fs/xfs/linux-2.6/xfs_aops.c

   1 /*
   2  * Copyright (c) 2000-2005 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 = (loff_t)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 kiocb    *iocb,
 153         loff_t          offset,
 154         ssize_t         size,
 155         void            *private)
 156 {
 157         struct inode    *inode = iocb->ki_filp->f_dentry->d_inode;
 158         ASSERT(!private || inode == (struct inode *)private);
 159
 160         /* private indicates an unwritten extent lay beneath this IO */
 161         if (private && size > 0) {
 162                 vnode_t *vp = LINVFS_GET_VP(inode);
 163                 int     error;
 164
 165                 VOP_BMAP(vp, offset, size, BMAPI_UNWRITTEN, NULL, NULL, error);
 166         }
 167 }
 168
 169 STATIC int
 170 xfs_map_blocks(
 171         struct inode            *inode,
 172         loff_t                  offset,
 173         ssize_t                 count,
 174         xfs_iomap_t             *mapp,
 175         int                     flags)
 176 {
 177         vnode_t                 *vp = LINVFS_GET_VP(inode);
 178         int                     error, nmaps = 1;
 179
 180         VOP_BMAP(vp, offset, count, flags, mapp, &nmaps, error);
 181         if (!error && (flags & (BMAPI_WRITE|BMAPI_ALLOCATE)))
 182                 VMODIFY(vp);
 183         return -error;
 184 }
 185
 186 /*
 187  * Finds the corresponding mapping in block @map array of the
 188  * given @offset within a @page.
 189  */
 190 STATIC xfs_iomap_t *
 191 xfs_offset_to_map(
 192         struct page             *page,
 193         xfs_iomap_t             *iomapp,
 194         unsigned long           offset)
 195 {
 196         loff_t                  full_offset;    /* offset from start of file */
 197
 198         ASSERT(offset < PAGE_CACHE_SIZE);
 199
 200         full_offset = page->index;              /* NB: using 64bit number */
 201         full_offset <<= PAGE_CACHE_SHIFT;       /* offset from file start */
 202         full_offset += offset;                  /* offset from page start */
 203
 204         if (full_offset < iomapp->iomap_offset)
 205                 return NULL;
 206         if (iomapp->iomap_offset + (iomapp->iomap_bsize -1) >= full_offset)
 207                 return iomapp;
 208         return NULL;
 209 }
 210
 211 STATIC void
 212 xfs_map_at_offset(
 213         struct page             *page,
 214         struct buffer_head      *bh,
 215         unsigned long           offset,
 216         int                     block_bits,
 217         xfs_iomap_t             *iomapp)
 218 {
 219         xfs_daddr_t             bn;
 220         loff_t                  delta;
 221         int                     sector_shift;
 222
 223         ASSERT(!(iomapp->iomap_flags & IOMAP_HOLE));
 224         ASSERT(!(iomapp->iomap_flags & IOMAP_DELAY));
 225         ASSERT(iomapp->iomap_bn != IOMAP_DADDR_NULL);
 226
 227         delta = page->index;
 228         delta <<= PAGE_CACHE_SHIFT;
 229         delta += offset;
 230         delta -= iomapp->iomap_offset;
 231         delta >>= block_bits;
 232
 233         sector_shift = block_bits - BBSHIFT;
 234         bn = iomapp->iomap_bn >> sector_shift;
 235         bn += delta;
 236         BUG_ON(!bn && !(iomapp->iomap_flags & IOMAP_REALTIME));
 237         ASSERT((bn << sector_shift) >= iomapp->iomap_bn);
 238
 239         lock_buffer(bh);
 240         bh->b_blocknr = bn;
 241         bh->b_bdev = iomapp->iomap_target->pbr_bdev;
 242         set_buffer_mapped(bh);
 243         clear_buffer_delay(bh);
 244 }
 245
 246 /*
 247  * Look for a page at index which is unlocked and contains our
 248  * unwritten extent flagged buffers at its head.  Returns page
 249  * locked and with an extra reference count, and length of the
 250  * unwritten extent component on this page that we can write,
 251  * in units of filesystem blocks.
 252  */
 253 STATIC struct page *
 254 xfs_probe_unwritten_page(
 255         struct address_space    *mapping,
 256         pgoff_t                 index,
 257         xfs_iomap_t             *iomapp,
 258         xfs_buf_t               *pb,
 259         unsigned long           max_offset,
 260         unsigned long           *fsbs,
 261         unsigned int            bbits)
 262 {
 263         struct page             *page;
 264
 265         page = find_trylock_page(mapping, index);
 266         if (!page)
 267                 return NULL;
 268         if (PageWriteback(page))
 269                 goto out;
 270
 271         if (page->mapping && page_has_buffers(page)) {
 272                 struct buffer_head      *bh, *head;
 273                 unsigned long           p_offset = 0;
 274
 275                 *fsbs = 0;
 276                 bh = head = page_buffers(page);
 277                 do {
 278                         if (!buffer_unwritten(bh) || !buffer_uptodate(bh))
 279                                 break;
 280                         if (!xfs_offset_to_map(page, iomapp, p_offset))
 281                                 break;
 282                         if (p_offset >= max_offset)
 283                                 break;
 284                         xfs_map_at_offset(page, bh, p_offset, bbits, iomapp);
 285                         set_buffer_unwritten_io(bh);
 286                         bh->b_private = pb;
 287                         p_offset += bh->b_size;
 288                         (*fsbs)++;
 289                 } while ((bh = bh->b_this_page) != head);
 290
 291                 if (p_offset)
 292                         return page;
 293         }
 294
 295 out:
 296         unlock_page(page);
 297         return NULL;
 298 }
 299
 300 /*
 301  * Look for a page at index which is unlocked and not mapped
 302  * yet - clustering for mmap write case.
