git.oblomov.eu 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
  52 #include <linux/capability.h>
  53 #include <linux/writeback.h>
  54
  55
  56 #if defined(XFS_RW_TRACE)
  57 void
  58 xfs_rw_enter_trace(
  59         int                     tag,
  60         xfs_iocore_t            *io,
  61         void                    *data,
  62         size_t                  segs,
  63         loff_t                  offset,
  64         int                     ioflags)
  65 {
  66         xfs_inode_t     *ip = XFS_IO_INODE(io);
  67
  68         if (ip->i_rwtrace == NULL)
  69                 return;
  70         ktrace_enter(ip->i_rwtrace,
  71                 (void *)(unsigned long)tag,
  72                 (void *)ip,
  73                 (void *)((unsigned long)((ip->i_d.di_size >> 32) & 0xffffffff)),
  74                 (void *)((unsigned long)(ip->i_d.di_size & 0xffffffff)),
  75                 (void *)data,
  76                 (void *)((unsigned long)segs),
  77                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
  78                 (void *)((unsigned long)(offset & 0xffffffff)),
  79                 (void *)((unsigned long)ioflags),
  80                 (void *)((unsigned long)((io->io_new_size >> 32) & 0xffffffff)),
  81                 (void *)((unsigned long)(io->io_new_size & 0xffffffff)),
  82                 (void *)((unsigned long)current_pid()),
  83                 (void *)NULL,
  84                 (void *)NULL,
  85                 (void *)NULL,
  86                 (void *)NULL);
  87 }
  88
  89 void
  90 xfs_inval_cached_trace(
  91         xfs_iocore_t    *io,
  92         xfs_off_t       offset,
  93         xfs_off_t       len,
  94         xfs_off_t       first,
  95         xfs_off_t       last)
  96 {
  97         xfs_inode_t     *ip = XFS_IO_INODE(io);
  98
  99         if (ip->i_rwtrace == NULL)
 100                 return;
 101         ktrace_enter(ip->i_rwtrace,
 102                 (void *)(__psint_t)XFS_INVAL_CACHED,
 103                 (void *)ip,
 104                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
 105                 (void *)((unsigned long)(offset & 0xffffffff)),
 106                 (void *)((unsigned long)((len >> 32) & 0xffffffff)),
 107                 (void *)((unsigned long)(len & 0xffffffff)),
 108                 (void *)((unsigned long)((first >> 32) & 0xffffffff)),
 109                 (void *)((unsigned long)(first & 0xffffffff)),
 110                 (void *)((unsigned long)((last >> 32) & 0xffffffff)),
 111                 (void *)((unsigned long)(last & 0xffffffff)),
 112                 (void *)((unsigned long)current_pid()),
 113                 (void *)NULL,
 114                 (void *)NULL,
 115                 (void *)NULL,
 116                 (void *)NULL,
 117                 (void *)NULL);
 118 }
 119 #endif
 120
 121 /*
 122  *      xfs_iozero
 123  *
 124  *      xfs_iozero clears the specified range of buffer supplied,
 125  *      and marks all the affected blocks as valid and modified.  If
 126  *      an affected block is not allocated, it will be allocated.  If
 127  *      an affected block is not completely overwritten, and is not
 128  *      valid before the operation, it will be read from disk before
 129  *      being partially zeroed.
