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_iocore_t            *io,
  62         void                    *data,
  63         size_t                  segs,
  64         loff_t                  offset,
  65         int                     ioflags)
  66 {
  67         xfs_inode_t     *ip = XFS_IO_INODE(io);
  68
  69         if (ip->i_rwtrace == NULL)
  70                 return;
  71         ktrace_enter(ip->i_rwtrace,
  72                 (void *)(unsigned long)tag,
  73                 (void *)ip,
  74                 (void *)((unsigned long)((ip->i_d.di_size >> 32) & 0xffffffff)),
  75                 (void *)((unsigned long)(ip->i_d.di_size & 0xffffffff)),
  76                 (void *)data,
  77                 (void *)((unsigned long)segs),
  78                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
  79                 (void *)((unsigned long)(offset & 0xffffffff)),
  80                 (void *)((unsigned long)ioflags),
  81                 (void *)((unsigned long)((io->io_new_size >> 32) & 0xffffffff)),
  82                 (void *)((unsigned long)(io->io_new_size & 0xffffffff)),
  83                 (void *)((unsigned long)current_pid()),
  84                 (void *)NULL,
  85                 (void *)NULL,
  86                 (void *)NULL,
  87                 (void *)NULL);
  88 }
  89
  90 void
  91 xfs_inval_cached_trace(
  92         xfs_iocore_t    *io,
  93         xfs_off_t       offset,
  94         xfs_off_t       len,
  95         xfs_off_t       first,
  96         xfs_off_t       last)
  97 {
  98         xfs_inode_t     *ip = XFS_IO_INODE(io);
  99
 100         if (ip->i_rwtrace == NULL)
 101                 return;
 102         ktrace_enter(ip->i_rwtrace,
 103                 (void *)(__psint_t)XFS_INVAL_CACHED,
 104                 (void *)ip,
 105                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
 106                 (void *)((unsigned long)(offset & 0xffffffff)),
 107                 (void *)((unsigned long)((len >> 32) & 0xffffffff)),
 108                 (void *)((unsigned long)(len & 0xffffffff)),
 109                 (void *)((unsigned long)((first >> 32) & 0xffffffff)),
 110                 (void *)((unsigned long)(first & 0xffffffff)),
 111                 (void *)((unsigned long)((last >> 32) & 0xffffffff)),
 112                 (void *)((unsigned long)(last & 0xffffffff)),
 113                 (void *)((unsigned long)current_pid()),
 114                 (void *)NULL,
 115                 (void *)NULL,
 116                 (void *)NULL,
 117                 (void *)NULL,
 118                 (void *)NULL);
 119 }
 120 #endif
 121
 122 /*
 123  *      xfs_iozero
 124  *
 125  *      xfs_iozero clears the specified range of buffer supplied,
 126  *      and marks all the affected blocks as valid and modified.  If
 127  *      an affected block is not allocated, it will be allocated.  If
 128  *      an affected block is not completely overwritten, and is not
 129  *      valid before the operation, it will be read from disk before
 130  *      being partially zeroed.
 131  */
 132 STATIC int
 133 xfs_iozero(
 134         struct inode            *ip,    /* inode                        */
 135         loff_t                  pos,    /* offset in file               */
 136         size_t                  count)  /* size of data to zero         */
 137 {
 138         struct page             *page;
 139         struct address_space    *mapping;
 140         int                     status;
 141
 142         mapping = ip->i_mapping;
 143         do {
 144                 unsigned offset, bytes;
 145                 void *fsdata;
 146
 147                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
 148                 bytes = PAGE_CACHE_SIZE - offset;
 149                 if (bytes > count)
 150                         bytes = count;
 151
 152                 status = pagecache_write_begin(NULL, mapping, pos, bytes,
 153                                         AOP_FLAG_UNINTERRUPTIBLE,
 154                                         &page, &fsdata);
 155                 if (status)
 156                         break;
 157
 158                 zero_user(page, offset, bytes);
 159
 160                 status = pagecache_write_end(NULL, mapping, pos, bytes, bytes,
 161                                         page, fsdata);
 162                 WARN_ON(status <= 0); /* can't return less than zero! */
 163                 pos += bytes;
 164                 count -= bytes;
 165                 status = 0;
 166         } while (count);
 167
 168         return (-status);
 169 }
 170
 171 ssize_t                 /* bytes read, or (-)  error */
 172 xfs_read(
 173         xfs_inode_t             *ip,
 174         struct kiocb            *iocb,
 175         const struct iovec      *iovp,
 176         unsigned int            segs,
 177         loff_t                  *offset,
 178         int                     ioflags)
 179 {
 180         struct file             *file = iocb->ki_filp;
 181         struct inode            *inode = file->f_mapping->host;
 182         bhv_vnode_t             *vp = XFS_ITOV(ip);
 183         xfs_mount_t             *mp = ip->i_mount;
 184         size_t                  size = 0;
 185         ssize_t                 ret = 0;
 186         xfs_fsize_t             n;
 187         unsigned long           seg;
 188
 189
 190         XFS_STATS_INC(xs_read_calls);
 191
 192         /* START copy & waste from filemap.c */
 193         for (seg = 0; seg < segs; seg++) {
 194                 const struct iovec *iv = &iovp[seg];
 195
 196                 /*
 197                  * If any segment has a negative length, or the cumulative
 198                  * length ever wraps negative then return -EINVAL.
