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                 zero_user_page(page, offset, bytes, KM_USER0);
 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_splice_read(
 291         bhv_desc_t              *bdp,
 292         struct file             *infilp,
 293         loff_t                  *ppos,
 294         struct pipe_inode_info  *pipe,
 295         size_t                  count,
 296         int                     flags,
 297         int                     ioflags,
 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(ip->i_mount))
 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                                         *ppos, count,
 317                                         FILP_DELAY_FLAG(infilp), &locktype);
 318                 if (error) {
 319                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 320                         return -error;
 321                 }
 322         }
 323         xfs_rw_enter_trace(XFS_SPLICE_READ_ENTER, &ip->i_iocore,
 324                            pipe, count, *ppos, ioflags);
 325         ret = generic_file_splice_read(infilp, ppos, pipe, count, flags);
 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_write(
 335         bhv_desc_t              *bdp,
 336         struct pipe_inode_info  *pipe,
 337         struct file             *outfilp,
 338         loff_t                  *ppos,
 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         xfs_iocore_t            *io = &ip->i_iocore;
 347         ssize_t                 ret;
 348         struct inode            *inode = outfilp->f_mapping->host;
 349         xfs_fsize_t             isize, new_size;
 350
 351         XFS_STATS_INC(xs_write_calls);
 352         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 353                 return -EIO;
 354
 355         xfs_ilock(ip, XFS_IOLOCK_EXCL);
 356
 357         if (DM_EVENT_ENABLED(BHV_TO_VNODE(bdp)->v_vfsp, ip, DM_EVENT_WRITE) &&
 358             (!(ioflags & IO_INVIS))) {
 359                 bhv_vrwlock_t locktype = VRWLOCK_WRITE;
 360                 int error;
 361
 362                 error = XFS_SEND_DATA(mp, DM_EVENT_WRITE, BHV_TO_VNODE(bdp),
 363                                         *ppos, count,
 364                                         FILP_DELAY_FLAG(outfilp), &locktype);
 365                 if (error) {
 366                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 367                         return -error;
 368                 }
 369         }
 370
 371         new_size = *ppos + count;
 372
 373         xfs_ilock(ip, XFS_ILOCK_EXCL);
 374         if (new_size > ip->i_size)
 375                 io->io_new_size = new_size;
 376         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 377
 378         xfs_rw_enter_trace(XFS_SPLICE_WRITE_ENTER, &ip->i_iocore,
 379                            pipe, count, *ppos, ioflags);
 380         ret = generic_file_splice_write(pipe, outfilp, ppos, count, flags);
 381         if (ret > 0)
 382                 XFS_STATS_ADD(xs_write_bytes, ret);
 383
 384         isize = i_size_read(inode);
 385         if (unlikely(ret < 0 && ret != -EFAULT && *ppos > isize))
 386                 *ppos = isize;
 387
 388         if (*ppos > ip->i_size) {
 389                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 390                 if (*ppos > ip->i_size)
 391                         ip->i_size = *ppos;
 392                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 393         }
 394
 395         if (io->io_new_size) {
 396                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 397                 io->io_new_size = 0;
 398                 if (ip->i_d.di_size > ip->i_size)
 399                         ip->i_d.di_size = ip->i_size;
 400                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 401         }
 402         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 403         return ret;
 404 }
 405
 406 /*
 407  * This routine is called to handle zeroing any space in the last
 408  * block of the file that is beyond the EOF.  We do this since the
 409  * size is being increased without writing anything to that block
 410  * and we don't want anyone to read the garbage on the disk.
 411  */
 412 STATIC int                              /* error (positive) */
 413 xfs_zero_last_block(
 414         struct inode    *ip,
 415         xfs_iocore_t    *io,
 416         xfs_fsize_t     offset,
 417         xfs_fsize_t     isize)
 418 {
 419         xfs_fileoff_t   last_fsb;
 420         xfs_mount_t     *mp = io->io_mount;
 421         int             nimaps;
 422         int             zero_offset;
 423         int             zero_len;
 424         int             error = 0;
 425         xfs_bmbt_irec_t imap;
 426
 427         ASSERT(ismrlocked(io->io_lock, MR_UPDATE) != 0);
 428
 429         zero_offset = XFS_B_FSB_OFFSET(mp, isize);
 430         if (zero_offset == 0) {
 431                 /*
 432                  * There are no extra bytes in the last block on disk to
 433                  * zero, so return.