 130  */
 131 STATIC int
 132 xfs_iozero(
 133         struct inode            *ip,    /* inode                        */
 134         loff_t                  pos,    /* offset in file               */
 135         size_t                  count)  /* size of data to zero         */
 136 {
 137         unsigned                bytes;
 138         struct page             *page;
 139         struct address_space    *mapping;
 140         int                     status;
 141
 142         mapping = ip->i_mapping;
 143         do {
 144                 unsigned long index, offset;
 145
 146                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
 147                 index = pos >> PAGE_CACHE_SHIFT;
 148                 bytes = PAGE_CACHE_SIZE - offset;
 149                 if (bytes > count)
 150                         bytes = count;
 151
 152                 status = -ENOMEM;
 153                 page = grab_cache_page(mapping, index);
 154                 if (!page)
 155                         break;
 156
 157                 status = mapping->a_ops->prepare_write(NULL, page, offset,
 158                                                         offset + bytes);
 159                 if (status)
 160                         goto unlock;
 161
 162                 memclear_highpage_flush(page, offset, bytes);
 163
 164                 status = mapping->a_ops->commit_write(NULL, page, offset,
 165                                                         offset + bytes);
 166                 if (!status) {
 167                         pos += bytes;
 168                         count -= bytes;
 169                 }
 170
 171 unlock:
 172                 unlock_page(page);
 173                 page_cache_release(page);
 174                 if (status)
 175                         break;
 176         } while (count);
 177
 178         return (-status);
 179 }
 180
 181 ssize_t                 /* bytes read, or (-)  error */
 182 xfs_read(
 183         bhv_desc_t              *bdp,
 184         struct kiocb            *iocb,
 185         const struct iovec      *iovp,
 186         unsigned int            segs,
 187         loff_t                  *offset,
 188         int                     ioflags,
 189         cred_t                  *credp)
 190 {
 191         struct file             *file = iocb->ki_filp;
 192         struct inode            *inode = file->f_mapping->host;
 193         size_t                  size = 0;
 194         ssize_t                 ret;
 195         xfs_fsize_t             n;
 196         xfs_inode_t             *ip;
 197         xfs_mount_t             *mp;
 198         bhv_vnode_t             *vp;
 199         unsigned long           seg;
 200
 201         ip = XFS_BHVTOI(bdp);
 202         vp = BHV_TO_VNODE(bdp);
 203         mp = ip->i_mount;
 204
 205         XFS_STATS_INC(xs_read_calls);
 206
 207         /* START copy & waste from filemap.c */
 208         for (seg = 0; seg < segs; seg++) {
 209                 const struct iovec *iv = &iovp[seg];
 210
 211                 /*
 212                  * If any segment has a negative length, or the cumulative
 213                  * length ever wraps negative then return -EINVAL.
 214                  */
 215                 size += iv->iov_len;
 216                 if (unlikely((ssize_t)(size|iv->iov_len) < 0))
 217                         return XFS_ERROR(-EINVAL);
 218         }
 219         /* END copy & waste from filemap.c */
 220
 221         if (unlikely(ioflags & IO_ISDIRECT)) {
 222                 xfs_buftarg_t   *target =
 223                         (ip->i_d.di_flags & XFS_DIFLAG_REALTIME) ?
 224                                 mp->m_rtdev_targp : mp->m_ddev_targp;
 225                 if ((*offset & target->bt_smask) ||
 226                     (size & target->bt_smask)) {
 227                         if (*offset == ip->i_d.di_size) {
 228                                 return (0);
 229                         }
 230                         return -XFS_ERROR(EINVAL);
 231                 }
 232         }
 233
 234         n = XFS_MAXIOFFSET(mp) - *offset;
 235         if ((n <= 0) || (size == 0))
 236                 return 0;
 237
 238         if (n < size)
 239                 size = n;
 240
 241         if (XFS_FORCED_SHUTDOWN(mp))
 242                 return -EIO;
 243
 244         if (unlikely(ioflags & IO_ISDIRECT))
 245                 mutex_lock(&inode->i_mutex);
 246         xfs_ilock(ip, XFS_IOLOCK_SHARED);
 247
 248         if (DM_EVENT_ENABLED(vp->v_vfsp, ip, DM_EVENT_READ) &&
 249             !(ioflags & IO_INVIS)) {
 250                 bhv_vrwlock_t locktype = VRWLOCK_READ;
 251                 int dmflags = FILP_DELAY_FLAG(file) | DM_SEM_FLAG_RD(ioflags);
 252
 253                 ret = -XFS_SEND_DATA(mp, DM_EVENT_READ,
 254                                         BHV_TO_VNODE(bdp), *offset, size,
 255                                         dmflags, &locktype);
 256                 if (ret) {
 257                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 258                         if (unlikely(ioflags & IO_ISDIRECT))
 259                                 mutex_unlock(&inode->i_mutex);
 260                         return ret;
 261                 }
 262         }
 263
 264         if (unlikely(ioflags & IO_ISDIRECT)) {
 265                 if (VN_CACHED(vp))
 266                         bhv_vop_flushinval_pages(vp, ctooff(offtoct(*offset)),
 267                                                  -1, FI_REMAPF_LOCKED);
 268                 mutex_unlock(&inode->i_mutex);
 269         }
 270
 271         xfs_rw_enter_trace(XFS_READ_ENTER, &ip->i_iocore,
 272                                 (void *)iovp, segs, *offset, ioflags);
 273
 274         iocb->ki_pos = *offset;
 275         ret = generic_file_aio_read(iocb, iovp, segs, *offset);
 276         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
 277                 ret = wait_on_sync_kiocb(iocb);
 278         if (ret > 0)
 279                 XFS_STATS_ADD(xs_read_bytes, ret);
 280
 281         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 282         return ret;
 283 }
 284
 285 ssize_t
 286 xfs_sendfile(
 287         bhv_desc_t              *bdp,
 288         struct file             *filp,
