fs/xfs/linux-2.6/xfs_aops.c

   1 /*
   2  * Copyright (c) 2000-2004 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 #include "xfs.h"
  34 #include "xfs_inum.h"
  35 #include "xfs_log.h"
  36 #include "xfs_sb.h"
  37 #include "xfs_dir.h"
  38 #include "xfs_dir2.h"
  39 #include "xfs_trans.h"
  40 #include "xfs_dmapi.h"
  41 #include "xfs_mount.h"
  42 #include "xfs_bmap_btree.h"
  43 #include "xfs_alloc_btree.h"
  44 #include "xfs_ialloc_btree.h"
  45 #include "xfs_alloc.h"
  46 #include "xfs_btree.h"
  47 #include "xfs_attr_sf.h"
  48 #include "xfs_dir_sf.h"
  49 #include "xfs_dir2_sf.h"
  50 #include "xfs_dinode.h"
  51 #include "xfs_inode.h"
  52 #include "xfs_error.h"
  53 #include "xfs_rw.h"
  54 #include "xfs_iomap.h"
  55 #include <linux/mpage.h>
  56 #include <linux/writeback.h>
  57
  58 STATIC void xfs_count_page_state(struct page *, int *, int *, int *);
  59 STATIC void xfs_convert_page(struct inode *, struct page *, xfs_iomap_t *,
  60                 struct writeback_control *wbc, void *, int, int);
  61
  62 #if defined(XFS_RW_TRACE)
  63 void
  64 xfs_page_trace(
  65         int             tag,
  66         struct inode    *inode,
  67         struct page     *page,
  68         int             mask)
  69 {
  70         xfs_inode_t     *ip;
  71         bhv_desc_t      *bdp;
  72         vnode_t         *vp = LINVFS_GET_VP(inode);
  73         loff_t          isize = i_size_read(inode);
  74         loff_t          offset = page->index << PAGE_CACHE_SHIFT;
  75         int             delalloc = -1, unmapped = -1, unwritten = -1;
  76
  77         if (page_has_buffers(page))
  78                 xfs_count_page_state(page, &delalloc, &unmapped, &unwritten);
  79
  80         bdp = vn_bhv_lookup(VN_BHV_HEAD(vp), &xfs_vnodeops);
  81         ip = XFS_BHVTOI(bdp);
  82         if (!ip->i_rwtrace)
  83                 return;
  84
  85         ktrace_enter(ip->i_rwtrace,
  86                 (void *)((unsigned long)tag),
  87                 (void *)ip,
  88                 (void *)inode,
  89                 (void *)page,
  90                 (void *)((unsigned long)mask),
  91                 (void *)((unsigned long)((ip->i_d.di_size >> 32) & 0xffffffff)),
  92                 (void *)((unsigned long)(ip->i_d.di_size & 0xffffffff)),
  93                 (void *)((unsigned long)((isize >> 32) & 0xffffffff)),
  94                 (void *)((unsigned long)(isize & 0xffffffff)),
  95                 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
  96                 (void *)((unsigned long)(offset & 0xffffffff)),
  97                 (void *)((unsigned long)delalloc),
  98                 (void *)((unsigned long)unmapped),
  99                 (void *)((unsigned long)unwritten),
 100                 (void *)NULL,
 101                 (void *)NULL);
 102 }
 103 #else
 104 #define xfs_page_trace(tag, inode, page, mask)
 105 #endif
 106
 107 void
 108 linvfs_unwritten_done(
 109         struct buffer_head      *bh,
 110         int                     uptodate)
 111 {
 112         xfs_buf_t               *pb = (xfs_buf_t *)bh->b_private;
 113
 114         ASSERT(buffer_unwritten(bh));
 115         bh->b_end_io = NULL;
 116         clear_buffer_unwritten(bh);
 117         if (!uptodate)
 118                 pagebuf_ioerror(pb, EIO);
 119         if (atomic_dec_and_test(&pb->pb_io_remaining) == 1) {
 120                 pagebuf_iodone(pb, 1, 1);
 121         }
 122         end_buffer_async_write(bh, uptodate);
 123 }
 124
 125 /*
 126  * Issue transactions to convert a buffer range from unwritten
 127  * to written extents (buffered IO).
 128  */
 129 STATIC void
 130 linvfs_unwritten_convert(
 131         xfs_buf_t       *bp)
 132 {
 133         vnode_t         *vp = XFS_BUF_FSPRIVATE(bp, vnode_t *);
 134         int             error;
 135
 136         BUG_ON(atomic_read(&bp->pb_hold) < 1);
 137         VOP_BMAP(vp, XFS_BUF_OFFSET(bp), XFS_BUF_SIZE(bp),
 138                         BMAPI_UNWRITTEN, NULL, NULL, error);
 139         XFS_BUF_SET_FSPRIVATE(bp, NULL);
 140         XFS_BUF_CLR_IODONE_FUNC(bp);
 141         XFS_BUF_UNDATAIO(bp);
 142         iput(LINVFS_GET_IP(vp));
 143         pagebuf_iodone(bp, 0, 0);
 144 }
 145
 146 /*
 147  * Issue transactions to convert a buffer range from unwritten
 148  * to written extents (direct IO).
 149  */
 150 STATIC void
 151 linvfs_unwritten_convert_direct(
 152         struct inode    *inode,
 153         loff_t          offset,
 154         ssize_t         size,
 155         void            *private)
 156 {
 157         ASSERT(!private || inode == (struct inode *)private);
 158
 159         /* private indicates an unwritten extent lay beneath this IO,
 160          * see linvfs_get_block_core.
