fs/xfs/linux-2.6/xfs_lrw.c

   1 /*
   2  * Copyright (c) 2000-2003 Silicon Graphics, Inc.  All Rights Reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms of version 2 of the GNU General Public License as
   6  * published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it would be useful, but
   9  * WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
  11  *
  12  * Further, this software is distributed without any warranty that it is
  13  * free of the rightful claim of any third person regarding infringement
  14  * or the like.  Any license provided herein, whether implied or
  15  * otherwise, applies only to this software file.  Patent licenses, if
  16  * any, provided herein do not apply to combinations of this program with
  17  * other software, or any other product whatsoever.
  18  *
  19  * You should have received a copy of the GNU General Public License along
  20  * with this program; if not, write the Free Software Foundation, Inc., 59
  21  * Temple Place - Suite 330, Boston MA 02111-1307, USA.
  22  *
  23  * Contact information: Silicon Graphics, Inc., 1600 Amphitheatre Pkwy,
  24  * Mountain View, CA  94043, or:
  25  *
  26  * http://www.sgi.com
  27  *
  28  * For further information regarding this notice, see:
  29  *
  30  * http://oss.sgi.com/projects/GenInfo/SGIGPLNoticeExplan/
  31  */
  32 /*
  33  *  fs/xfs/linux/xfs_lrw.c (Linux Read Write stuff)
  34  *
  35  */
  36
  37 #include "xfs.h"
  38
  39 #include "xfs_fs.h"
  40 #include "xfs_inum.h"
  41 #include "xfs_log.h"
  42 #include "xfs_trans.h"
  43 #include "xfs_sb.h"
  44 #include "xfs_ag.h"
  45 #include "xfs_dir.h"
  46 #include "xfs_dir2.h"
  47 #include "xfs_alloc.h"
  48 #include "xfs_dmapi.h"
  49 #include "xfs_quota.h"
  50 #include "xfs_mount.h"
  51 #include "xfs_alloc_btree.h"
  52 #include "xfs_bmap_btree.h"
  53 #include "xfs_ialloc_btree.h"
  54 #include "xfs_btree.h"
  55 #include "xfs_ialloc.h"
  56 #include "xfs_attr_sf.h"
  57 #include "xfs_dir_sf.h"
  58 #include "xfs_dir2_sf.h"
  59 #include "xfs_dinode.h"
  60 #include "xfs_inode.h"
  61 #include "xfs_bmap.h"
  62 #include "xfs_bit.h"
  63 #include "xfs_rtalloc.h"
  64 #include "xfs_error.h"
  65 #include "xfs_itable.h"
  66 #include "xfs_rw.h"
  67 #include "xfs_acl.h"
  68 #include "xfs_cap.h"
  69 #include "xfs_mac.h"
  70 #include "xfs_attr.h"
  71 #include "xfs_inode_item.h"
  72 #include "xfs_buf_item.h"
  73 #include "xfs_utils.h"
  74 #include "xfs_iomap.h"
  75
  76 #include <linux/capability.h>
  77 #include <linux/writeback.h>
  78
  79
  80 #if defined(XFS_RW_TRACE)
  81 void
  82 xfs_rw_enter_trace(
  83         int                     tag,
  84         xfs_iocore_t            *io,
  85         void                    *data,
  86         size_t                  segs,
  87         loff_t                  offset,
  88         int                     ioflags)
  89 {
  90         xfs_inode_t     *ip = XFS_IO_INODE(io);
  91
  92         if (ip->i_rwtrace == NULL)
  93                 return;
  94         ktrace_enter(ip->i_rwtrace,
  95                 (void *)(unsigned long)tag,
  96                 (void *)ip,
  97                 (void *)((unsigned long)((ip->i_d.di_size >> 32) & 0xffffffff)),
  98                 (void *)((unsigned long)(ip->i_d.di_size & 0xffffffff)),
  99                 (void *)data,
 100                 (void *)((unsigned long)segs),
 101                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
 102                 (void *)((unsigned long)(offset & 0xffffffff)),
 103                 (void *)((unsigned long)ioflags),
 104                 (void *)((unsigned long)((io->io_new_size >> 32) & 0xffffffff)),
 105                 (void *)((unsigned long)(io->io_new_size & 0xffffffff)),
 106                 (void *)NULL,
 107                 (void *)NULL,
 108                 (void *)NULL,
 109                 (void *)NULL,
 110                 (void *)NULL);
 111 }
 112
 113 void
 114 xfs_inval_cached_trace(
 115         xfs_iocore_t    *io,
 116         xfs_off_t       offset,
 117         xfs_off_t       len,
 118         xfs_off_t       first,
 119         xfs_off_t       last)
 120 {
 121         xfs_inode_t     *ip = XFS_IO_INODE(io);
 122
 123         if (ip->i_rwtrace == NULL)
 124                 return;
 125         ktrace_enter(ip->i_rwtrace,
 126                 (void *)(__psint_t)XFS_INVAL_CACHED,
 127                 (void *)ip,
 128                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
 129                 (void *)((unsigned long)(offset & 0xffffffff)),
 130                 (void *)((unsigned long)((len >> 32) & 0xffffffff)),
 131                 (void *)((unsigned long)(len & 0xffffffff)),
 132                 (void *)((unsigned long)((first >> 32) & 0xffffffff)),
 133                 (void *)((unsigned long)(first & 0xffffffff)),
 134                 (void *)((unsigned long)((last >> 32) & 0xffffffff)),
 135                 (void *)((unsigned long)(last & 0xffffffff)),
 136                 (void *)NULL,
 137                 (void *)NULL,
 138                 (void *)NULL,
 139                 (void *)NULL,
 140                 (void *)NULL,
 141                 (void *)NULL);
 142 }
 143 #endif
 144
 145 /*
 146  *      xfs_iozero
 147  *
 148  *      xfs_iozero clears the specified range of buffer supplied,
 149  *      and marks all the affected blocks as valid and modified.  If
 150  *      an affected block is not allocated, it will be allocated.  If
 151  *      an affected block is not completely overwritten, and is not
 152  *      valid before the operation, it will be read from disk before
 153  *      being partially zeroed.
