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_dir.h"
  27 #include "xfs_dir2.h"
  28 #include "xfs_alloc.h"
  29 #include "xfs_dmapi.h"
  30 #include "xfs_quota.h"
  31 #include "xfs_mount.h"
  32 #include "xfs_bmap_btree.h"
  33 #include "xfs_alloc_btree.h"
  34 #include "xfs_ialloc_btree.h"
  35 #include "xfs_dir_sf.h"
  36 #include "xfs_dir2_sf.h"
  37 #include "xfs_attr_sf.h"
  38 #include "xfs_dinode.h"
  39 #include "xfs_inode.h"
  40 #include "xfs_bmap.h"
  41 #include "xfs_btree.h"
  42 #include "xfs_ialloc.h"
  43 #include "xfs_rtalloc.h"
  44 #include "xfs_error.h"
  45 #include "xfs_itable.h"
  46 #include "xfs_rw.h"
  47 #include "xfs_acl.h"
  48 #include "xfs_cap.h"
  49 #include "xfs_mac.h"
  50 #include "xfs_attr.h"
  51 #include "xfs_inode_item.h"
  52 #include "xfs_buf_item.h"
  53 #include "xfs_utils.h"
  54 #include "xfs_iomap.h"
  55
  56 #include <linux/capability.h>
  57 #include <linux/writeback.h>
  58
  59
  60 #if defined(XFS_RW_TRACE)
  61 void
  62 xfs_rw_enter_trace(
  63         int                     tag,
  64         xfs_iocore_t            *io,
  65         void                    *data,
  66         size_t                  segs,
  67         loff_t                  offset,
  68         int                     ioflags)
  69 {
  70         xfs_inode_t     *ip = XFS_IO_INODE(io);
  71
  72         if (ip->i_rwtrace == NULL)
  73                 return;
  74         ktrace_enter(ip->i_rwtrace,
  75                 (void *)(unsigned long)tag,
  76                 (void *)ip,
  77                 (void *)((unsigned long)((ip->i_d.di_size >> 32) & 0xffffffff)),
  78                 (void *)((unsigned long)(ip->i_d.di_size & 0xffffffff)),
  79                 (void *)data,
  80                 (void *)((unsigned long)segs),
  81                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
  82                 (void *)((unsigned long)(offset & 0xffffffff)),
  83                 (void *)((unsigned long)ioflags),
  84                 (void *)((unsigned long)((io->io_new_size >> 32) & 0xffffffff)),
  85                 (void *)((unsigned long)(io->io_new_size & 0xffffffff)),
  86                 (void *)((unsigned long)current_pid()),
  87                 (void *)NULL,
  88                 (void *)NULL,
  89                 (void *)NULL,
  90                 (void *)NULL);
  91 }
  92
  93 void
  94 xfs_inval_cached_trace(
  95         xfs_iocore_t    *io,
  96         xfs_off_t       offset,
  97         xfs_off_t       len,
  98         xfs_off_t       first,
  99         xfs_off_t       last)
 100 {
 101         xfs_inode_t     *ip = XFS_IO_INODE(io);
 102
 103         if (ip->i_rwtrace == NULL)
 104                 return;
 105         ktrace_enter(ip->i_rwtrace,
 106                 (void *)(__psint_t)XFS_INVAL_CACHED,
 107                 (void *)ip,
 108                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
 109                 (void *)((unsigned long)(offset & 0xffffffff)),
 110                 (void *)((unsigned long)((len >> 32) & 0xffffffff)),
 111                 (void *)((unsigned long)(len & 0xffffffff)),
 112                 (void *)((unsigned long)((first >> 32) & 0xffffffff)),
 113                 (void *)((unsigned long)(first & 0xffffffff)),
 114                 (void *)((unsigned long)((last >> 32) & 0xffffffff)),
 115                 (void *)((unsigned long)(last & 0xffffffff)),
 116                 (void *)((unsigned long)current_pid()),
 117                 (void *)NULL,
 118                 (void *)NULL,
 119                 (void *)NULL,
 120                 (void *)NULL,
 121                 (void *)NULL);
 122 }
 123 #endif
 124
 125 /*
 126  *      xfs_iozero
 127  *
 128  *      xfs_iozero clears the specified range of buffer supplied,
 129  *      and marks all the affected blocks as valid and modified.  If
 130  *      an affected block is not allocated, it will be allocated.  If
 131  *      an affected block is not completely overwritten, and is not
 132  *      valid before the operation, it will be read from disk before
 133  *      being partially zeroed.