 161          */
 162         if (private && size > 0) {
 163                 vnode_t *vp = LINVFS_GET_VP(inode);
 164                 int     error;
 165
 166                 VOP_BMAP(vp, offset, size, BMAPI_UNWRITTEN, NULL, NULL, error);
 167         }
 168 }
 169
 170 STATIC int
 171 xfs_map_blocks(
 172         struct inode            *inode,
 173         loff_t                  offset,
 174         ssize_t                 count,
 175         xfs_iomap_t             *iomapp,
 176         int                     flags)
 177 {
 178         vnode_t                 *vp = LINVFS_GET_VP(inode);
 179         int                     error, niomaps = 1;
 180
 181         if (((flags & (BMAPI_DIRECT|BMAPI_SYNC)) == BMAPI_DIRECT) &&
 182             (offset >= i_size_read(inode)))
 183                 count = max_t(ssize_t, count, XFS_WRITE_IO_LOG);
 184 retry:
 185         VOP_BMAP(vp, offset, count, flags, iomapp, &niomaps, error);
 186         if ((error == EAGAIN) || (error == EIO))
 187                 return -error;
 188         if (unlikely((flags & (BMAPI_WRITE|BMAPI_DIRECT)) ==
 189                                         (BMAPI_WRITE|BMAPI_DIRECT) && niomaps &&
 190                                         (iomapp->iomap_flags & IOMAP_DELAY))) {
 191                 flags = BMAPI_ALLOCATE;
 192                 goto retry;
 193         }
 194         if (flags & (BMAPI_WRITE|BMAPI_ALLOCATE)) {
 195                 VMODIFY(vp);
 196         }
 197         return -error;
 198 }
 199
 200 /*
 201  * Finds the corresponding mapping in block @map array of the
 202  * given @offset within a @page.
 203  */
 204 STATIC xfs_iomap_t *
 205 xfs_offset_to_map(
 206         struct page             *page,
 207         xfs_iomap_t             *iomapp,
 208         unsigned long           offset)
 209 {
 210         loff_t                  full_offset;    /* offset from start of file */
 211
 212         ASSERT(offset < PAGE_CACHE_SIZE);
 213
 214         full_offset = page->index;              /* NB: using 64bit number */
 215         full_offset <<= PAGE_CACHE_SHIFT;       /* offset from file start */
 216         full_offset += offset;                  /* offset from page start */
 217
 218         if (full_offset < iomapp->iomap_offset)
 219                 return NULL;
 220         if (iomapp->iomap_offset + (iomapp->iomap_bsize -1) >= full_offset)
 221                 return iomapp;
 222         return NULL;
 223 }
 224
 225 STATIC void
 226 xfs_map_at_offset(
 227         struct page             *page,
 228         struct buffer_head      *bh,
 229         unsigned long           offset,
 230         int                     block_bits,
 231         xfs_iomap_t             *iomapp)
 232 {
 233         xfs_daddr_t             bn;
 234         loff_t                  delta;
 235         int                     sector_shift;
 236
 237         ASSERT(!(iomapp->iomap_flags & IOMAP_HOLE));
 238         ASSERT(!(iomapp->iomap_flags & IOMAP_DELAY));
 239         ASSERT(iomapp->iomap_bn != IOMAP_DADDR_NULL);
 240
 241         delta = page->index;
 242         delta <<= PAGE_CACHE_SHIFT;
 243         delta += offset;
 244         delta -= iomapp->iomap_offset;
 245         delta >>= block_bits;
 246
 247         sector_shift = block_bits - BBSHIFT;
 248         bn = iomapp->iomap_bn >> sector_shift;
 249         bn += delta;
 250         ASSERT((bn << sector_shift) >= iomapp->iomap_bn);
 251
 252         lock_buffer(bh);
 253         bh->b_blocknr = bn;
 254         bh->b_bdev = iomapp->iomap_target->pbr_bdev;
 255         set_buffer_mapped(bh);
 256         clear_buffer_delay(bh);
 257 }
 258
 259 /*
 260  * Look for a page at index which is unlocked and contains our
 261  * unwritten extent flagged buffers at its head.  Returns page
 262  * locked and with an extra reference count, and length of the
 263  * unwritten extent component on this page that we can write,
 264  * in units of filesystem blocks.
 265  */
 266 STATIC struct page *
 267 xfs_probe_unwritten_page(
 268         struct address_space    *mapping,
 269         pgoff_t                 index,
 270         xfs_iomap_t             *iomapp,
 271         xfs_buf_t               *pb,
 272         unsigned long           max_offset,
 273         unsigned long           *fsbs,
 274         unsigned int            bbits)
 275 {
 276         struct page             *page;
 277
 278         page = find_trylock_page(mapping, index);
 279         if (!page)
 280                 return 0;
 281         if (PageWriteback(page))
 282                 goto out;
 283
 284         if (page->mapping && page_has_buffers(page)) {
 285                 struct buffer_head      *bh, *head;
 286                 unsigned long           p_offset = 0;
 287
 288                 *fsbs = 0;
 289                 bh = head = page_buffers(page);
 290                 do {
 291                         if (!buffer_unwritten(bh))
 292                                 break;
 293                         if (!xfs_offset_to_map(page, iomapp, p_offset))
 294                                 break;
 295                         if (p_offset >= max_offset)
 296                                 break;
 297                         xfs_map_at_offset(page, bh, p_offset, bbits, iomapp);
 298                         set_buffer_unwritten_io(bh);
 299                         bh->b_private = pb;
 300                         p_offset += bh->b_size;
 301                         (*fsbs)++;
 302                 } while ((bh = bh->b_this_page) != head);
 303
 304                 if (p_offset)
 305                         return page;
 306         }
 307
 308 out:
 309         unlock_page(page);
 310         return NULL;
 311 }
 312
 313 /*
 314  * Look for a page at index which is unlocked and not mapped
 315  * yet - clustering for mmap write case.