 154  */
 155 STATIC int
 156 xfs_iozero(
 157         struct inode            *ip,    /* inode                        */
 158         loff_t                  pos,    /* offset in file               */
 159         size_t                  count,  /* size of data to zero         */
 160         loff_t                  end_size)       /* max file size to set */
 161 {
 162         unsigned                bytes;
 163         struct page             *page;
 164         struct address_space    *mapping;
 165         char                    *kaddr;
 166         int                     status;
 167
 168         mapping = ip->i_mapping;
 169         do {
 170                 unsigned long index, offset;
 171
 172                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
 173                 index = pos >> PAGE_CACHE_SHIFT;
 174                 bytes = PAGE_CACHE_SIZE - offset;
 175                 if (bytes > count)
 176                         bytes = count;
 177
 178                 status = -ENOMEM;
 179                 page = grab_cache_page(mapping, index);
 180                 if (!page)
 181                         break;
 182
 183                 kaddr = kmap(page);
 184                 status = mapping->a_ops->prepare_write(NULL, page, offset,
 185                                                         offset + bytes);
 186                 if (status) {
 187                         goto unlock;
 188                 }
 189
 190                 memset((void *) (kaddr + offset), 0, bytes);
 191                 flush_dcache_page(page);
 192                 status = mapping->a_ops->commit_write(NULL, page, offset,
 193                                                         offset + bytes);
 194                 if (!status) {
 195                         pos += bytes;
 196                         count -= bytes;
 197                         if (pos > i_size_read(ip))
 198                                 i_size_write(ip, pos < end_size ? pos : end_size);
 199                 }
 200
 201 unlock:
 202                 kunmap(page);
 203                 unlock_page(page);
 204                 page_cache_release(page);
 205                 if (status)
 206                         break;
 207         } while (count);
 208
 209         return (-status);
 210 }
 211
 212 /*
 213  * xfs_inval_cached_pages
 214  *
 215  * This routine is responsible for keeping direct I/O and buffered I/O
 216  * somewhat coherent.  From here we make sure that we're at least
 217  * temporarily holding the inode I/O lock exclusively and then call
 218  * the page cache to flush and invalidate any cached pages.  If there
 219  * are no cached pages this routine will be very quick.
 220  */
 221 void
 222 xfs_inval_cached_pages(
 223         vnode_t         *vp,
 224         xfs_iocore_t    *io,
 225         xfs_off_t       offset,
 226         int             write,
 227         int             relock)
 228 {
 229         if (VN_CACHED(vp)) {
 230                 xfs_inval_cached_trace(io, offset, -1, ctooff(offtoct(offset)), -1);
 231                 VOP_FLUSHINVAL_PAGES(vp, ctooff(offtoct(offset)), -1, FI_REMAPF_LOCKED);
 232         }
 233
 234 }
 235
 236 ssize_t                 /* bytes read, or (-)  error */
 237 xfs_read(
 238         bhv_desc_t              *bdp,
 239         struct kiocb            *iocb,
 240         const struct iovec      *iovp,
 241         unsigned int            segs,
 242         loff_t                  *offset,
 243         int                     ioflags,
 244         cred_t                  *credp)
 245 {
 246         struct file             *file = iocb->ki_filp;
 247         size_t                  size = 0;
 248         ssize_t                 ret;
 249         xfs_fsize_t             n;
 250         xfs_inode_t             *ip;
 251         xfs_mount_t             *mp;
 252         vnode_t                 *vp;
 253         unsigned long           seg;
 254
 255         ip = XFS_BHVTOI(bdp);
 256         vp = BHV_TO_VNODE(bdp);
 257         mp = ip->i_mount;
 258
 259         XFS_STATS_INC(xs_read_calls);
 260
 261         /* START copy & waste from filemap.c */
 262         for (seg = 0; seg < segs; seg++) {
 263                 const struct iovec *iv = &iovp[seg];
 264
 265                 /*
 266                  * If any segment has a negative length, or the cumulative
 267                  * length ever wraps negative then return -EINVAL.
 268                  */
 269                 size += iv->iov_len;
 270                 if (unlikely((ssize_t)(size|iv->iov_len) < 0))
 271                         return XFS_ERROR(-EINVAL);
 272         }
 273         /* END copy & waste from filemap.c */
 274
 275         if (ioflags & IO_ISDIRECT) {
 276                 xfs_buftarg_t   *target =
 277                         (ip->i_d.di_flags & XFS_DIFLAG_REALTIME) ?