 199                  */
 200                 size += iv->iov_len;
 201                 if (unlikely((ssize_t)(size|iv->iov_len) < 0))
 202                         return XFS_ERROR(-EINVAL);
 203         }
 204         /* END copy & waste from filemap.c */
 205
 206         if (unlikely(ioflags & IO_ISDIRECT)) {
 207                 xfs_buftarg_t   *target =
 208                         (ip->i_d.di_flags & XFS_DIFLAG_REALTIME) ?
 209                                 mp->m_rtdev_targp : mp->m_ddev_targp;
 210                 if ((*offset & target->bt_smask) ||
 211                     (size & target->bt_smask)) {
 212                         if (*offset == ip->i_size) {
 213                                 return (0);
 214                         }
 215                         return -XFS_ERROR(EINVAL);
 216                 }
 217         }
 218
 219         n = XFS_MAXIOFFSET(mp) - *offset;
 220         if ((n <= 0) || (size == 0))
 221                 return 0;
 222
 223         if (n < size)
 224                 size = n;
 225
 226         if (XFS_FORCED_SHUTDOWN(mp))
 227                 return -EIO;
 228
 229         if (unlikely(ioflags & IO_ISDIRECT))
 230                 mutex_lock(&inode->i_mutex);
 231         xfs_ilock(ip, XFS_IOLOCK_SHARED);
 232
 233         if (DM_EVENT_ENABLED(ip, DM_EVENT_READ) && !(ioflags & IO_INVIS)) {
 234                 bhv_vrwlock_t locktype = VRWLOCK_READ;
 235                 int dmflags = FILP_DELAY_FLAG(file) | DM_SEM_FLAG_RD(ioflags);
 236
 237                 ret = -XFS_SEND_DATA(mp, DM_EVENT_READ, vp, *offset, size,
 238                                         dmflags, &locktype);
 239                 if (ret) {
 240                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 241                         if (unlikely(ioflags & IO_ISDIRECT))
 242                                 mutex_unlock(&inode->i_mutex);
 243                         return ret;
 244                 }
 245         }
 246
 247         if (unlikely(ioflags & IO_ISDIRECT)) {
 248                 if (VN_CACHED(vp))
 249                         ret = xfs_flushinval_pages(ip,
 250                                         ctooff(offtoct(*offset)),
 251                                         -1, FI_REMAPF_LOCKED);
 252                 mutex_unlock(&inode->i_mutex);
 253                 if (ret) {
 254                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 255                         return ret;
 256                 }
 257         }
 258
 259         xfs_rw_enter_trace(XFS_READ_ENTER, &ip->i_iocore,
 260                                 (void *)iovp, segs, *offset, ioflags);
 261
 262         iocb->ki_pos = *offset;
 263         ret = generic_file_aio_read(iocb, iovp, segs, *offset);
 264         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
 265                 ret = wait_on_sync_kiocb(iocb);
 266         if (ret > 0)
 267                 XFS_STATS_ADD(xs_read_bytes, ret);
 268
 269         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 270         return ret;
 271 }
 272
 273 ssize_t
 274 xfs_splice_read(
 275         xfs_inode_t             *ip,
 276         struct file             *infilp,
 277         loff_t                  *ppos,
 278         struct pipe_inode_info  *pipe,
 279         size_t                  count,
 280         int                     flags,
 281         int                     ioflags)
 282 {
 283         bhv_vnode_t             *vp = XFS_ITOV(ip);
 284         xfs_mount_t             *mp = ip->i_mount;
 285         ssize_t                 ret;
 286
 287         XFS_STATS_INC(xs_read_calls);
 288         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 289                 return -EIO;
 290
 291         xfs_ilock(ip, XFS_IOLOCK_SHARED);
 292
 293         if (DM_EVENT_ENABLED(ip, DM_EVENT_READ) && !(ioflags & IO_INVIS)) {
 294                 bhv_vrwlock_t locktype = VRWLOCK_READ;
 295                 int error;
