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_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         unsigned                bytes;
 139         struct page             *page;
 140         struct address_space    *mapping;
 141         int                     status;
 142
 143         mapping = ip->i_mapping;
 144         do {
 145                 unsigned long index, offset;
 146
 147                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
 148                 index = pos >> PAGE_CACHE_SHIFT;
 149                 bytes = PAGE_CACHE_SIZE - offset;
 150                 if (bytes > count)
 151                         bytes = count;
 152
 153                 status = -ENOMEM;
 154                 page = grab_cache_page(mapping, index);
 155                 if (!page)
 156                         break;
 157
 158                 status = mapping->a_ops->prepare_write(NULL, page, offset,
 159                                                         offset + bytes);
 160                 if (status)
 161                         goto unlock;
 162
 163                 zero_user_page(page, offset, bytes, KM_USER0);
 164
 165                 status = mapping->a_ops->commit_write(NULL, page, offset,
 166                                                         offset + bytes);
 167                 if (!status) {
 168                         pos += bytes;
 169                         count -= bytes;
 170                 }
 171
 172 unlock:
 173                 unlock_page(page);
 174                 page_cache_release(page);
 175                 if (status)
 176                         break;
 177         } while (count);
 178
 179         return (-status);
 180 }
 181
 182 ssize_t                 /* bytes read, or (-)  error */
 183 xfs_read(
 184         xfs_inode_t             *ip,
 185         struct kiocb            *iocb,
 186         const struct iovec      *iovp,
 187         unsigned int            segs,
 188         loff_t                  *offset,
 189         int                     ioflags)
 190 {
 191         struct file             *file = iocb->ki_filp;
 192         struct inode            *inode = file->f_mapping->host;
 193         bhv_vnode_t             *vp = XFS_ITOV(ip);
 194         xfs_mount_t             *mp = ip->i_mount;
 195         size_t                  size = 0;
 196         ssize_t                 ret = 0;
 197         xfs_fsize_t             n;
 198         unsigned long           seg;
 199
 200
 201         XFS_STATS_INC(xs_read_calls);
 202
 203         /* START copy & waste from filemap.c */
 204         for (seg = 0; seg < segs; seg++) {
 205                 const struct iovec *iv = &iovp[seg];
 206
 207                 /*
 208                  * If any segment has a negative length, or the cumulative
 209                  * length ever wraps negative then return -EINVAL.
 210                  */
 211                 size += iv->iov_len;
 212                 if (unlikely((ssize_t)(size|iv->iov_len) < 0))
 213                         return XFS_ERROR(-EINVAL);
 214         }
 215         /* END copy & waste from filemap.c */
 216
 217         if (unlikely(ioflags & IO_ISDIRECT)) {
 218                 xfs_buftarg_t   *target =
 219                         (ip->i_d.di_flags & XFS_DIFLAG_REALTIME) ?
