SAFE public projects git trees. - safe/jmp/linux-2.6/blob - fs/xfs/linux-2.6/xfs_lrw.c

   1 /*
   2  * Copyright (c) 2000-2003,2005 Silicon Graphics, Inc.
   3  * All Rights Reserved.
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of the GNU General Public License as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it would be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write the Free Software Foundation,
  16  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  17  */
  18 #include "xfs.h"
  19 #include "xfs_fs.h"
  20 #include "xfs_bit.h"
  21 #include "xfs_log.h"
  22 #include "xfs_inum.h"
  23 #include "xfs_trans.h"
  24 #include "xfs_sb.h"
  25 #include "xfs_ag.h"
  26 #include "xfs_dir2.h"
  27 #include "xfs_alloc.h"
  28 #include "xfs_dmapi.h"
  29 #include "xfs_quota.h"
  30 #include "xfs_mount.h"
  31 #include "xfs_bmap_btree.h"
  32 #include "xfs_alloc_btree.h"
  33 #include "xfs_ialloc_btree.h"
  34 #include "xfs_dir2_sf.h"
  35 #include "xfs_attr_sf.h"
  36 #include "xfs_dinode.h"
  37 #include "xfs_inode.h"
  38 #include "xfs_bmap.h"
  39 #include "xfs_btree.h"
  40 #include "xfs_ialloc.h"
  41 #include "xfs_rtalloc.h"
  42 #include "xfs_error.h"
  43 #include "xfs_itable.h"
  44 #include "xfs_rw.h"
  45 #include "xfs_acl.h"
  46 #include "xfs_attr.h"
  47 #include "xfs_inode_item.h"
  48 #include "xfs_buf_item.h"
  49 #include "xfs_utils.h"
  50 #include "xfs_iomap.h"
  51 #include "xfs_vnodeops.h"
  52
  53 #include <linux/capability.h>
  54 #include <linux/writeback.h>
  55
  56
  57 #if defined(XFS_RW_TRACE)
  58 void
  59 xfs_rw_enter_trace(
  60         int                     tag,
  61         xfs_inode_t             *ip,
  62         void                    *data,
  63         size_t                  segs,
  64         loff_t                  offset,
  65         int                     ioflags)
  66 {
  67         if (ip->i_rwtrace == NULL)
  68                 return;
  69         ktrace_enter(ip->i_rwtrace,
  70                 (void *)(unsigned long)tag,
  71                 (void *)ip,
  72                 (void *)((unsigned long)((ip->i_d.di_size >> 32) & 0xffffffff)),
  73                 (void *)((unsigned long)(ip->i_d.di_size & 0xffffffff)),
  74                 (void *)data,
  75                 (void *)((unsigned long)segs),
  76                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
  77                 (void *)((unsigned long)(offset & 0xffffffff)),
  78                 (void *)((unsigned long)ioflags),
  79                 (void *)((unsigned long)((ip->i_new_size >> 32) & 0xffffffff)),
  80                 (void *)((unsigned long)(ip->i_new_size & 0xffffffff)),
  81                 (void *)((unsigned long)current_pid()),
  82                 (void *)NULL,
  83                 (void *)NULL,
  84                 (void *)NULL,
  85                 (void *)NULL);
  86 }
  87
  88 void
  89 xfs_inval_cached_trace(
  90         xfs_inode_t     *ip,
  91         xfs_off_t       offset,
  92         xfs_off_t       len,
  93         xfs_off_t       first,
  94         xfs_off_t       last)
  95 {
  96
  97         if (ip->i_rwtrace == NULL)
  98                 return;
  99         ktrace_enter(ip->i_rwtrace,
 100                 (void *)(__psint_t)XFS_INVAL_CACHED,
 101                 (void *)ip,
 102                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
 103                 (void *)((unsigned long)(offset & 0xffffffff)),
 104                 (void *)((unsigned long)((len >> 32) & 0xffffffff)),
 105                 (void *)((unsigned long)(len & 0xffffffff)),
 106                 (void *)((unsigned long)((first >> 32) & 0xffffffff)),
 107                 (void *)((unsigned long)(first & 0xffffffff)),
 108                 (void *)((unsigned long)((last >> 32) & 0xffffffff)),
 109                 (void *)((unsigned long)(last & 0xffffffff)),
 110                 (void *)((unsigned long)current_pid()),
 111                 (void *)NULL,
 112                 (void *)NULL,
 113                 (void *)NULL,
 114                 (void *)NULL,
 115                 (void *)NULL);
 116 }
 117 #endif
 118
 119 /*
 120  *      xfs_iozero
 121  *
 122  *      xfs_iozero clears the specified range of buffer supplied,
 123  *      and marks all the affected blocks as valid and modified.  If
 124  *      an affected block is not allocated, it will be allocated.  If
 125  *      an affected block is not completely overwritten, and is not
 126  *      valid before the operation, it will be read from disk before
 127  *      being partially zeroed.