 434                  */
 435                 return 0;
 436         }
 437
 438         last_fsb = XFS_B_TO_FSBT(mp, isize);
 439         nimaps = 1;
 440         error = XFS_BMAPI(mp, NULL, io, last_fsb, 1, 0, NULL, 0, &imap,
 441                           &nimaps, NULL, NULL);
 442         if (error) {
 443                 return error;
 444         }
 445         ASSERT(nimaps > 0);
 446         /*
 447          * If the block underlying isize is just a hole, then there
 448          * is nothing to zero.
 449          */
 450         if (imap.br_startblock == HOLESTARTBLOCK) {
 451                 return 0;
 452         }
 453         /*
 454          * Zero the part of the last block beyond the EOF, and write it
 455          * out sync.  We need to drop the ilock while we do this so we
 456          * don't deadlock when the buffer cache calls back to us.
 457          */
 458         XFS_IUNLOCK(mp, io, XFS_ILOCK_EXCL| XFS_EXTSIZE_RD);
 459
 460         zero_len = mp->m_sb.sb_blocksize - zero_offset;
 461         if (isize + zero_len > offset)
 462                 zero_len = offset - isize;
 463         error = xfs_iozero(ip, isize, zero_len);
 464
 465         XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 466         ASSERT(error >= 0);
 467         return error;
 468 }
 469
 470 /*
 471  * Zero any on disk space between the current EOF and the new,
 472  * larger EOF.  This handles the normal case of zeroing the remainder
 473  * of the last block in the file and the unusual case of zeroing blocks
 474  * out beyond the size of the file.  This second case only happens
 475  * with fixed size extents and when the system crashes before the inode
 476  * size was updated but after blocks were allocated.  If fill is set,
 477  * then any holes in the range are filled and zeroed.  If not, the holes
 478  * are left alone as holes.
 479  */
 480
 481 int                                     /* error (positive) */
 482 xfs_zero_eof(
 483         bhv_vnode_t     *vp,
 484         xfs_iocore_t    *io,
 485         xfs_off_t       offset,         /* starting I/O offset */
 486         xfs_fsize_t     isize)          /* current inode size */
 487 {
 488         struct inode    *ip = vn_to_inode(vp);
 489         xfs_fileoff_t   start_zero_fsb;
 490         xfs_fileoff_t   end_zero_fsb;
 491         xfs_fileoff_t   zero_count_fsb;
 492         xfs_fileoff_t   last_fsb;
 493         xfs_fileoff_t   zero_off;
 494         xfs_fsize_t     zero_len;
 495         xfs_mount_t     *mp = io->io_mount;
 496         int             nimaps;
 497         int             error = 0;
 498         xfs_bmbt_irec_t imap;
 499
 500         ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
 501         ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
 502         ASSERT(offset > isize);
 503
 504         /*
 505          * First handle zeroing the block on which isize resides.
 506          * We only zero a part of that block so it is handled specially.
 507          */
 508         error = xfs_zero_last_block(ip, io, offset, isize);
 509         if (error) {
 510                 ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
 511                 ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
 512                 return error;
 513         }
 514
 515         /*
 516          * Calculate the range between the new size and the old
 517          * where blocks needing to be zeroed may exist.  To get the
 518          * block where the last byte in the file currently resides,
 519          * we need to subtract one from the size and truncate back
 520          * to a block boundary.  We subtract 1 in case the size is
 521          * exactly on a block boundary.
 522          */
 523         last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
 524         start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
 525         end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
 526         ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
 527         if (last_fsb == end_zero_fsb) {
 528                 /*
 529                  * The size was only incremented on its last block.
 530                  * We took care of that above, so just return.