 220                                 mp->m_rtdev_targp : mp->m_ddev_targp;
 221                 if ((*offset & target->bt_smask) ||
 222                     (size & target->bt_smask)) {
 223                         if (*offset == ip->i_size) {
 224                                 return (0);
 225                         }
 226                         return -XFS_ERROR(EINVAL);
 227                 }
 228         }
 229
 230         n = XFS_MAXIOFFSET(mp) - *offset;
 231         if ((n <= 0) || (size == 0))
 232                 return 0;
 233
 234         if (n < size)
 235                 size = n;
 236
 237         if (XFS_FORCED_SHUTDOWN(mp))
 238                 return -EIO;
 239
 240         if (unlikely(ioflags & IO_ISDIRECT))
 241                 mutex_lock(&inode->i_mutex);
 242         xfs_ilock(ip, XFS_IOLOCK_SHARED);
 243
 244         if (DM_EVENT_ENABLED(ip, DM_EVENT_READ) && !(ioflags & IO_INVIS)) {
 245                 bhv_vrwlock_t locktype = VRWLOCK_READ;
 246                 int dmflags = FILP_DELAY_FLAG(file) | DM_SEM_FLAG_RD(ioflags);
 247
 248                 ret = -XFS_SEND_DATA(mp, DM_EVENT_READ, vp, *offset, size,
 249                                         dmflags, &locktype);
 250                 if (ret) {
 251                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 252                         if (unlikely(ioflags & IO_ISDIRECT))
 253                                 mutex_unlock(&inode->i_mutex);
 254                         return ret;
 255                 }
 256         }
 257
 258         if (unlikely(ioflags & IO_ISDIRECT)) {
 259                 if (VN_CACHED(vp))
 260                         ret = xfs_flushinval_pages(ip,
 261                                         ctooff(offtoct(*offset)),
 262                                         -1, FI_REMAPF_LOCKED);
 263                 mutex_unlock(&inode->i_mutex);
 264                 if (ret) {
 265                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 266                         return ret;
 267                 }
 268         }
 269
 270         xfs_rw_enter_trace(XFS_READ_ENTER, &ip->i_iocore,
 271                                 (void *)iovp, segs, *offset, ioflags);
 272
 273         iocb->ki_pos = *offset;
 274         ret = generic_file_aio_read(iocb, iovp, segs, *offset);
 275         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
 276                 ret = wait_on_sync_kiocb(iocb);
 277         if (ret > 0)
 278                 XFS_STATS_ADD(xs_read_bytes, ret);
 279
 280         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 281         return ret;
 282 }
 283
 284 ssize_t
 285 xfs_splice_read(
 286         xfs_inode_t             *ip,
 287         struct file             *infilp,
 288         loff_t                  *ppos,
 289         struct pipe_inode_info  *pipe,
 290         size_t                  count,
 291         int                     flags,
 292         int                     ioflags)
 293 {
 294         bhv_vnode_t             *vp = XFS_ITOV(ip);
 295         xfs_mount_t             *mp = ip->i_mount;
 296         ssize_t                 ret;
 297
 298         XFS_STATS_INC(xs_read_calls);
 299         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 300                 return -EIO;
 301
 302         xfs_ilock(ip, XFS_IOLOCK_SHARED);
 303
 304         if (DM_EVENT_ENABLED(ip, DM_EVENT_READ) && !(ioflags & IO_INVIS)) {
 305                 bhv_vrwlock_t locktype = VRWLOCK_READ;
 306                 int error;
 307
 308                 error = XFS_SEND_DATA(mp, DM_EVENT_READ, vp, *ppos, count,
 309                                         FILP_DELAY_FLAG(infilp), &locktype);
 310                 if (error) {
 311                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 312                         return -error;
 313                 }
 314         }
 315         xfs_rw_enter_trace(XFS_SPLICE_READ_ENTER, &ip->i_iocore,
 316                            pipe, count, *ppos, ioflags);
 317         ret = generic_file_splice_read(infilp, ppos, pipe, count, flags);
 318         if (ret > 0)
 319                 XFS_STATS_ADD(xs_read_bytes, ret);
 320
 321         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 322         return ret;
 323 }
 324
 325 ssize_t
 326 xfs_splice_write(
 327         xfs_inode_t             *ip,
 328         struct pipe_inode_info  *pipe,
 329         struct file             *outfilp,
 330         loff_t                  *ppos,
 331         size_t                  count,
 332         int                     flags,