 128  */
 129 STATIC int
 130 xfs_iozero(
 131         struct xfs_inode        *ip,    /* inode                        */
 132         loff_t                  pos,    /* offset in file               */
 133         size_t                  count)  /* size of data to zero         */
 134 {
 135         struct page             *page;
 136         struct address_space    *mapping;
 137         int                     status;
 138
 139         mapping = ip->i_vnode->i_mapping;
 140         do {
 141                 unsigned offset, bytes;
 142                 void *fsdata;
 143
 144                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
 145                 bytes = PAGE_CACHE_SIZE - offset;
 146                 if (bytes > count)
 147                         bytes = count;
 148
 149                 status = pagecache_write_begin(NULL, mapping, pos, bytes,
 150                                         AOP_FLAG_UNINTERRUPTIBLE,
 151                                         &page, &fsdata);
 152                 if (status)
 153                         break;
 154
 155                 zero_user_page(page, offset, bytes, KM_USER0);
 156
 157                 status = pagecache_write_end(NULL, mapping, pos, bytes, bytes,
 158                                         page, fsdata);
 159                 WARN_ON(status <= 0); /* can't return less than zero! */
 160                 pos += bytes;
 161                 count -= bytes;
 162                 status = 0;
 163         } while (count);
 164
 165         return (-status);
 166 }
 167
 168 ssize_t                 /* bytes read, or (-)  error */
 169 xfs_read(
 170         xfs_inode_t             *ip,
 171         struct kiocb            *iocb,
 172         const struct iovec      *iovp,
 173         unsigned int            segs,
 174         loff_t                  *offset,
 175         int                     ioflags)
 176 {
 177         struct file             *file = iocb->ki_filp;
 178         struct inode            *inode = file->f_mapping->host;
 179         bhv_vnode_t             *vp = XFS_ITOV(ip);
 180         xfs_mount_t             *mp = ip->i_mount;
 181         size_t                  size = 0;
 182         ssize_t                 ret = 0;
 183         xfs_fsize_t             n;
 184         unsigned long           seg;
 185
 186
 187         XFS_STATS_INC(xs_read_calls);
 188
 189         /* START copy & waste from filemap.c */
 190         for (seg = 0; seg < segs; seg++) {
 191                 const struct iovec *iv = &iovp[seg];
 192
 193                 /*
 194                  * If any segment has a negative length, or the cumulative
 195                  * length ever wraps negative then return -EINVAL.
 196                  */
 197                 size += iv->iov_len;
 198                 if (unlikely((ssize_t)(size|iv->iov_len) < 0))
 199                         return XFS_ERROR(-EINVAL);
 200         }
 201         /* END copy & waste from filemap.c */
 202
 203         if (unlikely(ioflags & IO_ISDIRECT)) {
 204                 xfs_buftarg_t   *target =
 205                         (ip->i_d.di_flags & XFS_DIFLAG_REALTIME) ?
