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