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 = 0;
 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_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                         ret = bhv_vop_flushinval_pages(vp, ctooff(offtoct(*offset)),
 267                                                  -1, FI_REMAPF_LOCKED);
 268                 mutex_unlock(&inode->i_mutex);
 269                 if (ret) {
 270                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 271                         return ret;
 272                 }
 273         }
 274
 275         xfs_rw_enter_trace(XFS_READ_ENTER, &ip->i_iocore,
 276                                 (void *)iovp, segs, *offset, ioflags);
 277
 278         iocb->ki_pos = *offset;
 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         return ret;
 287 }
 288
 289 ssize_t
 290 xfs_sendfile(
 291         bhv_desc_t              *bdp,
 292         struct file             *filp,
 293         loff_t                  *offset,
 294         int                     ioflags,
 295         size_t                  count,
 296         read_actor_t            actor,
 297         void                    *target,
 298         cred_t                  *credp)
 299 {
 300         xfs_inode_t             *ip = XFS_BHVTOI(bdp);
 301         xfs_mount_t             *mp = ip->i_mount;
 302         ssize_t                 ret;
 303
 304         XFS_STATS_INC(xs_read_calls);
 305         if (XFS_FORCED_SHUTDOWN(mp))
 306                 return -EIO;
 307
 308         xfs_ilock(ip, XFS_IOLOCK_SHARED);
 309
 310         if (DM_EVENT_ENABLED(BHV_TO_VNODE(bdp)->v_vfsp, ip, DM_EVENT_READ) &&
 311             (!(ioflags & IO_INVIS))) {
 312                 bhv_vrwlock_t locktype = VRWLOCK_READ;
 313                 int error;
 314
 315                 error = XFS_SEND_DATA(mp, DM_EVENT_READ, BHV_TO_VNODE(bdp),
 316                                       *offset, count,
 317                                       FILP_DELAY_FLAG(filp), &locktype);
 318                 if (error) {
 319                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 320                         return -error;
 321                 }
 322         }
 323         xfs_rw_enter_trace(XFS_SENDFILE_ENTER, &ip->i_iocore,
 324                    (void *)(unsigned long)target, count, *offset, ioflags);
 325         ret = generic_file_sendfile(filp, offset, count, actor, target);
 326         if (ret > 0)
 327                 XFS_STATS_ADD(xs_read_bytes, ret);
 328
 329         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 330         return ret;
 331 }
 332
 333 ssize_t
 334 xfs_splice_read(
 335         bhv_desc_t              *bdp,
 336         struct file             *infilp,
 337         loff_t                  *ppos,
 338         struct pipe_inode_info  *pipe,
 339         size_t                  count,
 340         int                     flags,
 341         int                     ioflags,
 342         cred_t                  *credp)
 343 {
 344         xfs_inode_t             *ip = XFS_BHVTOI(bdp);
 345         xfs_mount_t             *mp = ip->i_mount;
 346         ssize_t                 ret;
 347
 348         XFS_STATS_INC(xs_read_calls);
 349         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 350                 return -EIO;
 351
 352         xfs_ilock(ip, XFS_IOLOCK_SHARED);
 353
 354         if (DM_EVENT_ENABLED(BHV_TO_VNODE(bdp)->v_vfsp, ip, DM_EVENT_READ) &&
 355             (!(ioflags & IO_INVIS))) {
 356                 bhv_vrwlock_t locktype = VRWLOCK_READ;
 357                 int error;
 358
 359                 error = XFS_SEND_DATA(mp, DM_EVENT_READ, BHV_TO_VNODE(bdp),
 360                                         *ppos, count,
 361                                         FILP_DELAY_FLAG(infilp), &locktype);
 362                 if (error) {
 363                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 364                         return -error;
 365                 }
 366         }
 367         xfs_rw_enter_trace(XFS_SPLICE_READ_ENTER, &ip->i_iocore,
 368                            pipe, count, *ppos, ioflags);
 369         ret = generic_file_splice_read(infilp, ppos, pipe, count, flags);
 370         if (ret > 0)
 371                 XFS_STATS_ADD(xs_read_bytes, ret);