 296
 297                 error = XFS_SEND_DATA(mp, DM_EVENT_READ, vp, *ppos, count,
 298                                         FILP_DELAY_FLAG(infilp), &locktype);
 299                 if (error) {
 300                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 301                         return -error;
 302                 }
 303         }
 304         xfs_rw_enter_trace(XFS_SPLICE_READ_ENTER, &ip->i_iocore,
 305                            pipe, count, *ppos, ioflags);
 306         ret = generic_file_splice_read(infilp, ppos, pipe, count, flags);
 307         if (ret > 0)
 308                 XFS_STATS_ADD(xs_read_bytes, ret);
 309
 310         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 311         return ret;
 312 }
 313
 314 ssize_t
 315 xfs_splice_write(
 316         xfs_inode_t             *ip,
 317         struct pipe_inode_info  *pipe,
 318         struct file             *outfilp,
 319         loff_t                  *ppos,
 320         size_t                  count,
 321         int                     flags,
 322         int                     ioflags)
 323 {
 324         bhv_vnode_t             *vp = XFS_ITOV(ip);
 325         xfs_mount_t             *mp = ip->i_mount;
 326         xfs_iocore_t            *io = &ip->i_iocore;
 327         ssize_t                 ret;
 328         struct inode            *inode = outfilp->f_mapping->host;
 329         xfs_fsize_t             isize, new_size;
 330
 331         XFS_STATS_INC(xs_write_calls);
 332         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 333                 return -EIO;
 334
 335         xfs_ilock(ip, XFS_IOLOCK_EXCL);
 336
 337         if (DM_EVENT_ENABLED(ip, DM_EVENT_WRITE) && !(ioflags & IO_INVIS)) {
 338                 bhv_vrwlock_t locktype = VRWLOCK_WRITE;
 339                 int error;
 340
 341                 error = XFS_SEND_DATA(mp, DM_EVENT_WRITE, vp, *ppos, count,
 342                                         FILP_DELAY_FLAG(outfilp), &locktype);
 343                 if (error) {
 344                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 345                         return -error;
 346                 }
 347         }
 348
 349         new_size = *ppos + count;
 350
 351         xfs_ilock(ip, XFS_ILOCK_EXCL);
 352         if (new_size > ip->i_size)
 353                 io->io_new_size = new_size;
 354         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 355
 356         xfs_rw_enter_trace(XFS_SPLICE_WRITE_ENTER, &ip->i_iocore,
 357                            pipe, count, *ppos, ioflags);
 358         ret = generic_file_splice_write(pipe, outfilp, ppos, count, flags);
 359         if (ret > 0)
 360                 XFS_STATS_ADD(xs_write_bytes, ret);
 361
 362         isize = i_size_read(inode);
 363         if (unlikely(ret < 0 && ret != -EFAULT && *ppos > isize))
 364                 *ppos = isize;
 365
 366         if (*ppos > ip->i_size) {
 367                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 368                 if (*ppos > ip->i_size)
 369                         ip->i_size = *ppos;
 370                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 371         }
 372
 373         if (io->io_new_size) {
 374                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 375                 io->io_new_size = 0;
 376                 if (ip->i_d.di_size > ip->i_size)
 377                         ip->i_d.di_size = ip->i_size;
 378                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 379         }
 380         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 381         return ret;
 382 }
 383
 384 /*
 385  * This routine is called to handle zeroing any space in the last
 386  * block of the file that is beyond the EOF.  We do this since the
 387  * size is being increased without writing anything to that block
 388  * and we don't want anyone to read the garbage on the disk.