 278                                 mp->m_rtdev_targp : mp->m_ddev_targp;
 279                 if ((*offset & target->pbr_smask) ||
 280                     (size & target->pbr_smask)) {
 281                         if (*offset == ip->i_d.di_size) {
 282                                 return (0);
 283                         }
 284                         return -XFS_ERROR(EINVAL);
 285                 }
 286         }
 287
 288         n = XFS_MAXIOFFSET(mp) - *offset;
 289         if ((n <= 0) || (size == 0))
 290                 return 0;
 291
 292         if (n < size)
 293                 size = n;
 294
 295         if (XFS_FORCED_SHUTDOWN(mp)) {
 296                 return -EIO;
 297         }
 298
 299         xfs_ilock(ip, XFS_IOLOCK_SHARED);
 300
 301         if (DM_EVENT_ENABLED(vp->v_vfsp, ip, DM_EVENT_READ) &&
 302             !(ioflags & IO_INVIS)) {
 303                 vrwlock_t locktype = VRWLOCK_READ;
 304
 305                 ret = XFS_SEND_DATA(mp, DM_EVENT_READ,
 306                                         BHV_TO_VNODE(bdp), *offset, size,
 307                                         FILP_DELAY_FLAG(file), &locktype);
 308                 if (ret) {
 309                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 310                         return -ret;
 311                 }
 312         }
 313
 314         xfs_rw_enter_trace(XFS_READ_ENTER, &ip->i_iocore,
 315                                 (void *)iovp, segs, *offset, ioflags);
 316         ret = __generic_file_aio_read(iocb, iovp, segs, offset);
 317         if (ret == -EIOCBQUEUED)
 318                 ret = wait_on_sync_kiocb(iocb);
 319
 320         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 321
 322         if (ret > 0)
 323                 XFS_STATS_ADD(xs_read_bytes, ret);
 324
 325         if (likely(!(ioflags & IO_INVIS)))
 326                 xfs_ichgtime(ip, XFS_ICHGTIME_ACC);
 327
 328         return ret;
 329 }
 330
 331 ssize_t
 332 xfs_sendfile(
 333         bhv_desc_t              *bdp,
 334         struct file             *filp,
 335         loff_t                  *offset,
 336         int                     ioflags,
 337         size_t                  count,
 338         read_actor_t            actor,
 339         void                    *target,
 340         cred_t                  *credp)
 341 {
 342         ssize_t                 ret;
 343         xfs_fsize_t             n;
 344         xfs_inode_t             *ip;
 345         xfs_mount_t             *mp;
 346         vnode_t                 *vp;
 347
 348         ip = XFS_BHVTOI(bdp);
 349         vp = BHV_TO_VNODE(bdp);
 350         mp = ip->i_mount;
 351
 352         XFS_STATS_INC(xs_read_calls);
 353
 354         n = XFS_MAXIOFFSET(mp) - *offset;
 355         if ((n <= 0) || (count == 0))
 356                 return 0;
 357
 358         if (n < count)
 359                 count = n;
 360
 361         if (XFS_FORCED_SHUTDOWN(ip->i_mount))
 362                 return -EIO;
 363
 364         xfs_ilock(ip, XFS_IOLOCK_SHARED);
 365
 366         if (DM_EVENT_ENABLED(vp->v_vfsp, ip, DM_EVENT_READ) &&
 367             (!(ioflags & IO_INVIS))) {
 368                 vrwlock_t locktype = VRWLOCK_READ;
 369                 int error;
 370
 371                 error = XFS_SEND_DATA(mp, DM_EVENT_READ, BHV_TO_VNODE(bdp), *offset, count,
 372                                       FILP_DELAY_FLAG(filp), &locktype);
 373                 if (error) {
 374                         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 375                         return -error;
 376                 }
 377         }
 378         xfs_rw_enter_trace(XFS_SENDFILE_ENTER, &ip->i_iocore,
 379                    (void *)(unsigned long)target, count, *offset, ioflags);
 380         ret = generic_file_sendfile(filp, offset, count, actor, target);
 381
 382         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 383
 384         if (ret > 0)
 385                 XFS_STATS_ADD(xs_read_bytes, ret);
 386
 387         if (likely(!(ioflags & IO_INVIS)))
 388                 xfs_ichgtime(ip, XFS_ICHGTIME_ACC);
 389
 390         return ret;
 391 }
 392
 393 /*
 394  * This routine is called to handle zeroing any space in the last
 395  * block of the file that is beyond the EOF.  We do this since the
 396  * size is being increased without writing anything to that block
 397  * and we don't want anyone to read the garbage on the disk.