 372
 373         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 374         return ret;
 375 }
 376
 377 ssize_t
 378 xfs_splice_write(
 379         bhv_desc_t              *bdp,
 380         struct pipe_inode_info  *pipe,
 381         struct file             *outfilp,
 382         loff_t                  *ppos,
 383         size_t                  count,
 384         int                     flags,
 385         int                     ioflags,
 386         cred_t                  *credp)
 387 {
 388         xfs_inode_t             *ip = XFS_BHVTOI(bdp);
 389         xfs_mount_t             *mp = ip->i_mount;
 390         xfs_iocore_t            *io = &ip->i_iocore;
 391         ssize_t                 ret;
 392         struct inode            *inode = outfilp->f_mapping->host;
 393         xfs_fsize_t             isize, new_size;
 394
 395         XFS_STATS_INC(xs_write_calls);
 396         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 397                 return -EIO;
 398
 399         xfs_ilock(ip, XFS_IOLOCK_EXCL);
 400
 401         if (DM_EVENT_ENABLED(BHV_TO_VNODE(bdp)->v_vfsp, ip, DM_EVENT_WRITE) &&
 402             (!(ioflags & IO_INVIS))) {
 403                 bhv_vrwlock_t locktype = VRWLOCK_WRITE;
 404                 int error;
 405
 406                 error = XFS_SEND_DATA(mp, DM_EVENT_WRITE, BHV_TO_VNODE(bdp),
 407                                         *ppos, count,
 408                                         FILP_DELAY_FLAG(outfilp), &locktype);
 409                 if (error) {
 410                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 411                         return -error;
 412                 }
 413         }
 414
 415         new_size = *ppos + count;
 416
 417         xfs_ilock(ip, XFS_ILOCK_EXCL);
 418         if (new_size > ip->i_size)
 419                 io->io_new_size = new_size;
 420         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 421
 422         xfs_rw_enter_trace(XFS_SPLICE_WRITE_ENTER, &ip->i_iocore,
 423                            pipe, count, *ppos, ioflags);
 424         ret = generic_file_splice_write(pipe, outfilp, ppos, count, flags);
 425         if (ret > 0)
 426                 XFS_STATS_ADD(xs_write_bytes, ret);
 427
 428         isize = i_size_read(inode);
 429         if (unlikely(ret < 0 && ret != -EFAULT && *ppos > isize))
 430                 *ppos = isize;
 431
 432         if (*ppos > ip->i_size) {
 433                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 434                 if (*ppos > ip->i_size)
 435                         ip->i_size = *ppos;
 436                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 437         }
 438
 439         if (io->io_new_size) {
 440                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 441                 io->io_new_size = 0;
 442                 if (ip->i_d.di_size > ip->i_size)
 443                         ip->i_d.di_size = ip->i_size;
 444                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 445         }
 446         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 447         return ret;
 448 }
 449
 450 /*
 451  * This routine is called to handle zeroing any space in the last
 452  * block of the file that is beyond the EOF.  We do this since the
 453  * size is being increased without writing anything to that block
 454  * and we don't want anyone to read the garbage on the disk.
 455  */
 456 STATIC int                              /* error (positive) */
 457 xfs_zero_last_block(
 458         struct inode    *ip,
 459         xfs_iocore_t    *io,
 460         xfs_fsize_t     offset,
 461         xfs_fsize_t     isize)
 462 {
 463         xfs_fileoff_t   last_fsb;
 464         xfs_mount_t     *mp = io->io_mount;
 465         int             nimaps;
 466         int             zero_offset;
 467         int             zero_len;
 468         int             error = 0;
 469         xfs_bmbt_irec_t imap;
 470
 471         ASSERT(ismrlocked(io->io_lock, MR_UPDATE) != 0);
 472
 473         zero_offset = XFS_B_FSB_OFFSET(mp, isize);
 474         if (zero_offset == 0) {
 475                 /*
 476                  * There are no extra bytes in the last block on disk to
 477                  * zero, so return.