 333         int                     ioflags)
 334 {
 335         bhv_vnode_t             *vp = XFS_ITOV(ip);
 336         xfs_mount_t             *mp = ip->i_mount;
 337         xfs_iocore_t            *io = &ip->i_iocore;
 338         ssize_t                 ret;
 339         struct inode            *inode = outfilp->f_mapping->host;
 340         xfs_fsize_t             isize, new_size;
 341
 342         XFS_STATS_INC(xs_write_calls);
 343         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 344                 return -EIO;
 345
 346         xfs_ilock(ip, XFS_IOLOCK_EXCL);
 347
 348         if (DM_EVENT_ENABLED(ip, DM_EVENT_WRITE) && !(ioflags & IO_INVIS)) {
 349                 bhv_vrwlock_t locktype = VRWLOCK_WRITE;
 350                 int error;
 351
 352                 error = XFS_SEND_DATA(mp, DM_EVENT_WRITE, vp, *ppos, count,
 353                                         FILP_DELAY_FLAG(outfilp), &locktype);
 354                 if (error) {
 355                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 356                         return -error;
 357                 }
 358         }
 359
 360         new_size = *ppos + count;
 361
 362         xfs_ilock(ip, XFS_ILOCK_EXCL);
 363         if (new_size > ip->i_size)
 364                 io->io_new_size = new_size;
 365         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 366
 367         xfs_rw_enter_trace(XFS_SPLICE_WRITE_ENTER, &ip->i_iocore,
 368                            pipe, count, *ppos, ioflags);
 369         ret = generic_file_splice_write(pipe, outfilp, ppos, count, flags);
 370         if (ret > 0)
 371                 XFS_STATS_ADD(xs_write_bytes, ret);
 372
 373         isize = i_size_read(inode);
 374         if (unlikely(ret < 0 && ret != -EFAULT && *ppos > isize))
 375                 *ppos = isize;
 376
 377         if (*ppos > ip->i_size) {
 378                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 379                 if (*ppos > ip->i_size)
 380                         ip->i_size = *ppos;
 381                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 382         }
 383
 384         if (io->io_new_size) {
 385                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 386                 io->io_new_size = 0;
 387                 if (ip->i_d.di_size > ip->i_size)
 388                         ip->i_d.di_size = ip->i_size;
 389                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 390         }
 391         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 392         return ret;
 393 }
 394
 395 /*
 396  * This routine is called to handle zeroing any space in the last
 397  * block of the file that is beyond the EOF.  We do this since the
 398  * size is being increased without writing anything to that block
 399  * and we don't want anyone to read the garbage on the disk.
 400  */
 401 STATIC int                              /* error (positive) */
 402 xfs_zero_last_block(
 403         struct inode    *ip,
 404         xfs_iocore_t    *io,
 405         xfs_fsize_t     offset,
 406         xfs_fsize_t     isize)
 407 {
 408         xfs_fileoff_t   last_fsb;
 409         xfs_mount_t     *mp = io->io_mount;
 410         int             nimaps;
 411         int             zero_offset;
 412         int             zero_len;
 413         int             error = 0;
 414         xfs_bmbt_irec_t imap;
 415
 416         ASSERT(ismrlocked(io->io_lock, MR_UPDATE) != 0);
 417
 418         zero_offset = XFS_B_FSB_OFFSET(mp, isize);
 419         if (zero_offset == 0) {
 420                 /*
 421                  * There are no extra bytes in the last block on disk to
 422                  * zero, so return.
 423                  */
 424                 return 0;
 425         }
 426
 427         last_fsb = XFS_B_TO_FSBT(mp, isize);
 428         nimaps = 1;
 429         error = XFS_BMAPI(mp, NULL, io, last_fsb, 1, 0, NULL, 0, &imap,
 430                           &nimaps, NULL, NULL);
 431         if (error) {
 432                 return error;
 433         }
 434         ASSERT(nimaps > 0);
 435         /*
 436          * If the block underlying isize is just a hole, then there
 437          * is nothing to zero.
 438          */
 439         if (imap.br_startblock == HOLESTARTBLOCK) {
 440                 return 0;
 441         }
 442         /*
 443          * Zero the part of the last block beyond the EOF, and write it
 444          * out sync.  We need to drop the ilock while we do this so we
 445          * don't deadlock when the buffer cache calls back to us.