 206                                 mp->m_rtdev_targp : mp->m_ddev_targp;
 207                 if ((*offset & target->bt_smask) ||
 208                     (size & target->bt_smask)) {
 209                         if (*offset == ip->i_size) {
 210                                 return (0);
 211                         }
 212                         return -XFS_ERROR(EINVAL);
 213                 }
 214         }
 215
 216         n = XFS_MAXIOFFSET(mp) - *offset;
 217         if ((n <= 0) || (size == 0))
 218                 return 0;
 219
 220         if (n < size)
 221                 size = n;
 222
 223         if (XFS_FORCED_SHUTDOWN(mp))
 224                 return -EIO;
 225
 226         if (unlikely(ioflags & IO_ISDIRECT))
 227                 mutex_lock(&inode->i_mutex);
 228         xfs_ilock(ip, XFS_IOLOCK_SHARED);
 229
 230         if (DM_EVENT_ENABLED(ip, DM_EVENT_READ) && !(ioflags & IO_INVIS)) {
 231                 bhv_vrwlock_t locktype = VRWLOCK_READ;
 232                 int dmflags = FILP_DELAY_FLAG(file) | DM_SEM_FLAG_RD(ioflags);
 233
 234                 ret = -XFS_SEND_DATA(mp, DM_EVENT_READ, vp, *offset, size,
 235                                         dmflags, &locktype);
 236                 if (ret) {
 237                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 238                         if (unlikely(ioflags & IO_ISDIRECT))
 239                                 mutex_unlock(&inode->i_mutex);
 240                         return ret;
 241                 }
 242         }
 243
 244         if (unlikely(ioflags & IO_ISDIRECT)) {
 245                 if (VN_CACHED(vp))
 246                         ret = xfs_flushinval_pages(ip,
 247                                         ctooff(offtoct(*offset)),
 248                                         -1, FI_REMAPF_LOCKED);
 249                 mutex_unlock(&inode->i_mutex);
 250                 if (ret) {
 251                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 252                         return ret;
 253                 }
 254         }
 255
 256         xfs_rw_enter_trace(XFS_READ_ENTER, ip,
 257                                 (void *)iovp, segs, *offset, ioflags);
 258
 259         iocb->ki_pos = *offset;
 260         ret = generic_file_aio_read(iocb, iovp, segs, *offset);
 261         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
 262                 ret = wait_on_sync_kiocb(iocb);
 263         if (ret > 0)
 264                 XFS_STATS_ADD(xs_read_bytes, ret);
 265
 266         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 267         return ret;
 268 }
 269
 270 ssize_t
 271 xfs_splice_read(
 272         xfs_inode_t             *ip,
 273         struct file             *infilp,
 274         loff_t                  *ppos,
 275         struct pipe_inode_info  *pipe,
 276         size_t                  count,
 277         int                     flags,
 278         int                     ioflags)
 279 {
 280         bhv_vnode_t             *vp = XFS_ITOV(ip);
 281         xfs_mount_t             *mp = ip->i_mount;
 282         ssize_t                 ret;
 283
 284         XFS_STATS_INC(xs_read_calls);
 285         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 286                 return -EIO;
 287
 288         xfs_ilock(ip, XFS_IOLOCK_SHARED);
 289
 290         if (DM_EVENT_ENABLED(ip, DM_EVENT_READ) && !(ioflags & IO_INVIS)) {
 291                 bhv_vrwlock_t locktype = VRWLOCK_READ;
 292                 int error;
 293
 294                 error = XFS_SEND_DATA(mp, DM_EVENT_READ, vp, *ppos, count,
 295                                         FILP_DELAY_FLAG(infilp), &locktype);
 296                 if (error) {
 297                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 298                         return -error;
 299                 }
 300         }
 301         xfs_rw_enter_trace(XFS_SPLICE_READ_ENTER, ip,
 302                            pipe, count, *ppos, ioflags);
 303         ret = generic_file_splice_read(infilp, ppos, pipe, count, flags);
 304         if (ret > 0)
 305                 XFS_STATS_ADD(xs_read_bytes, ret);
 306
 307         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 308         return ret;
 309 }
 310
 311 ssize_t
 312 xfs_splice_write(
 313         xfs_inode_t             *ip,
 314         struct pipe_inode_info  *pipe,
 315         struct file             *outfilp,
 316         loff_t                  *ppos,
 317         size_t                  count,
 318         int                     flags,
 319         int                     ioflags)
 320 {
 321         bhv_vnode_t             *vp = XFS_ITOV(ip);
 322         xfs_mount_t             *mp = ip->i_mount;
 323         ssize_t                 ret;
 324         struct inode            *inode = outfilp->f_mapping->host;
 325         xfs_fsize_t             isize, new_size;
 326
 327         XFS_STATS_INC(xs_write_calls);
 328         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 329                 return -EIO;