 478                  */
 479                 return 0;
 480         }
 481
 482         last_fsb = XFS_B_TO_FSBT(mp, isize);
 483         nimaps = 1;
 484         error = XFS_BMAPI(mp, NULL, io, last_fsb, 1, 0, NULL, 0, &imap,
 485                           &nimaps, NULL, NULL);
 486         if (error) {
 487                 return error;
 488         }
 489         ASSERT(nimaps > 0);
 490         /*
 491          * If the block underlying isize is just a hole, then there
 492          * is nothing to zero.
 493          */
 494         if (imap.br_startblock == HOLESTARTBLOCK) {
 495                 return 0;
 496         }
 497         /*
 498          * Zero the part of the last block beyond the EOF, and write it
 499          * out sync.  We need to drop the ilock while we do this so we
 500          * don't deadlock when the buffer cache calls back to us.
 501          */
 502         XFS_IUNLOCK(mp, io, XFS_ILOCK_EXCL| XFS_EXTSIZE_RD);
 503
 504         zero_len = mp->m_sb.sb_blocksize - zero_offset;
 505         if (isize + zero_len > offset)
 506                 zero_len = offset - isize;
 507         error = xfs_iozero(ip, isize, zero_len);
 508
 509         XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 510         ASSERT(error >= 0);
 511         return error;
 512 }
 513
 514 /*
 515  * Zero any on disk space between the current EOF and the new,
 516  * larger EOF.  This handles the normal case of zeroing the remainder
 517  * of the last block in the file and the unusual case of zeroing blocks
 518  * out beyond the size of the file.  This second case only happens
 519  * with fixed size extents and when the system crashes before the inode
 520  * size was updated but after blocks were allocated.  If fill is set,
 521  * then any holes in the range are filled and zeroed.  If not, the holes
 522  * are left alone as holes.
 523  */
 524
 525 int                                     /* error (positive) */
 526 xfs_zero_eof(
 527         bhv_vnode_t     *vp,
 528         xfs_iocore_t    *io,
 529         xfs_off_t       offset,         /* starting I/O offset */
 530         xfs_fsize_t     isize)          /* current inode size */
 531 {
 532         struct inode    *ip = vn_to_inode(vp);
 533         xfs_fileoff_t   start_zero_fsb;
 534         xfs_fileoff_t   end_zero_fsb;
 535         xfs_fileoff_t   zero_count_fsb;
 536         xfs_fileoff_t   last_fsb;
 537         xfs_fileoff_t   zero_off;
 538         xfs_fsize_t     zero_len;
 539         xfs_mount_t     *mp = io->io_mount;
 540         int             nimaps;
 541         int             error = 0;
 542         xfs_bmbt_irec_t imap;
 543
 544         ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
 545         ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
 546         ASSERT(offset > isize);
 547
 548         /*
 549          * First handle zeroing the block on which isize resides.
 550          * We only zero a part of that block so it is handled specially.
 551          */
 552         error = xfs_zero_last_block(ip, io, offset, isize);
 553         if (error) {
 554                 ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
 555                 ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
 556                 return error;
 557         }
 558
 559         /*
 560          * Calculate the range between the new size and the old
 561          * where blocks needing to be zeroed may exist.  To get the
 562          * block where the last byte in the file currently resides,
 563          * we need to subtract one from the size and truncate back
 564          * to a block boundary.  We subtract 1 in case the size is
 565          * exactly on a block boundary.
 566          */
 567         last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
 568         start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
 569         end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
 570         ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
 571         if (last_fsb == end_zero_fsb) {
 572                 /*
 573                  * The size was only incremented on its last block.
 574                  * We took care of that above, so just return.