 303  */
 304 STATIC unsigned int
 305 xfs_probe_unmapped_page(
 306         struct address_space    *mapping,
 307         pgoff_t                 index,
 308         unsigned int            pg_offset)
 309 {
 310         struct page             *page;
 311         int                     ret = 0;
 312
 313         page = find_trylock_page(mapping, index);
 314         if (!page)
 315                 return 0;
 316         if (PageWriteback(page))
 317                 goto out;
 318
 319         if (page->mapping && PageDirty(page)) {
 320                 if (page_has_buffers(page)) {
 321                         struct buffer_head      *bh, *head;
 322
 323                         bh = head = page_buffers(page);
 324                         do {
 325                                 if (buffer_mapped(bh) || !buffer_uptodate(bh))
 326                                         break;
 327                                 ret += bh->b_size;
 328                                 if (ret >= pg_offset)
 329                                         break;
 330                         } while ((bh = bh->b_this_page) != head);
 331                 } else
 332                         ret = PAGE_CACHE_SIZE;
 333         }
 334
 335 out:
 336         unlock_page(page);
 337         return ret;
 338 }
 339
 340 STATIC unsigned int
 341 xfs_probe_unmapped_cluster(
 342         struct inode            *inode,
 343         struct page             *startpage,
 344         struct buffer_head      *bh,
 345         struct buffer_head      *head)
 346 {
 347         pgoff_t                 tindex, tlast, tloff;
 348         unsigned int            pg_offset, len, total = 0;
 349         struct address_space    *mapping = inode->i_mapping;
 350
 351         /* First sum forwards in this page */
 352         do {
 353                 if (buffer_mapped(bh))
 354                         break;
 355                 total += bh->b_size;
 356         } while ((bh = bh->b_this_page) != head);
 357
 358         /* If we reached the end of the page, sum forwards in
 359          * following pages.
 360          */
 361         if (bh == head) {
 362                 tlast = i_size_read(inode) >> PAGE_CACHE_SHIFT;
 363                 /* Prune this back to avoid pathological behavior */
 364                 tloff = min(tlast, startpage->index + 64);
 365                 for (tindex = startpage->index + 1; tindex < tloff; tindex++) {
 366                         len = xfs_probe_unmapped_page(mapping, tindex,
 367                                                         PAGE_CACHE_SIZE);
 368                         if (!len)
 369                                 return total;
 370                         total += len;
 371                 }
 372                 if (tindex == tlast &&
 373                     (pg_offset = i_size_read(inode) & (PAGE_CACHE_SIZE - 1))) {
 374                         total += xfs_probe_unmapped_page(mapping,
 375                                                         tindex, pg_offset);
 376                 }
 377         }
 378         return total;
 379 }
 380
 381 /*
 382  * Probe for a given page (index) in the inode and test if it is delayed
 383  * and without unwritten buffers.  Returns page locked and with an extra
 384  * reference count.
 385  */
 386 STATIC struct page *
 387 xfs_probe_delalloc_page(
 388         struct inode            *inode,
 389         pgoff_t                 index)
 390 {
 391         struct page             *page;
 392
 393         page = find_trylock_page(inode->i_mapping, index);
 394         if (!page)
 395                 return NULL;
 396         if (PageWriteback(page))
 397                 goto out;
 398
 399         if (page->mapping && page_has_buffers(page)) {
 400                 struct buffer_head      *bh, *head;
 401                 int                     acceptable = 0;
 402
 403                 bh = head = page_buffers(page);
 404                 do {
 405                         if (buffer_unwritten(bh)) {
 406                                 acceptable = 0;
 407                                 break;
 408                         } else if (buffer_delay(bh)) {
 409                                 acceptable = 1;
 410                         }
 411                 } while ((bh = bh->b_this_page) != head);
 412
 413                 if (acceptable)
 414                         return page;
 415         }
 416
 417 out:
 418         unlock_page(page);
 419         return NULL;
 420 }
 421
 422 STATIC int
 423 xfs_map_unwritten(
 424         struct inode            *inode,
 425         struct page             *start_page,
 426         struct buffer_head      *head,
 427         struct buffer_head      *curr,
 428         unsigned long           p_offset,
 429         int                     block_bits,
 430         xfs_iomap_t             *iomapp,
 431         struct writeback_control *wbc,
 432         int                     startio,
 433         int                     all_bh)
 434 {
 435         struct buffer_head      *bh = curr;
 436         xfs_iomap_t             *tmp;
 437         xfs_buf_t               *pb;
 438         loff_t                  offset, size;
 439         unsigned long           nblocks = 0;
 440
 441         offset = start_page->index;
 442         offset <<= PAGE_CACHE_SHIFT;
 443         offset += p_offset;
 444
 445         /* get an "empty" pagebuf to manage IO completion
 446          * Proper values will be set before returning */
 447         pb = pagebuf_lookup(iomapp->iomap_target, 0, 0, 0);
 448         if (!pb)
 449                 return -EAGAIN;
 450
 451         /* Take a reference to the inode to prevent it from
 452          * being reclaimed while we have outstanding unwritten
 453          * extent IO on it.