 134  */
 135 STATIC int
 136 xfs_iozero(
 137         struct inode            *ip,    /* inode                        */
 138         loff_t                  pos,    /* offset in file               */
 139         size_t                  count,  /* size of data to zero         */
 140         loff_t                  end_size)       /* max file size to set */
 141 {
 142         unsigned                bytes;
 143         struct page             *page;
 144         struct address_space    *mapping;
 145         char                    *kaddr;
 146         int                     status;
 147
 148         mapping = ip->i_mapping;
 149         do {
 150                 unsigned long index, offset;
 151
 152                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
 153                 index = pos >> PAGE_CACHE_SHIFT;
 154                 bytes = PAGE_CACHE_SIZE - offset;
 155                 if (bytes > count)
 156                         bytes = count;
 157
 158                 status = -ENOMEM;
 159                 page = grab_cache_page(mapping, index);
 160                 if (!page)
 161                         break;
 162
 163                 kaddr = kmap(page);
 164                 status = mapping->a_ops->prepare_write(NULL, page, offset,
 165                                                         offset + bytes);
 166                 if (status) {
 167                         goto unlock;
 168                 }
 169
 170                 memset((void *) (kaddr + offset), 0, bytes);
 171                 flush_dcache_page(page);
 172                 status = mapping->a_ops->commit_write(NULL, page, offset,
 173                                                         offset + bytes);
 174                 if (!status) {
 175                         pos += bytes;
 176                         count -= bytes;
 177                         if (pos > i_size_read(ip))
 178                                 i_size_write(ip, pos < end_size ? pos : end_size);
 179                 }
 180
 181 unlock:
 182                 kunmap(page);
 183                 unlock_page(page);
 184                 page_cache_release(page);
 185                 if (status)
 186                         break;
 187         } while (count);
 188
 189         return (-status);
 190 }
 191
 192 ssize_t                 /* bytes read, or (-)  error */
 193 xfs_read(
 194         bhv_desc_t              *bdp,
 195         struct kiocb            *iocb,
 196         const struct iovec      *iovp,
 197         unsigned int            segs,
 198         loff_t                  *offset,
 199         int                     ioflags,
 200         cred_t                  *credp)
 201 {
 202         struct file             *file = iocb->ki_filp;
 203         struct inode            *inode = file->f_mapping->host;
 204         size_t                  size = 0;
 205         ssize_t                 ret;
 206         xfs_fsize_t             n;
 207         xfs_inode_t             *ip;
 208         xfs_mount_t             *mp;
 209         bhv_vnode_t             *vp;
 210         unsigned long           seg;
 211
 212         ip = XFS_BHVTOI(bdp);
 213         vp = BHV_TO_VNODE(bdp);
 214         mp = ip->i_mount;
 215
 216         XFS_STATS_INC(xs_read_calls);
 217
 218         /* START copy & waste from filemap.c */
 219         for (seg = 0; seg < segs; seg++) {
 220                 const struct iovec *iv = &iovp[seg];
 221
 222                 /*
 223                  * If any segment has a negative length, or the cumulative
 224                  * length ever wraps negative then return -EINVAL.
 225                  */
 226                 size += iv->iov_len;
 227                 if (unlikely((ssize_t)(size|iv->iov_len) < 0))
 228                         return XFS_ERROR(-EINVAL);
 229         }
 230         /* END copy & waste from filemap.c */
 231
 232         if (unlikely(ioflags & IO_ISDIRECT)) {
 233                 xfs_buftarg_t   *target =
 234                         (ip->i_d.di_flags & XFS_DIFLAG_REALTIME) ?
 235                                 mp->m_rtdev_targp : mp->m_ddev_targp;
 236                 if ((*offset & target->bt_smask) ||
 237                     (size & target->bt_smask)) {
 238                         if (*offset == ip->i_d.di_size) {
 239                                 return (0);
 240                         }
 241                         return -XFS_ERROR(EINVAL);
 242                 }
 243         }
 244
 245         n = XFS_MAXIOFFSET(mp) - *offset;
 246         if ((n <= 0) || (size == 0))
 247                 return 0;
 248
 249         if (n < size)
 250                 size = n;
 251
 252         if (XFS_FORCED_SHUTDOWN(mp))
 253                 return -EIO;
 254
 255         if (unlikely(ioflags & IO_ISDIRECT))
 256                 mutex_lock(&inode->i_mutex);
 257         xfs_ilock(ip, XFS_IOLOCK_SHARED);
 258
 259         if (DM_EVENT_ENABLED(vp->v_vfsp, ip, DM_EVENT_READ) &&
 260             !(ioflags & IO_INVIS)) {
 261                 vrwlock_t locktype = VRWLOCK_READ;
 262                 int dmflags = FILP_DELAY_FLAG(file) | DM_SEM_FLAG_RD(ioflags);
 263
 264                 ret = -XFS_SEND_DATA(mp, DM_EVENT_READ,
 265                                         BHV_TO_VNODE(bdp), *offset, size,
 266                                         dmflags, &locktype);
 267                 if (ret) {
 268                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 269                         goto unlock_mutex;
 270                 }