 389  */
 390 STATIC int                              /* error (positive) */
 391 xfs_zero_last_block(
 392         struct inode    *ip,
 393         xfs_iocore_t    *io,
 394         xfs_fsize_t     offset,
 395         xfs_fsize_t     isize)
 396 {
 397         xfs_fileoff_t   last_fsb;
 398         xfs_mount_t     *mp = io->io_mount;
 399         int             nimaps;
 400         int             zero_offset;
 401         int             zero_len;
 402         int             error = 0;
 403         xfs_bmbt_irec_t imap;
 404
 405         ASSERT(ismrlocked(io->io_lock, MR_UPDATE) != 0);
 406
 407         zero_offset = XFS_B_FSB_OFFSET(mp, isize);
 408         if (zero_offset == 0) {
 409                 /*
 410                  * There are no extra bytes in the last block on disk to
 411                  * zero, so return.
 412                  */
 413                 return 0;
 414         }
 415
 416         last_fsb = XFS_B_TO_FSBT(mp, isize);
 417         nimaps = 1;
 418         error = XFS_BMAPI(mp, NULL, io, last_fsb, 1, 0, NULL, 0, &imap,
 419                           &nimaps, NULL, NULL);
 420         if (error) {
 421                 return error;
 422         }
 423         ASSERT(nimaps > 0);
 424         /*
 425          * If the block underlying isize is just a hole, then there
 426          * is nothing to zero.
 427          */
 428         if (imap.br_startblock == HOLESTARTBLOCK) {
 429                 return 0;
 430         }
 431         /*
 432          * Zero the part of the last block beyond the EOF, and write it
 433          * out sync.  We need to drop the ilock while we do this so we
 434          * don't deadlock when the buffer cache calls back to us.
 435          */
 436         XFS_IUNLOCK(mp, io, XFS_ILOCK_EXCL| XFS_EXTSIZE_RD);
 437
 438         zero_len = mp->m_sb.sb_blocksize - zero_offset;
 439         if (isize + zero_len > offset)
 440                 zero_len = offset - isize;
 441         error = xfs_iozero(ip, isize, zero_len);
 442
 443         XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 444         ASSERT(error >= 0);
 445         return error;
 446 }
 447
 448 /*
 449  * Zero any on disk space between the current EOF and the new,
 450  * larger EOF.  This handles the normal case of zeroing the remainder
 451  * of the last block in the file and the unusual case of zeroing blocks
 452  * out beyond the size of the file.  This second case only happens
 453  * with fixed size extents and when the system crashes before the inode
 454  * size was updated but after blocks were allocated.  If fill is set,
 455  * then any holes in the range are filled and zeroed.  If not, the holes
 456  * are left alone as holes.
 457  */
 458
 459 int                                     /* error (positive) */
 460 xfs_zero_eof(
 461         bhv_vnode_t     *vp,
 462         xfs_iocore_t    *io,
 463         xfs_off_t       offset,         /* starting I/O offset */
 464         xfs_fsize_t     isize)          /* current inode size */
 465 {
 466         struct inode    *ip = vn_to_inode(vp);
 467         xfs_fileoff_t   start_zero_fsb;
 468         xfs_fileoff_t   end_zero_fsb;
 469         xfs_fileoff_t   zero_count_fsb;
 470         xfs_fileoff_t   last_fsb;
 471         xfs_fileoff_t   zero_off;
 472         xfs_fsize_t     zero_len;
 473         xfs_mount_t     *mp = io->io_mount;
 474         int             nimaps;
 475         int             error = 0;
 476         xfs_bmbt_irec_t imap;
 477
 478         ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
 479         ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
 480         ASSERT(offset > isize);
 481
 482         /*
 483          * First handle zeroing the block on which isize resides.
 484          * We only zero a part of that block so it is handled specially.
 485          */
 486         error = xfs_zero_last_block(ip, io, offset, isize);
 487         if (error) {
 488                 ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
 489                 ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
 490                 return error;
 491         }
 492
 493         /*
 494          * Calculate the range between the new size and the old
 495          * where blocks needing to be zeroed may exist.  To get the
 496          * block where the last byte in the file currently resides,
 497          * we need to subtract one from the size and truncate back
 498          * to a block boundary.  We subtract 1 in case the size is
 499          * exactly on a block boundary.