 454          */
 455         if ((igrab(inode)) != inode) {
 456                 pagebuf_free(pb);
 457                 return -EAGAIN;
 458         }
 459
 460         /* Set the count to 1 initially, this will stop an I/O
 461          * completion callout which happens before we have started
 462          * all the I/O from calling pagebuf_iodone too early.
 463          */
 464         atomic_set(&pb->pb_io_remaining, 1);
 465
 466         /* First map forwards in the page consecutive buffers
 467          * covering this unwritten extent
 468          */
 469         do {
 470                 if (!buffer_unwritten(bh))
 471                         break;
 472                 tmp = xfs_offset_to_map(start_page, iomapp, p_offset);
 473                 if (!tmp)
 474                         break;
 475                 xfs_map_at_offset(start_page, bh, p_offset, block_bits, iomapp);
 476                 set_buffer_unwritten_io(bh);
 477                 bh->b_private = pb;
 478                 p_offset += bh->b_size;
 479                 nblocks++;
 480         } while ((bh = bh->b_this_page) != head);
 481
 482         atomic_add(nblocks, &pb->pb_io_remaining);
 483
 484         /* If we reached the end of the page, map forwards in any
 485          * following pages which are also covered by this extent.
 486          */
 487         if (bh == head) {
 488                 struct address_space    *mapping = inode->i_mapping;
 489                 pgoff_t                 tindex, tloff, tlast;
 490                 unsigned long           bs;
 491                 unsigned int            pg_offset, bbits = inode->i_blkbits;
 492                 struct page             *page;
 493
 494                 tlast = i_size_read(inode) >> PAGE_CACHE_SHIFT;
 495                 tloff = (iomapp->iomap_offset + iomapp->iomap_bsize) >> PAGE_CACHE_SHIFT;
 496                 tloff = min(tlast, tloff);
 497                 for (tindex = start_page->index + 1; tindex < tloff; tindex++) {
 498                         page = xfs_probe_unwritten_page(mapping,
 499                                                 tindex, iomapp, pb,
 500                                                 PAGE_CACHE_SIZE, &bs, bbits);
 501                         if (!page)
 502                                 break;
 503                         nblocks += bs;
 504                         atomic_add(bs, &pb->pb_io_remaining);
 505                         xfs_convert_page(inode, page, iomapp, wbc, pb,
 506                                                         startio, all_bh);
 507                         /* stop if converting the next page might add
 508                          * enough blocks that the corresponding byte
 509                          * count won't fit in our ulong page buf length */
 510                         if (nblocks >= ((ULONG_MAX - PAGE_SIZE) >> block_bits))
 511                                 goto enough;
 512                 }
 513
 514                 if (tindex == tlast &&
 515                     (pg_offset = (i_size_read(inode) & (PAGE_CACHE_SIZE - 1)))) {
 516                         page = xfs_probe_unwritten_page(mapping,
 517                                                         tindex, iomapp, pb,
 518                                                         pg_offset, &bs, bbits);
 519                         if (page) {
 520                                 nblocks += bs;
 521                                 atomic_add(bs, &pb->pb_io_remaining);
 522                                 xfs_convert_page(inode, page, iomapp, wbc, pb,
 523                                                         startio, all_bh);
 524                                 if (nblocks >= ((ULONG_MAX - PAGE_SIZE) >> block_bits))
 525                                         goto enough;
 526                         }
 527                 }
 528         }
 529
 530 enough:
 531         size = nblocks;         /* NB: using 64bit number here */
 532         size <<= block_bits;    /* convert fsb's to byte range */
 533
 534         XFS_BUF_DATAIO(pb);
 535         XFS_BUF_ASYNC(pb);
 536         XFS_BUF_SET_SIZE(pb, size);
 537         XFS_BUF_SET_COUNT(pb, size);
 538         XFS_BUF_SET_OFFSET(pb, offset);
 539         XFS_BUF_SET_FSPRIVATE(pb, LINVFS_GET_VP(inode));
 540         XFS_BUF_SET_IODONE_FUNC(pb, linvfs_unwritten_convert);
 541
 542         if (atomic_dec_and_test(&pb->pb_io_remaining) == 1) {
 543                 pagebuf_iodone(pb, 1, 1);
 544         }
 545
 546         return 0;
 547 }
 548
 549 STATIC void
 550 xfs_submit_page(
 551         struct page             *page,
 552         struct writeback_control *wbc,
 553         struct buffer_head      *bh_arr[],
 554         int                     bh_count,
 555         int                     probed_page,
 556         int                     clear_dirty)
 557 {
 558         struct buffer_head      *bh;
 559         int                     i;
 560
 561         BUG_ON(PageWriteback(page));
 562         if (bh_count)
 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         }
 584 }
 585
 586 /*
 587  * Allocate & map buffers for page given the extent map. Write it out.
 588  * except for the original page of a writepage, this is called on
 589  * delalloc/unwritten pages only, for the original page it is possible
 590  * that the page has no mapping at all.
