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             *mapp,
 176         int                     flags)
 177 {
 178         vnode_t                 *vp = LINVFS_GET_VP(inode);
 179         int                     error, nmaps = 1;
 180
 181         VOP_BMAP(vp, offset, count, flags, mapp, &nmaps, error);
 182         if (!error && (flags & (BMAPI_WRITE|BMAPI_ALLOCATE)))
 183                 VMODIFY(vp);
 184         return -error;
 185 }
 186
 187 /*
 188  * Finds the corresponding mapping in block @map array of the
 189  * given @offset within a @page.
 190  */
 191 STATIC xfs_iomap_t *
 192 xfs_offset_to_map(
 193         struct page             *page,
 194         xfs_iomap_t             *iomapp,
 195         unsigned long           offset)
 196 {
 197         loff_t                  full_offset;    /* offset from start of file */
 198
 199         ASSERT(offset < PAGE_CACHE_SIZE);
 200
 201         full_offset = page->index;              /* NB: using 64bit number */
 202         full_offset <<= PAGE_CACHE_SHIFT;       /* offset from file start */
 203         full_offset += offset;                  /* offset from page start */
 204
 205         if (full_offset < iomapp->iomap_offset)
 206                 return NULL;
 207         if (iomapp->iomap_offset + (iomapp->iomap_bsize -1) >= full_offset)
 208                 return iomapp;
 209         return NULL;
 210 }
 211
 212 STATIC void
 213 xfs_map_at_offset(
 214         struct page             *page,
 215         struct buffer_head      *bh,
 216         unsigned long           offset,
 217         int                     block_bits,
 218         xfs_iomap_t             *iomapp)
 219 {
 220         xfs_daddr_t             bn;
 221         loff_t                  delta;
 222         int                     sector_shift;
 223
 224         ASSERT(!(iomapp->iomap_flags & IOMAP_HOLE));
 225         ASSERT(!(iomapp->iomap_flags & IOMAP_DELAY));
 226         ASSERT(iomapp->iomap_bn != IOMAP_DADDR_NULL);
 227
 228         delta = page->index;
 229         delta <<= PAGE_CACHE_SHIFT;
 230         delta += offset;
 231         delta -= iomapp->iomap_offset;
 232         delta >>= block_bits;
 233
 234         sector_shift = block_bits - BBSHIFT;
 235         bn = iomapp->iomap_bn >> sector_shift;
 236         bn += delta;
 237         BUG_ON(!bn && !(iomapp->iomap_flags & IOMAP_REALTIME));
 238         ASSERT((bn << sector_shift) >= iomapp->iomap_bn);
 239
 240         lock_buffer(bh);
 241         bh->b_blocknr = bn;
 242         bh->b_bdev = iomapp->iomap_target->pbr_bdev;
 243         set_buffer_mapped(bh);
 244         clear_buffer_delay(bh);
 245 }
 246
 247 /*
 248  * Look for a page at index which is unlocked and contains our
 249  * unwritten extent flagged buffers at its head.  Returns page
 250  * locked and with an extra reference count, and length of the
 251  * unwritten extent component on this page that we can write,
 252  * in units of filesystem blocks.
 253  */
 254 STATIC struct page *
 255 xfs_probe_unwritten_page(
 256         struct address_space    *mapping,
 257         pgoff_t                 index,
 258         xfs_iomap_t             *iomapp,
 259         xfs_buf_t               *pb,
 260         unsigned long           max_offset,
 261         unsigned long           *fsbs,
 262         unsigned int            bbits)
 263 {
 264         struct page             *page;
 265
 266         page = find_trylock_page(mapping, index);
 267         if (!page)
 268                 return NULL;
 269         if (PageWriteback(page))
 270                 goto out;
 271
 272         if (page->mapping && page_has_buffers(page)) {
 273                 struct buffer_head      *bh, *head;
 274                 unsigned long           p_offset = 0;
 275
 276                 *fsbs = 0;
 277                 bh = head = page_buffers(page);
 278                 do {
 279                         if (!buffer_unwritten(bh) || !buffer_uptodate(bh))
 280                                 break;
 281                         if (!xfs_offset_to_map(page, iomapp, p_offset))
 282                                 break;
 283                         if (p_offset >= max_offset)
 284                                 break;
 285                         xfs_map_at_offset(page, bh, p_offset, bbits, iomapp);
 286                         set_buffer_unwritten_io(bh);
 287                         bh->b_private = pb;
 288                         p_offset += bh->b_size;
 289                         (*fsbs)++;
 290                 } while ((bh = bh->b_this_page) != head);
 291
 292                 if (p_offset)
 293                         return page;
 294         }
 295
 296 out:
 297         unlock_page(page);
 298         return NULL;
 299 }
 300
 301 /*
 302  * Look for a page at index which is unlocked and not mapped
 303  * yet - clustering for mmap write case.