 398  */
 399 STATIC int                              /* error (positive) */
 400 xfs_zero_last_block(
 401         struct inode    *ip,
 402         xfs_iocore_t    *io,
 403         xfs_off_t       offset,
 404         xfs_fsize_t     isize,
 405         xfs_fsize_t     end_size)
 406 {
 407         xfs_fileoff_t   last_fsb;
 408         xfs_mount_t     *mp;
 409         int             nimaps;
 410         int             zero_offset;
 411         int             zero_len;
 412         int             isize_fsb_offset;
 413         int             error = 0;
 414         xfs_bmbt_irec_t imap;
 415         loff_t          loff;
 416         size_t          lsize;
 417
 418         ASSERT(ismrlocked(io->io_lock, MR_UPDATE) != 0);
 419         ASSERT(offset > isize);
 420
 421         mp = io->io_mount;
 422
 423         isize_fsb_offset = XFS_B_FSB_OFFSET(mp, isize);
 424         if (isize_fsb_offset == 0) {
 425                 /*
 426                  * There are no extra bytes in the last block on disk to
 427                  * zero, so return.
 428                  */
 429                 return 0;
 430         }
 431
 432         last_fsb = XFS_B_TO_FSBT(mp, isize);
 433         nimaps = 1;
 434         error = XFS_BMAPI(mp, NULL, io, last_fsb, 1, 0, NULL, 0, &imap,
 435                           &nimaps, NULL);
 436         if (error) {
 437                 return error;
 438         }
 439         ASSERT(nimaps > 0);
 440         /*
 441          * If the block underlying isize is just a hole, then there
 442          * is nothing to zero.
 443          */
 444         if (imap.br_startblock == HOLESTARTBLOCK) {
 445                 return 0;
 446         }
 447         /*
 448          * Zero the part of the last block beyond the EOF, and write it
 449          * out sync.  We need to drop the ilock while we do this so we
 450          * don't deadlock when the buffer cache calls back to us.
 451          */
 452         XFS_IUNLOCK(mp, io, XFS_ILOCK_EXCL| XFS_EXTSIZE_RD);
 453         loff = XFS_FSB_TO_B(mp, last_fsb);
 454         lsize = XFS_FSB_TO_B(mp, 1);
 455
 456         zero_offset = isize_fsb_offset;
 457         zero_len = mp->m_sb.sb_blocksize - isize_fsb_offset;
 458
 459         error = xfs_iozero(ip, loff + zero_offset, zero_len, end_size);
 460
 461         XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 462         ASSERT(error >= 0);
 463         return error;
 464 }
 465
 466 /*
 467  * Zero any on disk space between the current EOF and the new,
 468  * larger EOF.  This handles the normal case of zeroing the remainder
 469  * of the last block in the file and the unusual case of zeroing blocks
 470  * out beyond the size of the file.  This second case only happens
 471  * with fixed size extents and when the system crashes before the inode
 472  * size was updated but after blocks were allocated.  If fill is set,
 473  * then any holes in the range are filled and zeroed.  If not, the holes
 474  * are left alone as holes.
 475  */
 476
 477 int                                     /* error (positive) */
 478 xfs_zero_eof(
 479         vnode_t         *vp,
 480         xfs_iocore_t    *io,
 481         xfs_off_t       offset,         /* starting I/O offset */
 482         xfs_fsize_t     isize,          /* current inode size */
 483         xfs_fsize_t     end_size)       /* terminal inode size */
 484 {
 485         struct inode    *ip = LINVFS_GET_IP(vp);
 486         xfs_fileoff_t   start_zero_fsb;
 487         xfs_fileoff_t   end_zero_fsb;
 488         xfs_fileoff_t   prev_zero_fsb;
 489         xfs_fileoff_t   zero_count_fsb;
 490         xfs_fileoff_t   last_fsb;
 491         xfs_extlen_t    buf_len_fsb;
 492         xfs_extlen_t    prev_zero_count;
 493         xfs_mount_t     *mp;
 494         int             nimaps;
 495         int             error = 0;
 496         xfs_bmbt_irec_t imap;
 497         loff_t          loff;
 498         size_t          lsize;
 499
 500         ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
 501         ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
 502
 503         mp = io->io_mount;
 504
 505         /*
 506          * First handle zeroing the block on which isize resides.
 507          * We only zero a part of that block so it is handled specially.
 508          */
 509         error = xfs_zero_last_block(ip, io, offset, isize, end_size);
 510         if (error) {
 511                 ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
 512                 ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
 513                 return error;
 514         }
 515
 516         /*
 517          * Calculate the range between the new size and the old
 518          * where blocks needing to be zeroed may exist.  To get the
 519          * block where the last byte in the file currently resides,
 520          * we need to subtract one from the size and truncate back
 521          * to a block boundary.  We subtract 1 in case the size is
 522          * exactly on a block boundary.
 523          */
 524         last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
 525         start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
 526         end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
 527         ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
 528         if (last_fsb == end_zero_fsb) {
 529                 /*
 530                  * The size was only incremented on its last block.
 531                  * We took care of that above, so just return.