 531                  */
 532                 return 0;
 533         }
 534
 535         ASSERT(start_zero_fsb <= end_zero_fsb);
 536         while (start_zero_fsb <= end_zero_fsb) {
 537                 nimaps = 1;
 538                 zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
 539                 error = XFS_BMAPI(mp, NULL, io, start_zero_fsb, zero_count_fsb,
 540                                   0, NULL, 0, &imap, &nimaps, NULL, NULL);
 541                 if (error) {
 542                         ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
 543                         ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
 544                         return error;
 545                 }
 546                 ASSERT(nimaps > 0);
 547
 548                 if (imap.br_state == XFS_EXT_UNWRITTEN ||
 549                     imap.br_startblock == HOLESTARTBLOCK) {
 550                         /*
 551                          * This loop handles initializing pages that were
 552                          * partially initialized by the code below this
 553                          * loop. It basically zeroes the part of the page
 554                          * that sits on a hole and sets the page as P_HOLE
 555                          * and calls remapf if it is a mapped file.
 556                          */
 557                         start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 558                         ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 559                         continue;
 560                 }
 561
 562                 /*
 563                  * There are blocks we need to zero.
 564                  * Drop the inode lock while we're doing the I/O.
 565                  * We'll still have the iolock to protect us.
 566                  */
 567                 XFS_IUNLOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 568
 569                 zero_off = XFS_FSB_TO_B(mp, start_zero_fsb);
 570                 zero_len = XFS_FSB_TO_B(mp, imap.br_blockcount);
 571
 572                 if ((zero_off + zero_len) > offset)
 573                         zero_len = offset - zero_off;
 574
 575                 error = xfs_iozero(ip, zero_off, zero_len);
 576                 if (error) {
 577                         goto out_lock;
 578                 }
 579
 580                 start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 581                 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 582
 583                 XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 584         }
 585
 586         return 0;
 587
 588 out_lock:
 589
 590         XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 591         ASSERT(error >= 0);
 592         return error;
 593 }
 594
 595 ssize_t                         /* bytes written, or (-) error */
 596 xfs_write(
 597         bhv_desc_t              *bdp,
 598         struct kiocb            *iocb,
 599         const struct iovec      *iovp,
 600         unsigned int            nsegs,
 601         loff_t                  *offset,
 602         int                     ioflags,
 603         cred_t                  *credp)
 604 {
 605         struct file             *file = iocb->ki_filp;
 606         struct address_space    *mapping = file->f_mapping;
 607         struct inode            *inode = mapping->host;
 608         unsigned long           segs = nsegs;
 609         xfs_inode_t             *xip;
 610         xfs_mount_t             *mp;
 611         ssize_t                 ret = 0, error = 0;
 612         xfs_fsize_t             isize, new_size;
 613         xfs_iocore_t            *io;
 614         bhv_vnode_t             *vp;
 615         int                     iolock;
 616         int                     eventsent = 0;
 617         bhv_vrwlock_t           locktype;
 618         size_t                  ocount = 0, count;
 619         loff_t                  pos;
 620         int                     need_i_mutex;
 621
 622         XFS_STATS_INC(xs_write_calls);
 623
 624         vp = BHV_TO_VNODE(bdp);
 625         xip = XFS_BHVTOI(bdp);
 626
 627         error = generic_segment_checks(iovp, &segs, &ocount, VERIFY_READ);
 628         if (error)
 629                 return error;
 630
 631         count = ocount;
 632         pos = *offset;
 633
 634         if (count == 0)
 635                 return 0;
 636
 637         io = &xip->i_iocore;
 638         mp = io->io_mount;
 639
 640         vfs_wait_for_freeze(vp->v_vfsp, SB_FREEZE_WRITE);
 641
 642         if (XFS_FORCED_SHUTDOWN(mp))
 643                 return -EIO;
 644
 645 relock:
 646         if (ioflags & IO_ISDIRECT) {
 647                 iolock = XFS_IOLOCK_SHARED;
 648                 locktype = VRWLOCK_WRITE_DIRECT;
 649                 need_i_mutex = 0;
 650         } else {
 651                 iolock = XFS_IOLOCK_EXCL;
 652                 locktype = VRWLOCK_WRITE;
 653                 need_i_mutex = 1;
 654                 mutex_lock(&inode->i_mutex);
 655         }
 656
 657         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
 658
 659 start:
 660         error = -generic_write_checks(file, &pos, &count,
 661                                         S_ISBLK(inode->i_mode));
 662         if (error) {
 663                 xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 664                 goto out_unlock_mutex;
 665         }
 666
 667         if ((DM_EVENT_ENABLED(vp->v_vfsp, xip, DM_EVENT_WRITE) &&
 668             !(ioflags & IO_INVIS) && !eventsent)) {
 669                 int             dmflags = FILP_DELAY_FLAG(file);
 670
 671                 if (need_i_mutex)
 672                         dmflags |= DM_FLAGS_IMUX;
 673
 674                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 675                 error = XFS_SEND_DATA(xip->i_mount, DM_EVENT_WRITE, vp,
 676                                       pos, count,
 677                                       dmflags, &locktype);
 678                 if (error) {
 679                         goto out_unlock_internal;
 680                 }
 681                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 682                 eventsent = 1;
 683
 684                 /*
 685                  * The iolock was dropped and reacquired in XFS_SEND_DATA
 686                  * so we have to recheck the size when appending.