 575                  */
 576                 return 0;
 577         }
 578
 579         ASSERT(start_zero_fsb <= end_zero_fsb);
 580         while (start_zero_fsb <= end_zero_fsb) {
 581                 nimaps = 1;
 582                 zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
 583                 error = XFS_BMAPI(mp, NULL, io, start_zero_fsb, zero_count_fsb,
 584                                   0, NULL, 0, &imap, &nimaps, NULL, NULL);
 585                 if (error) {
 586                         ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
 587                         ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
 588                         return error;
 589                 }
 590                 ASSERT(nimaps > 0);
 591
 592                 if (imap.br_state == XFS_EXT_UNWRITTEN ||
 593                     imap.br_startblock == HOLESTARTBLOCK) {
 594                         /*
 595                          * This loop handles initializing pages that were
 596                          * partially initialized by the code below this
 597                          * loop. It basically zeroes the part of the page
 598                          * that sits on a hole and sets the page as P_HOLE
 599                          * and calls remapf if it is a mapped file.
 600                          */
 601                         start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 602                         ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 603                         continue;
 604                 }
 605
 606                 /*
 607                  * There are blocks we need to zero.
 608                  * Drop the inode lock while we're doing the I/O.
 609                  * We'll still have the iolock to protect us.
 610                  */
 611                 XFS_IUNLOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 612
 613                 zero_off = XFS_FSB_TO_B(mp, start_zero_fsb);
 614                 zero_len = XFS_FSB_TO_B(mp, imap.br_blockcount);
 615
 616                 if ((zero_off + zero_len) > offset)
 617                         zero_len = offset - zero_off;
 618
 619                 error = xfs_iozero(ip, zero_off, zero_len);
 620                 if (error) {
 621                         goto out_lock;
 622                 }
 623
 624                 start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 625                 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 626
 627                 XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 628         }
 629
 630         return 0;
 631
 632 out_lock:
 633
 634         XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 635         ASSERT(error >= 0);
 636         return error;
 637 }
 638
 639 ssize_t                         /* bytes written, or (-) error */
 640 xfs_write(
 641         bhv_desc_t              *bdp,
 642         struct kiocb            *iocb,
 643         const struct iovec      *iovp,
 644         unsigned int            nsegs,
 645         loff_t                  *offset,
 646         int                     ioflags,
 647         cred_t                  *credp)
 648 {
 649         struct file             *file = iocb->ki_filp;
 650         struct address_space    *mapping = file->f_mapping;
 651         struct inode            *inode = mapping->host;
 652         unsigned long           segs = nsegs;
 653         xfs_inode_t             *xip;
 654         xfs_mount_t             *mp;
 655         ssize_t                 ret = 0, error = 0;
 656         xfs_fsize_t             isize, new_size;
 657         xfs_iocore_t            *io;
 658         bhv_vnode_t             *vp;
 659         int                     iolock;
 660         int                     eventsent = 0;
 661         bhv_vrwlock_t           locktype;
 662         size_t                  ocount = 0, count;
 663         loff_t                  pos;
 664         int                     need_i_mutex;
 665
 666         XFS_STATS_INC(xs_write_calls);
 667
 668         vp = BHV_TO_VNODE(bdp);
 669         xip = XFS_BHVTOI(bdp);
 670
 671         error = generic_segment_checks(iovp, &segs, &ocount, VERIFY_READ);
 672         if (error)
 673                 return error;
 674
 675         count = ocount;
 676         pos = *offset;
 677
 678         if (count == 0)
 679                 return 0;
 680
 681         io = &xip->i_iocore;
 682         mp = io->io_mount;
 683
 684         vfs_wait_for_freeze(vp->v_vfsp, SB_FREEZE_WRITE);
 685
 686         if (XFS_FORCED_SHUTDOWN(mp))
 687                 return -EIO;