 304  */
 305 STATIC unsigned int
 306 xfs_probe_unmapped_page(
 307         struct address_space    *mapping,
 308         pgoff_t                 index,
 309         unsigned int            pg_offset)
 310 {
 311         struct page             *page;
 312         int                     ret = 0;
 313
 314         page = find_trylock_page(mapping, index);
 315         if (!page)
 316                 return 0;
 317         if (PageWriteback(page))
 318                 goto out;
 319
 320         if (page->mapping && PageDirty(page)) {
 321                 if (page_has_buffers(page)) {
 322                         struct buffer_head      *bh, *head;
 323
 324                         bh = head = page_buffers(page);
 325                         do {
 326                                 if (buffer_mapped(bh) || !buffer_uptodate(bh))
 327                                         break;
 328                                 ret += bh->b_size;
 329                                 if (ret >= pg_offset)
 330                                         break;
 331                         } while ((bh = bh->b_this_page) != head);
 332                 } else
 333                         ret = PAGE_CACHE_SIZE;
 334         }
 335
 336 out:
 337         unlock_page(page);
 338         return ret;
 339 }
 340
 341 STATIC unsigned int
 342 xfs_probe_unmapped_cluster(
 343         struct inode            *inode,
 344         struct page             *startpage,
 345         struct buffer_head      *bh,
 346         struct buffer_head      *head)
 347 {
 348         pgoff_t                 tindex, tlast, tloff;
 349         unsigned int            pg_offset, len, total = 0;
 350         struct address_space    *mapping = inode->i_mapping;
 351
 352         /* First sum forwards in this page */
 353         do {
 354                 if (buffer_mapped(bh))
 355                         break;
 356                 total += bh->b_size;
 357         } while ((bh = bh->b_this_page) != head);
 358
 359         /* If we reached the end of the page, sum forwards in
 360          * following pages.
 361          */
 362         if (bh == head) {
 363                 tlast = i_size_read(inode) >> PAGE_CACHE_SHIFT;
 364                 /* Prune this back to avoid pathological behavior */
 365                 tloff = min(tlast, startpage->index + 64);
 366                 for (tindex = startpage->index + 1; tindex < tloff; tindex++) {
 367                         len = xfs_probe_unmapped_page(mapping, tindex,
 368                                                         PAGE_CACHE_SIZE);
 369                         if (!len)
 370                                 return total;
 371                         total += len;
 372                 }
 373                 if (tindex == tlast &&
 374                     (pg_offset = i_size_read(inode) & (PAGE_CACHE_SIZE - 1))) {
 375                         total += xfs_probe_unmapped_page(mapping,
 376                                                         tindex, pg_offset);
 377                 }
 378         }
 379         return total;
 380 }
 381
 382 /*
 383  * Probe for a given page (index) in the inode and test if it is delayed
 384  * and without unwritten buffers.  Returns page locked and with an extra
 385  * reference count.
 386  */
 387 STATIC struct page *
 388 xfs_probe_delalloc_page(
 389         struct inode            *inode,
 390         pgoff_t                 index)
 391 {
 392         struct page             *page;
 393
 394         page = find_trylock_page(inode->i_mapping, index);
 395         if (!page)
 396                 return NULL;
 397         if (PageWriteback(page))
 398                 goto out;
 399
 400         if (page->mapping && page_has_buffers(page)) {
 401                 struct buffer_head      *bh, *head;
 402                 int                     acceptable = 0;
 403
 404                 bh = head = page_buffers(page);
 405                 do {
 406                         if (buffer_unwritten(bh)) {
 407                                 acceptable = 0;
 408                                 break;
 409                         } else if (buffer_delay(bh)) {
 410                                 acceptable = 1;
 411                         }
 412                 } while ((bh = bh->b_this_page) != head);
 413
 414                 if (acceptable)
 415                         return page;
 416         }
 417
 418 out:
 419         unlock_page(page);
 420         return NULL;
 421 }
 422
 423 STATIC int
 424 xfs_map_unwritten(
 425         struct inode            *inode,
 426         struct page             *start_page,
 427         struct buffer_head      *head,
 428         struct buffer_head      *curr,
 429         unsigned long           p_offset,
 430         int                     block_bits,
 431         xfs_iomap_t             *iomapp,
 432         struct writeback_control *wbc,
 433         int                     startio,
 434         int                     all_bh)
 435 {
 436         struct buffer_head      *bh = curr;
 437         xfs_iomap_t             *tmp;
 438         xfs_buf_t               *pb;
 439         loff_t                  offset, size;
 440         unsigned long           nblocks = 0;
 441
 442         offset = start_page->index;
 443         offset <<= PAGE_CACHE_SHIFT;
 444         offset += p_offset;
 445
 446         /* get an "empty" pagebuf to manage IO completion
 447          * Proper values will be set before returning */
 448         pb = pagebuf_lookup(iomapp->iomap_target, 0, 0, 0);
 449         if (!pb)
 450                 return -EAGAIN;
 451
 452         /* Take a reference to the inode to prevent it from
 453          * being reclaimed while we have outstanding unwritten
 454          * extent IO on it.