 289         loff_t                  *offset,
 290         int                     ioflags,
 291         size_t                  count,
 292         read_actor_t            actor,
 293         void                    *target,
 294         cred_t                  *credp)
 295 {
 296         xfs_inode_t             *ip = XFS_BHVTOI(bdp);
 297         xfs_mount_t             *mp = ip->i_mount;
 298         ssize_t                 ret;
 299
 300         XFS_STATS_INC(xs_read_calls);
 301         if (XFS_FORCED_SHUTDOWN(mp))
 302                 return -EIO;
 303
 304         xfs_ilock(ip, XFS_IOLOCK_SHARED);
 305
 306         if (DM_EVENT_ENABLED(BHV_TO_VNODE(bdp)->v_vfsp, ip, DM_EVENT_READ) &&
 307             (!(ioflags & IO_INVIS))) {
 308                 bhv_vrwlock_t locktype = VRWLOCK_READ;
 309                 int error;
 310
 311                 error = XFS_SEND_DATA(mp, DM_EVENT_READ, BHV_TO_VNODE(bdp),
 312                                       *offset, count,
 313                                       FILP_DELAY_FLAG(filp), &locktype);
 314                 if (error) {
 315                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 316                         return -error;
 317                 }
 318         }
 319         xfs_rw_enter_trace(XFS_SENDFILE_ENTER, &ip->i_iocore,
 320                    (void *)(unsigned long)target, count, *offset, ioflags);
 321         ret = generic_file_sendfile(filp, offset, count, actor, target);
 322         if (ret > 0)
 323                 XFS_STATS_ADD(xs_read_bytes, ret);
 324
 325         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 326         return ret;
 327 }
 328
 329 ssize_t
 330 xfs_splice_read(
 331         bhv_desc_t              *bdp,
 332         struct file             *infilp,
 333         loff_t                  *ppos,
 334         struct pipe_inode_info  *pipe,
 335         size_t                  count,
 336         int                     flags,
 337         int                     ioflags,
 338         cred_t                  *credp)
 339 {
 340         xfs_inode_t             *ip = XFS_BHVTOI(bdp);
 341         xfs_mount_t             *mp = ip->i_mount;
 342         ssize_t                 ret;
 343
 344         XFS_STATS_INC(xs_read_calls);
 345         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 346                 return -EIO;
 347
 348         xfs_ilock(ip, XFS_IOLOCK_SHARED);
 349
 350         if (DM_EVENT_ENABLED(BHV_TO_VNODE(bdp)->v_vfsp, ip, DM_EVENT_READ) &&
 351             (!(ioflags & IO_INVIS))) {
 352                 bhv_vrwlock_t locktype = VRWLOCK_READ;
 353                 int error;
 354
 355                 error = XFS_SEND_DATA(mp, DM_EVENT_READ, BHV_TO_VNODE(bdp),
 356                                         *ppos, count,
 357                                         FILP_DELAY_FLAG(infilp), &locktype);
 358                 if (error) {
 359                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 360                         return -error;
 361                 }
 362         }
 363         xfs_rw_enter_trace(XFS_SPLICE_READ_ENTER, &ip->i_iocore,
 364                            pipe, count, *ppos, ioflags);
 365         ret = generic_file_splice_read(infilp, ppos, pipe, count, flags);
 366         if (ret > 0)
 367                 XFS_STATS_ADD(xs_read_bytes, ret);
 368
 369         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 370         return ret;
 371 }
 372
 373 ssize_t
 374 xfs_splice_write(
 375         bhv_desc_t              *bdp,
 376         struct pipe_inode_info  *pipe,
 377         struct file             *outfilp,
 378         loff_t                  *ppos,
 379         size_t                  count,
 380         int                     flags,
 381         int                     ioflags,
 382         cred_t                  *credp)
 383 {
 384         xfs_inode_t             *ip = XFS_BHVTOI(bdp);
 385         xfs_mount_t             *mp = ip->i_mount;
 386         ssize_t                 ret;
 387         struct inode            *inode = outfilp->f_mapping->host;
 388         xfs_fsize_t             isize;
 389
 390         XFS_STATS_INC(xs_write_calls);
 391         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 392                 return -EIO;
 393
 394         xfs_ilock(ip, XFS_IOLOCK_EXCL);
 395
 396         if (DM_EVENT_ENABLED(BHV_TO_VNODE(bdp)->v_vfsp, ip, DM_EVENT_WRITE) &&
 397             (!(ioflags & IO_INVIS))) {
 398                 bhv_vrwlock_t locktype = VRWLOCK_WRITE;
 399                 int error;
 400
 401                 error = XFS_SEND_DATA(mp, DM_EVENT_WRITE, BHV_TO_VNODE(bdp),
 402                                         *ppos, count,
 403                                         FILP_DELAY_FLAG(outfilp), &locktype);
 404                 if (error) {
 405                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 406                         return -error;
 407                 }
 408         }
 409         xfs_rw_enter_trace(XFS_SPLICE_WRITE_ENTER, &ip->i_iocore,
 410                            pipe, count, *ppos, ioflags);
 411         ret = generic_file_splice_write(pipe, outfilp, ppos, count, flags);
 412         if (ret > 0)
 413                 XFS_STATS_ADD(xs_write_bytes, ret);
 414
 415         isize = i_size_read(inode);
 416         if (unlikely(ret < 0 && ret != -EFAULT && *ppos > isize))
 417                 *ppos = isize;
 418
 419         if (*ppos > ip->i_d.di_size) {
 420                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 421                 if (*ppos > ip->i_d.di_size) {
 422                         ip->i_d.di_size = *ppos;
 423                         i_size_write(inode, *ppos);
 424                         ip->i_update_core = 1;
 425                         ip->i_update_size = 1;
 426                 }
 427                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 428         }
 429         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 430         return ret;
 431 }
 432
 433 /*
 434  * This routine is called to handle zeroing any space in the last
 435  * block of the file that is beyond the EOF.  We do this since the
 436  * size is being increased without writing anything to that block
 437  * and we don't want anyone to read the garbage on the disk.