 532                  */
 533                 return 0;
 534         }
 535
 536         ASSERT(start_zero_fsb <= end_zero_fsb);
 537         prev_zero_fsb = NULLFILEOFF;
 538         prev_zero_count = 0;
 539         while (start_zero_fsb <= end_zero_fsb) {
 540                 nimaps = 1;
 541                 zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
 542                 error = XFS_BMAPI(mp, NULL, io, start_zero_fsb, zero_count_fsb,
 543                                   0, NULL, 0, &imap, &nimaps, NULL);
 544                 if (error) {
 545                         ASSERT(ismrlocked(io->io_lock, MR_UPDATE));
 546                         ASSERT(ismrlocked(io->io_iolock, MR_UPDATE));
 547                         return error;
 548                 }
 549                 ASSERT(nimaps > 0);
 550
 551                 if (imap.br_state == XFS_EXT_UNWRITTEN ||
 552                     imap.br_startblock == HOLESTARTBLOCK) {
 553                         /*
 554                          * This loop handles initializing pages that were
 555                          * partially initialized by the code below this
 556                          * loop. It basically zeroes the part of the page
 557                          * that sits on a hole and sets the page as P_HOLE
 558                          * and calls remapf if it is a mapped file.
 559                          */
 560                         prev_zero_fsb = NULLFILEOFF;
 561                         prev_zero_count = 0;
 562                         start_zero_fsb = imap.br_startoff +
 563                                          imap.br_blockcount;
 564                         ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 565                         continue;
 566                 }
 567
 568                 /*
 569                  * There are blocks in the range requested.
 570                  * Zero them a single write at a time.  We actually
 571                  * don't zero the entire range returned if it is
 572                  * too big and simply loop around to get the rest.
 573                  * That is not the most efficient thing to do, but it
 574                  * is simple and this path should not be exercised often.
 575                  */
 576                 buf_len_fsb = XFS_FILBLKS_MIN(imap.br_blockcount,
 577                                               mp->m_writeio_blocks << 8);
 578                 /*
 579                  * Drop the inode lock while we're doing the I/O.
 580                  * We'll still have the iolock to protect us.
 581                  */
 582                 XFS_IUNLOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 583
 584                 loff = XFS_FSB_TO_B(mp, start_zero_fsb);
 585                 lsize = XFS_FSB_TO_B(mp, buf_len_fsb);
 586
 587                 error = xfs_iozero(ip, loff, lsize, end_size);
 588
 589                 if (error) {
 590                         goto out_lock;
 591                 }
 592
 593                 prev_zero_fsb = start_zero_fsb;
 594                 prev_zero_count = buf_len_fsb;
 595                 start_zero_fsb = imap.br_startoff + buf_len_fsb;
 596                 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
 597
 598                 XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 599         }
 600
 601         return 0;
 602
 603 out_lock:
 604
 605         XFS_ILOCK(mp, io, XFS_ILOCK_EXCL|XFS_EXTSIZE_RD);
 606         ASSERT(error >= 0);
 607         return error;
 608 }
 609
 610 ssize_t                         /* bytes written, or (-) error */
 611 xfs_write(
 612         bhv_desc_t              *bdp,
 613         struct kiocb            *iocb,
 614         const struct iovec      *iovp,
 615         unsigned int            nsegs,
 616         loff_t                  *offset,
 617         int                     ioflags,
 618         cred_t                  *credp)
 619 {
 620         struct file             *file = iocb->ki_filp;
 621         struct address_space    *mapping = file->f_mapping;
 622         struct inode            *inode = mapping->host;
 623         unsigned long           segs = nsegs;
 624         xfs_inode_t             *xip;
 625         xfs_mount_t             *mp;
 626         ssize_t                 ret = 0, error = 0;
 627         xfs_fsize_t             isize, new_size;
 628         xfs_iocore_t            *io;
 629         vnode_t                 *vp;
 630         unsigned long           seg;
 631         int                     iolock;
 632         int                     eventsent = 0;
 633         vrwlock_t               locktype;
 634         size_t                  ocount = 0, count;
 635         loff_t                  pos;
 636         int                     need_isem = 1, need_flush = 0;
 637
 638         XFS_STATS_INC(xs_write_calls);
 639
 640         vp = BHV_TO_VNODE(bdp);
 641         xip = XFS_BHVTOI(bdp);
 642
 643         for (seg = 0; seg < segs; seg++) {
 644                 const struct iovec *iv = &iovp[seg];
 645
 646                 /*
 647                  * If any segment has a negative length, or the cumulative
 648                  * length ever wraps negative then return -EINVAL.
 649                  */
 650                 ocount += iv->iov_len;
 651                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
 652                         return -EINVAL;
 653                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
 654                         continue;
 655                 if (seg == 0)
 656                         return -EFAULT;
 657                 segs = seg;
 658                 ocount -= iv->iov_len;  /* This segment is no good */
 659                 break;
 660         }
 661
 662         count = ocount;
 663         pos = *offset;
 664
 665         if (count == 0)
 666                 return 0;
 667
 668         io = &xip->i_iocore;
 669         mp = io->io_mount;
 670
 671         if (XFS_FORCED_SHUTDOWN(mp))
 672                 return -EIO;
 673
 674         if (ioflags & IO_ISDIRECT) {
 675                 xfs_buftarg_t   *target =
 676                         (xip->i_d.di_flags & XFS_DIFLAG_REALTIME) ?