 455          */
 456         if ((igrab(inode)) != inode) {
 457                 pagebuf_free(pb);
 458                 return -EAGAIN;
 459         }
 460
 461         /* Set the count to 1 initially, this will stop an I/O
 462          * completion callout which happens before we have started
 463          * all the I/O from calling pagebuf_iodone too early.
 464          */
 465         atomic_set(&pb->pb_io_remaining, 1);
 466
 467         /* First map forwards in the page consecutive buffers
 468          * covering this unwritten extent
 469          */
 470         do {
 471                 if (!buffer_unwritten(bh))
 472                         break;
 473                 tmp = xfs_offset_to_map(start_page, iomapp, p_offset);
 474                 if (!tmp)
 475                         break;
 476                 xfs_map_at_offset(start_page, bh, p_offset, block_bits, iomapp);
 477                 set_buffer_unwritten_io(bh);
 478                 bh->b_private = pb;
 479                 p_offset += bh->b_size;
 480                 nblocks++;
 481         } while ((bh = bh->b_this_page) != head);
 482
 483         atomic_add(nblocks, &pb->pb_io_remaining);
 484
 485         /* If we reached the end of the page, map forwards in any
 486          * following pages which are also covered by this extent.
 487          */
 488         if (bh == head) {
 489                 struct address_space    *mapping = inode->i_mapping;
 490                 pgoff_t                 tindex, tloff, tlast;
 491                 unsigned long           bs;
 492                 unsigned int            pg_offset, bbits = inode->i_blkbits;
 493                 struct page             *page;
 494
 495                 tlast = i_size_read(inode) >> PAGE_CACHE_SHIFT;
 496                 tloff = (iomapp->iomap_offset + iomapp->iomap_bsize) >> PAGE_CACHE_SHIFT;
 497                 tloff = min(tlast, tloff);
 498                 for (tindex = start_page->index + 1; tindex < tloff; tindex++) {
 499                         page = xfs_probe_unwritten_page(mapping,
 500                                                 tindex, iomapp, pb,
 501                                                 PAGE_CACHE_SIZE, &bs, bbits);
 502                         if (!page)
 503                                 break;
 504                         nblocks += bs;
 505                         atomic_add(bs, &pb->pb_io_remaining);
 506                         xfs_convert_page(inode, page, iomapp, wbc, pb,
 507                                                         startio, all_bh);
 508                         /* stop if converting the next page might add
 509                          * enough blocks that the corresponding byte
 510                          * count won't fit in our ulong page buf length */
 511                         if (nblocks >= ((ULONG_MAX - PAGE_SIZE) >> block_bits))
 512                                 goto enough;
 513                 }
 514
 515                 if (tindex == tlast &&
 516                     (pg_offset = (i_size_read(inode) & (PAGE_CACHE_SIZE - 1)))) {
 517                         page = xfs_probe_unwritten_page(mapping,
 518                                                         tindex, iomapp, pb,
 519                                                         pg_offset, &bs, bbits);
 520                         if (page) {
 521                                 nblocks += bs;
 522                                 atomic_add(bs, &pb->pb_io_remaining);
 523                                 xfs_convert_page(inode, page, iomapp, wbc, pb,
 524                                                         startio, all_bh);
 525                                 if (nblocks >= ((ULONG_MAX - PAGE_SIZE) >> block_bits))
 526                                         goto enough;
 527                         }
 528                 }
 529         }
 530
 531 enough:
 532         size = nblocks;         /* NB: using 64bit number here */
 533         size <<= block_bits;    /* convert fsb's to byte range */
 534
 535         XFS_BUF_DATAIO(pb);
 536         XFS_BUF_ASYNC(pb);
 537         XFS_BUF_SET_SIZE(pb, size);
 538         XFS_BUF_SET_COUNT(pb, size);
 539         XFS_BUF_SET_OFFSET(pb, offset);
 540         XFS_BUF_SET_FSPRIVATE(pb, LINVFS_GET_VP(inode));
 541         XFS_BUF_SET_IODONE_FUNC(pb, linvfs_unwritten_convert);
 542
 543         if (atomic_dec_and_test(&pb->pb_io_remaining) == 1) {
 544                 pagebuf_iodone(pb, 1, 1);
 545         }
 546
 547         return 0;
 548 }
 549
 550 STATIC void
 551 xfs_submit_page(
 552         struct page             *page,
 553         struct writeback_control *wbc,
 554         struct buffer_head      *bh_arr[],
 555         int                     bh_count,
 556         int                     probed_page,
 557         int                     clear_dirty)
 558 {
 559         struct buffer_head      *bh;
 560         int                     i;
 561
 562         BUG_ON(PageWriteback(page));
 563         set_page_writeback(page);
 564         if (clear_dirty)
 565                 clear_page_dirty(page);
 566         unlock_page(page);
 567
 568         if (bh_count) {
 569                 for (i = 0; i < bh_count; i++) {
 570                         bh = bh_arr[i];
 571                         mark_buffer_async_write(bh);
 572                         if (buffer_unwritten(bh))
 573                                 set_buffer_unwritten_io(bh);
 574                         set_buffer_uptodate(bh);
 575                         clear_buffer_dirty(bh);
 576                 }
 577
 578                 for (i = 0; i < bh_count; i++)
 579                         submit_bh(WRITE, bh_arr[i]);
 580
 581                 if (probed_page && clear_dirty)
 582                         wbc->nr_to_write--;     /* Wrote an "extra" page */
 583         } else {
 584                 end_page_writeback(page);
 585                 wbc->pages_skipped++;   /* We didn't write this page */
 586         }
 587 }
 588
 589 /*
 590  * Allocate & map buffers for page given the extent map. Write it out.
