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_cap.h"
  47 #include "xfs_mac.h"
  48 #include "xfs_attr.h"
  49 #include "xfs_inode_item.h"
  50 #include "xfs_buf_item.h"
  51 #include "xfs_utils.h"
  52 #include "xfs_iomap.h"
  53
  54 #include <linux/capability.h>
  55 #include <linux/writeback.h>
  56
  57
  58 #if defined(XFS_RW_TRACE)
  59 void
  60 xfs_rw_enter_trace(
  61         int                     tag,
  62         xfs_iocore_t            *io,
  63         void                    *data,
  64         size_t                  segs,
  65         loff_t                  offset,
  66         int                     ioflags)
  67 {
  68         xfs_inode_t     *ip = XFS_IO_INODE(io);
  69
  70         if (ip->i_rwtrace == NULL)
  71                 return;
  72         ktrace_enter(ip->i_rwtrace,
  73                 (void *)(unsigned long)tag,
  74                 (void *)ip,
  75                 (void *)((unsigned long)((ip->i_d.di_size >> 32) & 0xffffffff)),
  76                 (void *)((unsigned long)(ip->i_d.di_size & 0xffffffff)),
  77                 (void *)data,
  78                 (void *)((unsigned long)segs),
  79                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
  80                 (void *)((unsigned long)(offset & 0xffffffff)),
  81                 (void *)((unsigned long)ioflags),
  82                 (void *)((unsigned long)((io->io_new_size >> 32) & 0xffffffff)),
  83                 (void *)((unsigned long)(io->io_new_size & 0xffffffff)),
  84                 (void *)((unsigned long)current_pid()),
  85                 (void *)NULL,
  86                 (void *)NULL,
  87                 (void *)NULL,
  88                 (void *)NULL);
  89 }
  90
  91 void
  92 xfs_inval_cached_trace(
  93         xfs_iocore_t    *io,
  94         xfs_off_t       offset,
  95         xfs_off_t       len,
  96         xfs_off_t       first,
  97         xfs_off_t       last)
  98 {
  99         xfs_inode_t     *ip = XFS_IO_INODE(io);
 100
 101         if (ip->i_rwtrace == NULL)
 102                 return;
 103         ktrace_enter(ip->i_rwtrace,
 104                 (void *)(__psint_t)XFS_INVAL_CACHED,
 105                 (void *)ip,
 106                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
 107                 (void *)((unsigned long)(offset & 0xffffffff)),
 108                 (void *)((unsigned long)((len >> 32) & 0xffffffff)),
 109                 (void *)((unsigned long)(len & 0xffffffff)),
 110                 (void *)((unsigned long)((first >> 32) & 0xffffffff)),
 111                 (void *)((unsigned long)(first & 0xffffffff)),
 112                 (void *)((unsigned long)((last >> 32) & 0xffffffff)),
 113                 (void *)((unsigned long)(last & 0xffffffff)),
 114                 (void *)((unsigned long)current_pid()),
 115                 (void *)NULL,
 116                 (void *)NULL,
 117                 (void *)NULL,
 118                 (void *)NULL,
 119                 (void *)NULL);
 120 }
 121 #endif
 122
 123 /*
 124  *      xfs_iozero
 125  *
 126  *      xfs_iozero clears the specified range of buffer supplied,
 127  *      and marks all the affected blocks as valid and modified.  If
 128  *      an affected block is not allocated, it will be allocated.  If
 129  *      an affected block is not completely overwritten, and is not
 130  *      valid before the operation, it will be read from disk before
 131  *      being partially zeroed.