 591  */
 592 STATIC void
 593 xfs_convert_page(
 594         struct inode            *inode,
 595         struct page             *page,
 596         xfs_iomap_t             *iomapp,
 597         struct writeback_control *wbc,
 598         void                    *private,
 599         int                     startio,
 600         int                     all_bh)
 601 {
 602         struct buffer_head      *bh_arr[MAX_BUF_PER_PAGE], *bh, *head;
 603         xfs_iomap_t             *mp = iomapp, *tmp;
 604         unsigned long           offset, end_offset;
 605         int                     index = 0;
 606         int                     bbits = inode->i_blkbits;
 607         int                     len, page_dirty;
 608
 609         end_offset = (i_size_read(inode) & (PAGE_CACHE_SIZE - 1));
 610
 611         /*
 612          * page_dirty is initially a count of buffers on the page before
 613          * EOF and is decrememted as we move each into a cleanable state.
 614          */
 615         len = 1 << inode->i_blkbits;
 616         end_offset = max(end_offset, PAGE_CACHE_SIZE);
 617         end_offset = roundup(end_offset, len);
 618         page_dirty = end_offset / len;
 619
 620         offset = 0;
 621         bh = head = page_buffers(page);
 622         do {
 623                 if (offset >= end_offset)
 624                         break;
 625                 if (!(PageUptodate(page) || buffer_uptodate(bh)))
 626                         continue;
 627                 if (buffer_mapped(bh) && all_bh &&
 628                     !(buffer_unwritten(bh) || buffer_delay(bh))) {
 629                         if (startio) {
 630                                 lock_buffer(bh);
 631                                 bh_arr[index++] = bh;
 632                                 page_dirty--;
 633                         }
 634                         continue;
 635                 }
 636                 tmp = xfs_offset_to_map(page, mp, offset);
 637                 if (!tmp)
 638                         continue;
 639                 ASSERT(!(tmp->iomap_flags & IOMAP_HOLE));
 640                 ASSERT(!(tmp->iomap_flags & IOMAP_DELAY));
 641
 642                 /* If this is a new unwritten extent buffer (i.e. one
 643                  * that we haven't passed in private data for, we must
 644                  * now map this buffer too.
 645                  */
 646                 if (buffer_unwritten(bh) && !bh->b_end_io) {
 647                         ASSERT(tmp->iomap_flags & IOMAP_UNWRITTEN);
 648                         xfs_map_unwritten(inode, page, head, bh, offset,
 649                                         bbits, tmp, wbc, startio, all_bh);
 650                 } else if (! (buffer_unwritten(bh) && buffer_locked(bh))) {
 651                         xfs_map_at_offset(page, bh, offset, bbits, tmp);
 652                         if (buffer_unwritten(bh)) {
 653                                 set_buffer_unwritten_io(bh);
 654                                 bh->b_private = private;
 655                                 ASSERT(private);
 656                         }
 657                 }
 658                 if (startio) {
 659                         bh_arr[index++] = bh;
 660                 } else {
 661                         set_buffer_dirty(bh);
 662                         unlock_buffer(bh);
 663                         mark_buffer_dirty(bh);
 664                 }
 665                 page_dirty--;
 666         } while (offset += len, (bh = bh->b_this_page) != head);
 667
 668         if (startio && index) {
 669                 xfs_submit_page(page, wbc, bh_arr, index, 1, !page_dirty);
 670         } else {
 671                 unlock_page(page);
 672         }
 673 }
 674
 675 /*
 676  * Convert & write out a cluster of pages in the same extent as defined
 677  * by mp and following the start page.
 678  */
 679 STATIC void
 680 xfs_cluster_write(
 681         struct inode            *inode,
 682         pgoff_t                 tindex,
 683         xfs_iomap_t             *iomapp,
 684         struct writeback_control *wbc,
 685         int                     startio,
 686         int                     all_bh,
 687         pgoff_t                 tlast)
 688 {
 689         struct page             *page;
 690
 691         for (; tindex <= tlast; tindex++) {
 692                 page = xfs_probe_delalloc_page(inode, tindex);
 693                 if (!page)
 694                         break;
 695                 xfs_convert_page(inode, page, iomapp, wbc, NULL,
 696                                 startio, all_bh);
 697         }
 698 }
 699
 700 /*
 701  * Calling this without startio set means we are being asked to make a dirty
 702  * page ready for freeing it's buffers.  When called with startio set then
 703  * we are coming from writepage.
 704  *
 705  * When called with startio set it is important that we write the WHOLE
 706  * page if possible.
 707  * The bh->b_state's cannot know if any of the blocks or which block for
 708  * that matter are dirty due to mmap writes, and therefore bh uptodate is
 709  * only vaild if the page itself isn't completely uptodate.  Some layers
 710  * may clear the page dirty flag prior to calling write page, under the
 711  * assumption the entire page will be written out; by not writing out the
 712  * whole page the page can be reused before all valid dirty data is
 713  * written out.  Note: in the case of a page that has been dirty'd by
 714  * mapwrite and but partially setup by block_prepare_write the
 715  * bh->b_states's will not agree and only ones setup by BPW/BCW will have
 716  * valid state, thus the whole page must be written out thing.