 591  * except for the original page of a writepage, this is called on
 592  * delalloc/unwritten pages only, for the original page it is possible
 593  * that the page has no mapping at all.
 594  */
 595 STATIC void
 596 xfs_convert_page(
 597         struct inode            *inode,
 598         struct page             *page,
 599         xfs_iomap_t             *iomapp,
 600         struct writeback_control *wbc,
 601         void                    *private,
 602         int                     startio,
 603         int                     all_bh)
 604 {
 605         struct buffer_head      *bh_arr[MAX_BUF_PER_PAGE], *bh, *head;
 606         xfs_iomap_t             *mp = iomapp, *tmp;
 607         unsigned long           end, offset;
 608         pgoff_t                 end_index;
 609         int                     i = 0, index = 0;
 610         int                     bbits = inode->i_blkbits;
 611
 612         end_index = i_size_read(inode) >> PAGE_CACHE_SHIFT;
 613         if (page->index < end_index) {
 614                 end = PAGE_CACHE_SIZE;
 615         } else {
 616                 end = i_size_read(inode) & (PAGE_CACHE_SIZE-1);
 617         }
 618         bh = head = page_buffers(page);
 619         do {
 620                 offset = i << bbits;
 621                 if (offset >= end)
 622                         break;
 623                 if (!(PageUptodate(page) || buffer_uptodate(bh)))
 624                         continue;
 625                 if (buffer_mapped(bh) && all_bh &&
 626                     !(buffer_unwritten(bh) || buffer_delay(bh))) {
 627                         if (startio) {
 628                                 lock_buffer(bh);
 629                                 bh_arr[index++] = bh;
 630                         }
 631                         continue;
 632                 }
 633                 tmp = xfs_offset_to_map(page, mp, offset);
 634                 if (!tmp)
 635                         continue;
 636                 ASSERT(!(tmp->iomap_flags & IOMAP_HOLE));
 637                 ASSERT(!(tmp->iomap_flags & IOMAP_DELAY));
 638
 639                 /* If this is a new unwritten extent buffer (i.e. one
 640                  * that we haven't passed in private data for, we must
 641                  * now map this buffer too.
 642                  */
 643                 if (buffer_unwritten(bh) && !bh->b_end_io) {
 644                         ASSERT(tmp->iomap_flags & IOMAP_UNWRITTEN);
 645                         xfs_map_unwritten(inode, page, head, bh, offset,
 646                                         bbits, tmp, wbc, startio, all_bh);
 647                 } else if (! (buffer_unwritten(bh) && buffer_locked(bh))) {
 648                         xfs_map_at_offset(page, bh, offset, bbits, tmp);
 649                         if (buffer_unwritten(bh)) {
 650                                 set_buffer_unwritten_io(bh);
 651                                 bh->b_private = private;
 652                                 ASSERT(private);
 653                         }
 654                 }
 655                 if (startio) {
 656                         bh_arr[index++] = bh;
 657                 } else {
 658                         set_buffer_dirty(bh);
 659                         unlock_buffer(bh);
 660                         mark_buffer_dirty(bh);
 661                 }
 662         } while (i++, (bh = bh->b_this_page) != head);
 663
 664         if (startio) {
 665                 xfs_submit_page(page, wbc, bh_arr, index, 1, index == i);
 666         } else {
 667                 unlock_page(page);
 668         }
 669 }
 670
 671 /*
 672  * Convert & write out a cluster of pages in the same extent as defined
 673  * by mp and following the start page.
 674  */
 675 STATIC void
 676 xfs_cluster_write(
 677         struct inode            *inode,
 678         pgoff_t                 tindex,
 679         xfs_iomap_t             *iomapp,
 680         struct writeback_control *wbc,
 681         int                     startio,
 682         int                     all_bh,
 683         pgoff_t                 tlast)
 684 {
 685         struct page             *page;
 686
 687         for (; tindex <= tlast; tindex++) {
 688                 page = xfs_probe_delalloc_page(inode, tindex);
 689                 if (!page)
 690                         break;
 691                 xfs_convert_page(inode, page, iomapp, wbc, NULL,
 692                                 startio, all_bh);
 693         }
 694 }
 695
 696 /*
 697  * Calling this without startio set means we are being asked to make a dirty
 698  * page ready for freeing it's buffers.  When called with startio set then
 699  * we are coming from writepage.
 700  *
 701  * When called with startio set it is important that we write the WHOLE
 702  * page if possible.