 446          */
 447         XFS_IUNLOCK(mp, io, XFS_ILOCK_EXCL| XFS_EXTSIZE_RD);
 448
 449         zero_len = mp->m_sb.sb_blocksize - zero_offset;
 450         if (isize + zero_len > offset)
 451                 zero_len = offset - isize;
 452         error = xfs_iozero(ip, isize, zero_len);
 453
 454         XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 455         ASSERT(error >= 0);
 456         return error;
 457 }
 458
 459 /*
 460  * Zero any on disk space between the current EOF and the new,
 461  * larger EOF.  This handles the normal case of zeroing the remainder
 462  * of the last block in the file and the unusual case of zeroing blocks
 463  * out beyond the size of the file.  This second case only happens
 464  * with fixed size extents and when the system crashes before the inode
 465  * size was updated but after blocks were allocated.  If fill is set,
 466  * then any holes in the range are filled and zeroed.  If not, the holes
 467  * are left alone as holes.
 468  */
 469
 470 int                                     /* error (positive) */
 471 xfs_zero_eof(
 472         bhv_vnode_t     *vp,
 473         xfs_iocore_t    *io,
 474         xfs_off_t       offset,         /* starting I/O offset */
 475         xfs_fsize_t     isize)          /* current inode size */
 476 {
 477         struct inode    *ip = vn_to_inode(vp);
 478         xfs_fileoff_t   start_zero_fsb;
 479         xfs_fileoff_t   end_zero_fsb;
 480         xfs_fileoff_t   zero_count_fsb;
 481         xfs_fileoff_t   last_fsb;
 482         xfs_fileoff_t   zero_off;
 483         xfs_fsize_t     zero_len;
 484         xfs_mount_t     *mp = io->io_mount;
 485         int             nimaps;
 486         int             error = 0;
 487         xfs_bmbt_irec_t imap;
 488
 489         ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
 490         ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
 491         ASSERT(offset > isize);
 492
 493         /*
 494          * First handle zeroing the block on which isize resides.
 495          * We only zero a part of that block so it is handled specially.
 496          */
 497         error = xfs_zero_last_block(ip, io, offset, isize);
 498         if (error) {
 499                 ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
 500                 ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
 501                 return error;
 502         }
 503
 504         /*
 505          * Calculate the range between the new size and the old
 506          * where blocks needing to be zeroed may exist.  To get the
 507          * block where the last byte in the file currently resides,
 508          * we need to subtract one from the size and truncate back
 509          * to a block boundary.  We subtract 1 in case the size is
 510          * exactly on a block boundary.
 511          */
 512         last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
 513         start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
 514         end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
 515         ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
 516         if (last_fsb == end_zero_fsb) {
 517                 /*
 518                  * The size was only incremented on its last block.
 519                  * We took care of that above, so just return.
 520                  */
 521                 return 0;
 522         }
 523
 524         ASSERT(start_zero_fsb <= end_zero_fsb);
 525         while (start_zero_fsb <= end_zero_fsb) {
 526                 nimaps = 1;
 527                 zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
 528                 error = XFS_BMAPI(mp, NULL, io, start_zero_fsb, zero_count_fsb,
 529                                   0, NULL, 0, &imap, &nimaps, NULL, NULL);
 530                 if (error) {
 531                         ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
 532                         ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
 533                         return error;
 534                 }
 535                 ASSERT(nimaps > 0);
 536
 537                 if (imap.br_state == XFS_EXT_UNWRITTEN ||
 538                     imap.br_startblock == HOLESTARTBLOCK) {
 539                         /*
 540                          * This loop handles initializing pages that were
 541                          * partially initialized by the code below this
 542                          * loop. It basically zeroes the part of the page
 543                          * that sits on a hole and sets the page as P_HOLE
 544                          * and calls remapf if it is a mapped file.
 545                          */
 546                         start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 547                         ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 548                         continue;
 549                 }
 550
 551                 /*
 552                  * There are blocks we need to zero.
 553                  * Drop the inode lock while we're doing the I/O.
 554                  * We'll still have the iolock to protect us.