 271         }
 272
 273         if (unlikely((ioflags & IO_ISDIRECT) && VN_CACHED(vp)))
 274                 bhv_vop_flushinval_pages(vp, ctooff(offtoct(*offset)),
 275                                                 -1, FI_REMAPF_LOCKED);
 276
 277         xfs_rw_enter_trace(XFS_READ_ENTER, &ip->i_iocore,
 278                                 (void *)iovp, segs, *offset, ioflags);
 279         ret = __generic_file_aio_read(iocb, iovp, segs, offset);
 280         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
 281                 ret = wait_on_sync_kiocb(iocb);
 282         if (ret > 0)
 283                 XFS_STATS_ADD(xs_read_bytes, ret);
 284
 285         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 286
 287 unlock_mutex:
 288         if (unlikely(ioflags & IO_ISDIRECT))
 289                 mutex_unlock(&inode->i_mutex);
 290         return ret;
 291 }
 292
 293 ssize_t
 294 xfs_sendfile(
 295         bhv_desc_t              *bdp,
 296         struct file             *filp,
 297         loff_t                  *offset,
 298         int                     ioflags,
 299         size_t                  count,
 300         read_actor_t            actor,
 301         void                    *target,
 302         cred_t                  *credp)
 303 {
 304         xfs_inode_t             *ip = XFS_BHVTOI(bdp);
 305         xfs_mount_t             *mp = ip->i_mount;
 306         ssize_t                 ret;
 307
 308         XFS_STATS_INC(xs_read_calls);
 309         if (XFS_FORCED_SHUTDOWN(mp))
 310                 return -EIO;
 311
 312         xfs_ilock(ip, XFS_IOLOCK_SHARED);
 313
 314         if (DM_EVENT_ENABLED(BHV_TO_VNODE(bdp)->v_vfsp, ip, DM_EVENT_READ) &&
 315             (!(ioflags & IO_INVIS))) {
 316                 vrwlock_t locktype = VRWLOCK_READ;
 317                 int error;
 318
 319                 error = XFS_SEND_DATA(mp, DM_EVENT_READ, BHV_TO_VNODE(bdp),
 320                                       *offset, count,
 321                                       FILP_DELAY_FLAG(filp), &locktype);
 322                 if (error) {
 323                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 324                         return -error;
 325                 }
 326         }
 327         xfs_rw_enter_trace(XFS_SENDFILE_ENTER, &ip->i_iocore,
 328                    (void *)(unsigned long)target, count, *offset, ioflags);
 329         ret = generic_file_sendfile(filp, offset, count, actor, target);
 330         if (ret > 0)
 331                 XFS_STATS_ADD(xs_read_bytes, ret);
 332
 333         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 334         return ret;
 335 }
 336
 337 ssize_t
 338 xfs_splice_read(
 339         bhv_desc_t              *bdp,
 340         struct file             *infilp,
 341         loff_t                  *ppos,
 342         struct pipe_inode_info  *pipe,
 343         size_t                  count,
 344         int                     flags,
 345         int                     ioflags,
 346         cred_t                  *credp)
 347 {
 348         xfs_inode_t             *ip = XFS_BHVTOI(bdp);
 349         xfs_mount_t             *mp = ip->i_mount;
 350         ssize_t                 ret;
 351
 352         XFS_STATS_INC(xs_read_calls);
 353         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 354                 return -EIO;
 355
 356         xfs_ilock(ip, XFS_IOLOCK_SHARED);
 357
 358         if (DM_EVENT_ENABLED(BHV_TO_VNODE(bdp)->v_vfsp, ip, DM_EVENT_READ) &&
 359             (!(ioflags & IO_INVIS))) {
 360                 vrwlock_t locktype = VRWLOCK_READ;
 361                 int error;
 362
 363                 error = XFS_SEND_DATA(mp, DM_EVENT_READ, BHV_TO_VNODE(bdp),
 364                                         *ppos, count,
 365                                         FILP_DELAY_FLAG(infilp), &locktype);
 366                 if (error) {
 367                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 368                         return -error;
 369                 }
 370         }
 371         xfs_rw_enter_trace(XFS_SPLICE_READ_ENTER, &ip->i_iocore,
 372                            pipe, count, *ppos, ioflags);
 373         ret = generic_file_splice_read(infilp, ppos, pipe, count, flags);
 374         if (ret > 0)
 375                 XFS_STATS_ADD(xs_read_bytes, ret);
 376
 377         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 378         return ret;
 379 }
 380
 381 ssize_t
 382 xfs_splice_write(
 383         bhv_desc_t              *bdp,
 384         struct pipe_inode_info  *pipe,
 385         struct file             *outfilp,
 386         loff_t                  *ppos,
 387         size_t                  count,
 388         int                     flags,
 389         int                     ioflags,
 390         cred_t                  *credp)
 391 {
 392         xfs_inode_t             *ip = XFS_BHVTOI(bdp);
 393         xfs_mount_t             *mp = ip->i_mount;
 394         ssize_t                 ret;
 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                 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         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 422         return ret;
 423 }
 424
 425 /*
 426  * This routine is called to handle zeroing any space in the last
 427  * block of the file that is beyond the EOF.  We do this since the
 428  * size is being increased without writing anything to that block
 429  * and we don't want anyone to read the garbage on the disk.