 316  */
 317 STATIC unsigned int
 318 xfs_probe_unmapped_page(
 319         struct address_space    *mapping,
 320         pgoff_t                 index,
 321         unsigned int            pg_offset)
 322 {
 323         struct page             *page;
 324         int                     ret = 0;
 325
 326         page = find_trylock_page(mapping, index);
 327         if (!page)
 328                 return 0;
 329         if (PageWriteback(page))
 330                 goto out;
 331
 332         if (page->mapping && PageDirty(page)) {
 333                 if (page_has_buffers(page)) {
 334                         struct buffer_head      *bh, *head;
 335
 336                         bh = head = page_buffers(page);
 337                         do {
 338                                 if (buffer_mapped(bh) || !buffer_uptodate(bh))
 339                                         break;
 340                                 ret += bh->b_size;
 341                                 if (ret >= pg_offset)
 342                                         break;
 343                         } while ((bh = bh->b_this_page) != head);
 344                 } else
 345                         ret = PAGE_CACHE_SIZE;
 346         }
 347
 348 out:
 349         unlock_page(page);
 350         return ret;
 351 }
 352
 353 STATIC unsigned int
 354 xfs_probe_unmapped_cluster(
 355         struct inode            *inode,
 356         struct page             *startpage,
 357         struct buffer_head      *bh,
 358         struct buffer_head      *head)
 359 {
 360         pgoff_t                 tindex, tlast, tloff;
 361         unsigned int            pg_offset, len, total = 0;
 362         struct address_space    *mapping = inode->i_mapping;
 363
 364         /* First sum forwards in this page */
 365         do {
 366                 if (buffer_mapped(bh))
 367                         break;
 368                 total += bh->b_size;
 369         } while ((bh = bh->b_this_page) != head);
 370
 371         /* If we reached the end of the page, sum forwards in
 372          * following pages.
 373          */
 374         if (bh == head) {
 375                 tlast = i_size_read(inode) >> PAGE_CACHE_SHIFT;
 376                 /* Prune this back to avoid pathological behavior */
 377                 tloff = min(tlast, startpage->index + 64);
 378                 for (tindex = startpage->index + 1; tindex < tloff; tindex++) {
 379                         len = xfs_probe_unmapped_page(mapping, tindex,
 380                                                         PAGE_CACHE_SIZE);
 381                         if (!len)
 382                                 return total;
 383                         total += len;
 384                 }
 385                 if (tindex == tlast &&
 386                     (pg_offset = i_size_read(inode) & (PAGE_CACHE_SIZE - 1))) {
 387                         total += xfs_probe_unmapped_page(mapping,
 388                                                         tindex, pg_offset);
 389                 }
 390         }
 391         return total;
 392 }
 393
 394 /*
 395  * Probe for a given page (index) in the inode and test if it is delayed
 396  * and without unwritten buffers.  Returns page locked and with an extra
 397  * reference count.
 398  */
 399 STATIC struct page *
 400 xfs_probe_delalloc_page(
 401         struct inode            *inode,
 402         pgoff_t                 index)
 403 {
 404         struct page             *page;
 405
 406         page = find_trylock_page(inode->i_mapping, index);
 407         if (!page)
 408                 return NULL;
 409         if (PageWriteback(page))
 410                 goto out;
 411
 412         if (page->mapping && page_has_buffers(page)) {
 413                 struct buffer_head      *bh, *head;
 414                 int                     acceptable = 0;
 415
 416                 bh = head = page_buffers(page);
 417                 do {
 418                         if (buffer_unwritten(bh)) {
 419                                 acceptable = 0;
 420                                 break;
 421                         } else if (buffer_delay(bh)) {
 422                                 acceptable = 1;
 423                         }
 424                 } while ((bh = bh->b_this_page) != head);
 425
 426                 if (acceptable)
 427                         return page;
 428         }
 429
 430 out:
 431         unlock_page(page);
 432         return NULL;
 433 }
 434
 435 STATIC int
 436 xfs_map_unwritten(
 437         struct inode            *inode,
 438         struct page             *start_page,
 439         struct buffer_head      *head,
 440         struct buffer_head      *curr,
 441         unsigned long           p_offset,
 442         int                     block_bits,
 443         xfs_iomap_t             *iomapp,
 444         struct writeback_control *wbc,
 445         int                     startio,
 446         int                     all_bh)
 447 {
 448         struct buffer_head      *bh = curr;
 449         xfs_iomap_t             *tmp;
 450         xfs_buf_t               *pb;
 451         loff_t                  offset, size;
 452         unsigned long           nblocks = 0;
 453
 454         offset = start_page->index;
 455         offset <<= PAGE_CACHE_SHIFT;
 456         offset += p_offset;
 457
 458         /* get an "empty" pagebuf to manage IO completion
 459          * Proper values will be set before returning */
 460         pb = pagebuf_lookup(iomapp->iomap_target, 0, 0, 0);
 461         if (!pb)
 462                 return -EAGAIN;
 463
 464         /* Take a reference to the inode to prevent it from
 465          * being reclaimed while we have outstanding unwritten
 466          * extent IO on it.
 467          */
 468         if ((igrab(inode)) != inode) {
 469                 pagebuf_free(pb);
 470                 return -EAGAIN;
 471         }
 472
 473         /* Set the count to 1 initially, this will stop an I/O
 474          * completion callout which happens before we have started
 475          * all the I/O from calling pagebuf_iodone too early.
 476          */
 477         atomic_set(&pb->pb_io_remaining, 1);
 478
 479         /* First map forwards in the page consecutive buffers
 480          * covering this unwritten extent
 481          */
 482         do {
 483                 if (!buffer_unwritten(bh))
 484                         break;
 485                 tmp = xfs_offset_to_map(start_page, iomapp, p_offset);
 486                 if (!tmp)
 487                         break;
 488                 xfs_map_at_offset(start_page, bh, p_offset, block_bits, iomapp);
 489                 set_buffer_unwritten_io(bh);
 490                 bh->b_private = pb;
 491                 p_offset += bh->b_size;
 492                 nblocks++;
 493         } while ((bh = bh->b_this_page) != head);
 494
 495         atomic_add(nblocks, &pb->pb_io_remaining);
 496
 497         /* If we reached the end of the page, map forwards in any
 498          * following pages which are also covered by this extent.