 438  */
 439 STATIC int                              /* error (positive) */
 440 xfs_zero_last_block(
 441         struct inode    *ip,
 442         xfs_iocore_t    *io,
 443         xfs_fsize_t     offset,
 444         xfs_fsize_t     isize)
 445 {
 446         xfs_fileoff_t   last_fsb;
 447         xfs_mount_t     *mp = io->io_mount;
 448         int             nimaps;
 449         int             zero_offset;
 450         int             zero_len;
 451         int             error = 0;
 452         xfs_bmbt_irec_t imap;
 453
 454         ASSERT(ismrlocked(io->io_lock, MR_UPDATE) != 0);
 455
 456         zero_offset = XFS_B_FSB_OFFSET(mp, isize);
 457         if (zero_offset == 0) {
 458                 /*
 459                  * There are no extra bytes in the last block on disk to
 460                  * zero, so return.
 461                  */
 462                 return 0;
 463         }
 464
 465         last_fsb = XFS_B_TO_FSBT(mp, isize);
 466         nimaps = 1;
 467         error = XFS_BMAPI(mp, NULL, io, last_fsb, 1, 0, NULL, 0, &imap,
 468                           &nimaps, NULL, NULL);
 469         if (error) {
 470                 return error;
 471         }
 472         ASSERT(nimaps > 0);
 473         /*
 474          * If the block underlying isize is just a hole, then there
 475          * is nothing to zero.
 476          */
 477         if (imap.br_startblock == HOLESTARTBLOCK) {
 478                 return 0;
 479         }
 480         /*
 481          * Zero the part of the last block beyond the EOF, and write it
 482          * out sync.  We need to drop the ilock while we do this so we
 483          * don't deadlock when the buffer cache calls back to us.
 484          */
 485         XFS_IUNLOCK(mp, io, XFS_ILOCK_EXCL| XFS_EXTSIZE_RD);
 486
 487         zero_len = mp->m_sb.sb_blocksize - zero_offset;
 488         if (isize + zero_len > offset)
 489                 zero_len = offset - isize;
 490         error = xfs_iozero(ip, isize, zero_len);
 491
 492         XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 493         ASSERT(error >= 0);
 494         return error;
 495 }
 496
 497 /*
 498  * Zero any on disk space between the current EOF and the new,
 499  * larger EOF.  This handles the normal case of zeroing the remainder
 500  * of the last block in the file and the unusual case of zeroing blocks
 501  * out beyond the size of the file.  This second case only happens
 502  * with fixed size extents and when the system crashes before the inode
 503  * size was updated but after blocks were allocated.  If fill is set,
 504  * then any holes in the range are filled and zeroed.  If not, the holes
 505  * are left alone as holes.
 506  */
 507
 508 int                                     /* error (positive) */
 509 xfs_zero_eof(
 510         bhv_vnode_t     *vp,
 511         xfs_iocore_t    *io,
 512         xfs_off_t       offset,         /* starting I/O offset */
 513         xfs_fsize_t     isize)          /* current inode size */
 514 {
 515         struct inode    *ip = vn_to_inode(vp);
 516         xfs_fileoff_t   start_zero_fsb;
 517         xfs_fileoff_t   end_zero_fsb;
 518         xfs_fileoff_t   zero_count_fsb;
 519         xfs_fileoff_t   last_fsb;
 520         xfs_fileoff_t   zero_off;
 521         xfs_fsize_t     zero_len;
 522         xfs_mount_t     *mp = io->io_mount;
 523         int             nimaps;
 524         int             error = 0;
 525         xfs_bmbt_irec_t imap;
 526
 527         ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
 528         ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
 529         ASSERT(offset > isize);
 530
 531         /*
 532          * First handle zeroing the block on which isize resides.