 687                  * We will only "goto start;" once, since having sent the
 688                  * event prevents another call to XFS_SEND_DATA, which is
 689                  * what allows the size to change in the first place.
 690                  */
 691                 if ((file->f_flags & O_APPEND) && pos != xip->i_size)
 692                         goto start;
 693         }
 694
 695         if (ioflags & IO_ISDIRECT) {
 696                 xfs_buftarg_t   *target =
 697                         (xip->i_d.di_flags & XFS_DIFLAG_REALTIME) ?
 698                                 mp->m_rtdev_targp : mp->m_ddev_targp;
 699
 700                 if ((pos & target->bt_smask) || (count & target->bt_smask)) {
 701                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 702                         return XFS_ERROR(-EINVAL);
 703                 }
 704
 705                 if (!need_i_mutex && (VN_CACHED(vp) || pos > xip->i_size)) {
 706                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 707                         iolock = XFS_IOLOCK_EXCL;
 708                         locktype = VRWLOCK_WRITE;
 709                         need_i_mutex = 1;
 710                         mutex_lock(&inode->i_mutex);
 711                         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
 712                         goto start;
 713                 }
 714         }
 715
 716         new_size = pos + count;
 717         if (new_size > xip->i_size)
 718                 io->io_new_size = new_size;
 719
 720         if (likely(!(ioflags & IO_INVIS))) {
 721                 file_update_time(file);
 722                 xfs_ichgtime_fast(xip, inode,
 723                                   XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
 724         }
 725
 726         /*
 727          * If the offset is beyond the size of the file, we have a couple
 728          * of things to do. First, if there is already space allocated
 729          * we need to either create holes or zero the disk or ...
 730          *
 731          * If there is a page where the previous size lands, we need
 732          * to zero it out up to the new size.
 733          */
 734
 735         if (pos > xip->i_size) {
 736                 error = xfs_zero_eof(BHV_TO_VNODE(bdp), io, pos, xip->i_size);
 737                 if (error) {
 738                         xfs_iunlock(xip, XFS_ILOCK_EXCL);
 739                         goto out_unlock_internal;
 740                 }
 741         }
 742         xfs_iunlock(xip, XFS_ILOCK_EXCL);
 743
 744         /*
 745          * If we're writing the file then make sure to clear the
 746          * setuid and setgid bits if the process is not being run
 747          * by root.  This keeps people from modifying setuid and
 748          * setgid binaries.
 749          */
 750
 751         if (((xip->i_d.di_mode & S_ISUID) ||
 752             ((xip->i_d.di_mode & (S_ISGID | S_IXGRP)) ==
 753                 (S_ISGID | S_IXGRP))) &&
 754              !capable(CAP_FSETID)) {
 755                 error = xfs_write_clear_setuid(xip);
 756                 if (likely(!error))
 757                         error = -remove_suid(file->f_path.dentry);
 758                 if (unlikely(error)) {
 759                         goto out_unlock_internal;
 760                 }
 761         }
 762
 763 retry:
 764         /* We can write back this queue in page reclaim */
 765         current->backing_dev_info = mapping->backing_dev_info;
 766
 767         if ((ioflags & IO_ISDIRECT)) {
 768                 if (VN_CACHED(vp)) {
 769                         WARN_ON(need_i_mutex == 0);
 770                         xfs_inval_cached_trace(io, pos, -1,
 771                                         ctooff(offtoct(pos)), -1);
 772                         error = bhv_vop_flushinval_pages(vp, ctooff(offtoct(pos)),
 773                                         -1, FI_REMAPF_LOCKED);
 774                         if (error)
 775                                 goto out_unlock_internal;
 776                 }
 777
 778                 if (need_i_mutex) {
 779                         /* demote the lock now the cached pages are gone */
 780                         XFS_ILOCK_DEMOTE(mp, io, XFS_IOLOCK_EXCL);
 781                         mutex_unlock(&inode->i_mutex);
 782
 783                         iolock = XFS_IOLOCK_SHARED;
 784                         locktype = VRWLOCK_WRITE_DIRECT;
 785                         need_i_mutex = 0;
 786                 }
 787
 788                 xfs_rw_enter_trace(XFS_DIOWR_ENTER, io, (void *)iovp, segs,
 789                                 *offset, ioflags);
 790                 ret = generic_file_direct_write(iocb, iovp,
 791                                 &segs, pos, offset, count, ocount);
 792
 793                 /*
 794                  * direct-io write to a hole: fall through to buffered I/O
 795                  * for completing the rest of the request.