 499          */
 500         if (bh == head) {
 501                 struct address_space    *mapping = inode->i_mapping;
 502                 pgoff_t                 tindex, tloff, tlast;
 503                 unsigned long           bs;
 504                 unsigned int            pg_offset, bbits = inode->i_blkbits;
 505                 struct page             *page;
 506
 507                 tlast = i_size_read(inode) >> PAGE_CACHE_SHIFT;
 508                 tloff = (iomapp->iomap_offset + iomapp->iomap_bsize) >> PAGE_CACHE_SHIFT;
 509                 tloff = min(tlast, tloff);
 510                 for (tindex = start_page->index + 1; tindex < tloff; tindex++) {
 511                         page = xfs_probe_unwritten_page(mapping,
 512                                                 tindex, iomapp, pb,
 513                                                 PAGE_CACHE_SIZE, &bs, bbits);
 514                         if (!page)
 515                                 break;
 516                         nblocks += bs;
 517                         atomic_add(bs, &pb->pb_io_remaining);
 518                         xfs_convert_page(inode, page, iomapp, wbc, pb,
 519                                                         startio, all_bh);
 520                         /* stop if converting the next page might add
 521                          * enough blocks that the corresponding byte
 522                          * count won't fit in our ulong page buf length */
 523                         if (nblocks >= ((ULONG_MAX - PAGE_SIZE) >> block_bits))
 524                                 goto enough;
 525                 }
 526
 527                 if (tindex == tlast &&
 528                     (pg_offset = (i_size_read(inode) & (PAGE_CACHE_SIZE - 1)))) {
 529                         page = xfs_probe_unwritten_page(mapping,
 530                                                         tindex, iomapp, pb,
 531                                                         pg_offset, &bs, bbits);
 532                         if (page) {
 533                                 nblocks += bs;
 534                                 atomic_add(bs, &pb->pb_io_remaining);
 535                                 xfs_convert_page(inode, page, iomapp, wbc, pb,
 536                                                         startio, all_bh);
 537                                 if (nblocks >= ((ULONG_MAX - PAGE_SIZE) >> block_bits))
 538                                         goto enough;
 539                         }
 540                 }
 541         }
 542
 543 enough:
 544         size = nblocks;         /* NB: using 64bit number here */
 545         size <<= block_bits;    /* convert fsb's to byte range */
 546
 547         XFS_BUF_DATAIO(pb);
 548         XFS_BUF_ASYNC(pb);
 549         XFS_BUF_SET_SIZE(pb, size);
 550         XFS_BUF_SET_COUNT(pb, size);
 551         XFS_BUF_SET_OFFSET(pb, offset);
 552         XFS_BUF_SET_FSPRIVATE(pb, LINVFS_GET_VP(inode));
 553         XFS_BUF_SET_IODONE_FUNC(pb, linvfs_unwritten_convert);
 554
 555         if (atomic_dec_and_test(&pb->pb_io_remaining) == 1) {
 556                 pagebuf_iodone(pb, 1, 1);
 557         }
 558
 559         return 0;
 560 }
 561
 562 STATIC void
 563 xfs_submit_page(
 564         struct page             *page,
 565         struct buffer_head      *bh_arr[],
 566         int                     cnt)
 567 {
 568         struct buffer_head      *bh;
 569         int                     i;
 570
 571         BUG_ON(PageWriteback(page));
 572         set_page_writeback(page);
 573         clear_page_dirty(page);
 574         unlock_page(page);
 575
 576         if (cnt) {
 577                 for (i = 0; i < cnt; i++) {
 578                         bh = bh_arr[i];
 579                         mark_buffer_async_write(bh);
 580                         if (buffer_unwritten(bh))
 581                                 set_buffer_unwritten_io(bh);
 582                         set_buffer_uptodate(bh);
 583                         clear_buffer_dirty(bh);
 584                 }
 585
 586                 for (i = 0; i < cnt; i++)
 587                         submit_bh(WRITE, bh_arr[i]);
 588         } else
 589                 end_page_writeback(page);
 590 }
 591
 592 /*
 593  * Allocate & map buffers for page given the extent map. Write it out.
 594  * except for the original page of a writepage, this is called on
 595  * delalloc/unwritten pages only, for the original page it is possible
 596  * that the page has no mapping at all.
 597  */
 598 STATIC void
 599 xfs_convert_page(
 600         struct inode            *inode,
 601         struct page             *page,
 602         xfs_iomap_t             *iomapp,
 603         struct writeback_control *wbc,
 604         void                    *private,
 605         int                     startio,
 606         int                     all_bh)
 607 {
 608         struct buffer_head      *bh_arr[MAX_BUF_PER_PAGE], *bh, *head;
 609         xfs_iomap_t             *mp = iomapp, *tmp;
 610         unsigned long           end, offset;
 611         pgoff_t                 end_index;
 612         int                     i = 0, index = 0;
 613         int                     bbits = inode->i_blkbits;
 614
 615         end_index = i_size_read(inode) >> PAGE_CACHE_SHIFT;
 616         if (page->index < end_index) {
 617                 end = PAGE_CACHE_SIZE;
 618         } else {
 619                 end = i_size_read(inode) & (PAGE_CACHE_SIZE-1);
 620         }
 621         bh = head = page_buffers(page);
 622         do {
 623                 offset = i << bbits;
 624                 if (!(PageUptodate(page) || buffer_uptodate(bh)))
 625                         continue;
 626                 if (buffer_mapped(bh) && all_bh &&
 627                     !buffer_unwritten(bh) && !buffer_delay(bh)) {
 628                         if (startio && (offset < end)) {
 629                                 lock_buffer(bh);
 630                                 bh_arr[index++] = bh;
 631                         }
 632                         continue;
 633                 }
 634                 tmp = xfs_offset_to_map(page, mp, offset);
 635                 if (!tmp)
 636                         continue;
 637                 ASSERT(!(tmp->iomap_flags & IOMAP_HOLE));
 638                 ASSERT(!(tmp->iomap_flags & IOMAP_DELAY));
 639
 640                 /* If this is a new unwritten extent buffer (i.e. one
 641                  * that we haven't passed in private data for, we must
 642                  * now map this buffer too.