 533          * We only zero a part of that block so it is handled specially.
 534          */
 535         error = xfs_zero_last_block(ip, io, offset, isize);
 536         if (error) {
 537                 ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
 538                 ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
 539                 return error;
 540         }
 541
 542         /*
 543          * Calculate the range between the new size and the old
 544          * where blocks needing to be zeroed may exist.  To get the
 545          * block where the last byte in the file currently resides,
 546          * we need to subtract one from the size and truncate back
 547          * to a block boundary.  We subtract 1 in case the size is
 548          * exactly on a block boundary.
 549          */
 550         last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
 551         start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
 552         end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
 553         ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
 554         if (last_fsb == end_zero_fsb) {
 555                 /*
 556                  * The size was only incremented on its last block.
 557                  * We took care of that above, so just return.
 558                  */
 559                 return 0;
 560         }
 561
 562         ASSERT(start_zero_fsb <= end_zero_fsb);
 563         while (start_zero_fsb <= end_zero_fsb) {
 564                 nimaps = 1;
 565                 zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
 566                 error = XFS_BMAPI(mp, NULL, io, start_zero_fsb, zero_count_fsb,
 567                                   0, NULL, 0, &imap, &nimaps, NULL, NULL);
 568                 if (error) {
 569                         ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
 570                         ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
 571                         return error;
 572                 }
 573                 ASSERT(nimaps > 0);
 574
 575                 if (imap.br_state == XFS_EXT_UNWRITTEN ||
 576                     imap.br_startblock == HOLESTARTBLOCK) {
 577                         /*
 578                          * This loop handles initializing pages that were
 579                          * partially initialized by the code below this
 580                          * loop. It basically zeroes the part of the page
 581                          * that sits on a hole and sets the page as P_HOLE
 582                          * and calls remapf if it is a mapped file.
 583                          */
 584                         start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 585                         ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 586                         continue;
 587                 }
 588
 589                 /*
 590                  * There are blocks we need to zero.
 591                  * Drop the inode lock while we're doing the I/O.
 592                  * We'll still have the iolock to protect us.
 593                  */
 594                 XFS_IUNLOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 595
 596                 zero_off = XFS_FSB_TO_B(mp, start_zero_fsb);
 597                 zero_len = XFS_FSB_TO_B(mp, imap.br_blockcount);
 598
 599                 if ((zero_off + zero_len) > offset)
 600                         zero_len = offset - zero_off;
 601
 602                 error = xfs_iozero(ip, zero_off, zero_len);
 603                 if (error) {
 604                         goto out_lock;
 605                 }
 606
 607                 start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 608                 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 609
 610                 XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 611         }
 612
 613         return 0;
 614
 615 out_lock:
 616
 617         XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 618         ASSERT(error >= 0);
 619         return error;
 620 }
 621
 622 ssize_t                         /* bytes written, or (-) error */
 623 xfs_write(
 624         bhv_desc_t              *bdp,
 625         struct kiocb            *iocb,
 626         const struct iovec      *iovp,
 627         unsigned int            nsegs,
 628         loff_t                  *offset,
 629         int                     ioflags,
 630         cred_t                  *credp)
 631 {
 632         struct file             *file = iocb->ki_filp;
 633         struct address_space    *mapping = file->f_mapping;
 634         struct inode            *inode = mapping->host;
 635         unsigned long           segs = nsegs;
 636         xfs_inode_t             *xip;
 637         xfs_mount_t             *mp;
 638         ssize_t                 ret = 0, error = 0;
 639         xfs_fsize_t             isize, new_size;
 640         xfs_iocore_t            *io;
 641         bhv_vnode_t             *vp;
 642         unsigned long           seg;
 643         int                     iolock;
 644         int                     eventsent = 0;
 645         bhv_vrwlock_t           locktype;
 646         size_t                  ocount = 0, count;
 647         loff_t                  pos;
 648         int                     need_i_mutex = 1, need_flush = 0;
 649
 650         XFS_STATS_INC(xs_write_calls);
 651
 652         vp = BHV_TO_VNODE(bdp);
 653         xip = XFS_BHVTOI(bdp);
 654
 655         for (seg = 0; seg < segs; seg++) {
 656                 const struct iovec *iv = &iovp[seg];
 657
 658                 /*
 659                  * If any segment has a negative length, or the cumulative
 660                  * length ever wraps negative then return -EINVAL.