 132  */
 133 STATIC int
 134 xfs_iozero(
 135         struct inode            *ip,    /* inode                        */
 136         loff_t                  pos,    /* offset in file               */
 137         size_t                  count)  /* size of data to zero         */
 138 {
 139         unsigned                bytes;
 140         struct page             *page;
 141         struct address_space    *mapping;
 142         char                    *kaddr;
 143         int                     status;
 144
 145         mapping = ip->i_mapping;
 146         do {
 147                 unsigned long index, offset;
 148
 149                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
 150                 index = pos >> PAGE_CACHE_SHIFT;
 151                 bytes = PAGE_CACHE_SIZE - offset;
 152                 if (bytes > count)
 153                         bytes = count;
 154
 155                 status = -ENOMEM;
 156                 page = grab_cache_page(mapping, index);
 157                 if (!page)
 158                         break;
 159
 160                 kaddr = kmap(page);
 161                 status = mapping->a_ops->prepare_write(NULL, page, offset,
 162                                                         offset + bytes);
 163                 if (status) {
 164                         goto unlock;
 165                 }
 166
 167                 memset((void *) (kaddr + offset), 0, bytes);
 168                 flush_dcache_page(page);
 169                 status = mapping->a_ops->commit_write(NULL, page, offset,
 170                                                         offset + bytes);
 171                 if (!status) {
 172                         pos += bytes;
 173                         count -= bytes;
 174                 }
 175
 176 unlock:
 177                 kunmap(page);
 178                 unlock_page(page);
 179                 page_cache_release(page);
 180                 if (status)
 181                         break;
 182         } while (count);
 183
 184         return (-status);
 185 }
 186
 187 ssize_t                 /* bytes read, or (-)  error */
 188 xfs_read(
 189         bhv_desc_t              *bdp,
 190         struct kiocb            *iocb,
 191         const struct iovec      *iovp,
 192         unsigned int            segs,
 193         loff_t                  *offset,
 194         int                     ioflags,
 195         cred_t                  *credp)
 196 {
 197         struct file             *file = iocb->ki_filp;
 198         struct inode            *inode = file->f_mapping->host;
 199         size_t                  size = 0;
 200         ssize_t                 ret;
 201         xfs_fsize_t             n;
 202         xfs_inode_t             *ip;
 203         xfs_mount_t             *mp;
 204         bhv_vnode_t             *vp;
 205         unsigned long           seg;
 206
 207         ip = XFS_BHVTOI(bdp);
 208         vp = BHV_TO_VNODE(bdp);
 209         mp = ip->i_mount;
 210
 211         XFS_STATS_INC(xs_read_calls);
 212
 213         /* START copy & waste from filemap.c */
 214         for (seg = 0; seg < segs; seg++) {
 215                 const struct iovec *iv = &iovp[seg];
 216
 217                 /*
 218                  * If any segment has a negative length, or the cumulative
 219                  * length ever wraps negative then return -EINVAL.
 220                  */
 221                 size += iv->iov_len;
 222                 if (unlikely((ssize_t)(size|iv->iov_len) < 0))
 223                         return XFS_ERROR(-EINVAL);
 224         }
 225         /* END copy & waste from filemap.c */
 226
 227         if (unlikely(ioflags & IO_ISDIRECT)) {
 228                 xfs_buftarg_t   *target =
 229                         (ip->i_d.di_flags & XFS_DIFLAG_REALTIME) ?
 230                                 mp->m_rtdev_targp : mp->m_ddev_targp;
 231                 if ((*offset & target->bt_smask) ||
 232                     (size & target->bt_smask)) {
 233                         if (*offset == ip->i_d.di_size) {
 234                                 return (0);
 235                         }
 236                         return -XFS_ERROR(EINVAL);
 237                 }
 238         }
 239
 240         n = XFS_MAXIOFFSET(mp) - *offset;
 241         if ((n <= 0) || (size == 0))
 242                 return 0;
 243
 244         if (n < size)
 245                 size = n;
 246
 247         if (XFS_FORCED_SHUTDOWN(mp))
 248                 return -EIO;
 249
 250         if (unlikely(ioflags & IO_ISDIRECT))
 251                 mutex_lock(&inode->i_mutex);
 252         xfs_ilock(ip, XFS_IOLOCK_SHARED);
 253
 254         if (DM_EVENT_ENABLED(vp->v_vfsp, ip, DM_EVENT_READ) &&
 255             !(ioflags & IO_INVIS)) {
 256                 bhv_vrwlock_t locktype = VRWLOCK_READ;
 257                 int dmflags = FILP_DELAY_FLAG(file) | DM_SEM_FLAG_RD(ioflags);
 258
 259                 ret = -XFS_SEND_DATA(mp, DM_EVENT_READ,
 260                                         BHV_TO_VNODE(bdp), *offset, size,
 261                                         dmflags, &locktype);
 262                 if (ret) {