 717  */
 718
 719 STATIC int
 720 xfs_page_state_convert(
 721         struct inode    *inode,
 722         struct page     *page,
 723         struct writeback_control *wbc,
 724         int             startio,
 725         int             unmapped) /* also implies page uptodate */
 726 {
 727         struct buffer_head      *bh_arr[MAX_BUF_PER_PAGE], *bh, *head;
 728         xfs_iomap_t             *iomp, iomap;
 729         loff_t                  offset;
 730         unsigned long           p_offset = 0;
 731         __uint64_t              end_offset;
 732         pgoff_t                 end_index, last_index, tlast;
 733         int                     len, err, i, cnt = 0, uptodate = 1;
 734         int                     flags;
 735         int                     page_dirty;
 736
 737         /* wait for other IO threads? */
 738         flags = (startio && wbc->sync_mode != WB_SYNC_NONE) ? 0 : BMAPI_TRYLOCK;
 739
 740         /* Is this page beyond the end of the file? */
 741         offset = i_size_read(inode);
 742         end_index = offset >> PAGE_CACHE_SHIFT;
 743         last_index = (offset - 1) >> PAGE_CACHE_SHIFT;
 744         if (page->index >= end_index) {
 745                 if ((page->index >= end_index + 1) ||
 746                     !(i_size_read(inode) & (PAGE_CACHE_SIZE - 1))) {
 747                         err = -EIO;
 748                         goto error;
 749                 }
 750         }
 751
 752         end_offset = min_t(unsigned long long,
 753                         (loff_t)(page->index + 1) << PAGE_CACHE_SHIFT, offset);
 754         offset = (loff_t)page->index << PAGE_CACHE_SHIFT;
 755
 756         /*
 757          * page_dirty is initially a count of buffers on the page before
 758          * EOF and is decrememted as we move each into a cleanable state.
 759          */
 760         len = 1 << inode->i_blkbits;
 761         p_offset = max(p_offset, PAGE_CACHE_SIZE);
 762         p_offset = roundup(p_offset, len);
 763         page_dirty = p_offset / len;
 764
 765         iomp = NULL;
 766         p_offset = 0;
 767         bh = head = page_buffers(page);
 768
 769         do {
 770                 if (offset >= end_offset)
 771                         break;
 772                 if (!buffer_uptodate(bh))
 773                         uptodate = 0;
 774                 if (!(PageUptodate(page) || buffer_uptodate(bh)) && !startio)
 775                         continue;
 776
 777                 if (iomp) {
 778                         iomp = xfs_offset_to_map(page, &iomap, p_offset);
 779                 }
 780
 781                 /*
 782                  * First case, map an unwritten extent and prepare for
 783                  * extent state conversion transaction on completion.
 784                  */
 785                 if (buffer_unwritten(bh)) {
 786                         if (!startio)
 787                                 continue;
 788                         if (!iomp) {
 789                                 err = xfs_map_blocks(inode, offset, len, &iomap,
 790                                                 BMAPI_READ|BMAPI_IGNSTATE);
 791                                 if (err) {
 792                                         goto error;
 793                                 }
 794                                 iomp = xfs_offset_to_map(page, &iomap,
 795                                                                 p_offset);
 796                         }
 797                         if (iomp) {
 798                                 if (!bh->b_end_io) {
 799                                         err = xfs_map_unwritten(inode, page,
 800                                                         head, bh, p_offset,
 801                                                         inode->i_blkbits, iomp,
 802                                                         wbc, startio, unmapped);
 803                                         if (err) {
 804                                                 goto error;
 805                                         }
 806                                 } else {
 807                                         set_bit(BH_Lock, &bh->b_state);
 808                                 }
 809                                 BUG_ON(!buffer_locked(bh));
 810                                 bh_arr[cnt++] = bh;
 811                                 page_dirty--;
 812                         }
 813                 /*
 814                  * Second case, allocate space for a delalloc buffer.
 815                  * We can return EAGAIN here in the release page case.
 816                  */
 817                 } else if (buffer_delay(bh)) {
 818                         if (!iomp) {
 819                                 err = xfs_map_blocks(inode, offset, len, &iomap,
 820                                                 BMAPI_ALLOCATE | flags);
 821                                 if (err) {
 822                                         goto error;
 823                                 }
 824                                 iomp = xfs_offset_to_map(page, &iomap,
 825                                                                 p_offset);
 826                         }
 827                         if (iomp) {
 828                                 xfs_map_at_offset(page, bh, p_offset,
 829                                                 inode->i_blkbits, iomp);
 830                                 if (startio) {
 831                                         bh_arr[cnt++] = bh;
 832                                 } else {
 833                                         set_buffer_dirty(bh);
 834                                         unlock_buffer(bh);
 835                                         mark_buffer_dirty(bh);
 836                                 }
 837                                 page_dirty--;
 838                         }
 839                 } else if ((buffer_uptodate(bh) || PageUptodate(page)) &&
 840                            (unmapped || startio)) {
 841
 842                         if (!buffer_mapped(bh)) {
 843                                 int     size;
 844
 845                                 /*
 846                                  * Getting here implies an unmapped buffer
 847                                  * was found, and we are in a path where we
 848                                  * need to write the whole page out.