 643                  */
 644                 if (buffer_unwritten(bh) && !bh->b_end_io) {
 645                         ASSERT(tmp->iomap_flags & IOMAP_UNWRITTEN);
 646                         xfs_map_unwritten(inode, page, head, bh, offset,
 647                                         bbits, tmp, wbc, startio, all_bh);
 648                 } else if (! (buffer_unwritten(bh) && buffer_locked(bh))) {
 649                         xfs_map_at_offset(page, bh, offset, bbits, tmp);
 650                         if (buffer_unwritten(bh)) {
 651                                 set_buffer_unwritten_io(bh);
 652                                 bh->b_private = private;
 653                                 ASSERT(private);
 654                         }
 655                 }
 656                 if (startio && (offset < end)) {
 657                         bh_arr[index++] = bh;
 658                 } else {
 659                         set_buffer_dirty(bh);
 660                         unlock_buffer(bh);
 661                         mark_buffer_dirty(bh);
 662                 }
 663         } while (i++, (bh = bh->b_this_page) != head);
 664
 665         if (startio) {
 666                 wbc->nr_to_write--;
 667                 xfs_submit_page(page, bh_arr, index);
 668         } else {
 669                 unlock_page(page);
 670         }
 671 }
 672
 673 /*
 674  * Convert & write out a cluster of pages in the same extent as defined
 675  * by mp and following the start page.
 676  */
 677 STATIC void
 678 xfs_cluster_write(
 679         struct inode            *inode,
 680         pgoff_t                 tindex,
 681         xfs_iomap_t             *iomapp,
 682         struct writeback_control *wbc,
 683         int                     startio,
 684         int                     all_bh)
 685 {
 686         pgoff_t                 tlast;
 687         struct page             *page;
 688
 689         tlast = (iomapp->iomap_offset + iomapp->iomap_bsize) >> PAGE_CACHE_SHIFT;
 690         for (; tindex < tlast; tindex++) {
 691                 page = xfs_probe_delalloc_page(inode, tindex);
 692                 if (!page)
 693                         break;
 694                 xfs_convert_page(inode, page, iomapp, wbc, NULL,
 695                                 startio, all_bh);
 696         }
 697 }
 698
 699 /*
 700  * Calling this without startio set means we are being asked to make a dirty
 701  * page ready for freeing it's buffers.  When called with startio set then
 702  * we are coming from writepage.
 703  *
 704  * When called with startio set it is important that we write the WHOLE
 705  * page if possible.
 706  * The bh->b_state's cannot know if any of the blocks or which block for
 707  * that matter are dirty due to mmap writes, and therefore bh uptodate is
 708  * only vaild if the page itself isn't completely uptodate.  Some layers
 709  * may clear the page dirty flag prior to calling write page, under the
 710  * assumption the entire page will be written out; by not writing out the
 711  * whole page the page can be reused before all valid dirty data is
 712  * written out.  Note: in the case of a page that has been dirty'd by
 713  * mapwrite and but partially setup by block_prepare_write the
 714  * bh->b_states's will not agree and only ones setup by BPW/BCW will have
 715  * valid state, thus the whole page must be written out thing.
 716  */
 717
 718 STATIC int
 719 xfs_page_state_convert(
 720         struct inode    *inode,
 721         struct page     *page,
 722         struct writeback_control *wbc,
 723         int             startio,
 724         int             unmapped) /* also implies page uptodate */
 725 {
 726         struct buffer_head      *bh_arr[MAX_BUF_PER_PAGE], *bh, *head;
 727         xfs_iomap_t             *iomp, iomap;
 728         unsigned long           p_offset = 0;
 729         pgoff_t                 end_index;
 730         loff_t                  offset;
 731         unsigned long long      end_offset;
 732         int                     len, err, i, cnt = 0, uptodate = 1;
 733         int                     flags = startio ? 0 : BMAPI_TRYLOCK;
 734         int                     page_dirty = 1;
 735
 736
 737         /* Are we off the end of the file ? */
 738         end_index = i_size_read(inode) >> PAGE_CACHE_SHIFT;
 739         if (page->index >= end_index) {
 740                 if ((page->index >= end_index + 1) ||
 741                     !(i_size_read(inode) & (PAGE_CACHE_SIZE - 1))) {
 742                         err = -EIO;
 743                         goto error;
 744                 }
 745         }
 746
 747         offset = (loff_t)page->index << PAGE_CACHE_SHIFT;
 748         end_offset = min_t(unsigned long long,
 749                         offset + PAGE_CACHE_SIZE, i_size_read(inode));
 750
 751         bh = head = page_buffers(page);
 752         iomp = NULL;
 753
 754         len = bh->b_size;
 755         do {
 756                 if (offset >= end_offset)
 757                         break;
 758                 if (!buffer_uptodate(bh))
 759                         uptodate = 0;
 760                 if (!(PageUptodate(page) || buffer_uptodate(bh)) && !startio)
 761                         continue;
 762
 763                 if (iomp) {
 764                         iomp = xfs_offset_to_map(page, &iomap, p_offset);
 765                 }
 766
 767                 /*
 768                  * First case, map an unwritten extent and prepare for
 769                  * extent state conversion transaction on completion.
 770                  */
 771                 if (buffer_unwritten(bh)) {
 772                         if (!iomp) {
 773                                 err = xfs_map_blocks(inode, offset, len, &iomap,
 774                                                 BMAPI_READ|BMAPI_IGNSTATE);
 775                                 if (err) {
 776                                         goto error;
 777                                 }
 778                                 iomp = xfs_offset_to_map(page, &iomap,
 779                                                                 p_offset);
 780                         }
 781                         if (iomp && startio) {
 782                                 if (!bh->b_end_io) {
 783                                         err = xfs_map_unwritten(inode, page,
 784                                                         head, bh, p_offset,
 785                                                         inode->i_blkbits, iomp,
 786                                                         wbc, startio, unmapped);
 787                                         if (err) {
 788                                                 goto error;
 789                                         }
 790                                 }
 791                                 bh_arr[cnt++] = bh;
 792                                 page_dirty = 0;
 793                         }
 794                 /*
 795                  * Second case, allocate space for a delalloc buffer.