 500          */
 501         last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
 502         start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
 503         end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
 504         ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
 505         if (last_fsb == end_zero_fsb) {
 506                 /*
 507                  * The size was only incremented on its last block.
 508                  * We took care of that above, so just return.
 509                  */
 510                 return 0;
 511         }
 512
 513         ASSERT(start_zero_fsb <= end_zero_fsb);
 514         while (start_zero_fsb <= end_zero_fsb) {
 515                 nimaps = 1;
 516                 zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
 517                 error = XFS_BMAPI(mp, NULL, io, start_zero_fsb, zero_count_fsb,
 518                                   0, NULL, 0, &imap, &nimaps, NULL, NULL);
 519                 if (error) {
 520                         ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
 521                         ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
 522                         return error;
 523                 }
 524                 ASSERT(nimaps > 0);
 525
 526                 if (imap.br_state == XFS_EXT_UNWRITTEN ||
 527                     imap.br_startblock == HOLESTARTBLOCK) {
 528                         /*
 529                          * This loop handles initializing pages that were
 530                          * partially initialized by the code below this
 531                          * loop. It basically zeroes the part of the page
 532                          * that sits on a hole and sets the page as P_HOLE
 533                          * and calls remapf if it is a mapped file.
 534                          */
 535                         start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 536                         ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 537                         continue;
 538                 }
 539
 540                 /*
 541                  * There are blocks we need to zero.
 542                  * Drop the inode lock while we're doing the I/O.
 543                  * We'll still have the iolock to protect us.
 544                  */
 545                 XFS_IUNLOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 546
 547                 zero_off = XFS_FSB_TO_B(mp, start_zero_fsb);
 548                 zero_len = XFS_FSB_TO_B(mp, imap.br_blockcount);
 549
 550                 if ((zero_off + zero_len) > offset)
 551                         zero_len = offset - zero_off;
 552
 553                 error = xfs_iozero(ip, zero_off, zero_len);
 554                 if (error) {
 555                         goto out_lock;
 556                 }
 557
 558                 start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 559                 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 560
 561                 XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 562         }
 563
 564         return 0;
 565
 566 out_lock:
 567
 568         XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 569         ASSERT(error >= 0);
 570         return error;
 571 }
 572
 573 ssize_t                         /* bytes written, or (-) error */
 574 xfs_write(
 575         struct xfs_inode        *xip,
 576         struct kiocb            *iocb,
 577         const struct iovec      *iovp,
 578         unsigned int            nsegs,
 579         loff_t                  *offset,
 580         int                     ioflags)
 581 {
 582         struct file             *file = iocb->ki_filp;
 583         struct address_space    *mapping = file->f_mapping;
 584         struct inode            *inode = mapping->host;
 585         bhv_vnode_t             *vp = XFS_ITOV(xip);
 586         unsigned long           segs = nsegs;
 587         xfs_mount_t             *mp;
 588         ssize_t                 ret = 0, error = 0;
 589         xfs_fsize_t             isize, new_size;
 590         xfs_iocore_t            *io;
 591         int                     iolock;
 592         int                     eventsent = 0;
 593         bhv_vrwlock_t           locktype;
 594         size_t                  ocount = 0, count;
 595         loff_t                  pos;
 596         int                     need_i_mutex;
 597
 598         XFS_STATS_INC(xs_write_calls);
 599
 600         error = generic_segment_checks(iovp, &segs, &ocount, VERIFY_READ);
 601         if (error)
 602                 return error;
 603
 604         count = ocount;
 605         pos = *offset;
 606
 607         if (count == 0)
 608                 return 0;
 609
 610         io = &xip->i_iocore;
 611         mp = io->io_mount;
 612
 613         xfs_wait_for_freeze(mp, SB_FREEZE_WRITE);
 614
 615         if (XFS_FORCED_SHUTDOWN(mp))
 616                 return -EIO;
 617
 618 relock:
 619         if (ioflags & IO_ISDIRECT) {
 620                 iolock = XFS_IOLOCK_SHARED;
 621                 locktype = VRWLOCK_WRITE_DIRECT;
 622                 need_i_mutex = 0;
 623         } else {
 624                 iolock = XFS_IOLOCK_EXCL;
 625                 locktype = VRWLOCK_WRITE;
 626                 need_i_mutex = 1;
 627                 mutex_lock(&inode->i_mutex);
 628         }
 629
 630         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
 631
 632 start:
 633         error = -generic_write_checks(file, &pos, &count,
 634                                         S_ISBLK(inode->i_mode));
 635         if (error) {
 636                 xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 637                 goto out_unlock_mutex;
 638         }
 639
 640         if ((DM_EVENT_ENABLED(xip, DM_EVENT_WRITE) &&
 641             !(ioflags & IO_INVIS) && !eventsent)) {
 642                 int             dmflags = FILP_DELAY_FLAG(file);
 643
 644                 if (need_i_mutex)
 645                         dmflags |= DM_FLAGS_IMUX;
 646
 647                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 648                 error = XFS_SEND_DATA(xip->i_mount, DM_EVENT_WRITE, vp,
 649                                       pos, count,
 650                                       dmflags, &locktype);
 651                 if (error) {
 652                         goto out_unlock_internal;
 653                 }
 654                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 655                 eventsent = 1;
 656
 657                 /*
 658                  * The iolock was dropped and reacquired in XFS_SEND_DATA
 659                  * so we have to recheck the size when appending.