 430  */
 431 STATIC int                              /* error (positive) */
 432 xfs_zero_last_block(
 433         struct inode    *ip,
 434         xfs_iocore_t    *io,
 435         xfs_fsize_t     isize,
 436         xfs_fsize_t     end_size)
 437 {
 438         xfs_fileoff_t   last_fsb;
 439         xfs_mount_t     *mp = io->io_mount;
 440         int             nimaps;
 441         int             zero_offset;
 442         int             zero_len;
 443         int             error = 0;
 444         xfs_bmbt_irec_t imap;
 445         loff_t          loff;
 446
 447         ASSERT(ismrlocked(io->io_lock, MR_UPDATE) != 0);
 448
 449         zero_offset = XFS_B_FSB_OFFSET(mp, isize);
 450         if (zero_offset == 0) {
 451                 /*
 452                  * There are no extra bytes in the last block on disk to
 453                  * zero, so return.
 454                  */
 455                 return 0;
 456         }
 457
 458         last_fsb = XFS_B_TO_FSBT(mp, isize);
 459         nimaps = 1;
 460         error = XFS_BMAPI(mp, NULL, io, last_fsb, 1, 0, NULL, 0, &imap,
 461                           &nimaps, NULL, NULL);
 462         if (error) {
 463                 return error;
 464         }
 465         ASSERT(nimaps > 0);
 466         /*
 467          * If the block underlying isize is just a hole, then there
 468          * is nothing to zero.
 469          */
 470         if (imap.br_startblock == HOLESTARTBLOCK) {
 471                 return 0;
 472         }
 473         /*
 474          * Zero the part of the last block beyond the EOF, and write it
 475          * out sync.  We need to drop the ilock while we do this so we
 476          * don't deadlock when the buffer cache calls back to us.
 477          */
 478         XFS_IUNLOCK(mp, io, XFS_ILOCK_EXCL| XFS_EXTSIZE_RD);
 479
 480         loff = XFS_FSB_TO_B(mp, last_fsb);
 481         zero_len = mp->m_sb.sb_blocksize - zero_offset;
 482         error = xfs_iozero(ip, loff + zero_offset, zero_len, end_size);
 483
 484         XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 485         ASSERT(error >= 0);
 486         return error;
 487 }
 488
 489 /*
 490  * Zero any on disk space between the current EOF and the new,
 491  * larger EOF.  This handles the normal case of zeroing the remainder
 492  * of the last block in the file and the unusual case of zeroing blocks
 493  * out beyond the size of the file.  This second case only happens
 494  * with fixed size extents and when the system crashes before the inode
 495  * size was updated but after blocks were allocated.  If fill is set,
 496  * then any holes in the range are filled and zeroed.  If not, the holes
 497  * are left alone as holes.
 498  */
 499
 500 int                                     /* error (positive) */
 501 xfs_zero_eof(
 502         bhv_vnode_t     *vp,
 503         xfs_iocore_t    *io,
 504         xfs_off_t       offset,         /* starting I/O offset */
 505         xfs_fsize_t     isize,          /* current inode size */
 506         xfs_fsize_t     end_size)       /* terminal inode size */
 507 {
 508         struct inode    *ip = vn_to_inode(vp);
 509         xfs_fileoff_t   start_zero_fsb;
 510         xfs_fileoff_t   end_zero_fsb;
 511         xfs_fileoff_t   zero_count_fsb;
 512         xfs_fileoff_t   last_fsb;
 513         xfs_mount_t     *mp = io->io_mount;
 514         int             nimaps;
 515         int             error = 0;
 516         xfs_bmbt_irec_t imap;
 517
 518         ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
 519         ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
 520         ASSERT(offset > isize);
 521
 522         /*
 523          * First handle zeroing the block on which isize resides.
 524          * We only zero a part of that block so it is handled specially.
 525          */
 526         error = xfs_zero_last_block(ip, io, isize, end_size);
 527         if (error) {
 528                 ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
 529                 ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
 530                 return error;
 531         }
 532
 533         /*
 534          * Calculate the range between the new size and the old
 535          * where blocks needing to be zeroed may exist.  To get the
 536          * block where the last byte in the file currently resides,
 537          * we need to subtract one from the size and truncate back
 538          * to a block boundary.  We subtract 1 in case the size is
 539          * exactly on a block boundary.
 540          */
 541         last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
 542         start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
 543         end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
 544         ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
 545         if (last_fsb == end_zero_fsb) {
 546                 /*
 547                  * The size was only incremented on its last block.
 548                  * We took care of that above, so just return.