 555                  */
 556                 XFS_IUNLOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 557
 558                 zero_off = XFS_FSB_TO_B(mp, start_zero_fsb);
 559                 zero_len = XFS_FSB_TO_B(mp, imap.br_blockcount);
 560
 561                 if ((zero_off + zero_len) > offset)
 562                         zero_len = offset - zero_off;
 563
 564                 error = xfs_iozero(ip, zero_off, zero_len);
 565                 if (error) {
 566                         goto out_lock;
 567                 }
 568
 569                 start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 570                 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 571
 572                 XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 573         }
 574
 575         return 0;
 576
 577 out_lock:
 578
 579         XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 580         ASSERT(error >= 0);
 581         return error;
 582 }
 583
 584 ssize_t                         /* bytes written, or (-) error */
 585 xfs_write(
 586         struct xfs_inode        *xip,
 587         struct kiocb            *iocb,
 588         const struct iovec      *iovp,
 589         unsigned int            nsegs,
 590         loff_t                  *offset,
 591         int                     ioflags)
 592 {
 593         struct file             *file = iocb->ki_filp;
 594         struct address_space    *mapping = file->f_mapping;
 595         struct inode            *inode = mapping->host;
 596         bhv_vnode_t             *vp = XFS_ITOV(xip);
 597         unsigned long           segs = nsegs;
 598         xfs_mount_t             *mp;
 599         ssize_t                 ret = 0, error = 0;
 600         xfs_fsize_t             isize, new_size;
 601         xfs_iocore_t            *io;
 602         int                     iolock;
 603         int                     eventsent = 0;
 604         bhv_vrwlock_t           locktype;
 605         size_t                  ocount = 0, count;
 606         loff_t                  pos;
 607         int                     need_i_mutex;
 608
 609         XFS_STATS_INC(xs_write_calls);
 610
 611         error = generic_segment_checks(iovp, &segs, &ocount, VERIFY_READ);
 612         if (error)
 613                 return error;
 614
 615         count = ocount;
 616         pos = *offset;
 617
 618         if (count == 0)
 619                 return 0;
 620
 621         io = &xip->i_iocore;
 622         mp = io->io_mount;
 623
 624         xfs_wait_for_freeze(mp, SB_FREEZE_WRITE);
 625
 626         if (XFS_FORCED_SHUTDOWN(mp))
 627                 return -EIO;
 628
 629 relock:
 630         if (ioflags & IO_ISDIRECT) {
 631                 iolock = XFS_IOLOCK_SHARED;
 632                 locktype = VRWLOCK_WRITE_DIRECT;
 633                 need_i_mutex = 0;
 634         } else {
 635                 iolock = XFS_IOLOCK_EXCL;
 636                 locktype = VRWLOCK_WRITE;
 637                 need_i_mutex = 1;
 638                 mutex_lock(&inode->i_mutex);
 639         }
 640
 641         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
 642
 643 start:
 644         error = -generic_write_checks(file, &pos, &count,
 645                                         S_ISBLK(inode->i_mode));
 646         if (error) {
 647                 xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 648                 goto out_unlock_mutex;
 649         }
 650
 651         if ((DM_EVENT_ENABLED(xip, DM_EVENT_WRITE) &&
 652             !(ioflags & IO_INVIS) && !eventsent)) {
 653                 int             dmflags = FILP_DELAY_FLAG(file);
 654
 655                 if (need_i_mutex)
 656                         dmflags |= DM_FLAGS_IMUX;
 657
 658                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 659                 error = XFS_SEND_DATA(xip->i_mount, DM_EVENT_WRITE, vp,
 660                                       pos, count,
 661                                       dmflags, &locktype);
 662                 if (error) {
 663                         goto out_unlock_internal;
 664                 }
 665                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 666                 eventsent = 1;
 667
 668                 /*
 669                  * The iolock was dropped and reacquired in XFS_SEND_DATA
 670                  * so we have to recheck the size when appending.
 671                  * We will only "goto start;" once, since having sent the
 672                  * event prevents another call to XFS_SEND_DATA, which is
 673                  * what allows the size to change in the first place.