 263                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 264                         if (unlikely(ioflags & IO_ISDIRECT))
 265                                 mutex_unlock(&inode->i_mutex);
 266                         return ret;
 267                 }
 268         }
 269
 270         if (unlikely(ioflags & IO_ISDIRECT)) {
 271                 if (VN_CACHED(vp))
 272                         bhv_vop_flushinval_pages(vp, ctooff(offtoct(*offset)),
 273                                                  -1, FI_REMAPF_LOCKED);
 274                 mutex_unlock(&inode->i_mutex);
 275         }
 276
 277         xfs_rw_enter_trace(XFS_READ_ENTER, &ip->i_iocore,
 278                                 (void *)iovp, segs, *offset, ioflags);
 279
 280         iocb->ki_pos = *offset;
 281         ret = generic_file_aio_read(iocb, iovp, segs, *offset);
 282         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
 283                 ret = wait_on_sync_kiocb(iocb);
 284         if (ret > 0)
 285                 XFS_STATS_ADD(xs_read_bytes, ret);
 286
 287         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 288         return ret;
 289 }
 290
 291 ssize_t
 292 xfs_sendfile(
 293         bhv_desc_t              *bdp,
 294         struct file             *filp,
 295         loff_t                  *offset,
 296         int                     ioflags,
 297         size_t                  count,
 298         read_actor_t            actor,
 299         void                    *target,
 300         cred_t                  *credp)
 301 {
 302         xfs_inode_t             *ip = XFS_BHVTOI(bdp);
 303         xfs_mount_t             *mp = ip->i_mount;
 304         ssize_t                 ret;
 305
 306         XFS_STATS_INC(xs_read_calls);
 307         if (XFS_FORCED_SHUTDOWN(mp))
 308                 return -EIO;
 309
 310         xfs_ilock(ip, XFS_IOLOCK_SHARED);
 311
 312         if (DM_EVENT_ENABLED(BHV_TO_VNODE(bdp)->v_vfsp, ip, DM_EVENT_READ) &&
 313             (!(ioflags & IO_INVIS))) {
 314                 bhv_vrwlock_t locktype = VRWLOCK_READ;
 315                 int error;
 316
 317                 error = XFS_SEND_DATA(mp, DM_EVENT_READ, BHV_TO_VNODE(bdp),
 318                                       *offset, count,
 319                                       FILP_DELAY_FLAG(filp), &locktype);
 320                 if (error) {
 321                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 322                         return -error;
 323                 }
 324         }
 325         xfs_rw_enter_trace(XFS_SENDFILE_ENTER, &ip->i_iocore,
 326                    (void *)(unsigned long)target, count, *offset, ioflags);
 327         ret = generic_file_sendfile(filp, offset, count, actor, target);
 328         if (ret > 0)
 329                 XFS_STATS_ADD(xs_read_bytes, ret);
 330
 331         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 332         return ret;
 333 }
 334
 335 ssize_t
 336 xfs_splice_read(
 337         bhv_desc_t              *bdp,
 338         struct file             *infilp,
 339         loff_t                  *ppos,
 340         struct pipe_inode_info  *pipe,
 341         size_t                  count,
 342         int                     flags,
 343         int                     ioflags,
 344         cred_t                  *credp)
 345 {
 346         xfs_inode_t             *ip = XFS_BHVTOI(bdp);
 347         xfs_mount_t             *mp = ip->i_mount;
 348         ssize_t                 ret;
 349
 350         XFS_STATS_INC(xs_read_calls);
 351         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 352                 return -EIO;
 353
 354         xfs_ilock(ip, XFS_IOLOCK_SHARED);
 355
 356         if (DM_EVENT_ENABLED(BHV_TO_VNODE(bdp)->v_vfsp, ip, DM_EVENT_READ) &&
 357             (!(ioflags & IO_INVIS))) {
 358                 bhv_vrwlock_t locktype = VRWLOCK_READ;
 359                 int error;
 360
 361                 error = XFS_SEND_DATA(mp, DM_EVENT_READ, BHV_TO_VNODE(bdp),
 362                                         *ppos, count,
 363                                         FILP_DELAY_FLAG(infilp), &locktype);
 364                 if (error) {
 365                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 366                         return -error;
 367                 }
 368         }
 369         xfs_rw_enter_trace(XFS_SPLICE_READ_ENTER, &ip->i_iocore,
 370                            pipe, count, *ppos, ioflags);
 371         ret = generic_file_splice_read(infilp, ppos, pipe, count, flags);
 372         if (ret > 0)
 373                 XFS_STATS_ADD(xs_read_bytes, ret);
 374
 375         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 376         return ret;
 377 }
 378
 379 ssize_t
 380 xfs_splice_write(
 381         bhv_desc_t              *bdp,
 382         struct pipe_inode_info  *pipe,
 383         struct file             *outfilp,
 384         loff_t                  *ppos,
 385         size_t                  count,
 386         int                     flags,
 387         int                     ioflags,
 388         cred_t                  *credp)
 389 {