 661                  */
 662                 ocount += iv->iov_len;
 663                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
 664                         return -EINVAL;
 665                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
 666                         continue;
 667                 if (seg == 0)
 668                         return -EFAULT;
 669                 segs = seg;
 670                 ocount -= iv->iov_len;  /* This segment is no good */
 671                 break;
 672         }
 673
 674         count = ocount;
 675         pos = *offset;
 676
 677         if (count == 0)
 678                 return 0;
 679
 680         io = &xip->i_iocore;
 681         mp = io->io_mount;
 682
 683         vfs_wait_for_freeze(vp->v_vfsp, SB_FREEZE_WRITE);
 684
 685         if (XFS_FORCED_SHUTDOWN(mp))
 686                 return -EIO;
 687
 688         if (ioflags & IO_ISDIRECT) {
 689                 xfs_buftarg_t   *target =
 690                         (xip->i_d.di_flags & XFS_DIFLAG_REALTIME) ?
 691                                 mp->m_rtdev_targp : mp->m_ddev_targp;
 692
 693                 if ((pos & target->bt_smask) || (count & target->bt_smask))
 694                         return XFS_ERROR(-EINVAL);
 695
 696                 if (!VN_CACHED(vp) && pos < i_size_read(inode))
 697                         need_i_mutex = 0;
 698
 699                 if (VN_CACHED(vp))
 700                         need_flush = 1;
 701         }
 702
 703 relock:
 704         if (need_i_mutex) {
 705                 iolock = XFS_IOLOCK_EXCL;
 706                 locktype = VRWLOCK_WRITE;
 707
 708                 mutex_lock(&inode->i_mutex);
 709         } else {
 710                 iolock = XFS_IOLOCK_SHARED;
 711                 locktype = VRWLOCK_WRITE_DIRECT;
 712         }
 713
 714         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
 715
 716         isize = i_size_read(inode);
 717
 718         if (file->f_flags & O_APPEND)
 719                 *offset = isize;
 720
 721 start:
 722         error = -generic_write_checks(file, &pos, &count,
 723                                         S_ISBLK(inode->i_mode));
 724         if (error) {
 725                 xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 726                 goto out_unlock_mutex;
 727         }
 728
 729         new_size = pos + count;
 730         if (new_size > isize)
 731                 io->io_new_size = new_size;
 732
 733         if ((DM_EVENT_ENABLED(vp->v_vfsp, xip, DM_EVENT_WRITE) &&
 734             !(ioflags & IO_INVIS) && !eventsent)) {
 735                 loff_t          savedsize = pos;
 736                 int             dmflags = FILP_DELAY_FLAG(file);
 737
 738                 if (need_i_mutex)
 739                         dmflags |= DM_FLAGS_IMUX;
 740
 741                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 742                 error = XFS_SEND_DATA(xip->i_mount, DM_EVENT_WRITE, vp,
 743                                       pos, count,
 744                                       dmflags, &locktype);
 745                 if (error) {
 746                         xfs_iunlock(xip, iolock);
 747                         goto out_unlock_mutex;
 748                 }
 749                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 750                 eventsent = 1;
 751
 752                 /*
 753                  * The iolock was dropped and reacquired in XFS_SEND_DATA
 754                  * so we have to recheck the size when appending.
 755                  * We will only "goto start;" once, since having sent the
 756                  * event prevents another call to XFS_SEND_DATA, which is
 757                  * what allows the size to change in the first place.
 758                  */
 759                 if ((file->f_flags & O_APPEND) && savedsize != isize) {
 760                         pos = isize = xip->i_d.di_size;
 761                         goto start;
 762                 }
 763         }
 764
 765         if (likely(!(ioflags & IO_INVIS))) {
 766                 file_update_time(file);
 767                 xfs_ichgtime_fast(xip, inode,
 768                                   XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
 769         }
 770
 771         /*
 772          * If the offset is beyond the size of the file, we have a couple
 773          * of things to do. First, if there is already space allocated
 774          * we need to either create holes or zero the disk or ...
 775          *
 776          * If there is a page where the previous size lands, we need
 777          * to zero it out up to the new size.
 778          */
 779
 780         if (pos > isize) {
 781                 error = xfs_zero_eof(BHV_TO_VNODE(bdp), io, pos, isize);
 782                 if (error) {
 783                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 784                         goto out_unlock_mutex;
 785                 }
 786         }
 787         xfs_iunlock(xip, XFS_ILOCK_EXCL);
 788
 789         /*
 790          * If we're writing the file then make sure to clear the
 791          * setuid and setgid bits if the process is not being run
 792          * by root.  This keeps people from modifying setuid and
 793          * setgid binaries.