 688
 689 relock:
 690         if (ioflags & IO_ISDIRECT) {
 691                 iolock = XFS_IOLOCK_SHARED;
 692                 locktype = VRWLOCK_WRITE_DIRECT;
 693                 need_i_mutex = 0;
 694         } else {
 695                 iolock = XFS_IOLOCK_EXCL;
 696                 locktype = VRWLOCK_WRITE;
 697                 need_i_mutex = 1;
 698                 mutex_lock(&inode->i_mutex);
 699         }
 700
 701         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
 702
 703 start:
 704         error = -generic_write_checks(file, &pos, &count,
 705                                         S_ISBLK(inode->i_mode));
 706         if (error) {
 707                 xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 708                 goto out_unlock_mutex;
 709         }
 710
 711         if ((DM_EVENT_ENABLED(vp->v_vfsp, xip, DM_EVENT_WRITE) &&
 712             !(ioflags & IO_INVIS) && !eventsent)) {
 713                 int             dmflags = FILP_DELAY_FLAG(file);
 714
 715                 if (need_i_mutex)
 716                         dmflags |= DM_FLAGS_IMUX;
 717
 718                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 719                 error = XFS_SEND_DATA(xip->i_mount, DM_EVENT_WRITE, vp,
 720                                       pos, count,
 721                                       dmflags, &locktype);
 722                 if (error) {
 723                         goto out_unlock_internal;
 724                 }
 725                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 726                 eventsent = 1;
 727
 728                 /*
 729                  * The iolock was dropped and reacquired in XFS_SEND_DATA
 730                  * so we have to recheck the size when appending.
 731                  * We will only "goto start;" once, since having sent the
 732                  * event prevents another call to XFS_SEND_DATA, which is
 733                  * what allows the size to change in the first place.
 734                  */
 735                 if ((file->f_flags & O_APPEND) && pos != xip->i_size)
 736                         goto start;
 737         }
 738
 739         if (ioflags & IO_ISDIRECT) {
 740                 xfs_buftarg_t   *target =
 741                         (xip->i_d.di_flags & XFS_DIFLAG_REALTIME) ?
 742                                 mp->m_rtdev_targp : mp->m_ddev_targp;
 743
 744                 if ((pos & target->bt_smask) || (count & target->bt_smask)) {
 745                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 746                         return XFS_ERROR(-EINVAL);
 747                 }
 748
 749                 if (!need_i_mutex && (VN_CACHED(vp) || pos > xip->i_size)) {
 750                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 751                         iolock = XFS_IOLOCK_EXCL;
 752                         locktype = VRWLOCK_WRITE;
 753                         need_i_mutex = 1;
 754                         mutex_lock(&inode->i_mutex);
 755                         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
 756                         goto start;
 757                 }
 758         }
 759
 760         new_size = pos + count;
 761         if (new_size > xip->i_size)
 762                 io->io_new_size = new_size;
 763
 764         if (likely(!(ioflags & IO_INVIS))) {
 765                 file_update_time(file);
 766                 xfs_ichgtime_fast(xip, inode,
 767                                   XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
 768         }
 769
 770         /*
 771          * If the offset is beyond the size of the file, we have a couple
 772          * of things to do. First, if there is already space allocated
 773          * we need to either create holes or zero the disk or ...
 774          *
 775          * If there is a page where the previous size lands, we need
 776          * to zero it out up to the new size.
 777          */
 778
 779         if (pos > xip->i_size) {
 780                 error = xfs_zero_eof(BHV_TO_VNODE(bdp), io, pos, xip->i_size);
 781                 if (error) {
 782                         xfs_iunlock(xip, XFS_ILOCK_EXCL);
 783                         goto out_unlock_internal;
 784                 }
 785         }
 786         xfs_iunlock(xip, XFS_ILOCK_EXCL);
 787
 788         /*
 789          * If we're writing the file then make sure to clear the
 790          * setuid and setgid bits if the process is not being run
 791          * by root.  This keeps people from modifying setuid and
 792          * setgid binaries.