 796                  * We can return EAGAIN here in the release page case.
 797                  */
 798                 } else if (buffer_delay(bh)) {
 799                         if (!iomp) {
 800                                 err = xfs_map_blocks(inode, offset, len, &iomap,
 801                                                 BMAPI_ALLOCATE | flags);
 802                                 if (err) {
 803                                         goto error;
 804                                 }
 805                                 iomp = xfs_offset_to_map(page, &iomap,
 806                                                                 p_offset);
 807                         }
 808                         if (iomp) {
 809                                 xfs_map_at_offset(page, bh, p_offset,
 810                                                 inode->i_blkbits, iomp);
 811                                 if (startio) {
 812                                         bh_arr[cnt++] = bh;
 813                                 } else {
 814                                         set_buffer_dirty(bh);
 815                                         unlock_buffer(bh);
 816                                         mark_buffer_dirty(bh);
 817                                 }
 818                                 page_dirty = 0;
 819                         }
 820                 } else if ((buffer_uptodate(bh) || PageUptodate(page)) &&
 821                            (unmapped || startio)) {
 822
 823                         if (!buffer_mapped(bh)) {
 824                                 int     size;
 825
 826                                 /*
 827                                  * Getting here implies an unmapped buffer
 828                                  * was found, and we are in a path where we
 829                                  * need to write the whole page out.
 830                                  */
 831                                 if (!iomp) {
 832                                         size = xfs_probe_unmapped_cluster(
 833                                                         inode, page, bh, head);
 834                                         err = xfs_map_blocks(inode, offset,
 835                                                         size, &iomap,
 836                                                         BMAPI_WRITE|BMAPI_MMAP);
 837                                         if (err) {
 838                                                 goto error;
 839                                         }
 840                                         iomp = xfs_offset_to_map(page, &iomap,
 841                                                                      p_offset);
 842                                 }
 843                                 if (iomp) {
 844                                         xfs_map_at_offset(page,
 845                                                         bh, p_offset,
 846                                                         inode->i_blkbits, iomp);
 847                                         if (startio) {
 848                                                 bh_arr[cnt++] = bh;
 849                                         } else {
 850                                                 set_buffer_dirty(bh);
 851                                                 unlock_buffer(bh);
 852                                                 mark_buffer_dirty(bh);
 853                                         }
 854                                         page_dirty = 0;
 855                                 }
 856                         } else if (startio) {
 857                                 if (buffer_uptodate(bh) &&
 858                                     !test_and_set_bit(BH_Lock, &bh->b_state)) {
 859                                         bh_arr[cnt++] = bh;
 860                                         page_dirty = 0;
 861                                 }
 862                         }
 863                 }
 864         } while (offset += len, p_offset += len,
 865                 ((bh = bh->b_this_page) != head));
 866
 867         if (uptodate && bh == head)
 868                 SetPageUptodate(page);
 869
 870         if (startio)
 871                 xfs_submit_page(page, bh_arr, cnt);
 872
 873         if (iomp) {
 874                 xfs_cluster_write(inode, page->index + 1, iomp, wbc,
 875                                 startio, unmapped);
 876         }
 877
 878         return page_dirty;
 879
 880 error:
 881         for (i = 0; i < cnt; i++) {
 882                 unlock_buffer(bh_arr[i]);
 883         }
 884
 885         /*
 886          * If it's delalloc and we have nowhere to put it,
 887          * throw it away, unless the lower layers told
 888          * us to try again.
 889          */
 890         if (err != -EAGAIN) {
 891                 if (!unmapped) {
 892                         block_invalidatepage(page, 0);
 893                 }
 894                 ClearPageUptodate(page);
 895         }
 896         return err;
 897 }
 898
 899 STATIC int
 900 linvfs_get_block_core(
 901         struct inode            *inode,
 902         sector_t                iblock,
 903         unsigned long           blocks,
 904         struct buffer_head      *bh_result,
 905         int                     create,
 906         int                     direct,
 907         bmapi_flags_t           flags)
 908 {
 909         vnode_t                 *vp = LINVFS_GET_VP(inode);
 910         xfs_iomap_t             iomap;
 911         int                     retpbbm = 1;
 912         int                     error;
 913         ssize_t                 size;
 914         loff_t                  offset = (loff_t)iblock << inode->i_blkbits;
 915
 916         if (blocks)
 917                 size = blocks << inode->i_blkbits;
 918         else
 919                 size = 1 << inode->i_blkbits;
 920
 921         VOP_BMAP(vp, offset, size,
 922                 create ? flags : BMAPI_READ, &iomap, &retpbbm, error);
 923         if (error)
 924                 return -error;
 925
 926         if (retpbbm == 0)
 927                 return 0;
 928
 929         if (iomap.iomap_bn != IOMAP_DADDR_NULL) {
 930                 xfs_daddr_t             bn;
 931                 loff_t                  delta;
 932
 933                 /* For unwritten extents do not report a disk address on
 934                  * the read case (treat as if we're reading into a hole).
 935                  */
 936                 if (create || !(iomap.iomap_flags & IOMAP_UNWRITTEN)) {
 937                         delta = offset - iomap.iomap_offset;
 938                         delta >>= inode->i_blkbits;
 939
 940                         bn = iomap.iomap_bn >> (inode->i_blkbits - BBSHIFT);
 941                         bn += delta;
 942
 943                         bh_result->b_blocknr = bn;
 944                         bh_result->b_bdev = iomap.iomap_target->pbr_bdev;
 945                         set_buffer_mapped(bh_result);
 946                 }
 947                 if (create && (iomap.iomap_flags & IOMAP_UNWRITTEN)) {
 948                         if (direct)
 949                                 bh_result->b_private = inode;
 950                         set_buffer_unwritten(bh_result);
 951                         set_buffer_delay(bh_result);
 952                 }
 953         }
 954
 955         /* If this is a realtime file, data might be on a new device */
 956         bh_result->b_bdev = iomap.iomap_target->pbr_bdev;
 957
 958         /* If we previously allocated a block out beyond eof and
 959          * we are now coming back to use it then we will need to
 960          * flag it as new even if it has a disk address.