 703  * The bh->b_state's cannot know if any of the blocks or which block for
 704  * that matter are dirty due to mmap writes, and therefore bh uptodate is
 705  * only vaild if the page itself isn't completely uptodate.  Some layers
 706  * may clear the page dirty flag prior to calling write page, under the
 707  * assumption the entire page will be written out; by not writing out the
 708  * whole page the page can be reused before all valid dirty data is
 709  * written out.  Note: in the case of a page that has been dirty'd by
 710  * mapwrite and but partially setup by block_prepare_write the
 711  * bh->b_states's will not agree and only ones setup by BPW/BCW will have
 712  * valid state, thus the whole page must be written out thing.
 713  */
 714
 715 STATIC int
 716 xfs_page_state_convert(
 717         struct inode    *inode,
 718         struct page     *page,
 719         struct writeback_control *wbc,
 720         int             startio,
 721         int             unmapped) /* also implies page uptodate */
 722 {
 723         struct buffer_head      *bh_arr[MAX_BUF_PER_PAGE], *bh, *head;
 724         xfs_iomap_t             *iomp, iomap;
 725         loff_t                  offset;
 726         unsigned long           p_offset = 0;
 727         __uint64_t              end_offset;
 728         pgoff_t                 end_index, last_index, tlast;
 729         int                     len, err, i, cnt = 0, uptodate = 1;
 730         int                     flags = startio ? 0 : BMAPI_TRYLOCK;
 731         int                     page_dirty = 1;
 732         int                     delalloc = 0;
 733
 734
 735         /* Are we off the end of the file ? */
 736         offset = i_size_read(inode);
 737         end_index = offset >> PAGE_CACHE_SHIFT;
 738         last_index = (offset - 1) >> 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 (!startio)
 773                                 continue;
 774                         if (!iomp) {
 775                                 err = xfs_map_blocks(inode, offset, len, &iomap,
 776                                                 BMAPI_READ|BMAPI_IGNSTATE);
 777                                 if (err) {
 778                                         goto error;
 779                                 }
 780                                 iomp = xfs_offset_to_map(page, &iomap,
 781                                                                 p_offset);
 782                         }
 783                         if (iomp) {
 784                                 if (!bh->b_end_io) {
 785                                         err = xfs_map_unwritten(inode, page,
 786                                                         head, bh, p_offset,
 787                                                         inode->i_blkbits, iomp,
 788                                                         wbc, startio, unmapped);
 789                                         if (err) {
 790                                                 goto error;
 791                                         }
 792                                 } else {
 793                                         set_bit(BH_Lock, &bh->b_state);
 794                                 }
 795                                 BUG_ON(!buffer_locked(bh));
 796                                 bh_arr[cnt++] = bh;
 797                                 page_dirty = 0;
 798                         }
 799                 /*
 800                  * Second case, allocate space for a delalloc buffer.
 801                  * We can return EAGAIN here in the release page case.
 802                  */
 803                 } else if (buffer_delay(bh)) {
 804                         if (!iomp) {
 805                                 delalloc = 1;
 806                                 err = xfs_map_blocks(inode, offset, len, &iomap,
 807                                                 BMAPI_ALLOCATE | flags);
 808                                 if (err) {
 809                                         goto error;
 810                                 }
 811                                 iomp = xfs_offset_to_map(page, &iomap,
 812                                                                 p_offset);
 813                         }
 814                         if (iomp) {
 815                                 xfs_map_at_offset(page, bh, p_offset,
 816                                                 inode->i_blkbits, iomp);
 817                                 if (startio) {
 818                                         bh_arr[cnt++] = bh;
 819                                 } else {
 820                                         set_buffer_dirty(bh);
 821                                         unlock_buffer(bh);
 822                                         mark_buffer_dirty(bh);
 823                                 }
 824                                 page_dirty = 0;
 825                         }
 826                 } else if ((buffer_uptodate(bh) || PageUptodate(page)) &&
 827                            (unmapped || startio)) {
 828
 829                         if (!buffer_mapped(bh)) {
 830                                 int     size;
 831
 832                                 /*
 833                                  * Getting here implies an unmapped buffer
 834                                  * was found, and we are in a path where we
 835                                  * need to write the whole page out.