 330
 331         xfs_ilock(ip, XFS_IOLOCK_EXCL);
 332
 333         if (DM_EVENT_ENABLED(ip, DM_EVENT_WRITE) && !(ioflags & IO_INVIS)) {
 334                 bhv_vrwlock_t locktype = VRWLOCK_WRITE;
 335                 int error;
 336
 337                 error = XFS_SEND_DATA(mp, DM_EVENT_WRITE, vp, *ppos, count,
 338                                         FILP_DELAY_FLAG(outfilp), &locktype);
 339                 if (error) {
 340                         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 341                         return -error;
 342                 }
 343         }
 344
 345         new_size = *ppos + count;
 346
 347         xfs_ilock(ip, XFS_ILOCK_EXCL);
 348         if (new_size > ip->i_size)
 349                 ip->i_new_size = new_size;
 350         xfs_iunlock(ip, XFS_ILOCK_EXCL);
 351
 352         xfs_rw_enter_trace(XFS_SPLICE_WRITE_ENTER, ip,
 353                            pipe, count, *ppos, ioflags);
 354         ret = generic_file_splice_write(pipe, outfilp, ppos, count, flags);
 355         if (ret > 0)
 356                 XFS_STATS_ADD(xs_write_bytes, ret);
 357
 358         isize = i_size_read(inode);
 359         if (unlikely(ret < 0 && ret != -EFAULT && *ppos > isize))
 360                 *ppos = isize;
 361
 362         if (*ppos > ip->i_size) {
 363                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 364                 if (*ppos > ip->i_size)
 365                         ip->i_size = *ppos;
 366                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 367         }
 368
 369         if (ip->i_new_size) {
 370                 xfs_ilock(ip, XFS_ILOCK_EXCL);
 371                 ip->i_new_size = 0;
 372                 if (ip->i_d.di_size > ip->i_size)
 373                         ip->i_d.di_size = ip->i_size;
 374                 xfs_iunlock(ip, XFS_ILOCK_EXCL);
 375         }
 376         xfs_iunlock(ip, XFS_IOLOCK_EXCL);
 377         return ret;
 378 }
 379
 380 /*
 381  * This routine is called to handle zeroing any space in the last
 382  * block of the file that is beyond the EOF.  We do this since the
 383  * size is being increased without writing anything to that block
 384  * and we don't want anyone to read the garbage on the disk.
 385  */
 386 STATIC int                              /* error (positive) */
 387 xfs_zero_last_block(
 388         xfs_inode_t     *ip,
 389         xfs_fsize_t     offset,
 390         xfs_fsize_t     isize)
 391 {
 392         xfs_fileoff_t   last_fsb;
 393         xfs_mount_t     *mp = ip->i_mount;
 394         int             nimaps;
 395         int             zero_offset;
 396         int             zero_len;
 397         int             error = 0;
 398         xfs_bmbt_irec_t imap;
 399
 400         ASSERT(ismrlocked(&ip->i_lock, MR_UPDATE) != 0);
 401
 402         zero_offset = XFS_B_FSB_OFFSET(mp, isize);
 403         if (zero_offset == 0) {
 404                 /*
 405                  * There are no extra bytes in the last block on disk to
 406                  * zero, so return.
 407                  */
 408                 return 0;
 409         }
 410
 411         last_fsb = XFS_B_TO_FSBT(mp, isize);
 412         nimaps = 1;
 413         error = xfs_bmapi(NULL, ip, last_fsb, 1, 0, NULL, 0, &imap,
 414                           &nimaps, NULL, NULL);
 415         if (error) {
 416                 return error;
 417         }
 418         ASSERT(nimaps > 0);
 419         /*
 420          * If the block underlying isize is just a hole, then there
 421          * is nothing to zero.
 422          */
 423         if (imap.br_startblock == HOLESTARTBLOCK) {
 424                 return 0;
 425         }
 426         /*
 427          * Zero the part of the last block beyond the EOF, and write it
 428          * out sync.  We need to drop the ilock while we do this so we
 429          * don't deadlock when the buffer cache calls back to us.
 430          */
 431         xfs_iunlock(ip, XFS_ILOCK_EXCL| XFS_EXTSIZE_RD);
 432
 433         zero_len = mp->m_sb.sb_blocksize - zero_offset;
 434         if (isize + zero_len > offset)
 435                 zero_len = offset - isize;
 436         error = xfs_iozero(ip, isize, zero_len);
 437
 438         xfs_ilock(ip, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 439         ASSERT(error >= 0);
 440         return error;
 441 }
 442
 443 /*
 444  * Zero any on disk space between the current EOF and the new,
 445  * larger EOF.  This handles the normal case of zeroing the remainder
 446  * of the last block in the file and the unusual case of zeroing blocks
 447  * out beyond the size of the file.  This second case only happens
 448  * with fixed size extents and when the system crashes before the inode
 449  * size was updated but after blocks were allocated.  If fill is set,
 450  * then any holes in the range are filled and zeroed.  If not, the holes
 451  * are left alone as holes.