 961          */
 962         if (create &&
 963             ((!buffer_mapped(bh_result) && !buffer_uptodate(bh_result)) ||
 964              (offset >= i_size_read(inode)) || (iomap.iomap_flags & IOMAP_NEW))) {
 965                 set_buffer_new(bh_result);
 966         }
 967
 968         if (iomap.iomap_flags & IOMAP_DELAY) {
 969                 if (unlikely(direct))
 970                         BUG();
 971                 if (create) {
 972                         set_buffer_mapped(bh_result);
 973                         set_buffer_uptodate(bh_result);
 974                 }
 975                 bh_result->b_bdev = iomap.iomap_target->pbr_bdev;
 976                 set_buffer_delay(bh_result);
 977         }
 978
 979         if (blocks) {
 980                 loff_t iosize;
 981                 iosize = (iomap.iomap_bsize - iomap.iomap_delta);
 982                 bh_result->b_size =
 983                     (ssize_t)min(iosize, (loff_t)(blocks << inode->i_blkbits));
 984         }
 985
 986         return 0;
 987 }
 988
 989 int
 990 linvfs_get_block(
 991         struct inode            *inode,
 992         sector_t                iblock,
 993         struct buffer_head      *bh_result,
 994         int                     create)
 995 {
 996         return linvfs_get_block_core(inode, iblock, 0, bh_result,
 997                                         create, 0, BMAPI_WRITE);
 998 }
 999
1000 STATIC int
1001 linvfs_get_block_sync(
1002         struct inode            *inode,
1003         sector_t                iblock,
1004         struct buffer_head      *bh_result,
1005         int                     create)
1006 {
1007         return linvfs_get_block_core(inode, iblock, 0, bh_result,
1008                                         create, 0, BMAPI_SYNC|BMAPI_WRITE);
1009 }
1010
1011 STATIC int
1012 linvfs_get_blocks_direct(
1013         struct inode            *inode,
1014         sector_t                iblock,
1015         unsigned long           max_blocks,
1016         struct buffer_head      *bh_result,
1017         int                     create)
1018 {
1019         return linvfs_get_block_core(inode, iblock, max_blocks, bh_result,
1020                                         create, 1, BMAPI_WRITE|BMAPI_DIRECT);
1021 }
1022
1023 STATIC ssize_t
1024 linvfs_direct_IO(
1025         int                     rw,
1026         struct kiocb            *iocb,
1027         const struct iovec      *iov,
1028         loff_t                  offset,
1029         unsigned long           nr_segs)
1030 {
1031         struct file     *file = iocb->ki_filp;
1032         struct inode    *inode = file->f_mapping->host;
1033         vnode_t         *vp = LINVFS_GET_VP(inode);
1034         xfs_iomap_t     iomap;
1035         int             maps = 1;
1036         int             error;
1037
1038         VOP_BMAP(vp, offset, 0, BMAPI_DEVICE, &iomap, &maps, error);
1039         if (error)
1040                 return -error;
1041
1042         return blockdev_direct_IO_no_locking(rw, iocb, inode,
1043                 iomap.iomap_target->pbr_bdev,
1044                 iov, offset, nr_segs,
1045                 linvfs_get_blocks_direct,
1046                 linvfs_unwritten_convert_direct);
1047 }
1048
1049
1050 STATIC sector_t
1051 linvfs_bmap(
1052         struct address_space    *mapping,
1053         sector_t                block)
1054 {
1055         struct inode            *inode = (struct inode *)mapping->host;
1056         vnode_t                 *vp = LINVFS_GET_VP(inode);
1057         int                     error;
1058
1059         vn_trace_entry(vp, "linvfs_bmap", (inst_t *)__return_address);
1060
1061         VOP_RWLOCK(vp, VRWLOCK_READ);
1062         VOP_FLUSH_PAGES(vp, (xfs_off_t)0, -1, 0, FI_REMAPF, error);
1063         VOP_RWUNLOCK(vp, VRWLOCK_READ);
1064         return generic_block_bmap(mapping, block, linvfs_get_block);
1065 }
1066
1067 STATIC int
1068 linvfs_readpage(
1069         struct file             *unused,
1070         struct page             *page)
1071 {
1072         return mpage_readpage(page, linvfs_get_block);
1073 }
1074
1075 STATIC int
1076 linvfs_readpages(
1077         struct file             *unused,
1078         struct address_space    *mapping,
1079         struct list_head        *pages,
1080         unsigned                nr_pages)
1081 {
1082         return mpage_readpages(mapping, pages, nr_pages, linvfs_get_block);
1083 }
1084
1085 STATIC void
1086 xfs_count_page_state(
1087         struct page             *page,
1088         int                     *delalloc,
1089         int                     *unmapped,
1090         int                     *unwritten)
1091 {
1092         struct buffer_head      *bh, *head;
1093
1094         *delalloc = *unmapped = *unwritten = 0;
1095
1096         bh = head = page_buffers(page);
1097         do {
1098                 if (buffer_uptodate(bh) && !buffer_mapped(bh))
1099                         (*unmapped) = 1;
1100                 else if (buffer_unwritten(bh) && !buffer_delay(bh))
1101                         clear_buffer_unwritten(bh);
1102                 else if (buffer_unwritten(bh))
1103                         (*unwritten) = 1;
1104                 else if (buffer_delay(bh))
1105                         (*delalloc) = 1;
1106         } while ((bh = bh->b_this_page) != head);
1107 }
1108
1109
1110 /*
1111  * writepage: Called from one of two places:
1112  *
1113  * 1. we are flushing a delalloc buffer head.
1114  *
1115  * 2. we are writing out a dirty page. Typically the page dirty
1116  *    state is cleared before we get here. In this case is it
1117  *    conceivable we have no buffer heads.