 390         xfs_inode_t             *ip = XFS_BHVTOI(bdp);
 391         xfs_mount_t             *mp = ip->i_mount;
 392         ssize_t                 ret;
 393         struct inode            *inode = outfilp->f_mapping->host;
 394         xfs_fsize_t             isize;
 395
 396         XFS_STATS_INC(xs_write_calls);
 397         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 398                 return -EIO;
 399
 400         xfs_ilock(ip, XFS_IOLOCK_EXCL);
 401
 402         if (DM_EVENT_ENABLED(BHV_TO_VNODE(bdp)->v_vfsp, ip, DM_EVENT_WRITE) &&
 403             (!(ioflags & IO_INVIS))) {
 404                 bhv_vrwlock_t locktype = VRWLOCK_WRITE;
 405                 int error;
 406
 407                 error = XFS_SEND_DATA(mp, DM_EVENT_WRITE, BHV_TO_VNODE(bdp),
 408                                         *ppos, count,
 409                                         FILP_DELAY_FLAG(outfilp), &locktype);
 410                 if (error) {
 411                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 412                         return -error;
 413                 }
 414         }
 415         xfs_rw_enter_trace(XFS_SPLICE_WRITE_ENTER, &ip->i_iocore,
 416                            pipe, count, *ppos, ioflags);
 417         ret = generic_file_splice_write(pipe, outfilp, ppos, count, flags);
 418         if (ret > 0)
 419                 XFS_STATS_ADD(xs_write_bytes, ret);
 420
 421         isize = i_size_read(inode);
 422         if (unlikely(ret < 0 && ret != -EFAULT && *ppos > isize))
 423                 *ppos = isize;
 424
 425         if (*ppos > ip->i_d.di_size) {
 426                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 427                 if (*ppos > ip->i_d.di_size) {
 428                         ip->i_d.di_size = *ppos;
 429                         i_size_write(inode, *ppos);
 430                         ip->i_update_core = 1;
 431                         ip->i_update_size = 1;
 432                 }
 433                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 434         }
 435         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 436         return ret;
 437 }
 438
 439 /*
 440  * This routine is called to handle zeroing any space in the last
 441  * block of the file that is beyond the EOF.  We do this since the
 442  * size is being increased without writing anything to that block
 443  * and we don't want anyone to read the garbage on the disk.
 444  */
 445 STATIC int                              /* error (positive) */
 446 xfs_zero_last_block(
 447         struct inode    *ip,
 448         xfs_iocore_t    *io,
 449         xfs_fsize_t     offset,
 450         xfs_fsize_t     isize)
 451 {
 452         xfs_fileoff_t   last_fsb;
 453         xfs_mount_t     *mp = io->io_mount;
 454         int             nimaps;
 455         int             zero_offset;
 456         int             zero_len;
 457         int             error = 0;
 458         xfs_bmbt_irec_t imap;
 459
 460         ASSERT(ismrlocked(io->io_lock, MR_UPDATE) != 0);
 461
 462         zero_offset = XFS_B_FSB_OFFSET(mp, isize);
 463         if (zero_offset == 0) {
 464                 /*
 465                  * There are no extra bytes in the last block on disk to
 466                  * zero, so return.
 467                  */
 468                 return 0;
 469         }
 470
 471         last_fsb = XFS_B_TO_FSBT(mp, isize);
 472         nimaps = 1;
 473         error = XFS_BMAPI(mp, NULL, io, last_fsb, 1, 0, NULL, 0, &imap,
 474                           &nimaps, NULL, NULL);
 475         if (error) {
 476                 return error;
 477         }
 478         ASSERT(nimaps > 0);
 479         /*
 480          * If the block underlying isize is just a hole, then there
 481          * is nothing to zero.
 482          */
 483         if (imap.br_startblock == HOLESTARTBLOCK) {
 484                 return 0;
 485         }
 486         /*
 487          * Zero the part of the last block beyond the EOF, and write it
 488          * out sync.  We need to drop the ilock while we do this so we
 489          * don't deadlock when the buffer cache calls back to us.
 490          */
 491         XFS_IUNLOCK(mp, io, XFS_ILOCK_EXCL| XFS_EXTSIZE_RD);
 492
 493         zero_len = mp->m_sb.sb_blocksize - zero_offset;
 494         if (isize + zero_len > offset)
 495                 zero_len = offset - isize;
 496         error = xfs_iozero(ip, isize, zero_len);
 497
 498         XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 499         ASSERT(error >= 0);
 500         return error;
 501 }
 502
 503 /*
 504  * Zero any on disk space between the current EOF and the new,
 505  * larger EOF.  This handles the normal case of zeroing the remainder
 506  * of the last block in the file and the unusual case of zeroing blocks
 507  * out beyond the size of the file.  This second case only happens
 508  * with fixed size extents and when the system crashes before the inode
 509  * size was updated but after blocks were allocated.  If fill is set,
 510  * then any holes in the range are filled and zeroed.  If not, the holes
 511  * are left alone as holes.