 677                                 mp->m_rtdev_targp : mp->m_ddev_targp;
 678
 679                 if ((pos & target->pbr_smask) || (count & target->pbr_smask))
 680                         return XFS_ERROR(-EINVAL);
 681
 682                 if (!VN_CACHED(vp) && pos < i_size_read(inode))
 683                         need_isem = 0;
 684
 685                 if (VN_CACHED(vp))
 686                         need_flush = 1;
 687         }
 688
 689 relock:
 690         if (need_isem) {
 691                 iolock = XFS_IOLOCK_EXCL;
 692                 locktype = VRWLOCK_WRITE;
 693
 694                 down(&inode->i_sem);
 695         } else {
 696                 iolock = XFS_IOLOCK_SHARED;
 697                 locktype = VRWLOCK_WRITE_DIRECT;
 698         }
 699
 700         xfs_ilock(xip, XFS_ILOCK_EXCL|iolock);
 701
 702         isize = i_size_read(inode);
 703
 704         if (file->f_flags & O_APPEND)
 705                 *offset = isize;
 706
 707 start:
 708         error = -generic_write_checks(file, &pos, &count,
 709                                         S_ISBLK(inode->i_mode));
 710         if (error) {
 711                 xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 712                 goto out_unlock_isem;
 713         }
 714
 715         new_size = pos + count;
 716         if (new_size > isize)
 717                 io->io_new_size = new_size;
 718
 719         if ((DM_EVENT_ENABLED(vp->v_vfsp, xip, DM_EVENT_WRITE) &&
 720             !(ioflags & IO_INVIS) && !eventsent)) {
 721                 loff_t          savedsize = pos;
 722                 int             dmflags = FILP_DELAY_FLAG(file);
 723
 724                 if (need_isem)
 725                         dmflags |= DM_FLAGS_ISEM;
 726
 727                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 728                 error = XFS_SEND_DATA(xip->i_mount, DM_EVENT_WRITE, vp,
 729                                       pos, count,
 730                                       dmflags, &locktype);
 731                 if (error) {
 732                         xfs_iunlock(xip, iolock);
 733                         goto out_unlock_isem;
 734                 }
 735                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 736                 eventsent = 1;
 737
 738                 /*
 739                  * The iolock was dropped and reaquired in XFS_SEND_DATA
 740                  * so we have to recheck the size when appending.
 741                  * We will only "goto start;" once, since having sent the
 742                  * event prevents another call to XFS_SEND_DATA, which is
 743                  * what allows the size to change in the first place.
 744                  */
 745                 if ((file->f_flags & O_APPEND) && savedsize != isize) {
 746                         pos = isize = xip->i_d.di_size;
 747                         goto start;
 748                 }
 749         }
 750
 751         /*
 752          * On Linux, generic_file_write updates the times even if
 753          * no data is copied in so long as the write had a size.
 754          *
 755          * We must update xfs' times since revalidate will overcopy xfs.
 756          */
 757         if (!(ioflags & IO_INVIS)) {
 758                 xfs_ichgtime(xip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
 759                 inode_update_time(inode, 1);
 760         }
 761
 762         /*
 763          * If the offset is beyond the size of the file, we have a couple
 764          * of things to do. First, if there is already space allocated
 765          * we need to either create holes or zero the disk or ...
 766          *
 767          * If there is a page where the previous size lands, we need
 768          * to zero it out up to the new size.
 769          */
 770
 771         if (pos > isize) {
 772                 error = xfs_zero_eof(BHV_TO_VNODE(bdp), io, pos,
 773                                         isize, pos + count);
 774                 if (error) {
 775                         xfs_iunlock(xip, XFS_ILOCK_EXCL|iolock);
 776                         goto out_unlock_isem;
 777                 }
 778         }
 779         xfs_iunlock(xip, XFS_ILOCK_EXCL);
 780
 781         /*
 782          * If we're writing the file then make sure to clear the
 783          * setuid and setgid bits if the process is not being run
 784          * by root.  This keeps people from modifying setuid and
 785          * setgid binaries.
 786          */
 787
 788         if (((xip->i_d.di_mode & S_ISUID) ||
 789             ((xip->i_d.di_mode & (S_ISGID | S_IXGRP)) ==
 790                 (S_ISGID | S_IXGRP))) &&
 791              !capable(CAP_FSETID)) {
 792                 error = xfs_write_clear_setuid(xip);
 793                 if (likely(!error))
 794                         error = -remove_suid(file->f_dentry);
 795                 if (unlikely(error)) {
 796                         xfs_iunlock(xip, iolock);
 797                         goto out_unlock_isem;
 798                 }
 799         }
 800
 801 retry:
 802         /* We can write back this queue in page reclaim */
 803         current->backing_dev_info = mapping->backing_dev_info;
 804
 805         if ((ioflags & IO_ISDIRECT)) {
 806                 if (need_flush) {
 807                         xfs_inval_cached_trace(io, pos, -1,
 808                                         ctooff(offtoct(pos)), -1);
 809                         VOP_FLUSHINVAL_PAGES(vp, ctooff(offtoct(pos)),
 810                                         -1, FI_REMAPF_LOCKED);
 811                 }
 812
 813                 if (need_isem) {
 814                         /* demote the lock now the cached pages are gone */
 815                         XFS_ILOCK_DEMOTE(mp, io, XFS_IOLOCK_EXCL);
 816                         up(&inode->i_sem);
 817
 818                         iolock = XFS_IOLOCK_SHARED;
 819                         locktype = VRWLOCK_WRITE_DIRECT;
 820                         need_isem = 0;
 821                 }
 822
 823                 xfs_rw_enter_trace(XFS_DIOWR_ENTER, io, (void *)iovp, segs,
 824                                 *offset, ioflags);
 825                 ret = generic_file_direct_write(iocb, iovp,
 826                                 &segs, pos, offset, count, ocount);
 827
 828                 /*
 829                  * direct-io write to a hole: fall through to buffered I/O
 830                  * for completing the rest of the request.
