *xip,
 561         struct kiocb            *iocb,
 562         const struct iovec      *iovp,
 563         unsigned int            nsegs,
 564         loff_t                  *offset,
 565         int                     ioflags)
 566 {
 567         struct file             *file = iocb->ki_filp;
 568         struct address_space    *mapping = file->f_mapping;
 569         struct inode            *inode = mapping->host;
 570         unsigned long           segs = nsegs;
 571         xfs_mount_t             *mp;
 572         ssize_t                 ret = 0, error = 0;
 573         xfs_fsize_t             isize, new_size;
 574         int                     iolock;
 575         int                     eventsent = 0;
 576         size_t                  ocount = 0, count;
 577         loff_t                  pos;
 578         int                     need_i_mutex;
 579
 580         XFS_STATS_INC(xs_write_calls);
 581
 582         error = generic_segment_checks(iovp, &segs, &ocount, VERIFY_READ);
 583         if (error)
 584                 return error;
 585
 586         count = ocount;
 587         pos = *offset;
 588
 589         if (count == 0)
 590                 return 0;
 591
 592         mp = xip->i_mount;
 593
 594         xfs_wait_for_freeze(mp, SB_FREEZE_WRITE);
 595
 596         if (XFS_FORCED_SHUTDOWN(mp))
 597                 return -EIO;
 598
 599 relock:
 600         if (ioflags & IO_ISDIRECT) {
 601                 iolock = XFS_IOLOCK_SHARED;
 602                 need_i_mutex = 0;
 603         } else {
 604                 iolock = XFS_IOLOCK_EXCL;
 605                 need_i_mutex = 1;
 606                 mutex_lock(&inode->i_mutex);
 607         }
 608
 609         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
 610
 611 start:
 612         error = -generic_write_checks(file, &pos, &count,
 613                                         S_ISBLK(inode->i_mode));
 614         if (error) {
 615                 xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 616                 goto out_unlock_mutex;
 617         }
 618
 619         if ((DM_EVENT_ENABLED(xip, DM_EVENT_WRITE) &&
 620             !(ioflags & IO_INVIS) && !eventsent)) {
 621                 int             dmflags = FILP_DELAY_FLAG(file);
 622
 623                 if (need_i_mutex)
 624                         dmflags |= DM_FLAGS_IMUX;
 625
 626                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 627                 error = XFS_SEND_DATA(xip->i_mount, DM_EVENT_WRITE, xip,
 628                                       pos, count, dmflags, &iolock);
 629                 if (error) {
 630                         goto out_unlock_internal;
 631                 }
 632                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 633                 eventsent = 1;
 634
 635                 /*
 636                  * The iolock was dropped and reacquired in XFS_SEND_DATA
 637                  * so we have to recheck the size when appending.
 638                  * We will only "goto start;" once, since having sent the
 639                  * event prevents another call to XFS_SEND_DATA, which is
 640                  * what allows the size to change in the first place.
 641                  */
 642                 if ((file->f_flags & O_APPEND) && pos != xip->i_size)
 643                         goto start;
 644         }
 645
 646         if (ioflags & IO_ISDIRECT) {
 647                 xfs_buftarg_t   *target =
 648                         XFS_IS_REALTIME_INODE(xip) ?
 649                                 mp->m_rtdev_targp : mp->m_ddev_targp;
 650
 651                 if ((pos & target->bt_smask) || (count & target->bt_smask)) {
 652                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 653                         return XFS_ERROR(-EINVAL);
 654                 }
 655
 656                 if (!need_i_mutex && (mapping->nrpages || pos > xip->i_size)) {
 657                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 658                         iolock = XFS_IOLOCK_EXCL;
 659                         need_i_mutex = 1;
 660                         mutex_lock(&inode->i_mutex);
 661                         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
 662                         goto start;
 663                 }
 664         }
 665
 666         new_size = pos + count;
 667         if (new_size > xip->i_size)
 668                 xip->i_new_size = new_size;
 669
 670         if (likely(!(ioflags & IO_INVIS)))
 671                 xfs_ichgtime(xip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
 672
 673         /*
 674          * If the offset is beyond the size of the file, we have a couple
 675          * of things to do. First, if there is already space allocated
 676          * we need to either create holes or zero the disk or ...
 677          *
 678          * If there is a page where the previous size lands, we need
 679          * to zero it out up to the new size.