 674                  */
 675                 if ((file->f_flags & O_APPEND) && pos != xip->i_size)
 676                         goto start;
 677         }
 678
 679         if (ioflags & IO_ISDIRECT) {
 680                 xfs_buftarg_t   *target =
 681                         (xip->i_d.di_flags & XFS_DIFLAG_REALTIME) ?
 682                                 mp->m_rtdev_targp : mp->m_ddev_targp;
 683
 684                 if ((pos & target->bt_smask) || (count & target->bt_smask)) {
 685                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 686                         return XFS_ERROR(-EINVAL);
 687                 }
 688
 689                 if (!need_i_mutex && (VN_CACHED(vp) || pos > xip->i_size)) {
 690                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 691                         iolock = XFS_IOLOCK_EXCL;
 692                         locktype = VRWLOCK_WRITE;
 693                         need_i_mutex = 1;
 694                         mutex_lock(&inode->i_mutex);
 695                         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
 696                         goto start;
 697                 }
 698         }
 699
 700         new_size = pos + count;
 701         if (new_size > xip->i_size)
 702                 io->io_new_size = new_size;
 703
 704         if (likely(!(ioflags & IO_INVIS))) {
 705                 file_update_time(file);
 706                 xfs_ichgtime_fast(xip, inode,
 707                                   XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
 708         }
 709
 710         /*
 711          * If the offset is beyond the size of the file, we have a couple
 712          * of things to do. First, if there is already space allocated
 713          * we need to either create holes or zero the disk or ...
 714          *
 715          * If there is a page where the previous size lands, we need
 716          * to zero it out up to the new size.
 717          */
 718
 719         if (pos > xip->i_size) {
 720                 error = xfs_zero_eof(vp, io, pos, xip->i_size);
 721                 if (error) {
 722                         xfs_iunlock(xip, XFS_ILOCK_EXCL);
 723                         goto out_unlock_internal;
 724                 }
 725         }
 726         xfs_iunlock(xip, XFS_ILOCK_EXCL);
 727
 728         /*
 729          * If we're writing the file then make sure to clear the
 730          * setuid and setgid bits if the process is not being run
 731          * by root.  This keeps people from modifying setuid and
 732          * setgid binaries.
 733          */
 734
 735         if (((xip->i_d.di_mode & S_ISUID) ||
 736             ((xip->i_d.di_mode & (S_ISGID | S_IXGRP)) ==
 737                 (S_ISGID | S_IXGRP))) &&
 738              !capable(CAP_FSETID)) {
 739                 error = xfs_write_clear_setuid(xip);
 740                 if (likely(!error))
 741                         error = -remove_suid(file->f_path.dentry);
 742                 if (unlikely(error)) {
 743                         goto out_unlock_internal;
 744                 }
 745         }
 746
 747 retry:
 748         /* We can write back this queue in page reclaim */
 749         current->backing_dev_info = mapping->backing_dev_info;
 750
 751         if ((ioflags & IO_ISDIRECT)) {
 752                 if (VN_CACHED(vp)) {
 753                         WARN_ON(need_i_mutex == 0);
 754                         xfs_inval_cached_trace(io, pos, -1,
 755                                         ctooff(offtoct(pos)), -1);
 756                         error = xfs_flushinval_pages(xip,
 757                                         ctooff(offtoct(pos)),
 758                                         -1, FI_REMAPF_LOCKED);
 759                         if (error)
 760                                 goto out_unlock_internal;
 761                 }
 762
 763                 if (need_i_mutex) {
 764                         /* demote the lock now the cached pages are gone */
 765                         XFS_ILOCK_DEMOTE(mp, io, XFS_IOLOCK_EXCL);
 766                         mutex_unlock(&inode->i_mutex);
 767
 768                         iolock = XFS_IOLOCK_SHARED;
 769                         locktype = VRWLOCK_WRITE_DIRECT;
 770                         need_i_mutex = 0;
 771                 }
 772
 773                 xfs_rw_enter_trace(XFS_DIOWR_ENTER, io, (void *)iovp, segs,
 774                                 *offset, ioflags);
 775                 ret = generic_file_direct_write(iocb, iovp,
 776                                 &segs, pos, offset, count, ocount);
 777
 778                 /*
 779                  * direct-io write to a hole: fall through to buffered I/O
 780                  * for completing the rest of the request.