 793          */
 794
 795         if (((xip->i_d.di_mode & S_ISUID) ||
 796             ((xip->i_d.di_mode & (S_ISGID | S_IXGRP)) ==
 797                 (S_ISGID | S_IXGRP))) &&
 798              !capable(CAP_FSETID)) {
 799                 error = xfs_write_clear_setuid(xip);
 800                 if (likely(!error))
 801                         error = -remove_suid(file->f_path.dentry);
 802                 if (unlikely(error)) {
 803                         goto out_unlock_internal;
 804                 }
 805         }
 806
 807 retry:
 808         /* We can write back this queue in page reclaim */
 809         current->backing_dev_info = mapping->backing_dev_info;
 810
 811         if ((ioflags & IO_ISDIRECT)) {
 812                 if (VN_CACHED(vp)) {
 813                         WARN_ON(need_i_mutex == 0);
 814                         xfs_inval_cached_trace(io, pos, -1,
 815                                         ctooff(offtoct(pos)), -1);
 816                         error = bhv_vop_flushinval_pages(vp, ctooff(offtoct(pos)),
 817                                         -1, FI_REMAPF_LOCKED);
 818                         if (error)
 819                                 goto out_unlock_internal;
 820                 }
 821
 822                 if (need_i_mutex) {
 823                         /* demote the lock now the cached pages are gone */
 824                         XFS_ILOCK_DEMOTE(mp, io, XFS_IOLOCK_EXCL);
 825                         mutex_unlock(&inode->i_mutex);
 826
 827                         iolock = XFS_IOLOCK_SHARED;
 828                         locktype = VRWLOCK_WRITE_DIRECT;
 829                         need_i_mutex = 0;
 830                 }
 831
 832                 xfs_rw_enter_trace(XFS_DIOWR_ENTER, io, (void *)iovp, segs,
 833                                 *offset, ioflags);
 834                 ret = generic_file_direct_write(iocb, iovp,
 835                                 &segs, pos, offset, count, ocount);
 836
 837                 /*
 838                  * direct-io write to a hole: fall through to buffered I/O
 839                  * for completing the rest of the request.
 840                  */
 841                 if (ret >= 0 && ret != count) {
 842                         XFS_STATS_ADD(xs_write_bytes, ret);
 843
 844                         pos += ret;
 845                         count -= ret;
 846
 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 (need_i_mutex)
 874                         mutex_lock(&inode->i_mutex);
 875                 xfs_rwlock(bdp, locktype);
 876                 if (error)
 877                         goto out_unlock_internal;
 878                 pos = xip->i_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_size) {
 888                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 889                 if (*offset > xip->i_size)
 890                         xip->i_size = *offset;
 891                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 892         }
 893
 894         error = -ret;
 895         if (ret <= 0)
 896                 goto out_unlock_internal;
 897
 898         XFS_STATS_ADD(xs_write_bytes, ret);
 899
 900         /* Handle various SYNC-type writes */
 901         if ((file->f_flags & O_SYNC) || IS_SYNC(inode)) {
 902                 error = xfs_write_sync_logforce(mp, xip);
 903                 if (error)
 904                         goto out_unlock_internal;
 905
 906                 xfs_rwunlock(bdp, locktype);
 907                 if (need_i_mutex)
 908                         mutex_unlock(&inode->i_mutex);
 909
 910                 error = sync_page_range(inode, mapping, pos, ret);
 911                 if (!error)
 912                         error = -ret;
 913                 if (need_i_mutex)
 914                         mutex_lock(&inode->i_mutex);
 915                 xfs_rwlock(bdp, locktype);
 916         }
 917
 918  out_unlock_internal:
 919         if (io->io_new_size) {
 920                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 921                 io->io_new_size = 0;
 922                 /*
 923                  * If this was a direct or synchronous I/O that failed (such
 924                  * as ENOSPC) then part of the I/O may have been written to
 925                  * disk before the error occured.  In this case the on-disk
 926                  * file size may have been adjusted beyond the in-memory file
 927                  * size and now needs to be truncated back.