 512  */
 513
 514 int                                     /* error (positive) */
 515 xfs_zero_eof(
 516         bhv_vnode_t     *vp,
 517         xfs_iocore_t    *io,
 518         xfs_off_t       offset,         /* starting I/O offset */
 519         xfs_fsize_t     isize)          /* current inode size */
 520 {
 521         struct inode    *ip = vn_to_inode(vp);
 522         xfs_fileoff_t   start_zero_fsb;
 523         xfs_fileoff_t   end_zero_fsb;
 524         xfs_fileoff_t   zero_count_fsb;
 525         xfs_fileoff_t   last_fsb;
 526         xfs_fileoff_t   zero_off;
 527         xfs_fsize_t     zero_len;
 528         xfs_mount_t     *mp = io->io_mount;
 529         int             nimaps;
 530         int             error = 0;
 531         xfs_bmbt_irec_t imap;
 532
 533         ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
 534         ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
 535         ASSERT(offset > isize);
 536
 537         /*
 538          * First handle zeroing the block on which isize resides.
 539          * We only zero a part of that block so it is handled specially.
 540          */
 541         error = xfs_zero_last_block(ip, io, offset, isize);
 542         if (error) {
 543                 ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
 544                 ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
 545                 return error;
 546         }
 547
 548         /*
 549          * Calculate the range between the new size and the old
 550          * where blocks needing to be zeroed may exist.  To get the
 551          * block where the last byte in the file currently resides,
 552          * we need to subtract one from the size and truncate back
 553          * to a block boundary.  We subtract 1 in case the size is
 554          * exactly on a block boundary.
 555          */
 556         last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
 557         start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
 558         end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
 559         ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
 560         if (last_fsb == end_zero_fsb) {
 561                 /*
 562                  * The size was only incremented on its last block.
 563                  * We took care of that above, so just return.
 564                  */
 565                 return 0;
 566         }
 567
 568         ASSERT(start_zero_fsb <= end_zero_fsb);
 569         while (start_zero_fsb <= end_zero_fsb) {
 570                 nimaps = 1;
 571                 zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
 572                 error = XFS_BMAPI(mp, NULL, io, start_zero_fsb, zero_count_fsb,
 573                                   0, NULL, 0, &imap, &nimaps, NULL, NULL);
 574                 if (error) {
 575                         ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
 576                         ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
 577                         return error;
 578                 }
 579                 ASSERT(nimaps > 0);
 580
 581                 if (imap.br_state == XFS_EXT_UNWRITTEN ||
 582                     imap.br_startblock == HOLESTARTBLOCK) {
 583                         /*
 584                          * This loop handles initializing pages that were
 585                          * partially initialized by the code below this
 586                          * loop. It basically zeroes the part of the page
 587                          * that sits on a hole and sets the page as P_HOLE
 588                          * and calls remapf if it is a mapped file.
 589                          */
 590                         start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 591                         ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 592                         continue;
 593                 }
 594
 595                 /*
 596                  * There are blocks we need to zero.
 597                  * Drop the inode lock while we're doing the I/O.
 598                  * We'll still have the iolock to protect us.
 599                  */
 600                 XFS_IUNLOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 601
 602                 zero_off = XFS_FSB_TO_B(mp, start_zero_fsb);
 603                 zero_len = XFS_FSB_TO_B(mp, imap.br_blockcount);
 604
 605                 if ((zero_off + zero_len) > offset)
 606                         zero_len = offset - zero_off;
 607
 608                 error = xfs_iozero(ip, zero_off, zero_len);
 609                 if (error) {
 610                         goto out_lock;
 611                 }
 612
 613                 start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 614                 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 615
 616                 XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 617         }
 618
 619         return 0;
 620
 621 out_lock:
 622
 623         XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 624         ASSERT(error >= 0);
 625         return error;
 626 }
 627
 628 ssize_t                         /* bytes written, or (-) error */
 629 xfs_write(
 630         bhv_desc_t              *bdp,
 631         struct kiocb            *iocb,
 632         const struct iovec      *iovp,
 633         unsigned int            nsegs,
 634         loff_t                  *offset,
 635         int                     ioflags,
 636         cred_t                  *credp)
 637 {
 638         struct file             *file = iocb->ki_filp;
 639         struct address_space    *mapping = file->f_mapping;
 640         struct inode            *inode = mapping->host;
 641         unsigned long           segs = nsegs;
 642         xfs_inode_t             *xip;
 643         xfs_mount_t             *mp;
 644         ssize_t                 ret = 0, error = 0;
 645         xfs_fsize_t             isize, new_size;
 646         xfs_iocore_t            *io;
 647         bhv_vnode_t             *vp;
 648         unsigned long           seg;
 649         int                     iolock;
 650         int                     eventsent = 0;
 651         bhv_vrwlock_t           locktype;
 652         size_t                  ocount = 0, count;
 653         loff_t                  pos;
 654         int                     need_i_mutex = 1, need_flush = 0;
 655
 656         XFS_STATS_INC(xs_write_calls);
 657
 658         vp = BHV_TO_VNODE(bdp);
 659         xip = XFS_BHVTOI(bdp);
 660
 661         for (seg = 0; seg < segs; seg++) {
 662                 const struct iovec *iv = &iovp[seg];
 663
 664                 /*
 665                  * If any segment has a negative length, or the cumulative
 666                  * length ever wraps negative then return -EINVAL.