 680          */
 681
 682         if (pos > xip->i_size) {
 683                 error = xfs_zero_eof(xip, pos, xip->i_size);
 684                 if (error) {
 685                         xfs_iunlock(xip, XFS_ILOCK_EXCL);
 686                         goto out_unlock_internal;
 687                 }
 688         }
 689         xfs_iunlock(xip, XFS_ILOCK_EXCL);
 690
 691         /*
 692          * If we're writing the file then make sure to clear the
 693          * setuid and setgid bits if the process is not being run
 694          * by root.  This keeps people from modifying setuid and
 695          * setgid binaries.
 696          */
 697
 698         if (((xip->i_d.di_mode & S_ISUID) ||
 699             ((xip->i_d.di_mode & (S_ISGID | S_IXGRP)) ==
 700                 (S_ISGID | S_IXGRP))) &&
 701              !capable(CAP_FSETID)) {
 702                 error = xfs_write_clear_setuid(xip);
 703                 if (likely(!error))
 704                         error = -file_remove_suid(file);
 705                 if (unlikely(error)) {
 706                         goto out_unlock_internal;
 707                 }
 708         }
 709
 710         /* We can write back this queue in page reclaim */
 711         current->backing_dev_info = mapping->backing_dev_info;
 712
 713         if ((ioflags & IO_ISDIRECT)) {
 714                 if (mapping->nrpages) {
 715                         WARN_ON(need_i_mutex == 0);
 716                         xfs_inval_cached_trace(xip, pos, -1,
 717                                         (pos & PAGE_CACHE_MASK), -1);
 718                         error = xfs_flushinval_pages(xip,
 719                                         (pos & PAGE_CACHE_MASK),
 720                                         -1, FI_REMAPF_LOCKED);
 721                         if (error)
 722                                 goto out_unlock_internal;
 723                 }
 724
 725                 if (need_i_mutex) {
 726                         /* demote the lock now the cached pages are gone */
 727                         xfs_ilock_demote(xip, XFS_IOLOCK_EXCL);
 728                         mutex_unlock(&inode->i_mutex);
 729
 730                         iolock = XFS_IOLOCK_SHARED;
 731                         need_i_mutex = 0;
 732                 }
 733
 734                 xfs_rw_enter_trace(XFS_DIOWR_ENTER, xip, (void *)iovp, segs,
 735                                 *offset, ioflags);
 736                 ret = generic_file_direct_write(iocb, iovp,
 737                                 &segs, pos, offset, count, ocount);
 738
 739                 /*
 740                  * direct-io write to a hole: fall through to buffered I/O
 741                  * for completing the rest of the request.
 742                  */
 743                 if (ret >= 0 && ret != count) {
 744                         XFS_STATS_ADD(xs_write_bytes, ret);
 745
 746                         pos += ret;
 747                         count -= ret;
 748
 749                         ioflags &= ~IO_ISDIRECT;
 750                         xfs_iunlock(xip, iolock);
 751                         goto relock;
 752                 }
 753         } else {
 754                 int enospc = 0;
 755                 ssize_t ret2 = 0;
 756
 757 write_retry:
 758                 xfs_rw_enter_trace(XFS_WRITE_ENTER, xip, (void *)iovp, segs,
 759                                 *offset, ioflags);
 760                 ret2 = generic_file_buffered_write(iocb, iovp, segs,
 761                                 pos, offset, count, ret);
 762                 /*
 763                  * if we just got an ENOSPC, flush the inode now we
 764                  * aren't holding any page locks and retry *once*
 765                  */
 766                 if (ret2 == -ENOSPC && !enospc) {
 767                         error = xfs_flush_pages(xip, 0, -1, 0, FI_NONE);
 768                         if (error)
 769                                 goto out_unlock_internal;
 770                         enospc = 1;
 771                         goto write_retry;
 772                 }
 773                 ret = ret2;
 774         }
 775
 776         current->backing_dev_info = NULL;
 777
 778         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
 779                 ret = wait_on_sync_kiocb(iocb);
 780
 781         isize = i_size_read(inode);
 782         if (unlikely(ret < 0 && ret != -EFAULT && *offset > isize))
 783                 *offset = isize;
 784
 785         if (*offset > xip->i_size) {
 786                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 787                 if (*offset > xip->i_size)
 788                         xip->i_size = *offset;
 789                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 790         }
 791
 792         if (ret == -ENOSPC &&
 793             DM_EVENT_ENABLED(xip, DM_EVENT_NOSPACE) && !(ioflags & IO_INVIS)) {
 794                 xfs_iunlock(xip, iolock);
 795                 if (need_i_mutex)