 549                  */
 550                 return 0;
 551         }
 552
 553         ASSERT(start_zero_fsb <= end_zero_fsb);
 554         while (start_zero_fsb <= end_zero_fsb) {
 555                 nimaps = 1;
 556                 zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
 557                 error = XFS_BMAPI(mp, NULL, io, start_zero_fsb, zero_count_fsb,
 558                                   0, NULL, 0, &imap, &nimaps, NULL, NULL);
 559                 if (error) {
 560                         ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
 561                         ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
 562                         return error;
 563                 }
 564                 ASSERT(nimaps > 0);
 565
 566                 if (imap.br_state == XFS_EXT_UNWRITTEN ||
 567                     imap.br_startblock == HOLESTARTBLOCK) {
 568                         /*
 569                          * This loop handles initializing pages that were
 570                          * partially initialized by the code below this
 571                          * loop. It basically zeroes the part of the page
 572                          * that sits on a hole and sets the page as P_HOLE
 573                          * and calls remapf if it is a mapped file.
 574                          */
 575                         start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 576                         ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 577                         continue;
 578                 }
 579
 580                 /*
 581                  * There are blocks we need to zero.
 582                  * Drop the inode lock while we're doing the I/O.
 583                  * We'll still have the iolock to protect us.
 584                  */
 585                 XFS_IUNLOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 586
 587                 error = xfs_iozero(ip,
 588                                    XFS_FSB_TO_B(mp, start_zero_fsb),
 589                                    XFS_FSB_TO_B(mp, imap.br_blockcount),
 590                                    end_size);
 591                 if (error) {
 592                         goto out_lock;
 593                 }
 594
 595                 start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 596                 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 597
 598                 XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 599         }
 600
 601         return 0;
 602
 603 out_lock:
 604
 605         XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 606         ASSERT(error >= 0);
 607         return error;
 608 }
 609
 610 ssize_t                         /* bytes written, or (-) error */
 611 xfs_write(
 612         bhv_desc_t              *bdp,
 613         struct kiocb            *iocb,
 614         const struct iovec      *iovp,
 615         unsigned int            nsegs,
 616         loff_t                  *offset,
 617         int                     ioflags,
 618         cred_t                  *credp)
 619 {
 620         struct file             *file = iocb->ki_filp;
 621         struct address_space    *mapping = file->f_mapping;
 622         struct inode            *inode = mapping->host;
 623         unsigned long           segs = nsegs;
 624         xfs_inode_t             *xip;
 625         xfs_mount_t             *mp;
 626         ssize_t                 ret = 0, error = 0;
 627         xfs_fsize_t             isize, new_size;
 628         xfs_iocore_t            *io;
 629         bhv_vnode_t             *vp;
 630         unsigned long           seg;
 631         int                     iolock;
 632         int                     eventsent = 0;
 633         vrwlock_t               locktype;
 634         size_t                  ocount = 0, count;
 635         loff_t                  pos;
 636         int                     need_i_mutex = 1, need_flush = 0;
 637
 638         XFS_STATS_INC(xs_write_calls);
 639
 640         vp = BHV_TO_VNODE(bdp);
 641         xip = XFS_BHVTOI(bdp);
 642
 643         for (seg = 0; seg < segs; seg++) {
 644                 const struct iovec *iv = &iovp[seg];
 645
 646                 /*
 647                  * If any segment has a negative length, or the cumulative
 648                  * length ever wraps negative then return -EINVAL.
 649                  */
 650                 ocount += iv->iov_len;
 651                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
 652                         return -EINVAL;
 653                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
 654                         continue;
 655                 if (seg == 0)
 656                         return -EFAULT;
 657                 segs = seg;
 658                 ocount -= iv->iov_len;  /* This segment is no good */
 659                 break;
 660         }
 661
 662         count = ocount;
 663         pos = *offset;
 664
 665         if (count == 0)
 666                 return 0;
 667
 668         io = &xip->i_iocore;
 669         mp = io->io_mount;
 670
 671         if (XFS_FORCED_SHUTDOWN(mp))
 672                 return -EIO;
 673
 674         fs_check_frozen(vp->v_vfsp, SB_FREEZE_WRITE);
 675
 676         if (ioflags & IO_ISDIRECT) {
 677                 xfs_buftarg_t   *target =
 678                         (xip->i_d.di_flags & XFS_DIFLAG_REALTIME) ?