 660                  * We will only "goto start;" once, since having sent the
 661                  * event prevents another call to XFS_SEND_DATA, which is
 662                  * what allows the size to change in the first place.
 663                  */
 664                 if ((file->f_flags & O_APPEND) && pos != xip->i_size)
 665                         goto start;
 666         }
 667
 668         if (ioflags & IO_ISDIRECT) {
 669                 xfs_buftarg_t   *target =
 670                         (xip->i_d.di_flags & XFS_DIFLAG_REALTIME) ?
 671                                 mp->m_rtdev_targp : mp->m_ddev_targp;
 672
 673                 if ((pos & target->bt_smask) || (count & target->bt_smask)) {
 674                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 675                         return XFS_ERROR(-EINVAL);
 676                 }
 677
 678                 if (!need_i_mutex && (VN_CACHED(vp) || pos > xip->i_size)) {
 679                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 680                         iolock = XFS_IOLOCK_EXCL;
 681                         locktype = VRWLOCK_WRITE;
 682                         need_i_mutex = 1;
 683                         mutex_lock(&inode->i_mutex);
 684                         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
 685                         goto start;
 686                 }
 687         }
 688
 689         new_size = pos + count;
 690         if (new_size > xip->i_size)
 691                 io->io_new_size = new_size;
 692
 693         if (likely(!(ioflags & IO_INVIS))) {
 694                 file_update_time(file);
 695                 xfs_ichgtime_fast(xip, inode,
 696                                   XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
 697         }
 698
 699         /*
 700          * If the offset is beyond the size of the file, we have a couple
 701          * of things to do. First, if there is already space allocated
 702          * we need to either create holes or zero the disk or ...
 703          *
 704          * If there is a page where the previous size lands, we need
 705          * to zero it out up to the new size.
 706          */
 707
 708         if (pos > xip->i_size) {
 709                 error = xfs_zero_eof(vp, io, pos, xip->i_size);
 710                 if (error) {
 711                         xfs_iunlock(xip, XFS_ILOCK_EXCL);
 712                         goto out_unlock_internal;
 713                 }
 714         }
 715         xfs_iunlock(xip, XFS_ILOCK_EXCL);
 716
 717         /*
 718          * If we're writing the file then make sure to clear the
 719          * setuid and setgid bits if the process is not being run
 720          * by root.  This keeps people from modifying setuid and
 721          * setgid binaries.