 796                         mutex_unlock(&inode->i_mutex);
 797                 error = XFS_SEND_NAMESP(xip->i_mount, DM_EVENT_NOSPACE, xip,
 798                                 DM_RIGHT_NULL, xip, DM_RIGHT_NULL, NULL, NULL,
 799                                 0, 0, 0); /* Delay flag intentionally  unused */
 800                 if (need_i_mutex)
 801                         mutex_lock(&inode->i_mutex);
 802                 xfs_ilock(xip, iolock);
 803                 if (error)
 804                         goto out_unlock_internal;
 805                 goto start;
 806         }
 807
 808         error = -ret;
 809         if (ret <= 0)
 810                 goto out_unlock_internal;
 811
 812         XFS_STATS_ADD(xs_write_bytes, ret);
 813
 814         /* Handle various SYNC-type writes */
 815         if ((file->f_flags & O_SYNC) || IS_SYNC(inode)) {
 816                 int error2;
 817
 818                 xfs_iunlock(xip, iolock);
 819                 if (need_i_mutex)
 820                         mutex_unlock(&inode->i_mutex);
 821                 error2 = sync_page_range(inode, mapping, pos, ret);
 822                 if (!error)
 823                         error = error2;
 824                 if (need_i_mutex)
 825                         mutex_lock(&inode->i_mutex);
 826                 xfs_ilock(xip, iolock);
 827                 error2 = xfs_write_sync_logforce(mp, xip);
 828                 if (!error)
 829                         error = error2;
 830         }
 831
 832  out_unlock_internal:
 833         if (xip->i_new_size) {
 834                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 835                 xip->i_new_size = 0;
 836                 /*
 837                  * If this was a direct or synchronous I/O that failed (such
 838                  * as ENOSPC) then part of the I/O may have been written to
 839                  * disk before the error occured.  In this case the on-disk
 840                  * file size may have been adjusted beyond the in-memory file
 841                  * size and now needs to be truncated back.
 842                  */
 843                 if (xip->i_d.di_size > xip->i_size)
 844                         xip->i_d.di_size = xip->i_size;
 845                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 846         }
 847         xfs_iunlock(xip, iolock);
 848  out_unlock_mutex:
 849         if (need_i_mutex)
 850                 mutex_unlock(&inode->i_mutex);
 851         return -error;
 852 }
 853
 854 /*
 855  * All xfs metadata buffers except log state machine buffers
 856  * get this attached as their b_bdstrat callback function.
 857  * This is so that we can catch a buffer
 858  * after prematurely unpinning it to forcibly shutdown the filesystem.
 859  */
 860 int
 861 xfs_bdstrat_cb(struct xfs_buf *bp)
 862 {
 863         if (XFS_FORCED_SHUTDOWN(bp->b_mount)) {
 864                 xfs_buftrace("XFS__BDSTRAT IOERROR", bp);
 865                 /*
 866                  * Metadata write that didn't get logged but
 867                  * written delayed anyway. These aren't associated
 868                  * with a transaction, and can be ignored.
 869                  */
 870                 if (XFS_BUF_IODONE_FUNC(bp) == NULL &&
 871                     (XFS_BUF_ISREAD(bp)) == 0)
 872                         return (xfs_bioerror_relse(bp));
 873                 else
 874                         return (xfs_bioerror(bp));
 875         }
 876
 877         xfs_buf_iorequest(bp);
 878         return 0;
 879 }
 880
 881 /*
 882  * Wrapper around bdstrat so that we can stop data from going to disk in case
 883  * we are shutting down the filesystem.  Typically user data goes thru this
 884  * path; one of the exceptions is the superblock.
 885  */
 886 void
 887 xfsbdstrat(
 888         struct xfs_mount        *mp,
 889         struct xfs_buf          *bp)
 890 {
 891         ASSERT(mp);
 892         if (!XFS_FORCED_SHUTDOWN(mp)) {
 893                 xfs_buf_iorequest(bp);
 894                 return;
 895         }
 896
 897         xfs_buftrace("XFSBDSTRAT IOERROR", bp);
 898         xfs_bioerror_relse(bp);
 899 }
 900
 901 /*
 902  * If the underlying (data/log/rt) device is readonly, there are some
 903  * operations that cannot proceed.
 904  */
 905 int
 906 xfs_dev_is_read_only(
 907         xfs_mount_t             *mp,
 908         char                    *message)
 909 {
 910         if (xfs_readonly_buftarg(mp->m_ddev_targp) ||
 911             xfs_readonly_buftarg(mp->m_logdev_targp) ||
 912             (mp->m_rtdev_targp && xfs_readonly_buftarg(mp->m_rtdev_targp))) {
 913                 cmn_err(CE_NOTE,
 914                         "XFS: %s required on read-only device.", message);
 915                 cmn_err(CE_NOTE,
 916                         "XFS: write access unavailable, cannot proceed.");
 917                 return EROFS;
 918         }
 919         return 0;
 920 }