 679                                 mp->m_rtdev_targp : mp->m_ddev_targp;
 680
 681                 if ((pos & target->bt_smask) || (count & target->bt_smask))
 682                         return XFS_ERROR(-EINVAL);
 683
 684                 if (!VN_CACHED(vp) && pos < i_size_read(inode))
 685                         need_i_mutex = 0;
 686
 687                 if (VN_CACHED(vp))
 688                         need_flush = 1;
 689         }
 690
 691 relock:
 692         if (need_i_mutex) {
 693                 iolock = XFS_IOLOCK_EXCL;
 694                 locktype = VRWLOCK_WRITE;
 695
 696                 mutex_lock(&inode->i_mutex);
 697         } else {
 698                 iolock = XFS_IOLOCK_SHARED;
 699                 locktype = VRWLOCK_WRITE_DIRECT;
 700         }
 701
 702         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
 703
 704         isize = i_size_read(inode);
 705
 706         if (file->f_flags & O_APPEND)
 707                 *offset = isize;
 708
 709 start:
 710         error = -generic_write_checks(file, &pos, &count,
 711                                         S_ISBLK(inode->i_mode));
 712         if (error) {
 713                 xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 714                 goto out_unlock_mutex;
 715         }
 716
 717         new_size = pos + count;
 718         if (new_size > isize)
 719                 io->io_new_size = new_size;
 720
 721         if ((DM_EVENT_ENABLED(vp->v_vfsp, xip, DM_EVENT_WRITE) &&
 722             !(ioflags & IO_INVIS) && !eventsent)) {
 723                 loff_t          savedsize = pos;
 724                 int             dmflags = FILP_DELAY_FLAG(file);
 725
 726                 if (need_i_mutex)
 727                         dmflags |= DM_FLAGS_IMUX;
 728
 729                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 730                 error = XFS_SEND_DATA(xip->i_mount, DM_EVENT_WRITE, vp,
 731                                       pos, count,
 732                                       dmflags, &locktype);
 733                 if (error) {
 734                         xfs_iunlock(xip, iolock);
 735                         goto out_unlock_mutex;
 736                 }
 737                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 738                 eventsent = 1;
 739
 740                 /*
 741                  * The iolock was dropped and reacquired in XFS_SEND_DATA
 742                  * so we have to recheck the size when appending.
 743                  * We will only "goto start;" once, since having sent the
 744                  * event prevents another call to XFS_SEND_DATA, which is
 745                  * what allows the size to change in the first place.
 746                  */
 747                 if ((file->f_flags & O_APPEND) && savedsize != isize) {
 748                         pos = isize = xip->i_d.di_size;
 749                         goto start;
 750                 }
 751         }
 752
 753         if (likely(!(ioflags & IO_INVIS))) {
 754                 file_update_time(file);
 755                 xfs_ichgtime_fast(xip, inode,
 756                                   XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
 757         }
 758
 759         /*
 760          * If the offset is beyond the size of the file, we have a couple
 761          * of things to do. First, if there is already space allocated
 762          * we need to either create holes or zero the disk or ...
 763          *
 764          * If there is a page where the previous size lands, we need
 765          * to zero it out up to the new size.
 766          */
 767
 768         if (pos > isize) {
 769                 error = xfs_zero_eof(BHV_TO_VNODE(bdp), io, pos,
 770                                         isize, pos + count);
 771                 if (error) {
 772                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 773                         goto out_unlock_mutex;
 774                 }
 775         }
 776         xfs_iunlock(xip, XFS_ILOCK_EXCL);
 777
 778         /*
 779          * If we're writing the file then make sure to clear the
 780          * setuid and setgid bits if the process is not being run
 781          * by root.  This keeps people from modifying setuid and
 782          * setgid binaries.
 783          */
 784
 785         if (((xip->i_d.di_mode & S_ISUID) ||
 786             ((xip->i_d.di_mode & (S_ISGID | S_IXGRP)) ==
 787                 (S_ISGID | S_IXGRP))) &&
 788              !capable(CAP_FSETID)) {
 789                 error = xfs_write_clear_setuid(xip);
 790                 if (likely(!error))
 791                         error = -remove_suid(file->f_dentry);
 792                 if (unlikely(error)) {
 793                         xfs_iunlock(xip, iolock);
 794                         goto out_unlock_mutex;
 795                 }
 796         }
 797
 798 retry:
 799         /* We can write back this queue in page reclaim */
 800         current->backing_dev_info = mapping->backing_dev_info;
 801
 802         if ((ioflags & IO_ISDIRECT)) {
 803                 if (need_flush) {
 804                         xfs_inval_cached_trace(io, pos, -1,
 805                                         ctooff(offtoct(pos)), -1);
 806                         bhv_vop_flushinval_pages(vp, ctooff(offtoct(pos)),
 807                                         -1, FI_REMAPF_LOCKED);
 808                 }
 809
 810                 if (need_i_mutex) {
 811                         /* demote the lock now the cached pages are gone */
 812                         XFS_ILOCK_DEMOTE(mp, io, XFS_IOLOCK_EXCL);
 813                         mutex_unlock(&inode->i_mutex);
 814
 815                         iolock = XFS_IOLOCK_SHARED;
 816                         locktype = VRWLOCK_WRITE_DIRECT;
 817                         need_i_mutex = 0;
 818                 }
 819
 820                 xfs_rw_enter_trace(XFS_DIOWR_ENTER, io, (void *)iovp, segs,
 821                                 *offset, ioflags);
 822                 ret = generic_file_direct_write(iocb, iovp,
 823                                 &segs, pos, offset, count, ocount);
 824
 825                 /*
 826                  * direct-io write to a hole: fall through to buffered I/O
 827                  * for completing the rest of the request.