 452  */
 453
 454 int                                     /* error (positive) */
 455 xfs_zero_eof(
 456         xfs_inode_t     *ip,
 457         xfs_off_t       offset,         /* starting I/O offset */
 458         xfs_fsize_t     isize)          /* current inode size */
 459 {
 460         xfs_mount_t     *mp = ip->i_mount;
 461         xfs_fileoff_t   start_zero_fsb;
 462         xfs_fileoff_t   end_zero_fsb;
 463         xfs_fileoff_t   zero_count_fsb;
 464         xfs_fileoff_t   last_fsb;
 465         xfs_fileoff_t   zero_off;
 466         xfs_fsize_t     zero_len;
 467         int             nimaps;
 468         int             error = 0;
 469         xfs_bmbt_irec_t imap;
 470
 471         ASSERT(ismrlocked(&ip->i_lock, MR_UPDATE));
 472         ASSERT(ismrlocked(&ip->i_iolock, MR_UPDATE));
 473         ASSERT(offset > isize);
 474
 475         /*
 476          * First handle zeroing the block on which isize resides.
 477          * We only zero a part of that block so it is handled specially.
 478          */
 479         error = xfs_zero_last_block(ip, offset, isize);
 480         if (error) {
 481                 ASSERT(ismrlocked(&ip->i_lock, MR_UPDATE));
 482                 ASSERT(ismrlocked(&ip->i_iolock, MR_UPDATE));
 483                 return error;
 484         }
 485
 486         /*
 487          * Calculate the range between the new size and the old
 488          * where blocks needing to be zeroed may exist.  To get the
 489          * block where the last byte in the file currently resides,
 490          * we need to subtract one from the size and truncate back
 491          * to a block boundary.  We subtract 1 in case the size is
 492          * exactly on a block boundary.
 493          */
 494         last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
 495         start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
 496         end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
 497         ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
 498         if (last_fsb == end_zero_fsb) {
 499                 /*
 500                  * The size was only incremented on its last block.
 501                  * We took care of that above, so just return.
 502                  */
 503                 return 0;
 504         }
 505
 506         ASSERT(start_zero_fsb <= end_zero_fsb);
 507         while (start_zero_fsb <= end_zero_fsb) {
 508                 nimaps = 1;
 509                 zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
 510                 error = xfs_bmapi(NULL, ip, start_zero_fsb, zero_count_fsb,
 511                                   0, NULL, 0, &imap, &nimaps, NULL, NULL);
 512                 if (error) {
 513                         ASSERT(ismrlocked(&ip->i_lock, MR_UPDATE));
 514                         ASSERT(ismrlocked(&ip->i_iolock, MR_UPDATE));
 515                         return error;
 516                 }
 517                 ASSERT(nimaps > 0);
 518
 519                 if (imap.br_state == XFS_EXT_UNWRITTEN ||
 520                     imap.br_startblock == HOLESTARTBLOCK) {
 521                         /*
 522                          * This loop handles initializing pages that were
 523                          * partially initialized by the code below this
 524                          * loop. It basically zeroes the part of the page
 525                          * that sits on a hole and sets the page as P_HOLE
 526                          * and calls remapf if it is a mapped file.
 527                          */
 528                         start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 529                         ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 530                         continue;
 531                 }
 532
 533                 /*
 534                  * There are blocks we need to zero.
 535                  * Drop the inode lock while we're doing the I/O.
 536                  * We'll still have the iolock to protect us.
 537                  */
 538                 xfs_iunlock(ip, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 539
 540                 zero_off = XFS_FSB_TO_B(mp, start_zero_fsb);
 541                 zero_len = XFS_FSB_TO_B(mp, imap.br_blockcount);
 542
 543                 if ((zero_off + zero_len) > offset)
 544                         zero_len = offset - zero_off;
 545
 546                 error = xfs_iozero(ip, zero_off, zero_len);
 547                 if (error) {
 548                         goto out_lock;
 549                 }
 550
 551                 start_zero_fsb = imap.br_startoff + imap.br_blockcount;
 552                 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 553
 554                 xfs_ilock(ip, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 555         }
 556
 557         return 0;
 558
 559 out_lock:
 560         xfs_ilock(ip, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 561         ASSERT(error >= 0);
 562         return error;
 563 }
 564
 565 ssize_t                         /* bytes written, or (-) error */
 566 xfs_write(
 567         struct xfs_inode        *xip,
 568         struct kiocb            *iocb,
 569         const struct iovec      *iovp,
 570         unsigned int            nsegs,
 571         loff_t                  *offset,
 572         int                     ioflags)
 573 {
 574         struct file             *file = iocb->ki_filp;
 575         struct address_space    *mapping = file->f_mapping;
 576         struct inode            *inode = mapping->host;