 722          */
 723
 724         if (((xip->i_d.di_mode & S_ISUID) ||
 725             ((xip->i_d.di_mode & (S_ISGID | S_IXGRP)) ==
 726                 (S_ISGID | S_IXGRP))) &&
 727              !capable(CAP_FSETID)) {
 728                 error = xfs_write_clear_setuid(xip);
 729                 if (likely(!error))
 730                         error = -remove_suid(file->f_path.dentry);
 731                 if (unlikely(error)) {
 732                         goto out_unlock_internal;
 733                 }
 734         }
 735
 736 retry:
 737         /* We can write back this queue in page reclaim */
 738         current->backing_dev_info = mapping->backing_dev_info;
 739
 740         if ((ioflags & IO_ISDIRECT)) {
 741                 if (VN_CACHED(vp)) {
 742                         WARN_ON(need_i_mutex == 0);
 743                         xfs_inval_cached_trace(io, pos, -1,
 744                                         ctooff(offtoct(pos)), -1);
 745                         error = xfs_flushinval_pages(xip,
 746                                         ctooff(offtoct(pos)),
 747                                         -1, FI_REMAPF_LOCKED);
 748                         if (error)
 749                                 goto out_unlock_internal;
 750                 }
 751
 752                 if (need_i_mutex) {
 753                         /* demote the lock now the cached pages are gone */
 754                         XFS_ILOCK_DEMOTE(mp, io, XFS_IOLOCK_EXCL);
 755                         mutex_unlock(&inode->i_mutex);
 756
 757                         iolock = XFS_IOLOCK_SHARED;
 758                         locktype = VRWLOCK_WRITE_DIRECT;
 759                         need_i_mutex = 0;
 760                 }
 761
 762                 xfs_rw_enter_trace(XFS_DIOWR_ENTER, io, (void *)iovp, segs,
 763                                 *offset, ioflags);
 764                 ret = generic_file_direct_write(iocb, iovp,
 765                                 &segs, pos, offset, count, ocount);
 766
 767                 /*
 768                  * direct-io write to a hole: fall through to buffered I/O
 769                  * for completing the rest of the request.
 770                  */
 771                 if (ret >= 0 && ret != count) {
 772                         XFS_STATS_ADD(xs_write_bytes, ret);
 773
 774                         pos += ret;
 775                         count -= ret;
 776
 777                         ioflags &= ~IO_ISDIRECT;
 778                         xfs_iunlock(xip, iolock);
 779                         goto relock;
 780                 }
 781         } else {
 782                 xfs_rw_enter_trace(XFS_WRITE_ENTER, io, (void *)iovp, segs,
 783                                 *offset, ioflags);
 784                 ret = generic_file_buffered_write(iocb, iovp, segs,
 785                                 pos, offset, count, ret);
 786         }
 787
 788         current->backing_dev_info = NULL;
 789
 790         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
 791                 ret = wait_on_sync_kiocb(iocb);
 792
 793         if (ret == -ENOSPC &&
 794             DM_EVENT_ENABLED(xip, DM_EVENT_NOSPACE) && !(ioflags & IO_INVIS)) {
 795                 xfs_rwunlock(xip, locktype);
 796                 if (need_i_mutex)
 797                         mutex_unlock(&inode->i_mutex);
 798                 error = XFS_SEND_NAMESP(xip->i_mount, DM_EVENT_NOSPACE, vp,
 799                                 DM_RIGHT_NULL, vp, DM_RIGHT_NULL, NULL, NULL,
 800                                 0, 0, 0); /* Delay flag intentionally  unused */
 801                 if (need_i_mutex)
 802                         mutex_lock(&inode->i_mutex);
 803                 xfs_rwlock(xip, locktype);
 804                 if (error)
 805                         goto out_unlock_internal;
 806                 pos = xip->i_size;
 807                 ret = 0;
 808                 goto retry;
 809         }
 810
 811         isize = i_size_read(inode);
 812         if (unlikely(ret < 0 && ret != -EFAULT && *offset > isize))
 813                 *offset = isize;
 814
 815         if (*offset > xip->i_size) {
 816                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 817                 if (*offset > xip->i_size)
 818                         xip->i_size = *offset;
 819                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 820         }
 821
 822         error = -ret;
 823         if (ret <= 0)
 824                 goto out_unlock_internal;
 825
 826         XFS_STATS_ADD(xs_write_bytes, ret);
 827
 828         /* Handle various SYNC-type writes */
 829         if ((file->f_flags & O_SYNC) || IS_SYNC(inode)) {
 830                 int error2;
 831                 xfs_rwunlock(xip, locktype);
 832                 if (need_i_mutex)
 833                         mutex_unlock(&inode->i_mutex);
 834                 error2 = sync_page_range(inode, mapping, pos, ret);
 835                 if (!error)
 836                         error = error2;
 837                 if (need_i_mutex)
 838                         mutex_lock(&inode->i_mutex);
 839                 xfs_rwlock(xip, locktype);
 840                 error2 = xfs_write_sync_logforce(mp, xip);
 841                 if (!error)
 842                         error = error2;
 843         }
 844
 845  out_unlock_internal:
 846         if (io->io_new_size) {
 847                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 848                 io->io_new_size = 0;
 849                 /*
 850                  * If this was a direct or synchronous I/O that failed (such
 851                  * as ENOSPC) then part of the I/O may have been written to
 852                  * disk before the error occured.  In this case the on-disk
 853                  * file size may have been adjusted beyond the in-memory file
 854                  * size and now needs to be truncated back.