 836                                  */
 837                                 if (!iomp) {
 838                                         size = xfs_probe_unmapped_cluster(
 839                                                         inode, page, bh, head);
 840                                         err = xfs_map_blocks(inode, offset,
 841                                                         size, &iomap,
 842                                                         BMAPI_WRITE|BMAPI_MMAP);
 843                                         if (err) {
 844                                                 goto error;
 845                                         }
 846                                         iomp = xfs_offset_to_map(page, &iomap,
 847                                                                      p_offset);
 848                                 }
 849                                 if (iomp) {
 850                                         xfs_map_at_offset(page,
 851                                                         bh, p_offset,
 852                                                         inode->i_blkbits, iomp);
 853                                         if (startio) {
 854                                                 bh_arr[cnt++] = bh;
 855                                         } else {
 856                                                 set_buffer_dirty(bh);
 857                                                 unlock_buffer(bh);
 858                                                 mark_buffer_dirty(bh);
 859                                         }
 860                                         page_dirty = 0;
 861                                 }
 862                         } else if (startio) {
 863                                 if (buffer_uptodate(bh) &&
 864                                     !test_and_set_bit(BH_Lock, &bh->b_state)) {
 865                                         bh_arr[cnt++] = bh;
 866                                         page_dirty = 0;
 867                                 }
 868                         }
 869                 }
 870         } while (offset += len, p_offset += len,
 871                 ((bh = bh->b_this_page) != head));
 872
 873         if (uptodate && bh == head)
 874                 SetPageUptodate(page);
 875
 876         if (startio)
 877                 xfs_submit_page(page, wbc, bh_arr, cnt, 0, 1);
 878
 879         if (iomp) {
 880                 tlast = (iomp->iomap_offset + iomp->iomap_bsize - 1) >>
 881                                         PAGE_CACHE_SHIFT;
 882                 if (delalloc && (tlast > last_index))
 883                         tlast = last_index;
 884                 xfs_cluster_write(inode, page->index + 1, iomp, wbc,
 885                                         startio, unmapped, tlast);
 886         }
 887
 888         return page_dirty;
 889
 890 error:
 891         for (i = 0; i < cnt; i++) {
 892                 unlock_buffer(bh_arr[i]);
 893         }
 894
 895         /*
 896          * If it's delalloc and we have nowhere to put it,
 897          * throw it away, unless the lower layers told
 898          * us to try again.
 899          */
 900         if (err != -EAGAIN) {
 901                 if (!unmapped) {
 902                         block_invalidatepage(page, 0);
 903                 }
 904                 ClearPageUptodate(page);
 905         }
 906         return err;
 907 }
 908
 909 STATIC int
 910 linvfs_get_block_core(
 911         struct inode            *inode,
 912         sector_t                iblock,
 913         unsigned long           blocks,
 914         struct buffer_head      *bh_result,
 915         int                     create,
 916         int                     direct,
 917         bmapi_flags_t           flags)
 918 {
 919         vnode_t                 *vp = LINVFS_GET_VP(inode);
 920         xfs_iomap_t             iomap;
 921         int                     retpbbm = 1;
 922         int                     error;
 923         ssize_t                 size;
 924         loff_t                  offset = (loff_t)iblock << inode->i_blkbits;
 925
 926         if (blocks)
 927                 size = blocks << inode->i_blkbits;
 928         else
 929                 size = 1 << inode->i_blkbits;
 930
 931         VOP_BMAP(vp, offset, size,
 932                 create ? flags : BMAPI_READ, &iomap, &retpbbm, error);
 933         if (error)
 934                 return -error;
 935
 936         if (retpbbm == 0)
 937                 return 0;
 938
 939         if (iomap.iomap_bn != IOMAP_DADDR_NULL) {
 940                 xfs_daddr_t             bn;
 941                 loff_t                  delta;
 942
 943                 /* For unwritten extents do not report a disk address on
 944                  * the read case (treat as if we're reading into a hole).
 945                  */
 946                 if (create || !(iomap.iomap_flags & IOMAP_UNWRITTEN)) {
 947                         delta = offset - iomap.iomap_offset;
 948                         delta >>= inode->i_blkbits;
 949
 950                         bn = iomap.iomap_bn >> (inode->i_blkbits - BBSHIFT);
 951                         bn += delta;
 952                         BUG_ON(!bn && !(iomap.iomap_flags & IOMAP_REALTIME));
 953                         bh_result->b_blocknr = bn;
 954                         set_buffer_mapped(bh_result);
 955                 }
 956                 if (create && (iomap.iomap_flags & IOMAP_UNWRITTEN)) {
 957                         if (direct)
 958                                 bh_result->b_private = inode;
 959                         set_buffer_unwritten(bh_result);
 960                         set_buffer_delay(bh_result);
 961                 }
 962         }
 963
 964         /* If this is a realtime file, data might be on a new device */
 965         bh_result->b_bdev = iomap.iomap_target->pbr_bdev;
 966
 967         /* If we previously allocated a block out beyond eof and
 968          * we are now coming back to use it then we will need to
 969          * flag it as new even if it has a disk address.