 849                                  */
 850                                 if (!iomp) {
 851                                         size = xfs_probe_unmapped_cluster(
 852                                                         inode, page, bh, head);
 853                                         err = xfs_map_blocks(inode, offset,
 854                                                         size, &iomap,
 855                                                         BMAPI_WRITE|BMAPI_MMAP);
 856                                         if (err) {
 857                                                 goto error;
 858                                         }
 859                                         iomp = xfs_offset_to_map(page, &iomap,
 860                                                                      p_offset);
 861                                 }
 862                                 if (iomp) {
 863                                         xfs_map_at_offset(page,
 864                                                         bh, p_offset,
 865                                                         inode->i_blkbits, iomp);
 866                                         if (startio) {
 867                                                 bh_arr[cnt++] = bh;
 868                                         } else {
 869                                                 set_buffer_dirty(bh);
 870                                                 unlock_buffer(bh);
 871                                                 mark_buffer_dirty(bh);
 872                                         }
 873                                         page_dirty--;
 874                                 }
 875                         } else if (startio) {
 876                                 if (buffer_uptodate(bh) &&
 877                                     !test_and_set_bit(BH_Lock, &bh->b_state)) {
 878                                         bh_arr[cnt++] = bh;
 879                                         page_dirty--;
 880                                 }
 881                         }
 882                 }
 883         } while (offset += len, p_offset += len,
 884                 ((bh = bh->b_this_page) != head));
 885
 886         if (uptodate && bh == head)
 887                 SetPageUptodate(page);
 888
 889         if (startio) {
 890                 xfs_submit_page(page, wbc, bh_arr, cnt, 0, !page_dirty);
 891         }
 892
 893         if (iomp) {
 894                 offset = (iomp->iomap_offset + iomp->iomap_bsize - 1) >>
 895                                         PAGE_CACHE_SHIFT;
 896                 tlast = min_t(pgoff_t, offset, last_index);
 897                 xfs_cluster_write(inode, page->index + 1, iomp, wbc,
 898                                         startio, unmapped, tlast);
 899         }
 900
 901         return page_dirty;
 902
 903 error:
 904         for (i = 0; i < cnt; i++) {
 905                 unlock_buffer(bh_arr[i]);
 906         }
 907
 908         /*
 909          * If it's delalloc and we have nowhere to put it,
 910          * throw it away, unless the lower layers told
 911          * us to try again.
 912          */
 913         if (err != -EAGAIN) {
 914                 if (!unmapped) {
 915                         block_invalidatepage(page, 0);
 916                 }
 917                 ClearPageUptodate(page);
 918         }
 919         return err;
 920 }
 921
 922 STATIC int
 923 __linvfs_get_block(
 924         struct inode            *inode,
 925         sector_t                iblock,
 926         unsigned long           blocks,
 927         struct buffer_head      *bh_result,
 928         int                     create,
 929         int                     direct,
 930         bmapi_flags_t           flags)
 931 {
 932         vnode_t                 *vp = LINVFS_GET_VP(inode);
 933         xfs_iomap_t             iomap;
 934         int                     retpbbm = 1;
 935         int                     error;
 936         ssize_t                 size;
 937         loff_t                  offset = (loff_t)iblock << inode->i_blkbits;
 938
 939         if (blocks)
 940                 size = blocks << inode->i_blkbits;
 941         else
 942                 size = 1 << inode->i_blkbits;
 943
 944         VOP_BMAP(vp, offset, size,
 945                 create ? flags : BMAPI_READ, &iomap, &retpbbm, error);
 946         if (error)
 947                 return -error;
 948
 949         if (retpbbm == 0)
 950                 return 0;
 951
 952         if (iomap.iomap_bn != IOMAP_DADDR_NULL) {
 953                 xfs_daddr_t             bn;
 954                 loff_t                  delta;
 955
 956                 /* For unwritten extents do not report a disk address on
 957                  * the read case (treat as if we're reading into a hole).
 958                  */
 959                 if (create || !(iomap.iomap_flags & IOMAP_UNWRITTEN)) {
 960                         delta = offset - iomap.iomap_offset;
 961                         delta >>= inode->i_blkbits;
 962
 963                         bn = iomap.iomap_bn >> (inode->i_blkbits - BBSHIFT);
 964                         bn += delta;
 965                         BUG_ON(!bn && !(iomap.iomap_flags & IOMAP_REALTIME));
 966                         bh_result->b_blocknr = bn;
 967                         set_buffer_mapped(bh_result);
 968                 }
 969                 if (create && (iomap.iomap_flags & IOMAP_UNWRITTEN)) {
 970                         if (direct)
 971                                 bh_result->b_private = inode;
 972                         set_buffer_unwritten(bh_result);
 973                         set_buffer_delay(bh_result);
 974                 }
 975         }
 976
 977         /* If this is a realtime file, data might be on a new device */
 978         bh_result->b_bdev = iomap.iomap_target->pbr_bdev;
 979
 980         /* If we previously allocated a block out beyond eof and
 981          * we are now coming back to use it then we will need to
 982          * flag it as new even if it has a disk address.