 831                  */
 832                 if (ret >= 0 && ret != count) {
 833                         XFS_STATS_ADD(xs_write_bytes, ret);
 834
 835                         pos += ret;
 836                         count -= ret;
 837
 838                         need_isem = 1;
 839                         ioflags &= ~IO_ISDIRECT;
 840                         xfs_iunlock(xip, iolock);
 841                         goto relock;
 842                 }
 843         } else {
 844                 xfs_rw_enter_trace(XFS_WRITE_ENTER, io, (void *)iovp, segs,
 845                                 *offset, ioflags);
 846                 ret = generic_file_buffered_write(iocb, iovp, segs,
 847                                 pos, offset, count, ret);
 848         }
 849
 850         current->backing_dev_info = NULL;
 851
 852         if (ret == -EIOCBQUEUED)
 853                 ret = wait_on_sync_kiocb(iocb);
 854
 855         if ((ret == -ENOSPC) &&
 856             DM_EVENT_ENABLED(vp->v_vfsp, xip, DM_EVENT_NOSPACE) &&
 857             !(ioflags & IO_INVIS)) {
 858
 859                 xfs_rwunlock(bdp, locktype);
 860                 error = XFS_SEND_NAMESP(xip->i_mount, DM_EVENT_NOSPACE, vp,
 861                                 DM_RIGHT_NULL, vp, DM_RIGHT_NULL, NULL, NULL,
 862                                 0, 0, 0); /* Delay flag intentionally  unused */
 863                 if (error)
 864                         goto out_unlock_isem;
 865                 xfs_rwlock(bdp, locktype);
 866                 pos = xip->i_d.di_size;
 867                 goto retry;
 868         }
 869
 870         if (*offset > xip->i_d.di_size) {
 871                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 872                 if (*offset > xip->i_d.di_size) {
 873                         xip->i_d.di_size = *offset;
 874                         i_size_write(inode, *offset);
 875                         xip->i_update_core = 1;
 876                         xip->i_update_size = 1;
 877                 }
 878                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 879         }
 880
 881         error = -ret;
 882         if (ret <= 0)
 883                 goto out_unlock_internal;
 884
 885         XFS_STATS_ADD(xs_write_bytes, ret);
 886
 887         /* Handle various SYNC-type writes */
 888         if ((file->f_flags & O_SYNC) || IS_SYNC(inode)) {
 889                 /*
 890                  * If we're treating this as O_DSYNC and we have not updated the
 891                  * size, force the log.
 892                  */
 893                 if (!(mp->m_flags & XFS_MOUNT_OSYNCISOSYNC) &&
 894                     !(xip->i_update_size)) {
 895                         xfs_inode_log_item_t    *iip = xip->i_itemp;
 896
 897                         /*
 898                          * If an allocation transaction occurred
 899                          * without extending the size, then we have to force
 900                          * the log up the proper point to ensure that the
 901                          * allocation is permanent.  We can't count on
 902                          * the fact that buffered writes lock out direct I/O
 903                          * writes - the direct I/O write could have extended
 904                          * the size nontransactionally, then finished before
 905                          * we started.  xfs_write_file will think that the file
 906                          * didn't grow but the update isn't safe unless the
 907                          * size change is logged.
 908                          *
 909                          * Force the log if we've committed a transaction
 910                          * against the inode or if someone else has and
 911                          * the commit record hasn't gone to disk (e.g.
 912                          * the inode is pinned).  This guarantees that
 913                          * all changes affecting the inode are permanent
 914                          * when we return.
 915                          */
 916                         if (iip && iip->ili_last_lsn) {
 917                                 xfs_log_force(mp, iip->ili_last_lsn,
 918                                                 XFS_LOG_FORCE | XFS_LOG_SYNC);
 919                         } else if (xfs_ipincount(xip) > 0) {
 920                                 xfs_log_force(mp, (xfs_lsn_t)0,
 921                                                 XFS_LOG_FORCE | XFS_LOG_SYNC);
 922                         }
 923
 924                 } else {
 925                         xfs_trans_t     *tp;
 926
 927                         /*
 928                          * O_SYNC or O_DSYNC _with_ a size update are handled
 929                          * the same way.
 930                          *
 931                          * If the write was synchronous then we need to make
 932                          * sure that the inode modification time is permanent.
 933                          * We'll have updated the timestamp above, so here
 934                          * we use a synchronous transaction to log the inode.
 935                          * It's not fast, but it's necessary.
 936                          *
 937                          * If this a dsync write and the size got changed
 938                          * non-transactionally, then we need to ensure that
 939                          * the size change gets logged in a synchronous
 940                          * transaction.