 794          */
 795
 796         if (((xip->i_d.di_mode & S_ISUID) ||
 797             ((xip->i_d.di_mode & (S_ISGID | S_IXGRP)) ==
 798                 (S_ISGID | S_IXGRP))) &&
 799              !capable(CAP_FSETID)) {
 800                 error = xfs_write_clear_setuid(xip);
 801                 if (likely(!error))
 802                         error = -remove_suid(file->f_path.dentry);
 803                 if (unlikely(error)) {
 804                         xfs_iunlock(xip, iolock);
 805                         goto out_unlock_mutex;
 806                 }
 807         }
 808
 809 retry:
 810         /* We can write back this queue in page reclaim */
 811         current->backing_dev_info = mapping->backing_dev_info;
 812
 813         if ((ioflags & IO_ISDIRECT)) {
 814                 if (need_flush) {
 815                         xfs_inval_cached_trace(io, pos, -1,
 816                                         ctooff(offtoct(pos)), -1);
 817                         bhv_vop_flushinval_pages(vp, ctooff(offtoct(pos)),
 818                                         -1, FI_REMAPF_LOCKED);
 819                 }
 820
 821                 if (need_i_mutex) {
 822                         /* demote the lock now the cached pages are gone */
 823                         XFS_ILOCK_DEMOTE(mp, io, XFS_IOLOCK_EXCL);
 824                         mutex_unlock(&inode->i_mutex);
 825
 826                         iolock = XFS_IOLOCK_SHARED;
 827                         locktype = VRWLOCK_WRITE_DIRECT;
 828                         need_i_mutex = 0;
 829                 }
 830
 831                 xfs_rw_enter_trace(XFS_DIOWR_ENTER, io, (void *)iovp, segs,
 832                                 *offset, ioflags);
 833                 ret = generic_file_direct_write(iocb, iovp,
 834                                 &segs, pos, offset, count, ocount);
 835
 836                 /*
 837                  * direct-io write to a hole: fall through to buffered I/O
 838                  * for completing the rest of the request.
 839                  */
 840                 if (ret >= 0 && ret != count) {
 841                         XFS_STATS_ADD(xs_write_bytes, ret);
 842
 843                         pos += ret;
 844                         count -= ret;
 845
 846                         need_i_mutex = 1;
 847                         ioflags &= ~IO_ISDIRECT;
 848                         xfs_iunlock(xip, iolock);
 849                         goto relock;
 850                 }
 851         } else {
 852                 xfs_rw_enter_trace(XFS_WRITE_ENTER, io, (void *)iovp, segs,
 853                                 *offset, ioflags);
 854                 ret = generic_file_buffered_write(iocb, iovp, segs,
 855                                 pos, offset, count, ret);
 856         }
 857
 858         current->backing_dev_info = NULL;
 859
 860         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
 861                 ret = wait_on_sync_kiocb(iocb);
 862
 863         if ((ret == -ENOSPC) &&
 864             DM_EVENT_ENABLED(vp->v_vfsp, xip, DM_EVENT_NOSPACE) &&
 865             !(ioflags & IO_INVIS)) {
 866
 867                 xfs_rwunlock(bdp, locktype);
 868                 if (need_i_mutex)
 869                         mutex_unlock(&inode->i_mutex);
 870                 error = XFS_SEND_NAMESP(xip->i_mount, DM_EVENT_NOSPACE, vp,
 871                                 DM_RIGHT_NULL, vp, DM_RIGHT_NULL, NULL, NULL,
 872                                 0, 0, 0); /* Delay flag intentionally  unused */
 873                 if (error)
 874                         goto out_nounlocks;
 875                 if (need_i_mutex)
 876                         mutex_lock(&inode->i_mutex);
 877                 xfs_rwlock(bdp, locktype);
 878                 pos = xip->i_d.di_size;
 879                 ret = 0;
 880                 goto retry;
 881         }
 882
 883         isize = i_size_read(inode);
 884         if (unlikely(ret < 0 && ret != -EFAULT && *offset > isize))
 885                 *offset = isize;
 886
 887         if (*offset > xip->i_d.di_size) {
 888                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 889                 if (*offset > xip->i_d.di_size) {
 890                         xip->i_d.di_size = *offset;
 891                         i_size_write(inode, *offset);
 892                         xip->i_update_core = 1;
 893                         xip->i_update_size = 1;
 894                 }
 895                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 896         }
 897
 898         error = -ret;
 899         if (ret <= 0)
 900                 goto out_unlock_internal;
 901
 902         XFS_STATS_ADD(xs_write_bytes, ret);
 903
 904         /* Handle various SYNC-type writes */
 905         if ((file->f_flags & O_SYNC) || IS_SYNC(inode)) {
 906                 error = xfs_write_sync_logforce(mp, xip);
 907                 if (error)
 908                         goto out_unlock_internal;
 909
 910                 xfs_rwunlock(bdp, locktype);
 911                 if (need_i_mutex)
 912                         mutex_unlock(&inode->i_mutex);
 913
 914                 error = sync_page_range(inode, mapping, pos, ret);
 915                 if (!error)
 916                         error = ret;
 917                 return error;
 918         }
 919
 920  out_unlock_internal:
 921         xfs_rwunlock(bdp, locktype);
 922  out_unlock_mutex:
 923         if (need_i_mutex)
 924                 mutex_unlock(&inode->i_mutex);
 925  out_nounlocks:
 926         return -error;
 927 }
 928
 929 /*
 930  * All xfs metadata buffers except log state machine buffers
 931  * get this attached as their b_bdstrat callback function.