 577         bhv_vnode_t             *vp = XFS_ITOV(xip);
 578         unsigned long           segs = nsegs;
 579         xfs_mount_t             *mp;
 580         ssize_t                 ret = 0, error = 0;
 581         xfs_fsize_t             isize, new_size;
 582         int                     iolock;
 583         int                     eventsent = 0;
 584         bhv_vrwlock_t           locktype;
 585         size_t                  ocount = 0, count;
 586         loff_t                  pos;
 587         int                     need_i_mutex;
 588
 589         XFS_STATS_INC(xs_write_calls);
 590
 591         error = generic_segment_checks(iovp, &segs, &ocount, VERIFY_READ);
 592         if (error)
 593                 return error;
 594
 595         count = ocount;
 596         pos = *offset;
 597
 598         if (count == 0)
 599                 return 0;
 600
 601         mp = xip->i_mount;
 602
 603         xfs_wait_for_freeze(mp, SB_FREEZE_WRITE);
 604
 605         if (XFS_FORCED_SHUTDOWN(mp))
 606                 return -EIO;
 607
 608 relock:
 609         if (ioflags & IO_ISDIRECT) {
 610                 iolock = XFS_IOLOCK_SHARED;
 611                 locktype = VRWLOCK_WRITE_DIRECT;
 612                 need_i_mutex = 0;
 613         } else {
 614                 iolock = XFS_IOLOCK_EXCL;
 615                 locktype = VRWLOCK_WRITE;
 616                 need_i_mutex = 1;
 617                 mutex_lock(&inode->i_mutex);
 618         }
 619
 620         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
 621
 622 start:
 623         error = -generic_write_checks(file, &pos, &count,
 624                                         S_ISBLK(inode->i_mode));
 625         if (error) {
 626                 xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 627                 goto out_unlock_mutex;
 628         }
 629
 630         if ((DM_EVENT_ENABLED(xip, DM_EVENT_WRITE) &&
 631             !(ioflags & IO_INVIS) && !eventsent)) {
 632                 int             dmflags = FILP_DELAY_FLAG(file);
 633
 634                 if (need_i_mutex)
 635                         dmflags |= DM_FLAGS_IMUX;
 636
 637                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 638                 error = XFS_SEND_DATA(xip->i_mount, DM_EVENT_WRITE, vp,
 639                                       pos, count,
 640                                       dmflags, &locktype);
 641                 if (error) {
 642                         goto out_unlock_internal;
 643                 }
 644                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 645                 eventsent = 1;
 646
 647                 /*
 648                  * The iolock was dropped and reacquired in XFS_SEND_DATA
 649                  * so we have to recheck the size when appending.
 650                  * We will only "goto start;" once, since having sent the
 651                  * event prevents another call to XFS_SEND_DATA, which is
 652                  * what allows the size to change in the first place.
 653                  */
 654                 if ((file->f_flags & O_APPEND) && pos != xip->i_size)
 655                         goto start;
 656         }
 657
 658         if (ioflags & IO_ISDIRECT) {
 659                 xfs_buftarg_t   *target =
 660                         (xip->i_d.di_flags & XFS_DIFLAG_REALTIME) ?
 661                                 mp->m_rtdev_targp : mp->m_ddev_targp;
 662
 663                 if ((pos & target->bt_smask) || (count & target->bt_smask)) {
 664                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 665                         return XFS_ERROR(-EINVAL);
 666                 }
 667
 668                 if (!need_i_mutex && (VN_CACHED(vp) || pos > xip->i_size)) {
 669                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 670                         iolock = XFS_IOLOCK_EXCL;
 671                         locktype = VRWLOCK_WRITE;
 672                         need_i_mutex = 1;
 673                         mutex_lock(&inode->i_mutex);
 674                         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
 675                         goto start;
 676                 }
 677         }
 678
 679         new_size = pos + count;
 680         if (new_size > xip->i_size)
 681                 xip->i_new_size = new_size;
 682
 683         if (likely(!(ioflags & IO_INVIS))) {
 684                 file_update_time(file);
 685                 xfs_ichgtime_fast(xip, inode,
 686                                   XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
 687         }
 688
 689         /*
 690          * If the offset is beyond the size of the file, we have a couple
 691          * of things to do. First, if there is already space allocated
 692          * we need to either create holes or zero the disk or ...
 693          *
 694          * If there is a page where the previous size lands, we need
 695          * to zero it out up to the new size.
 696          */
 697
 698         if (pos > xip->i_size) {
 699                 error = xfs_zero_eof(xip, pos, xip->i_size);
 700                 if (error) {
 701                         xfs_iunlock(xip, XFS_ILOCK_EXCL);
 702                         goto out_unlock_internal;
 703                 }
 704         }
 705         xfs_iunlock(xip, XFS_ILOCK_EXCL);
 706
 707         /*
 708          * If we're writing the file then make sure to clear the
 709          * setuid and setgid bits if the process is not being run
 710          * by root.  This keeps people from modifying setuid and
 711          * setgid binaries.