 941                          */
 942
 943                         tp = xfs_trans_alloc(mp, XFS_TRANS_WRITE_SYNC);
 944                         if ((error = xfs_trans_reserve(tp, 0,
 945                                                       XFS_SWRITE_LOG_RES(mp),
 946                                                       0, 0, 0))) {
 947                                 /* Transaction reserve failed */
 948                                 xfs_trans_cancel(tp, 0);
 949                         } else {
 950                                 /* Transaction reserve successful */
 951                                 xfs_ilock(xip, XFS_ILOCK_EXCL);
 952                                 xfs_trans_ijoin(tp, xip, XFS_ILOCK_EXCL);
 953                                 xfs_trans_ihold(tp, xip);
 954                                 xfs_trans_log_inode(tp, xip, XFS_ILOG_CORE);
 955                                 xfs_trans_set_sync(tp);
 956                                 error = xfs_trans_commit(tp, 0, NULL);
 957                                 xfs_iunlock(xip, XFS_ILOCK_EXCL);
 958                                 if (error)
 959                                         goto out_unlock_internal;
 960                         }
 961                 }
 962
 963                 xfs_rwunlock(bdp, locktype);
 964                 if (need_isem)
 965                         up(&inode->i_sem);
 966
 967                 error = sync_page_range(inode, mapping, pos, ret);
 968                 if (!error)
 969                         error = ret;
 970                 return error;
 971         }
 972
 973  out_unlock_internal:
 974         xfs_rwunlock(bdp, locktype);
 975  out_unlock_isem:
 976         if (need_isem)
 977                 up(&inode->i_sem);
 978         return -error;
 979 }
 980
 981 /*
 982  * All xfs metadata buffers except log state machine buffers
 983  * get this attached as their b_bdstrat callback function.
 984  * This is so that we can catch a buffer
 985  * after prematurely unpinning it to forcibly shutdown the filesystem.
 986  */
 987 int
 988 xfs_bdstrat_cb(struct xfs_buf *bp)
 989 {
 990         xfs_mount_t     *mp;
 991
 992         mp = XFS_BUF_FSPRIVATE3(bp, xfs_mount_t *);
 993         if (!XFS_FORCED_SHUTDOWN(mp)) {
 994                 pagebuf_iorequest(bp);
 995                 return 0;
 996         } else {
 997                 xfs_buftrace("XFS__BDSTRAT IOERROR", bp);
 998                 /*
 999                  * Metadata write that didn't get logged but
1000                  * written delayed anyway. These aren't associated
1001                  * with a transaction, and can be ignored.
1002                  */
1003                 if (XFS_BUF_IODONE_FUNC(bp) == NULL &&
1004                     (XFS_BUF_ISREAD(bp)) == 0)
1005                         return (xfs_bioerror_relse(bp));
1006                 else
1007                         return (xfs_bioerror(bp));
1008         }
1009 }
1010
1011
1012 int
1013 xfs_bmap(bhv_desc_t     *bdp,
1014         xfs_off_t       offset,
1015         ssize_t         count,
1016         int             flags,
1017         xfs_iomap_t     *iomapp,
1018         int             *niomaps)
1019 {
1020         xfs_inode_t     *ip = XFS_BHVTOI(bdp);
1021         xfs_iocore_t    *io = &ip->i_iocore;
1022
1023         ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFREG);
1024         ASSERT(((ip->i_d.di_flags & XFS_DIFLAG_REALTIME) != 0) ==
1025                ((ip->i_iocore.io_flags & XFS_IOCORE_RT) != 0));
1026
1027         return xfs_iomap(io, offset, count, flags, iomapp, niomaps);
1028 }
1029
1030 /*
1031  * Wrapper around bdstrat so that we can stop data
1032  * from going to disk in case we are shutting down the filesystem.
1033  * Typically user data goes thru this path; one of the exceptions
1034  * is the superblock.
1035  */
1036 int
1037 xfsbdstrat(
1038         struct xfs_mount        *mp,
1039         struct xfs_buf          *bp)
1040 {
1041         ASSERT(mp);
1042         if (!XFS_FORCED_SHUTDOWN(mp)) {
1043                 /* Grio redirection would go here
1044                  * if (XFS_BUF_IS_GRIO(bp)) {
1045                  */
1046
1047                 pagebuf_iorequest(bp);
1048                 return 0;
1049         }
1050
1051         xfs_buftrace("XFSBDSTRAT IOERROR", bp);
1052         return (xfs_bioerror_relse(bp));
1053 }
1054
1055 /*
1056  * If the underlying (data/log/rt) device is readonly, there are some
1057  * operations that cannot proceed.
1058  */
1059 int
1060 xfs_dev_is_read_only(
1061         xfs_mount_t             *mp,
1062         char                    *message)
1063 {
1064         if (xfs_readonly_buftarg(mp->m_ddev_targp) ||
1065             xfs_readonly_buftarg(mp->m_logdev_targp) ||
1066             (mp->m_rtdev_targp && xfs_readonly_buftarg(mp->m_rtdev_targp))) {
1067                 cmn_err(CE_NOTE,
1068                         "XFS: %s required on read-only device.", message);
1069                 cmn_err(CE_NOTE,
1070                         "XFS: write access unavailable, cannot proceed.");
1071                 return EROFS;
1072         }
1073         return 0;
1074 }