 928                  */
 929                 if (xip->i_d.di_size > xip->i_size)
 930                         xip->i_d.di_size = xip->i_size;
 931                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 932         }
 933         xfs_rwunlock(bdp, locktype);
 934  out_unlock_mutex:
 935         if (need_i_mutex)
 936                 mutex_unlock(&inode->i_mutex);
 937         return -error;
 938 }
 939
 940 /*
 941  * All xfs metadata buffers except log state machine buffers
 942  * get this attached as their b_bdstrat callback function.
 943  * This is so that we can catch a buffer
 944  * after prematurely unpinning it to forcibly shutdown the filesystem.
 945  */
 946 int
 947 xfs_bdstrat_cb(struct xfs_buf *bp)
 948 {
 949         xfs_mount_t     *mp;
 950
 951         mp = XFS_BUF_FSPRIVATE3(bp, xfs_mount_t *);
 952         if (!XFS_FORCED_SHUTDOWN(mp)) {
 953                 xfs_buf_iorequest(bp);
 954                 return 0;
 955         } else {
 956                 xfs_buftrace("XFS__BDSTRAT IOERROR", bp);
 957                 /*
 958                  * Metadata write that didn't get logged but
 959                  * written delayed anyway. These aren't associated
 960                  * with a transaction, and can be ignored.
 961                  */
 962                 if (XFS_BUF_IODONE_FUNC(bp) == NULL &&
 963                     (XFS_BUF_ISREAD(bp)) == 0)
 964                         return (xfs_bioerror_relse(bp));
 965                 else
 966                         return (xfs_bioerror(bp));
 967         }
 968 }
 969
 970
 971 int
 972 xfs_bmap(bhv_desc_t     *bdp,
 973         xfs_off_t       offset,
 974         ssize_t         count,
 975         int             flags,
 976         xfs_iomap_t     *iomapp,
 977         int             *niomaps)
 978 {
 979         xfs_inode_t     *ip = XFS_BHVTOI(bdp);
 980         xfs_iocore_t    *io = &ip->i_iocore;
 981
 982         ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFREG);
 983         ASSERT(((ip->i_d.di_flags & XFS_DIFLAG_REALTIME) != 0) ==
 984                ((ip->i_iocore.io_flags & XFS_IOCORE_RT) != 0));
 985
 986         return xfs_iomap(io, offset, count, flags, iomapp, niomaps);
 987 }
 988
 989 /*
 990  * Wrapper around bdstrat so that we can stop data
 991  * from going to disk in case we are shutting down the filesystem.
 992  * Typically user data goes thru this path; one of the exceptions
 993  * is the superblock.
 994  */
 995 int
 996 xfsbdstrat(
 997         struct xfs_mount        *mp,
 998         struct xfs_buf          *bp)
 999 {
1000         ASSERT(mp);
1001         if (!XFS_FORCED_SHUTDOWN(mp)) {
1002                 /* Grio redirection would go here
1003                  * if (XFS_BUF_IS_GRIO(bp)) {
1004                  */
1005
1006                 xfs_buf_iorequest(bp);
1007                 return 0;
1008         }
1009
1010         xfs_buftrace("XFSBDSTRAT IOERROR", bp);
1011         return (xfs_bioerror_relse(bp));
1012 }
1013
1014 /*
1015  * If the underlying (data/log/rt) device is readonly, there are some
1016  * operations that cannot proceed.
1017  */
1018 int
1019 xfs_dev_is_read_only(
1020         xfs_mount_t             *mp,
1021         char                    *message)
1022 {
1023         if (xfs_readonly_buftarg(mp->m_ddev_targp) ||
1024             xfs_readonly_buftarg(mp->m_logdev_targp) ||
1025             (mp->m_rtdev_targp && xfs_readonly_buftarg(mp->m_rtdev_targp))) {
1026                 cmn_err(CE_NOTE,
1027                         "XFS: %s required on read-only device.", message);
1028                 cmn_err(CE_NOTE,
1029                         "XFS: write access unavailable, cannot proceed.");
1030                 return EROFS;
1031         }
1032         return 0;
1033 }