 712          */
 713
 714         if (((xip->i_d.di_mode & S_ISUID) ||
 715             ((xip->i_d.di_mode & (S_ISGID | S_IXGRP)) ==
 716                 (S_ISGID | S_IXGRP))) &&
 717              !capable(CAP_FSETID)) {
 718                 error = xfs_write_clear_setuid(xip);
 719                 if (likely(!error))
 720                         error = -remove_suid(file->f_path.dentry);
 721                 if (unlikely(error)) {
 722                         goto out_unlock_internal;
 723                 }
 724         }
 725
 726 retry:
 727         /* We can write back this queue in page reclaim */
 728         current->backing_dev_info = mapping->backing_dev_info;
 729
 730         if ((ioflags & IO_ISDIRECT)) {
 731                 if (VN_CACHED(vp)) {
 732                         WARN_ON(need_i_mutex == 0);
 733                         xfs_inval_cached_trace(xip, pos, -1,
 734                                         ctooff(offtoct(pos)), -1);
 735                         error = xfs_flushinval_pages(xip,
 736                                         ctooff(offtoct(pos)),
 737                                         -1, FI_REMAPF_LOCKED);
 738                         if (error)
 739                                 goto out_unlock_internal;
 740                 }
 741
 742                 if (need_i_mutex) {
 743                         /* demote the lock now the cached pages are gone */
 744                         xfs_ilock_demote(xip, XFS_IOLOCK_EXCL);
 745                         mutex_unlock(&inode->i_mutex);
 746
 747                         iolock = XFS_IOLOCK_SHARED;
 748                         locktype = VRWLOCK_WRITE_DIRECT;
 749                         need_i_mutex = 0;
 750                 }
 751
 752                 xfs_rw_enter_trace(XFS_DIOWR_ENTER, xip, (void *)iovp, segs,
 753                                 *offset, ioflags);
 754                 ret = generic_file_direct_write(iocb, iovp,
 755                                 &segs, pos, offset, count, ocount);
 756
 757                 /*
 758                  * direct-io write to a hole: fall through to buffered I/O
 759                  * for completing the rest of the request.
 760                  */
 761                 if (ret >= 0 && ret != count) {
 762                         XFS_STATS_ADD(xs_write_bytes, ret);
 763
 764                         pos += ret;
 765                         count -= ret;
 766
 767                         ioflags &= ~IO_ISDIRECT;
 768                         xfs_iunlock(xip, iolock);
 769                         goto relock;
 770                 }
 771         } else {
 772                 xfs_rw_enter_trace(XFS_WRITE_ENTER, xip, (void *)iovp, segs,
 773                                 *offset, ioflags);
 774                 ret = generic_file_buffered_write(iocb, iovp, segs,
 775                                 pos, offset, count, ret);
 776         }
 777
 778         current->backing_dev_info = NULL;
 779
 780         if (ret == -EIOCBQUEUED && !(ioflags & IO_ISAIO))
 781                 ret = wait_on_sync_kiocb(iocb);
 782
 783         if (ret == -ENOSPC &&
 784             DM_EVENT_ENABLED(xip, DM_EVENT_NOSPACE) && !(ioflags & IO_INVIS)) {
 785                 xfs_rwunlock(xip, locktype);
 786                 if (need_i_mutex)
 787                         mutex_unlock(&inode->i_mutex);
 788                 error = XFS_SEND_NAMESP(xip->i_mount, DM_EVENT_NOSPACE, vp,
 789                                 DM_RIGHT_NULL, vp, DM_RIGHT_NULL, NULL, NULL,
 790                                 0, 0, 0); /* Delay flag intentionally  unused */
 791                 if (need_i_mutex)
 792                         mutex_lock(&inode->i_mutex);
 793                 xfs_rwlock(xip, locktype);
 794                 if (error)
 795                         goto out_unlock_internal;
 796                 pos = xip->i_size;
 797                 ret = 0;
 798                 goto retry;
 799         }
 800
 801         isize = i_size_read(inode);
 802         if (unlikely(ret < 0 && ret != -EFAULT && *offset > isize))
 803                 *offset = isize;
 804
 805         if (*offset > xip->i_size) {
 806                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 807                 if (*offset > xip->i_size)
 808                         xip->i_size = *offset;
 809                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 810         }
 811
 812         error = -ret;
 813         if (ret <= 0)
 814                 goto out_unlock_internal;
 815
 816         XFS_STATS_ADD(xs_write_bytes, ret);
 817
 818         /* Handle various SYNC-type writes */
 819         if ((file->f_flags & O_SYNC) || IS_SYNC(inode)) {
 820                 int error2;
 821                 xfs_rwunlock(xip, locktype);
 822                 if (need_i_mutex)
 823                         mutex_unlock(&inode->i_mutex);
 824                 error2 = sync_page_range(inode, mapping, pos, ret);
 825                 if (!error)
 826                         error = error2;
 827                 if (need_i_mutex)
 828                         mutex_lock(&inode->i_mutex);
 829                 xfs_rwlock(xip, locktype);
 830                 error2 = xfs_write_sync_logforce(mp, xip);
 831                 if (!error)
 832                         error = error2;
 833         }
 834
 835  out_unlock_internal:
 836         if (xip->i_new_size) {
 837                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 838                 xip->i_new_size = 0;
 839                 /*
 840                  * If this was a direct or synchronous I/O that failed (such
 841                  * as ENOSPC) then part of the I/O may have been written to
 842                  * disk before the error occured.  In this case the on-disk
 843                  * file size may have been adjusted beyond the in-memory file
 844                  * size and now needs to be truncated back.