 667                  */
 668                 ocount += iv->iov_len;
 669                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
 670                         return -EINVAL;
 671                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
 672                         continue;
 673                 if (seg == 0)
 674                         return -EFAULT;
 675                 segs = seg;
 676                 ocount -= iv->iov_len;  /* This segment is no good */
 677                 break;
 678         }
 679
 680         count = ocount;
 681         pos = *offset;
 682
 683         if (count == 0)
 684                 return 0;
 685
 686         io = &xip->i_iocore;
 687         mp = io->io_mount;
 688
 689         vfs_wait_for_freeze(vp->v_vfsp, SB_FREEZE_WRITE);
 690
 691         if (XFS_FORCED_SHUTDOWN(mp))
 692                 return -EIO;
 693
 694         if (ioflags & IO_ISDIRECT) {
 695                 xfs_buftarg_t   *target =
 696                         (xip->i_d.di_flags & XFS_DIFLAG_REALTIME) ?
 697                                 mp->m_rtdev_targp : mp->m_ddev_targp;
 698
 699                 if ((pos & target->bt_smask) || (count & target->bt_smask))
 700                         return XFS_ERROR(-EINVAL);
 701
 702                 if (!VN_CACHED(vp) && pos < i_size_read(inode))
 703                         need_i_mutex = 0;
 704
 705                 if (VN_CACHED(vp))
 706                         need_flush = 1;
 707         }
 708
 709 relock:
 710         if (need_i_mutex) {
 711                 iolock = XFS_IOLOCK_EXCL;
 712                 locktype = VRWLOCK_WRITE;
 713
 714                 mutex_lock(&inode->i_mutex);
 715         } else {
 716                 iolock = XFS_IOLOCK_SHARED;
 717                 locktype = VRWLOCK_WRITE_DIRECT;
 718         }
 719
 720         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
 721
 722         isize = i_size_read(inode);
 723
 724         if (file->f_flags & O_APPEND)
 725                 *offset = isize;
 726
 727 start:
 728         error = -generic_write_checks(file, &pos, &count,
 729                                         S_ISBLK(inode->i_mode));
 730         if (error) {
 731                 xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 732                 goto out_unlock_mutex;
 733         }
 734
 735         new_size = pos + count;
 736         if (new_size > isize)
 737                 io->io_new_size = new_size;
 738
 739         if ((DM_EVENT_ENABLED(vp->v_vfsp, xip, DM_EVENT_WRITE) &&
 740             !(ioflags & IO_INVIS) && !eventsent)) {
 741                 loff_t          savedsize = pos;
 742                 int             dmflags = FILP_DELAY_FLAG(file);
 743
 744                 if (need_i_mutex)
 745                         dmflags |= DM_FLAGS_IMUX;
 746
 747                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 748                 error = XFS_SEND_DATA(xip->i_mount, DM_EVENT_WRITE, vp,
 749                                       pos, count,
 750                                       dmflags, &locktype);
 751                 if (error) {
 752                         xfs_iunlock(xip, iolock);
 753                         goto out_unlock_mutex;
 754                 }
 755                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 756                 eventsent = 1;
 757
 758                 /*
 759                  * The iolock was dropped and reacquired in XFS_SEND_DATA
 760                  * so we have to recheck the size when appending.
 761                  * We will only "goto start;" once, since having sent the
 762                  * event prevents another call to XFS_SEND_DATA, which is
 763                  * what allows the size to change in the first place.
 764                  */
 765                 if ((file->f_flags & O_APPEND) && savedsize != isize) {
 766                         pos = isize = xip->i_d.di_size;
 767                         goto start;
 768                 }
 769         }
 770
 771         if (likely(!(ioflags & IO_INVIS))) {
 772                 file_update_time(file);
 773                 xfs_ichgtime_fast(xip, inode,
 774                                   XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
 775         }
 776
 777         /*
 778          * If the offset is beyond the size of the file, we have a couple
 779          * of things to do. First, if there is already space allocated
 780          * we need to either create holes or zero the disk or ...
 781          *
 782          * If there is a page where the previous size lands, we need
 783          * to zero it out up to the new size.
 784          */
 785
 786         if (pos > isize) {
 787                 error = xfs_zero_eof(BHV_TO_VNODE(bdp), io, pos, isize);
 788                 if (error) {
 789                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 790                         goto out_unlock_mutex;
 791                 }
 792         }
 793         xfs_iunlock(xip, XFS_ILOCK_EXCL);
 794
 795         /*
 796          * If we're writing the file then make sure to clear the
 797          * setuid and setgid bits if the process is not being run
 798          * by root.  This keeps people from modifying setuid and
 799          * setgid binaries.