 970          */
 971         if (create &&
 972             ((!buffer_mapped(bh_result) && !buffer_uptodate(bh_result)) ||
 973              (offset >= i_size_read(inode)) || (iomap.iomap_flags & IOMAP_NEW))) {
 974                 set_buffer_new(bh_result);
 975         }
 976
 977         if (iomap.iomap_flags & IOMAP_DELAY) {
 978                 BUG_ON(direct);
 979                 if (create) {
 980                         set_buffer_mapped(bh_result);
 981                         set_buffer_uptodate(bh_result);
 982                 }
 983                 set_buffer_delay(bh_result);
 984         }
 985
 986         if (blocks) {
 987                 bh_result->b_size = (ssize_t)min(
 988                         (loff_t)(iomap.iomap_bsize - iomap.iomap_delta),
 989                         (loff_t)(blocks << inode->i_blkbits));
 990         }
 991
 992         return 0;
 993 }
 994
 995 int
 996 linvfs_get_block(
 997         struct inode            *inode,
 998         sector_t                iblock,
 999         struct buffer_head      *bh_result,
1000         int                     create)
1001 {
1002         return linvfs_get_block_core(inode, iblock, 0, bh_result,
1003                                         create, 0, BMAPI_WRITE);
1004 }
1005
1006 STATIC int
1007 linvfs_get_blocks_direct(
1008         struct inode            *inode,
1009         sector_t                iblock,
1010         unsigned long           max_blocks,
1011         struct buffer_head      *bh_result,
1012         int                     create)
1013 {
1014         return linvfs_get_block_core(inode, iblock, max_blocks, bh_result,
1015                                         create, 1, BMAPI_WRITE|BMAPI_DIRECT);
1016 }
1017
1018 STATIC ssize_t
1019 linvfs_direct_IO(
1020         int                     rw,
1021         struct kiocb            *iocb,
1022         const struct iovec      *iov,
1023         loff_t                  offset,
1024         unsigned long           nr_segs)
1025 {
1026         struct file     *file = iocb->ki_filp;
1027         struct inode    *inode = file->f_mapping->host;
1028         vnode_t         *vp = LINVFS_GET_VP(inode);
1029         xfs_iomap_t     iomap;
1030         int             maps = 1;
1031         int             error;
1032
1033         VOP_BMAP(vp, offset, 0, BMAPI_DEVICE, &iomap, &maps, error);
1034         if (error)
1035                 return -error;
1036
1037         return blockdev_direct_IO_own_locking(rw, iocb, inode,
1038                 iomap.iomap_target->pbr_bdev,
1039                 iov, offset, nr_segs,
1040                 linvfs_get_blocks_direct,
1041                 linvfs_unwritten_convert_direct);
1042 }
1043
1044
1045 STATIC sector_t
1046 linvfs_bmap(
1047         struct address_space    *mapping,
1048         sector_t                block)
1049 {
1050         struct inode            *inode = (struct inode *)mapping->host;
1051         vnode_t                 *vp = LINVFS_GET_VP(inode);
1052         int                     error;
1053
1054         vn_trace_entry(vp, "linvfs_bmap", (inst_t *)__return_address);
1055
1056         VOP_RWLOCK(vp, VRWLOCK_READ);
1057         VOP_FLUSH_PAGES(vp, (xfs_off_t)0, -1, 0, FI_REMAPF, error);
1058         VOP_RWUNLOCK(vp, VRWLOCK_READ);
1059         return generic_block_bmap(mapping, block, linvfs_get_block);
1060 }
1061
1062 STATIC int
1063 linvfs_readpage(
1064         struct file             *unused,
1065         struct page             *page)
1066 {
1067         return mpage_readpage(page, linvfs_get_block);
1068 }
1069
1070 STATIC int
1071 linvfs_readpages(
1072         struct file             *unused,
1073         struct address_space    *mapping,
1074         struct list_head        *pages,
1075         unsigned                nr_pages)
1076 {
1077         return mpage_readpages(mapping, pages, nr_pages, linvfs_get_block);
1078 }
1079
1080 STATIC void
1081 xfs_count_page_state(
1082         struct page             *page,
1083         int                     *delalloc,
1084         int                     *unmapped,
1085         int                     *unwritten)
1086 {
1087         struct buffer_head      *bh, *head;
1088
1089         *delalloc = *unmapped = *unwritten = 0;
1090
1091         bh = head = page_buffers(page);
1092         do {
1093                 if (buffer_uptodate(bh) && !buffer_mapped(bh))
1094                         (*unmapped) = 1;
1095                 else if (buffer_unwritten(bh) && !buffer_delay(bh))
1096                         clear_buffer_unwritten(bh);
1097                 else if (buffer_unwritten(bh))
1098                         (*unwritten) = 1;
1099                 else if (buffer_delay(bh))
1100                         (*delalloc) = 1;
1101         } while ((bh = bh->b_this_page) != head);
1102 }
1103
1104
1105 /*
1106  * writepage: Called from one of two places:
1107  *
1108  * 1. we are flushing a delalloc buffer head.
1109  *
1110  * 2. we are writing out a dirty page. Typically the page dirty
1111  *    state is cleared before we get here. In this case is it
1112  *    conceivable we have no buffer heads.
1113  *
1114  * For delalloc space on the page we need to allocate space and
1115  * flush it. For unmapped buffer heads on the page we should
1116  * allocate space if the page is uptodate. For any other dirty
1117  * buffer heads on the page we should flush them.
1118  *
1119  * If we detect that a transaction would be required to flush
1120  * the page, we have to check the process flags first, if we
1121  * are already in a transaction or disk I/O during allocations
1122  * is off, we need to fail the writepage and redirty the page.