 983          */
 984         if (create &&
 985             ((!buffer_mapped(bh_result) && !buffer_uptodate(bh_result)) ||
 986              (offset >= i_size_read(inode)) || (iomap.iomap_flags & IOMAP_NEW))) {
 987                 set_buffer_new(bh_result);
 988         }
 989
 990         if (iomap.iomap_flags & IOMAP_DELAY) {
 991                 BUG_ON(direct);
 992                 if (create) {
 993                         set_buffer_uptodate(bh_result);
 994                         set_buffer_mapped(bh_result);
 995                         set_buffer_delay(bh_result);
 996                 }
 997         }
 998
 999         if (blocks) {
1000                 bh_result->b_size = (ssize_t)min(
1001                         (loff_t)(iomap.iomap_bsize - iomap.iomap_delta),
1002                         (loff_t)(blocks << inode->i_blkbits));
1003         }
1004
1005         return 0;
1006 }
1007
1008 int
1009 linvfs_get_block(
1010         struct inode            *inode,
1011         sector_t                iblock,
1012         struct buffer_head      *bh_result,
1013         int                     create)
1014 {
1015         return __linvfs_get_block(inode, iblock, 0, bh_result,
1016                                         create, 0, BMAPI_WRITE);
1017 }
1018
1019 STATIC int
1020 linvfs_get_blocks_direct(
1021         struct inode            *inode,
1022         sector_t                iblock,
1023         unsigned long           max_blocks,
1024         struct buffer_head      *bh_result,
1025         int                     create)
1026 {
1027         return __linvfs_get_block(inode, iblock, max_blocks, bh_result,
1028                                         create, 1, BMAPI_WRITE|BMAPI_DIRECT);
1029 }
1030
1031 STATIC ssize_t
1032 linvfs_direct_IO(
1033         int                     rw,
1034         struct kiocb            *iocb,
1035         const struct iovec      *iov,
1036         loff_t                  offset,
1037         unsigned long           nr_segs)
1038 {
1039         struct file     *file = iocb->ki_filp;
1040         struct inode    *inode = file->f_mapping->host;
1041         vnode_t         *vp = LINVFS_GET_VP(inode);
1042         xfs_iomap_t     iomap;
1043         int             maps = 1;
1044         int             error;
1045
1046         VOP_BMAP(vp, offset, 0, BMAPI_DEVICE, &iomap, &maps, error);
1047         if (error)
1048                 return -error;
1049
1050         return blockdev_direct_IO_own_locking(rw, iocb, inode,
1051                 iomap.iomap_target->pbr_bdev,
1052                 iov, offset, nr_segs,
1053                 linvfs_get_blocks_direct,
1054                 linvfs_unwritten_convert_direct);
1055 }
1056
1057
1058 STATIC sector_t
1059 linvfs_bmap(
1060         struct address_space    *mapping,
1061         sector_t                block)
1062 {
1063         struct inode            *inode = (struct inode *)mapping->host;
1064         vnode_t                 *vp = LINVFS_GET_VP(inode);
1065         int                     error;
1066
1067         vn_trace_entry(vp, "linvfs_bmap", (inst_t *)__return_address);
1068
1069         VOP_RWLOCK(vp, VRWLOCK_READ);
1070         VOP_FLUSH_PAGES(vp, (xfs_off_t)0, -1, 0, FI_REMAPF, error);
1071         VOP_RWUNLOCK(vp, VRWLOCK_READ);
1072         return generic_block_bmap(mapping, block, linvfs_get_block);
1073 }
1074
1075 STATIC int
1076 linvfs_readpage(
1077         struct file             *unused,
1078         struct page             *page)
1079 {
1080         return mpage_readpage(page, linvfs_get_block);
1081 }
1082
1083 STATIC int
1084 linvfs_readpages(
1085         struct file             *unused,
1086         struct address_space    *mapping,
1087         struct list_head        *pages,
1088         unsigned                nr_pages)
1089 {
1090         return mpage_readpages(mapping, pages, nr_pages, linvfs_get_block);
1091 }
1092
1093 STATIC void
1094 xfs_count_page_state(
1095         struct page             *page,
1096         int                     *delalloc,
1097         int                     *unmapped,
1098         int                     *unwritten)
1099 {
1100         struct buffer_head      *bh, *head;
1101
1102         *delalloc = *unmapped = *unwritten = 0;
1103
1104         bh = head = page_buffers(page);
1105         do {
1106                 if (buffer_uptodate(bh) && !buffer_mapped(bh))
1107                         (*unmapped) = 1;
1108                 else if (buffer_unwritten(bh) && !buffer_delay(bh))
1109                         clear_buffer_unwritten(bh);
1110                 else if (buffer_unwritten(bh))
1111                         (*unwritten) = 1;
1112                 else if (buffer_delay(bh))
1113                         (*delalloc) = 1;
1114         } while ((bh = bh->b_this_page) != head);
1115 }
1116
1117
1118 /*
1119  * writepage: Called from one of two places:
1120  *
1121  * 1. we are flushing a delalloc buffer head.
1122  *
1123  * 2. we are writing out a dirty page. Typically the page dirty
1124  *    state is cleared before we get here. In this case is it
1125  *    conceivable we have no buffer heads.
1126  *
1127  * For delalloc space on the page we need to allocate space and
1128  * flush it. For unmapped buffer heads on the page we should
1129  * allocate space if the page is uptodate. For any other dirty
1130  * buffer heads on the page we should flush them.
1131  *
1132  * If we detect that a transaction would be required to flush
1133  * the page, we have to check the process flags first, if we
1134  * are already in a transaction or disk I/O during allocations
1135  * is off, we need to fail the writepage and redirty the page.