 828                  */
 829                 if (ret >= 0 && ret != count) {
 830                         XFS_STATS_ADD(xs_write_bytes, ret);
 831
 832                         pos += ret;
 833                         count -= ret;
 834
 835                         need_i_mutex = 1;
 836                         ioflags &= ~IO_ISDIRECT;
 837                         xfs_iunlock(xip, iolock);
 838                         goto relock;
 839                 }
 840         } else {
 841                 xfs_rw_enter_trace(XFS_WRITE_ENTER, io, (void *)iovp, segs,
 842                                 *offset, ioflags);
 843                 ret = generic_file_buffered_write(iocb, iovp, segs,
 844                                 pos, offset, count, ret);
 845         }
 846
 847         current->backing_dev_info = NULL;
 848
 849         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
 850                 ret = wait_on_sync_kiocb(iocb);
 851
 852         if ((ret == -ENOSPC) &&
 853             DM_EVENT_ENABLED(vp->v_vfsp, xip, DM_EVENT_NOSPACE) &&
 854             !(ioflags & IO_INVIS)) {
 855
 856                 xfs_rwunlock(bdp, locktype);
 857                 if (need_i_mutex)
 858                         mutex_unlock(&inode->i_mutex);
 859                 error = XFS_SEND_NAMESP(xip->i_mount, DM_EVENT_NOSPACE, vp,
 860                                 DM_RIGHT_NULL, vp, DM_RIGHT_NULL, NULL, NULL,
 861                                 0, 0, 0); /* Delay flag intentionally  unused */
 862                 if (error)
 863                         goto out_nounlocks;
 864                 if (need_i_mutex)
 865                         mutex_lock(&inode->i_mutex);
 866                 xfs_rwlock(bdp, locktype);
 867                 pos = xip->i_d.di_size;
 868                 ret = 0;
 869                 goto retry;
 870         }
 871
 872         isize = i_size_read(inode);
 873         if (unlikely(ret < 0 && ret != -EFAULT && *offset > isize))
 874                 *offset = isize;
 875
 876         if (*offset > xip->i_d.di_size) {
 877                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 878                 if (*offset > xip->i_d.di_size) {
 879                         xip->i_d.di_size = *offset;
 880                         i_size_write(inode, *offset);
 881                         xip->i_update_core = 1;
 882                         xip->i_update_size = 1;
 883                 }
 884                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 885         }
 886
 887         error = -ret;
 888         if (ret <= 0)
 889                 goto out_unlock_internal;
 890
 891         XFS_STATS_ADD(xs_write_bytes, ret);
 892
 893         /* Handle various SYNC-type writes */
 894         if ((file->f_flags & O_SYNC) || IS_SYNC(inode)) {
 895                 /*
 896                  * If we're treating this as O_DSYNC and we have not updated the
 897                  * size, force the log.
 898                  */
 899                 if (!(mp->m_flags & XFS_MOUNT_OSYNCISOSYNC) &&
 900                     !(xip->i_update_size)) {
 901                         xfs_inode_log_item_t    *iip = xip->i_itemp;
 902
 903                         /*
 904                          * If an allocation transaction occurred
 905                          * without extending the size, then we have to force
 906                          * the log up the proper point to ensure that the
 907                          * allocation is permanent.  We can't count on
 908                          * the fact that buffered writes lock out direct I/O
 909                          * writes - the direct I/O write could have extended
 910                          * the size nontransactionally, then finished before
 911                          * we started.  xfs_write_file will think that the file
 912                          * didn't grow but the update isn't safe unless the
 913                          * size change is logged.
 914                          *
 915                          * Force the log if we've committed a transaction
 916                          * against the inode or if someone else has and
 917                          * the commit record hasn't gone to disk (e.g.
 918                          * the inode is pinned).  This guarantees that
 919                          * all changes affecting the inode are permanent
 920                          * when we return.
 921                          */
 922                         if (iip && iip->ili_last_lsn) {
 923                                 xfs_log_force(mp, iip->ili_last_lsn,
 924                                                 XFS_LOG_FORCE | XFS_LOG_SYNC);
 925                         } else if (xfs_ipincount(xip) > 0) {
 926                                 xfs_log_force(mp, (xfs_lsn_t)0,
 927                                                 XFS_LOG_FORCE | XFS_LOG_SYNC);
 928                         }
 929
 930                 } else {
 931                         xfs_trans_t     *tp;
 932
 933                         /*
 934                          * O_SYNC or O_DSYNC _with_ a size update are handled
 935                          * the same way.
 936                          *
 937                          * If the write was synchronous then we need to make
 938                          * sure that the inode modification time is permanent.
 939                          * We'll have updated the timestamp above, so here
 940                          * we use a synchronous transaction to log the inode.
 941                          * It's not fast, but it's necessary.
 942                          *
 943                          * If this a dsync write and the size got changed
 944                          * non-transactionally, then we need to ensure that
 945                          * the size change gets logged in a synchronous
 946                          * transaction.