1123  */
1124
1125 STATIC int
1126 linvfs_writepage(
1127         struct page             *page,
1128         struct writeback_control *wbc)
1129 {
1130         int                     error;
1131         int                     need_trans;
1132         int                     delalloc, unmapped, unwritten;
1133         struct inode            *inode = page->mapping->host;
1134
1135         xfs_page_trace(XFS_WRITEPAGE_ENTER, inode, page, 0);
1136
1137         /*
1138          * We need a transaction if:
1139          *  1. There are delalloc buffers on the page
1140          *  2. The page is uptodate and we have unmapped buffers
1141          *  3. The page is uptodate and we have no buffers
1142          *  4. There are unwritten buffers on the page
1143          */
1144
1145         if (!page_has_buffers(page)) {
1146                 unmapped = 1;
1147                 need_trans = 1;
1148         } else {
1149                 xfs_count_page_state(page, &delalloc, &unmapped, &unwritten);
1150                 if (!PageUptodate(page))
1151                         unmapped = 0;
1152                 need_trans = delalloc + unmapped + unwritten;
1153         }
1154
1155         /*
1156          * If we need a transaction and the process flags say
1157          * we are already in a transaction, or no IO is allowed
1158          * then mark the page dirty again and leave the page
1159          * as is.
1160          */
1161         if (PFLAGS_TEST_FSTRANS() && need_trans)
1162                 goto out_fail;
1163
1164         /*
1165          * Delay hooking up buffer heads until we have
1166          * made our go/no-go decision.
1167          */
1168         if (!page_has_buffers(page))
1169                 create_empty_buffers(page, 1 << inode->i_blkbits, 0);
1170
1171         /*
1172          * Convert delayed allocate, unwritten or unmapped space
1173          * to real space and flush out to disk.
1174          */
1175         error = xfs_page_state_convert(inode, page, wbc, 1, unmapped);
1176         if (error == -EAGAIN)
1177                 goto out_fail;
1178         if (unlikely(error < 0))
1179                 goto out_unlock;
1180
1181         return 0;
1182
1183 out_fail:
1184         redirty_page_for_writepage(wbc, page);
1185         unlock_page(page);
1186         return 0;
1187 out_unlock:
1188         unlock_page(page);
1189         return error;
1190 }
1191
1192 /*
1193  * Called to move a page into cleanable state - and from there
1194  * to be released. Possibly the page is already clean. We always
1195  * have buffer heads in this call.
1196  *
1197  * Returns 0 if the page is ok to release, 1 otherwise.
1198  *
1199  * Possible scenarios are:
1200  *
1201  * 1. We are being called to release a page which has been written
1202  *    to via regular I/O. buffer heads will be dirty and possibly
1203  *    delalloc. If no delalloc buffer heads in this case then we
1204  *    can just return zero.
1205  *
1206  * 2. We are called to release a page which has been written via
1207  *    mmap, all we need to do is ensure there is no delalloc
1208  *    state in the buffer heads, if not we can let the caller
1209  *    free them and we should come back later via writepage.
1210  */
1211 STATIC int
1212 linvfs_release_page(
1213         struct page             *page,
1214         int                     gfp_mask)
1215 {
1216         struct inode            *inode = page->mapping->host;
1217         int                     dirty, delalloc, unmapped, unwritten;
1218         struct writeback_control wbc = {
1219                 .sync_mode = WB_SYNC_ALL,
1220                 .nr_to_write = 1,
1221         };
1222
1223         xfs_page_trace(XFS_RELEASEPAGE_ENTER, inode, page, gfp_mask);
1224
1225         xfs_count_page_state(page, &delalloc, &unmapped, &unwritten);
1226         if (!delalloc && !unwritten)
1227                 goto free_buffers;
1228
1229         if (!(gfp_mask & __GFP_FS))
1230                 return 0;
1231
1232         /* If we are already inside a transaction or the thread cannot
1233          * do I/O, we cannot release this page.
1234          */
1235         if (PFLAGS_TEST_FSTRANS())
1236                 return 0;
1237
1238         /*
1239          * Convert delalloc space to real space, do not flush the
1240          * data out to disk, that will be done by the caller.
1241          * Never need to allocate space here - we will always
1242          * come back to writepage in that case.
1243          */
1244         dirty = xfs_page_state_convert(inode, page, &wbc, 0, 0);
1245         if (dirty == 0 && !unwritten)
1246                 goto free_buffers;
1247         return 0;
1248
1249 free_buffers:
1250         return try_to_free_buffers(page);
1251 }
1252
1253 STATIC int
1254 linvfs_prepare_write(
1255         struct file             *file,
1256         struct page             *page,
1257         unsigned int            from,
1258         unsigned int            to)
1259 {
1260         return block_prepare_write(page, from, to, linvfs_get_block);
1261 }
1262
1263 struct address_space_operations linvfs_aops = {
1264         .readpage               = linvfs_readpage,
1265         .readpages              = linvfs_readpages,
1266         .writepage              = linvfs_writepage,
1267         .sync_page              = block_sync_page,
1268         .releasepage            = linvfs_release_page,
1269         .prepare_write          = linvfs_prepare_write,
1270         .commit_write           = generic_commit_write,
1271         .bmap                   = linvfs_bmap,
1272         .direct_IO              = linvfs_direct_IO,
1273 };