 800          */
 801
 802         if (((xip->i_d.di_mode & S_ISUID) ||
 803             ((xip->i_d.di_mode & (S_ISGID | S_IXGRP)) ==
 804                 (S_ISGID | S_IXGRP))) &&
 805              !capable(CAP_FSETID)) {
 806                 error = xfs_write_clear_setuid(xip);
 807                 if (likely(!error))
 808                         error = -remove_suid(file->f_path.dentry);
 809                 if (unlikely(error)) {
 810                         xfs_iunlock(xip, iolock);
 811                         goto out_unlock_mutex;
 812                 }
 813         }
 814
 815 retry:
 816         /* We can write back this queue in page reclaim */
 817         current->backing_dev_info = mapping->backing_dev_info;
 818
 819         if ((ioflags & IO_ISDIRECT)) {
 820                 if (need_flush) {
 821                         xfs_inval_cached_trace(io, pos, -1,
 822                                         ctooff(offtoct(pos)), -1);
 823                         bhv_vop_flushinval_pages(vp, ctooff(offtoct(pos)),
 824                                         -1, FI_REMAPF_LOCKED);
 825                 }
 826
 827                 if (need_i_mutex) {
 828                         /* demote the lock now the cached pages are gone */
 829                         XFS_ILOCK_DEMOTE(mp, io, XFS_IOLOCK_EXCL);
 830                         mutex_unlock(&inode->i_mutex);
 831
 832                         iolock = XFS_IOLOCK_SHARED;
 833                         locktype = VRWLOCK_WRITE_DIRECT;
 834                         need_i_mutex = 0;
 835                 }
 836
 837                 xfs_rw_enter_trace(XFS_DIOWR_ENTER, io, (void *)iovp, segs,
 838                                 *offset, ioflags);
 839                 ret = generic_file_direct_write(iocb, iovp,
 840                                 &segs, pos, offset, count, ocount);
 841
 842                 /*
 843                  * direct-io write to a hole: fall through to buffered I/O
 844                  * for completing the rest of the request.
 845                  */
 846                 if (ret >= 0 && ret != count) {
 847                         XFS_STATS_ADD(xs_write_bytes, ret);
 848
 849                         pos += ret;
 850                         count -= ret;
 851
 852                         need_i_mutex = 1;
 853                         ioflags &= ~IO_ISDIRECT;
 854                         xfs_iunlock(xip, iolock);
 855                         goto relock;
 856                 }
 857         } else {
 858                 xfs_rw_enter_trace(XFS_WRITE_ENTER, io, (void *)iovp, segs,
 859                                 *offset, ioflags);
 860                 ret = generic_file_buffered_write(iocb, iovp, segs,
 861                                 pos, offset, count, ret);
 862         }
 863
 864         current->backing_dev_info = NULL;
 865
 866         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
 867                 ret = wait_on_sync_kiocb(iocb);
 868
 869         if ((ret == -ENOSPC) &&
 870             DM_EVENT_ENABLED(vp->v_vfsp, xip, DM_EVENT_NOSPACE) &&
 871             !(ioflags & IO_INVIS)) {
 872
 873                 xfs_rwunlock(bdp, locktype);
 874                 if (need_i_mutex)
 875                         mutex_unlock(&inode->i_mutex);
 876                 error = XFS_SEND_NAMESP(xip->i_mount, DM_EVENT_NOSPACE, vp,
 877                                 DM_RIGHT_NULL, vp, DM_RIGHT_NULL, NULL, NULL,
 878                                 0, 0, 0); /* Delay flag intentionally  unused */
 879                 if (error)
 880                         goto out_nounlocks;
 881                 if (need_i_mutex)
 882                         mutex_lock(&inode->i_mutex);
 883                 xfs_rwlock(bdp, locktype);
 884                 pos = xip->i_d.di_size;
 885                 ret = 0;
 886                 goto retry;
 887         }
 888
 889         isize = i_size_read(inode);
 890         if (unlikely(ret < 0 && ret != -EFAULT && *offset > isize))
 891                 *offset = isize;
 892
 893         if (*offset > xip->i_d.di_size) {
 894                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 895                 if (*offset > xip->i_d.di_size) {
 896                         xip->i_d.di_size = *offset;
 897                         i_size_write(inode, *offset);
 898                         xip->i_update_core = 1;
 899                         xip->i_update_size = 1;
 900                 }
 901                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 902         }
 903
 904         error = -ret;
 905         if (ret <= 0)
 906                 goto out_unlock_internal;
 907
 908         XFS_STATS_ADD(xs_write_bytes, ret);
 909
 910         /* Handle various SYNC-type writes */
 911         if ((file->f_flags & O_SYNC) || IS_SYNC(inode)) {
 912                 error = xfs_write_sync_logforce(mp, xip);
 913                 if (error)
 914                         goto out_unlock_internal;
 915
 916                 xfs_rwunlock(bdp, locktype);
 917                 if (need_i_mutex)
 918                         mutex_unlock(&inode->i_mutex);
 919
 920                 error = sync_page_range(inode, mapping, pos, ret);
 921                 if (!error)
 922                         error = ret;
 923                 return error;
 924         }
 925
 926  out_unlock_internal:
 927         xfs_rwunlock(bdp, locktype);
 928  out_unlock_mutex:
 929         if (need_i_mutex)
 930                 mutex_unlock(&inode->i_mutex);
 931  out_nounlocks:
 932         return -error;
 933 }
 934
 935 /*
 936  * All xfs metadata buffers except log state machine buffers
 937  * get this attached as their b_bdstrat callback function.