 781                  */
 782                 if (ret >= 0 && ret != count) {
 783                         XFS_STATS_ADD(xs_write_bytes, ret);
 784
 785                         pos += ret;
 786                         count -= ret;
 787
 788                         ioflags &= ~IO_ISDIRECT;
 789                         xfs_iunlock(xip, iolock);
 790                         goto relock;
 791                 }
 792         } else {
 793                 xfs_rw_enter_trace(XFS_WRITE_ENTER, io, (void *)iovp, segs,
 794                                 *offset, ioflags);
 795                 ret = generic_file_buffered_write(iocb, iovp, segs,
 796                                 pos, offset, count, ret);
 797         }
 798
 799         current->backing_dev_info = NULL;
 800
 801         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
 802                 ret = wait_on_sync_kiocb(iocb);
 803
 804         if (ret == -ENOSPC &&
 805             DM_EVENT_ENABLED(xip, DM_EVENT_NOSPACE) && !(ioflags & IO_INVIS)) {
 806                 xfs_rwunlock(xip, locktype);
 807                 if (need_i_mutex)
 808                         mutex_unlock(&inode->i_mutex);
 809                 error = XFS_SEND_NAMESP(xip->i_mount, DM_EVENT_NOSPACE, vp,
 810                                 DM_RIGHT_NULL, vp, DM_RIGHT_NULL, NULL, NULL,
 811                                 0, 0, 0); /* Delay flag intentionally  unused */
 812                 if (need_i_mutex)
 813                         mutex_lock(&inode->i_mutex);
 814                 xfs_rwlock(xip, locktype);
 815                 if (error)
 816                         goto out_unlock_internal;
 817                 pos = xip->i_size;
 818                 ret = 0;
 819                 goto retry;
 820         }
 821
 822         isize = i_size_read(inode);
 823         if (unlikely(ret < 0 && ret != -EFAULT && *offset > isize))
 824                 *offset = isize;
 825
 826         if (*offset > xip->i_size) {
 827                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 828                 if (*offset > xip->i_size)
 829                         xip->i_size = *offset;
 830                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 831         }
 832
 833         error = -ret;
 834         if (ret <= 0)
 835                 goto out_unlock_internal;
 836
 837         XFS_STATS_ADD(xs_write_bytes, ret);
 838
 839         /* Handle various SYNC-type writes */
 840         if ((file->f_flags & O_SYNC) || IS_SYNC(inode)) {
 841                 error = xfs_write_sync_logforce(mp, xip);
 842                 if (error)
 843                         goto out_unlock_internal;
 844
 845                 xfs_rwunlock(xip, locktype);
 846                 if (need_i_mutex)
 847                         mutex_unlock(&inode->i_mutex);
 848
 849                 error = sync_page_range(inode, mapping, pos, ret);
 850                 if (!error)
 851                         error = -ret;
 852                 if (need_i_mutex)
 853                         mutex_lock(&inode->i_mutex);
 854                 xfs_rwlock(xip, locktype);
 855         }
 856
 857  out_unlock_internal:
 858         if (io->io_new_size) {
 859                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 860                 io->io_new_size = 0;
 861                 /*
 862                  * If this was a direct or synchronous I/O that failed (such
 863                  * as ENOSPC) then part of the I/O may have been written to
 864                  * disk before the error occured.  In this case the on-disk
 865                  * file size may have been adjusted beyond the in-memory file
 866                  * size and now needs to be truncated back.