 796                  */
 797                 if (ret >= 0 && ret != count) {
 798                         XFS_STATS_ADD(xs_write_bytes, ret);
 799
 800                         pos += ret;
 801                         count -= ret;
 802
 803                         ioflags &= ~IO_ISDIRECT;
 804                         xfs_iunlock(xip, iolock);
 805                         goto relock;
 806                 }
 807         } else {
 808                 xfs_rw_enter_trace(XFS_WRITE_ENTER, io, (void *)iovp, segs,
 809                                 *offset, ioflags);
 810                 ret = generic_file_buffered_write(iocb, iovp, segs,
 811                                 pos, offset, count, ret);
 812         }
 813
 814         current->backing_dev_info = NULL;
 815
 816         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
 817                 ret = wait_on_sync_kiocb(iocb);
 818
 819         if ((ret == -ENOSPC) &&
 820             DM_EVENT_ENABLED(vp->v_vfsp, xip, DM_EVENT_NOSPACE) &&
 821             !(ioflags & IO_INVIS)) {
 822
 823                 xfs_rwunlock(bdp, locktype);
 824                 if (need_i_mutex)
 825                         mutex_unlock(&inode->i_mutex);
 826                 error = XFS_SEND_NAMESP(xip->i_mount, DM_EVENT_NOSPACE, vp,
 827                                 DM_RIGHT_NULL, vp, DM_RIGHT_NULL, NULL, NULL,
 828                                 0, 0, 0); /* Delay flag intentionally  unused */
 829                 if (need_i_mutex)
 830                         mutex_lock(&inode->i_mutex);
 831                 xfs_rwlock(bdp, locktype);
 832                 if (error)
 833                         goto out_unlock_internal;
 834                 pos = xip->i_size;
 835                 ret = 0;
 836                 goto retry;
 837         }
 838
 839         isize = i_size_read(inode);
 840         if (unlikely(ret < 0 && ret != -EFAULT && *offset > isize))
 841                 *offset = isize;
 842
 843         if (*offset > xip->i_size) {
 844                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 845                 if (*offset > xip->i_size)
 846                         xip->i_size = *offset;
 847                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 848         }
 849
 850         error = -ret;
 851         if (ret <= 0)
 852                 goto out_unlock_internal;
 853
 854         XFS_STATS_ADD(xs_write_bytes, ret);
 855
 856         /* Handle various SYNC-type writes */
 857         if ((file->f_flags & O_SYNC) || IS_SYNC(inode)) {
 858                 error = xfs_write_sync_logforce(mp, xip);
 859                 if (error)
 860                         goto out_unlock_internal;
 861
 862                 xfs_rwunlock(bdp, locktype);
 863                 if (need_i_mutex)
 864                         mutex_unlock(&inode->i_mutex);
 865
 866                 error = sync_page_range(inode, mapping, pos, ret);
 867                 if (!error)
 868                         error = -ret;
 869                 if (need_i_mutex)
 870                         mutex_lock(&inode->i_mutex);
 871                 xfs_rwlock(bdp, locktype);
 872         }
 873
 874  out_unlock_internal:
 875         if (io->io_new_size) {
 876                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 877                 io->io_new_size = 0;
 878                 /*
 879                  * If this was a direct or synchronous I/O that failed (such
 880                  * as ENOSPC) then part of the I/O may have been written to
 881                  * disk before the error occured.  In this case the on-disk
 882                  * file size may have been adjusted beyond the in-memory file
 883                  * size and now needs to be truncated back.