1118  *
1119  * For delalloc space on the page we need to allocate space and
1120  * flush it. For unmapped buffer heads on the page we should
1121  * allocate space if the page is uptodate. For any other dirty
1122  * buffer heads on the page we should flush them.
1123  *
1124  * If we detect that a transaction would be required to flush
1125  * the page, we have to check the process flags first, if we
1126  * are already in a transaction or disk I/O during allocations
1127  * is off, we need to fail the writepage and redirty the page.
1128  */
1129
1130 STATIC int
1131 linvfs_writepage(
1132         struct page             *page,
1133         struct writeback_control *wbc)
1134 {
1135         int                     error;
1136         int                     need_trans;
1137         int                     delalloc, unmapped, unwritten;
1138         struct inode            *inode = page->mapping->host;
1139
1140         xfs_page_trace(XFS_WRITEPAGE_ENTER, inode, page, 0);
1141
1142         /*
1143          * We need a transaction if:
1144          *  1. There are delalloc buffers on the page
1145          *  2. The page is uptodate and we have unmapped buffers
1146          *  3. The page is uptodate and we have no buffers
1147          *  4. There are unwritten buffers on the page
1148          */
1149
1150         if (!page_has_buffers(page)) {
1151                 unmapped = 1;
1152                 need_trans = 1;
1153         } else {
1154                 xfs_count_page_state(page, &delalloc, &unmapped, &unwritten);
1155                 if (!PageUptodate(page))
1156                         unmapped = 0;
1157                 need_trans = delalloc + unmapped + unwritten;
1158         }
1159
1160         /*
1161          * If we need a transaction and the process flags say
1162          * we are already in a transaction, or no IO is allowed
1163          * then mark the page dirty again and leave the page
1164          * as is.
1165          */
1166         if (PFLAGS_TEST_FSTRANS() && need_trans)
1167                 goto out_fail;
1168
1169         /*
1170          * Delay hooking up buffer heads until we have
1171          * made our go/no-go decision.
1172          */
1173         if (!page_has_buffers(page))
1174                 create_empty_buffers(page, 1 << inode->i_blkbits, 0);
1175
1176         /*
1177          * Convert delayed allocate, unwritten or unmapped space
1178          * to real space and flush out to disk.
1179          */
1180         error = xfs_page_state_convert(inode, page, wbc, 1, unmapped);
1181         if (error == -EAGAIN)
1182                 goto out_fail;
1183         if (unlikely(error < 0))
1184                 goto out_unlock;
1185
1186         return 0;
1187
1188 out_fail:
1189         set_page_dirty(page);
1190         unlock_page(page);
1191         return 0;
1192 out_unlock:
1193         unlock_page(page);
1194         return error;
1195 }
1196
1197 /*
1198  * Called to move a page into cleanable state - and from there
1199  * to be released. Possibly the page is already clean. We always
1200  * have buffer heads in this call.
1201  *
1202  * Returns 0 if the page is ok to release, 1 otherwise.
1203  *
1204  * Possible scenarios are:
1205  *
1206  * 1. We are being called to release a page which has been written
1207  *    to via regular I/O. buffer heads will be dirty and possibly
1208  *    delalloc. If no delalloc buffer heads in this case then we
1209  *    can just return zero.
1210  *
1211  * 2. We are called to release a page which has been written via
1212  *    mmap, all we need to do is ensure there is no delalloc
1213  *    state in the buffer heads, if not we can let the caller
1214  *    free them and we should come back later via writepage.
1215  */
1216 STATIC int
1217 linvfs_release_page(
1218         struct page             *page,
1219         int                     gfp_mask)
1220 {
1221         struct inode            *inode = page->mapping->host;
1222         int                     dirty, delalloc, unmapped, unwritten;
1223         struct writeback_control wbc = {
1224                 .sync_mode = WB_SYNC_ALL,
1225                 .nr_to_write = 1,
1226         };
1227
1228         xfs_page_trace(XFS_RELEASEPAGE_ENTER, inode, page, gfp_mask);
1229
1230         xfs_count_page_state(page, &delalloc, &unmapped, &unwritten);
1231         if (!delalloc && !unwritten)
1232                 goto free_buffers;
1233
1234         if (!(gfp_mask & __GFP_FS))
1235                 return 0;
1236
1237         /* If we are already inside a transaction or the thread cannot
1238          * do I/O, we cannot release this page.
1239          */
1240         if (PFLAGS_TEST_FSTRANS())
1241                 return 0;
1242
1243         /*
1244          * Convert delalloc space to real space, do not flush the
1245          * data out to disk, that will be done by the caller.
1246          * Never need to allocate space here - we will always
1247          * come back to writepage in that case.
1248          */
1249         dirty = xfs_page_state_convert(inode, page, &wbc, 0, 0);
1250         if (dirty == 0 && !unwritten)
1251                 goto free_buffers;
1252         return 0;
1253
1254 free_buffers:
1255         return try_to_free_buffers(page);
1256 }
1257
1258 STATIC int
1259 linvfs_prepare_write(
1260         struct file             *file,
1261         struct page             *page,
1262         unsigned int            from,
1263         unsigned int            to)
1264 {
1265         if (file && (file->f_flags & O_SYNC)) {
1266                 return block_prepare_write(page, from, to,
1267                                                 linvfs_get_block_sync);
1268         } else {
1269                 return block_prepare_write(page, from, to,
1270                                                 linvfs_get_block);
1271         }
1272 }
1273
1274 struct address_space_operations linvfs_aops = {
1275         .readpage               = linvfs_readpage,
1276         .readpages              = linvfs_readpages,
1277         .writepage              = linvfs_writepage,
1278         .sync_page              = block_sync_page,
1279         .releasepage            = linvfs_release_page,
1280         .prepare_write          = linvfs_prepare_write,
1281         .commit_write           = generic_commit_write,
1282         .bmap                   = linvfs_bmap,
1283         .direct_IO              = linvfs_direct_IO,
1284 };