 845                  */
 846                 if (xip->i_d.di_size > xip->i_size)
 847                         xip->i_d.di_size = xip->i_size;
 848                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 849         }
 850         xfs_rwunlock(xip, locktype);
 851  out_unlock_mutex:
 852         if (need_i_mutex)
 853                 mutex_unlock(&inode->i_mutex);
 854         return -error;
 855 }
 856
 857 /*
 858  * All xfs metadata buffers except log state machine buffers
 859  * get this attached as their b_bdstrat callback function.
 860  * This is so that we can catch a buffer
 861  * after prematurely unpinning it to forcibly shutdown the filesystem.
 862  */
 863 int
 864 xfs_bdstrat_cb(struct xfs_buf *bp)
 865 {
 866         xfs_mount_t     *mp;
 867
 868         mp = XFS_BUF_FSPRIVATE3(bp, xfs_mount_t *);
 869         if (!XFS_FORCED_SHUTDOWN(mp)) {
 870                 xfs_buf_iorequest(bp);
 871                 return 0;
 872         } else {
 873                 xfs_buftrace("XFS__BDSTRAT IOERROR", bp);
 874                 /*
 875                  * Metadata write that didn't get logged but
 876                  * written delayed anyway. These aren't associated
 877                  * with a transaction, and can be ignored.
 878                  */
 879                 if (XFS_BUF_IODONE_FUNC(bp) == NULL &&
 880                     (XFS_BUF_ISREAD(bp)) == 0)
 881                         return (xfs_bioerror_relse(bp));
 882                 else
 883                         return (xfs_bioerror(bp));
 884         }
 885 }
 886
 887 /*
 888  * Wrapper around bdstrat so that we can stop data
 889  * from going to disk in case we are shutting down the filesystem.
 890  * Typically user data goes thru this path; one of the exceptions
 891  * is the superblock.
 892  */
 893 int
 894 xfsbdstrat(
 895         struct xfs_mount        *mp,
 896         struct xfs_buf          *bp)
 897 {
 898         ASSERT(mp);
 899         if (!XFS_FORCED_SHUTDOWN(mp)) {
 900                 /* Grio redirection would go here
 901                  * if (XFS_BUF_IS_GRIO(bp)) {
 902                  */
 903
 904                 xfs_buf_iorequest(bp);
 905                 return 0;
 906         }
 907
 908         xfs_buftrace("XFSBDSTRAT IOERROR", bp);
 909         return (xfs_bioerror_relse(bp));
 910 }
 911
 912 /*
 913  * If the underlying (data/log/rt) device is readonly, there are some
 914  * operations that cannot proceed.
 915  */
 916 int
 917 xfs_dev_is_read_only(
 918         xfs_mount_t             *mp,
 919         char                    *message)
 920 {
 921         if (xfs_readonly_buftarg(mp->m_ddev_targp) ||
 922             xfs_readonly_buftarg(mp->m_logdev_targp) ||
 923             (mp->m_rtdev_targp && xfs_readonly_buftarg(mp->m_rtdev_targp))) {
 924                 cmn_err(CE_NOTE,
 925                         "XFS: %s required on read-only device.", message);
 926                 cmn_err(CE_NOTE,
 927                         "XFS: write access unavailable, cannot proceed.");
 928                 return EROFS;
 929         }
 930         return 0;
 931 }