 884                  */
 885                 if (xip->i_d.di_size > xip->i_size)
 886                         xip->i_d.di_size = xip->i_size;
 887                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 888         }
 889         xfs_rwunlock(bdp, locktype);
 890  out_unlock_mutex:
 891         if (need_i_mutex)
 892                 mutex_unlock(&inode->i_mutex);
 893         return -error;
 894 }
 895
 896 /*
 897  * All xfs metadata buffers except log state machine buffers
 898  * get this attached as their b_bdstrat callback function.
 899  * This is so that we can catch a buffer
 900  * after prematurely unpinning it to forcibly shutdown the filesystem.
 901  */
 902 int
 903 xfs_bdstrat_cb(struct xfs_buf *bp)
 904 {
 905         xfs_mount_t     *mp;
 906
 907         mp = XFS_BUF_FSPRIVATE3(bp, xfs_mount_t *);
 908         if (!XFS_FORCED_SHUTDOWN(mp)) {
 909                 xfs_buf_iorequest(bp);
 910                 return 0;
 911         } else {
 912                 xfs_buftrace("XFS__BDSTRAT IOERROR", bp);
 913                 /*
 914                  * Metadata write that didn't get logged but
 915                  * written delayed anyway. These aren't associated
 916                  * with a transaction, and can be ignored.
 917                  */
 918                 if (XFS_BUF_IODONE_FUNC(bp) == NULL &&
 919                     (XFS_BUF_ISREAD(bp)) == 0)
 920                         return (xfs_bioerror_relse(bp));
 921                 else
 922                         return (xfs_bioerror(bp));
 923         }
 924 }
 925
 926
 927 int
 928 xfs_bmap(bhv_desc_t     *bdp,
 929         xfs_off_t       offset,
 930         ssize_t         count,
 931         int             flags,
 932         xfs_iomap_t     *iomapp,
 933         int             *niomaps)
 934 {
 935         xfs_inode_t     *ip = XFS_BHVTOI(bdp);
 936         xfs_iocore_t    *io = &ip->i_iocore;
 937
 938         ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFREG);
 939         ASSERT(((ip->i_d.di_flags & XFS_DIFLAG_REALTIME) != 0) ==
 940                ((ip->i_iocore.io_flags & XFS_IOCORE_RT) != 0));
 941
 942         return xfs_iomap(io, offset, count, flags, iomapp, niomaps);
 943 }
 944
 945 /*
 946  * Wrapper around bdstrat so that we can stop data
 947  * from going to disk in case we are shutting down the filesystem.
 948  * Typically user data goes thru this path; one of the exceptions
 949  * is the superblock.
 950  */
 951 int
 952 xfsbdstrat(
 953         struct xfs_mount        *mp,
 954         struct xfs_buf          *bp)
 955 {
 956         ASSERT(mp);
 957         if (!XFS_FORCED_SHUTDOWN(mp)) {
 958                 /* Grio redirection would go here
 959                  * if (XFS_BUF_IS_GRIO(bp)) {
 960                  */
 961
 962                 xfs_buf_iorequest(bp);
 963                 return 0;
 964         }
 965
 966         xfs_buftrace("XFSBDSTRAT IOERROR", bp);
 967         return (xfs_bioerror_relse(bp));
 968 }
 969
 970 /*
 971  * If the underlying (data/log/rt) device is readonly, there are some
 972  * operations that cannot proceed.
 973  */
 974 int
 975 xfs_dev_is_read_only(
 976         xfs_mount_t             *mp,
 977         char                    *message)
 978 {
 979         if (xfs_readonly_buftarg(mp->m_ddev_targp) ||
 980             xfs_readonly_buftarg(mp->m_logdev_targp) ||
 981             (mp->m_rtdev_targp && xfs_readonly_buftarg(mp->m_rtdev_targp))) {
 982                 cmn_err(CE_NOTE,
 983                         "XFS: %s required on read-only device.", message);
 984                 cmn_err(CE_NOTE,
 985                         "XFS: write access unavailable, cannot proceed.");
 986                 return EROFS;
 987         }
 988         return 0;
 989 }