 947                          */
 948
 949                         tp = xfs_trans_alloc(mp, XFS_TRANS_WRITE_SYNC);
 950                         if ((error = xfs_trans_reserve(tp, 0,
 951                                                       XFS_SWRITE_LOG_RES(mp),
 952                                                       0, 0, 0))) {
 953                                 /* Transaction reserve failed */
 954                                 xfs_trans_cancel(tp, 0);
 955                         } else {
 956                                 /* Transaction reserve successful */
 957                                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 958                                 xfs_trans_ijoin(tp, xip, XFS_ILOCK_EXCL);
 959                                 xfs_trans_ihold(tp, xip);
 960                                 xfs_trans_log_inode(tp, xip, XFS_ILOG_CORE);
 961                                 xfs_trans_set_sync(tp);
 962                                 error = xfs_trans_commit(tp, 0, NULL);
 963                                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 964                         }
 965                         if (error)
 966                                 goto out_unlock_internal;
 967                 }
 968
 969                 xfs_rwunlock(bdp, locktype);
 970                 if (need_i_mutex)
 971                         mutex_unlock(&inode->i_mutex);
 972
 973                 error = sync_page_range(inode, mapping, pos, ret);
 974                 if (!error)
 975                         error = ret;
 976                 return error;
 977         }
 978
 979  out_unlock_internal:
 980         xfs_rwunlock(bdp, locktype);
 981  out_unlock_mutex:
 982         if (need_i_mutex)
 983                 mutex_unlock(&inode->i_mutex);
 984  out_nounlocks:
 985         return -error;
 986 }
 987
 988 /*
 989  * All xfs metadata buffers except log state machine buffers
 990  * get this attached as their b_bdstrat callback function.
 991  * This is so that we can catch a buffer
 992  * after prematurely unpinning it to forcibly shutdown the filesystem.
 993  */
 994 int
 995 xfs_bdstrat_cb(struct xfs_buf *bp)
 996 {
 997         xfs_mount_t     *mp;
 998
 999         mp = XFS_BUF_FSPRIVATE3(bp, xfs_mount_t *);
1000         if (!XFS_FORCED_SHUTDOWN(mp)) {
1001                 xfs_buf_iorequest(bp);
1002                 return 0;
1003         } else {
1004                 xfs_buftrace("XFS__BDSTRAT IOERROR", bp);
1005                 /*
1006                  * Metadata write that didn't get logged but
1007                  * written delayed anyway. These aren't associated
1008                  * with a transaction, and can be ignored.
1009                  */
1010                 if (XFS_BUF_IODONE_FUNC(bp) == NULL &&
1011                     (XFS_BUF_ISREAD(bp)) == 0)
1012                         return (xfs_bioerror_relse(bp));
1013                 else
1014                         return (xfs_bioerror(bp));
1015         }
1016 }
1017
1018
1019 int
1020 xfs_bmap(bhv_desc_t     *bdp,
1021         xfs_off_t       offset,
1022         ssize_t         count,
1023         int             flags,
1024         xfs_iomap_t     *iomapp,
1025         int             *niomaps)
1026 {
1027         xfs_inode_t     *ip = XFS_BHVTOI(bdp);
1028         xfs_iocore_t    *io = &ip->i_iocore;
1029
1030         ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFREG);
1031         ASSERT(((ip->i_d.di_flags & XFS_DIFLAG_REALTIME) != 0) ==
1032                ((ip->i_iocore.io_flags & XFS_IOCORE_RT) != 0));
1033
1034         return xfs_iomap(io, offset, count, flags, iomapp, niomaps);
1035 }
1036
1037 /*
1038  * Wrapper around bdstrat so that we can stop data
1039  * from going to disk in case we are shutting down the filesystem.
1040  * Typically user data goes thru this path; one of the exceptions
1041  * is the superblock.
1042  */
1043 int
1044 xfsbdstrat(
1045         struct xfs_mount        *mp,
1046         struct xfs_buf          *bp)
1047 {
1048         ASSERT(mp);
1049         if (!XFS_FORCED_SHUTDOWN(mp)) {
1050                 /* Grio redirection would go here
1051                  * if (XFS_BUF_IS_GRIO(bp)) {
1052                  */
1053
1054                 xfs_buf_iorequest(bp);
1055                 return 0;
1056         }
1057
1058         xfs_buftrace("XFSBDSTRAT IOERROR", bp);
1059         return (xfs_bioerror_relse(bp));
1060 }
1061
1062 /*
1063  * If the underlying (data/log/rt) device is readonly, there are some
1064  * operations that cannot proceed.
1065  */
1066 int
1067 xfs_dev_is_read_only(
1068         xfs_mount_t             *mp,
1069         char                    *message)
1070 {
1071         if (xfs_readonly_buftarg(mp->m_ddev_targp) ||
1072             xfs_readonly_buftarg(mp->m_logdev_targp) ||
1073             (mp->m_rtdev_targp && xfs_readonly_buftarg(mp->m_rtdev_targp))) {
1074                 cmn_err(CE_NOTE,
1075                         "XFS: %s required on read-only device.", message);
1076                 cmn_err(CE_NOTE,
1077                         "XFS: write access unavailable, cannot proceed.");
1078                 return EROFS;
1079         }
1080         return 0;
1081 }