 855                  */
 856                 if (xip->i_d.di_size > xip->i_size)
 857                         xip->i_d.di_size = xip->i_size;
 858                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 859         }
 860         xfs_rwunlock(xip, locktype);
 861  out_unlock_mutex:
 862         if (need_i_mutex)
 863                 mutex_unlock(&inode->i_mutex);
 864         return -error;
 865 }
 866
 867 /*
 868  * All xfs metadata buffers except log state machine buffers
 869  * get this attached as their b_bdstrat callback function.
 870  * This is so that we can catch a buffer
 871  * after prematurely unpinning it to forcibly shutdown the filesystem.
 872  */
 873 int
 874 xfs_bdstrat_cb(struct xfs_buf *bp)
 875 {
 876         xfs_mount_t     *mp;
 877
 878         mp = XFS_BUF_FSPRIVATE3(bp, xfs_mount_t *);
 879         if (!XFS_FORCED_SHUTDOWN(mp)) {
 880                 xfs_buf_iorequest(bp);
 881                 return 0;
 882         } else {
 883                 xfs_buftrace("XFS__BDSTRAT IOERROR", bp);
 884                 /*
 885                  * Metadata write that didn't get logged but
 886                  * written delayed anyway. These aren't associated
 887                  * with a transaction, and can be ignored.
 888                  */
 889                 if (XFS_BUF_IODONE_FUNC(bp) == NULL &&
 890                     (XFS_BUF_ISREAD(bp)) == 0)
 891                         return (xfs_bioerror_relse(bp));
 892                 else
 893                         return (xfs_bioerror(bp));
 894         }
 895 }
 896
 897
 898 int
 899 xfs_bmap(
 900         xfs_inode_t     *ip,
 901         xfs_off_t       offset,
 902         ssize_t         count,
 903         int             flags,
 904         xfs_iomap_t     *iomapp,
 905         int             *niomaps)
 906 {
 907         xfs_iocore_t    *io = &ip->i_iocore;
 908
 909         ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFREG);
 910         ASSERT(((ip->i_d.di_flags & XFS_DIFLAG_REALTIME) != 0) ==
 911                ((ip->i_iocore.io_flags & XFS_IOCORE_RT) != 0));
 912
 913         return xfs_iomap(io, offset, count, flags, iomapp, niomaps);
 914 }
 915
 916 /*
 917  * Wrapper around bdstrat so that we can stop data
 918  * from going to disk in case we are shutting down the filesystem.
 919  * Typically user data goes thru this path; one of the exceptions
 920  * is the superblock.
 921  */
 922 int
 923 xfsbdstrat(
 924         struct xfs_mount        *mp,
 925         struct xfs_buf          *bp)
 926 {
 927         ASSERT(mp);
 928         if (!XFS_FORCED_SHUTDOWN(mp)) {
 929                 /* Grio redirection would go here
 930                  * if (XFS_BUF_IS_GRIO(bp)) {
 931                  */
 932
 933                 xfs_buf_iorequest(bp);
 934                 return 0;
 935         }
 936
 937         xfs_buftrace("XFSBDSTRAT IOERROR", bp);
 938         return (xfs_bioerror_relse(bp));
 939 }
 940
 941 /*
 942  * If the underlying (data/log/rt) device is readonly, there are some
 943  * operations that cannot proceed.
 944  */
 945 int
 946 xfs_dev_is_read_only(
 947         xfs_mount_t             *mp,
 948         char                    *message)
 949 {
 950         if (xfs_readonly_buftarg(mp->m_ddev_targp) ||
 951             xfs_readonly_buftarg(mp->m_logdev_targp) ||
 952             (mp->m_rtdev_targp && xfs_readonly_buftarg(mp->m_rtdev_targp))) {
 953                 cmn_err(CE_NOTE,
 954                         "XFS: %s required on read-only device.", message);
 955                 cmn_err(CE_NOTE,
 956                         "XFS: write access unavailable, cannot proceed.");
 957                 return EROFS;
 958         }
 959         return 0;
 960 }