 938  * This is so that we can catch a buffer
 939  * after prematurely unpinning it to forcibly shutdown the filesystem.
 940  */
 941 int
 942 xfs_bdstrat_cb(struct xfs_buf *bp)
 943 {
 944         xfs_mount_t     *mp;
 945
 946         mp = XFS_BUF_FSPRIVATE3(bp, xfs_mount_t *);
 947         if (!XFS_FORCED_SHUTDOWN(mp)) {
 948                 xfs_buf_iorequest(bp);
 949                 return 0;
 950         } else {
 951                 xfs_buftrace("XFS__BDSTRAT IOERROR", bp);
 952                 /*
 953                  * Metadata write that didn't get logged but
 954                  * written delayed anyway. These aren't associated
 955                  * with a transaction, and can be ignored.
 956                  */
 957                 if (XFS_BUF_IODONE_FUNC(bp) == NULL &&
 958                     (XFS_BUF_ISREAD(bp)) == 0)
 959                         return (xfs_bioerror_relse(bp));
 960                 else
 961                         return (xfs_bioerror(bp));
 962         }
 963 }
 964
 965
 966 int
 967 xfs_bmap(bhv_desc_t     *bdp,
 968         xfs_off_t       offset,
 969         ssize_t         count,
 970         int             flags,
 971         xfs_iomap_t     *iomapp,
 972         int             *niomaps)
 973 {
 974         xfs_inode_t     *ip = XFS_BHVTOI(bdp);
 975         xfs_iocore_t    *io = &ip->i_iocore;
 976
 977         ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFREG);
 978         ASSERT(((ip->i_d.di_flags & XFS_DIFLAG_REALTIME) != 0) ==
 979                ((ip->i_iocore.io_flags & XFS_IOCORE_RT) != 0));
 980
 981         return xfs_iomap(io, offset, count, flags, iomapp, niomaps);
 982 }
 983
 984 /*
 985  * Wrapper around bdstrat so that we can stop data
 986  * from going to disk in case we are shutting down the filesystem.
 987  * Typically user data goes thru this path; one of the exceptions
 988  * is the superblock.
 989  */
 990 int
 991 xfsbdstrat(
 992         struct xfs_mount        *mp,
 993         struct xfs_buf          *bp)
 994 {
 995         ASSERT(mp);
 996         if (!XFS_FORCED_SHUTDOWN(mp)) {
 997                 /* Grio redirection would go here
 998                  * if (XFS_BUF_IS_GRIO(bp)) {
 999                  */
1000
1001                 xfs_buf_iorequest(bp);
1002                 return 0;
1003         }
1004
1005         xfs_buftrace("XFSBDSTRAT IOERROR", bp);
1006         return (xfs_bioerror_relse(bp));
1007 }
1008
1009 /*
1010  * If the underlying (data/log/rt) device is readonly, there are some
1011  * operations that cannot proceed.
1012  */
1013 int
1014 xfs_dev_is_read_only(
1015         xfs_mount_t             *mp,
1016         char                    *message)
1017 {
1018         if (xfs_readonly_buftarg(mp->m_ddev_targp) ||
1019             xfs_readonly_buftarg(mp->m_logdev_targp) ||
1020             (mp->m_rtdev_targp && xfs_readonly_buftarg(mp->m_rtdev_targp))) {
1021                 cmn_err(CE_NOTE,
1022                         "XFS: %s required on read-only device.", message);
1023                 cmn_err(CE_NOTE,
1024                         "XFS: write access unavailable, cannot proceed.");
1025                 return EROFS;
1026         }
1027         return 0;
1028 }