1136  */
1137
1138 STATIC int
1139 linvfs_writepage(
1140         struct page             *page,
1141         struct writeback_control *wbc)
1142 {
1143         int                     error;
1144         int                     need_trans;
1145         int                     delalloc, unmapped, unwritten;
1146         struct inode            *inode = page->mapping->host;
1147
1148         xfs_page_trace(XFS_WRITEPAGE_ENTER, inode, page, 0);
1149
1150         /*
1151          * We need a transaction if:
1152          *  1. There are delalloc buffers on the page
1153          *  2. The page is uptodate and we have unmapped buffers
1154          *  3. The page is uptodate and we have no buffers
1155          *  4. There are unwritten buffers on the page
1156          */
1157
1158         if (!page_has_buffers(page)) {
1159                 unmapped = 1;
1160                 need_trans = 1;
1161         } else {
1162                 xfs_count_page_state(page, &delalloc, &unmapped, &unwritten);
1163                 if (!PageUptodate(page))
1164                         unmapped = 0;
1165                 need_trans = delalloc + unmapped + unwritten;
1166         }
1167
1168         /*
1169          * If we need a transaction and the process flags say
1170          * we are already in a transaction, or no IO is allowed
1171          * then mark the page dirty again and leave the page
1172          * as is.
1173          */
1174         if (PFLAGS_TEST_FSTRANS() && need_trans)
1175                 goto out_fail;
1176
1177         /*
1178          * Delay hooking up buffer heads until we have
1179          * made our go/no-go decision.
1180          */
1181         if (!page_has_buffers(page))
1182                 create_empty_buffers(page, 1 << inode->i_blkbits, 0);
1183
1184         /*
1185          * Convert delayed allocate, unwritten or unmapped space
1186          * to real space and flush out to disk.
1187          */
1188         error = xfs_page_state_convert(inode, page, wbc, 1, unmapped);
1189         if (error == -EAGAIN)
1190                 goto out_fail;
1191         if (unlikely(error < 0))
1192                 goto out_unlock;
1193
1194         return 0;
1195
1196 out_fail:
1197         redirty_page_for_writepage(wbc, page);
1198         unlock_page(page);
1199         return 0;
1200 out_unlock:
1201         unlock_page(page);
1202         return error;
1203 }
1204
1205 /*
1206  * Called to move a page into cleanable state - and from there
1207  * to be released. Possibly the page is already clean. We always
1208  * have buffer heads in this call.
1209  *
1210  * Returns 0 if the page is ok to release, 1 otherwise.
1211  *
1212  * Possible scenarios are:
1213  *
1214  * 1. We are being called to release a page which has been written
1215  *    to via regular I/O. buffer heads will be dirty and possibly
1216  *    delalloc. If no delalloc buffer heads in this case then we
1217  *    can just return zero.
1218  *
1219  * 2. We are called to release a page which has been written via
1220  *    mmap, all we need to do is ensure there is no delalloc
1221  *    state in the buffer heads, if not we can let the caller
1222  *    free them and we should come back later via writepage.
1223  */
1224 STATIC int
1225 linvfs_release_page(
1226         struct page             *page,
1227         int                     gfp_mask)
1228 {
1229         struct inode            *inode = page->mapping->host;
1230         int                     dirty, delalloc, unmapped, unwritten;
1231         struct writeback_control wbc = {
1232                 .sync_mode = WB_SYNC_ALL,
1233                 .nr_to_write = 1,
1234         };
1235
1236         xfs_page_trace(XFS_RELEASEPAGE_ENTER, inode, page, gfp_mask);
1237
1238         xfs_count_page_state(page, &delalloc, &unmapped, &unwritten);
1239         if (!delalloc && !unwritten)
1240                 goto free_buffers;
1241
1242         if (!(gfp_mask & __GFP_FS))
1243                 return 0;
1244
1245         /* If we are already inside a transaction or the thread cannot
1246          * do I/O, we cannot release this page.
1247          */
1248         if (PFLAGS_TEST_FSTRANS())
1249                 return 0;
1250
1251         /*
1252          * Convert delalloc space to real space, do not flush the
1253          * data out to disk, that will be done by the caller.
1254          * Never need to allocate space here - we will always
1255          * come back to writepage in that case.
1256          */
1257         dirty = xfs_page_state_convert(inode, page, &wbc, 0, 0);
1258         if (dirty == 0 && !unwritten)
1259                 goto free_buffers;
1260         return 0;
1261
1262 free_buffers:
1263         return try_to_free_buffers(page);
1264 }
1265
1266 STATIC int
1267 linvfs_prepare_write(
1268         struct file             *file,
1269         struct page             *page,
1270         unsigned int            from,
1271         unsigned int            to)
1272 {
1273         return block_prepare_write(page, from, to, linvfs_get_block);
1274 }
1275
1276 struct address_space_operations linvfs_aops = {
1277         .readpage               = linvfs_readpage,
1278         .readpages              = linvfs_readpages,
1279         .writepage              = linvfs_writepage,
1280         .sync_page              = block_sync_page,
1281         .releasepage            = linvfs_release_page,
1282         .prepare_write          = linvfs_prepare_write,
1283         .commit_write           = generic_commit_write,
1284         .bmap                   = linvfs_bmap,
1285         .direct_IO              = linvfs_direct_IO,
1286 };