 932  * This is so that we can catch a buffer
 933  * after prematurely unpinning it to forcibly shutdown the filesystem.
 934  */
 935 int
 936 xfs_bdstrat_cb(struct xfs_buf *bp)
 937 {
 938         xfs_mount_t     *mp;
 939
 940         mp = XFS_BUF_FSPRIVATE3(bp, xfs_mount_t *);
 941         if (!XFS_FORCED_SHUTDOWN(mp)) {
 942                 xfs_buf_iorequest(bp);
 943                 return 0;
 944         } else {
 945                 xfs_buftrace("XFS__BDSTRAT IOERROR", bp);
 946                 /*
 947                  * Metadata write that didn't get logged but
 948                  * written delayed anyway. These aren't associated
 949                  * with a transaction, and can be ignored.
 950                  */
 951                 if (XFS_BUF_IODONE_FUNC(bp) == NULL &&
 952                     (XFS_BUF_ISREAD(bp)) == 0)
 953                         return (xfs_bioerror_relse(bp));
 954                 else
 955                         return (xfs_bioerror(bp));
 956         }
 957 }
 958
 959
 960 int
 961 xfs_bmap(bhv_desc_t     *bdp,
 962         xfs_off_t       offset,
 963         ssize_t         count,
 964         int             flags,
 965         xfs_iomap_t     *iomapp,
 966         int             *niomaps)
 967 {
 968         xfs_inode_t     *ip = XFS_BHVTOI(bdp);
 969         xfs_iocore_t    *io = &ip->i_iocore;
 970
 971         ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFREG);
 972         ASSERT(((ip->i_d.di_flags & XFS_DIFLAG_REALTIME) != 0) ==
 973                ((ip->i_iocore.io_flags & XFS_IOCORE_RT) != 0));
 974
 975         return xfs_iomap(io, offset, count, flags, iomapp, niomaps);
 976 }
 977
 978 /*
 979  * Wrapper around bdstrat so that we can stop data
 980  * from going to disk in case we are shutting down the filesystem.
 981  * Typically user data goes thru this path; one of the exceptions
 982  * is the superblock.
 983  */
 984 int
 985 xfsbdstrat(
 986         struct xfs_mount        *mp,
 987         struct xfs_buf          *bp)
 988 {
 989         ASSERT(mp);
 990         if (!XFS_FORCED_SHUTDOWN(mp)) {
 991                 /* Grio redirection would go here
 992                  * if (XFS_BUF_IS_GRIO(bp)) {
 993                  */
 994
 995                 xfs_buf_iorequest(bp);
 996                 return 0;
 997         }
 998
 999         xfs_buftrace("XFSBDSTRAT IOERROR", bp);
1000         return (xfs_bioerror_relse(bp));
1001 }
1002
1003 /*
1004  * If the underlying (data/log/rt) device is readonly, there are some
1005  * operations that cannot proceed.
1006  */
1007 int
1008 xfs_dev_is_read_only(
1009         xfs_mount_t             *mp,
1010         char                    *message)
1011 {
1012         if (xfs_readonly_buftarg(mp->m_ddev_targp) ||
1013             xfs_readonly_buftarg(mp->m_logdev_targp) ||
1014             (mp->m_rtdev_targp && xfs_readonly_buftarg(mp->m_rtdev_targp))) {
1015                 cmn_err(CE_NOTE,
1016                         "XFS: %s required on read-only device.", message);
1017                 cmn_err(CE_NOTE,
1018                         "XFS: write access unavailable, cannot proceed.");
1019                 return EROFS;
1020         }
1021         return 0;
1022 }