 867                  */
 868                 if (xip->i_d.di_size > xip->i_size)
 869                         xip->i_d.di_size = xip->i_size;
 870                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 871         }
 872         xfs_rwunlock(xip, locktype);
 873  out_unlock_mutex:
 874         if (need_i_mutex)
 875                 mutex_unlock(&inode->i_mutex);
 876         return -error;
 877 }
 878
 879 /*
 880  * All xfs metadata buffers except log state machine buffers
 881  * get this attached as their b_bdstrat callback function.
 882  * This is so that we can catch a buffer
 883  * after prematurely unpinning it to forcibly shutdown the filesystem.
 884  */
 885 int
 886 xfs_bdstrat_cb(struct xfs_buf *bp)
 887 {
 888         xfs_mount_t     *mp;
 889
 890         mp = XFS_BUF_FSPRIVATE3(bp, xfs_mount_t *);
 891         if (!XFS_FORCED_SHUTDOWN(mp)) {
 892                 xfs_buf_iorequest(bp);
 893                 return 0;
 894         } else {
 895                 xfs_buftrace("XFS__BDSTRAT IOERROR", bp);
 896                 /*
 897                  * Metadata write that didn't get logged but
 898                  * written delayed anyway. These aren't associated
 899                  * with a transaction, and can be ignored.
 900                  */
 901                 if (XFS_BUF_IODONE_FUNC(bp) == NULL &&
 902                     (XFS_BUF_ISREAD(bp)) == 0)
 903                         return (xfs_bioerror_relse(bp));
 904                 else
 905                         return (xfs_bioerror(bp));
 906         }
 907 }
 908
 909
 910 int
 911 xfs_bmap(
 912         xfs_inode_t     *ip,
 913         xfs_off_t       offset,
 914         ssize_t         count,
 915         int             flags,
 916         xfs_iomap_t     *iomapp,
 917         int             *niomaps)
 918 {
 919         xfs_iocore_t    *io = &ip->i_iocore;
 920
 921         ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFREG);
 922         ASSERT(((ip->i_d.di_flags & XFS_DIFLAG_REALTIME) != 0) ==
 923                ((ip->i_iocore.io_flags & XFS_IOCORE_RT) != 0));
 924
 925         return xfs_iomap(io, offset, count, flags, iomapp, niomaps);
 926 }
 927
 928 /*
 929  * Wrapper around bdstrat so that we can stop data
 930  * from going to disk in case we are shutting down the filesystem.
 931  * Typically user data goes thru this path; one of the exceptions
 932  * is the superblock.
 933  */
 934 int
 935 xfsbdstrat(
 936         struct xfs_mount        *mp,
 937         struct xfs_buf          *bp)
 938 {
 939         ASSERT(mp);
 940         if (!XFS_FORCED_SHUTDOWN(mp)) {
 941                 /* Grio redirection would go here
 942                  * if (XFS_BUF_IS_GRIO(bp)) {
 943                  */
 944
 945                 xfs_buf_iorequest(bp);
 946                 return 0;
 947         }
 948
 949         xfs_buftrace("XFSBDSTRAT IOERROR", bp);
 950         return (xfs_bioerror_relse(bp));
 951 }
 952
 953 /*
 954  * If the underlying (data/log/rt) device is readonly, there are some
 955  * operations that cannot proceed.
 956  */
 957 int
 958 xfs_dev_is_read_only(
 959         xfs_mount_t             *mp,
 960         char                    *message)
 961 {
 962         if (xfs_readonly_buftarg(mp->m_ddev_targp) ||
 963             xfs_readonly_buftarg(mp->m_logdev_targp) ||
 964             (mp->m_rtdev_targp && xfs_readonly_buftarg(mp->m_rtdev_targp))) {
 965                 cmn_err(CE_NOTE,
 966                         "XFS: %s required on read-only device.", message);
 967                 cmn_err(CE_NOTE,
 968                         "XFS: write access unavailable, cannot proceed.");
 969                 return EROFS;
 970         }
 971         return 0;
 972 }