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         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 0;
 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 buffer_head      *bh_arr[],
 553         int                     cnt)
 554 {
 555         struct buffer_head      *bh;
 556         int                     i;
 557
 558         BUG_ON(PageWriteback(page));
 559         set_page_writeback(page);
 560         clear_page_dirty(page);
 561         unlock_page(page);
 562
 563         if (cnt) {
 564                 for (i = 0; i < cnt; i++) {
 565                         bh = bh_arr[i];
 566                         mark_buffer_async_write(bh);
 567                         if (buffer_unwritten(bh))
 568                                 set_buffer_unwritten_io(bh);
 569                         set_buffer_uptodate(bh);
 570                         clear_buffer_dirty(bh);
 571                 }
 572
 573                 for (i = 0; i < cnt; i++)
 574                         submit_bh(WRITE, bh_arr[i]);
 575         } else
 576                 end_page_writeback(page);
 577 }
 578
 579 /*
 580  * Allocate & map buffers for page given the extent map. Write it out.
 581  * except for the original page of a writepage, this is called on
 582  * delalloc/unwritten pages only, for the original page it is possible
 583  * that the page has no mapping at all.
 584  */
 585 STATIC void
 586 xfs_convert_page(
 587         struct inode            *inode,
 588         struct page             *page,
 589         xfs_iomap_t             *iomapp,
 590         struct writeback_control *wbc,
 591         void                    *private,
 592         int                     startio,
 593         int                     all_bh)
 594 {
 595         struct buffer_head      *bh_arr[MAX_BUF_PER_PAGE], *bh, *head;
 596         xfs_iomap_t             *mp = iomapp, *tmp;
 597         unsigned long           end, offset;
 598         pgoff_t                 end_index;
 599         int                     i = 0, index = 0;
 600         int                     bbits = inode->i_blkbits;
 601
 602         end_index = i_size_read(inode) >> PAGE_CACHE_SHIFT;
 603         if (page->index < end_index) {
 604                 end = PAGE_CACHE_SIZE;
 605         } else {
 606                 end = i_size_read(inode) & (PAGE_CACHE_SIZE-1);
 607         }
 608         bh = head = page_buffers(page);
 609         do {
 610                 offset = i << bbits;
 611                 if (!(PageUptodate(page) || buffer_uptodate(bh)))
 612                         continue;
 613                 if (buffer_mapped(bh) && all_bh &&
 614                     !buffer_unwritten(bh) && !buffer_delay(bh)) {
 615                         if (startio && (offset < end)) {
 616                                 lock_buffer(bh);
 617                                 bh_arr[index++] = bh;
 618                         }
 619                         continue;
 620                 }
 621                 tmp = xfs_offset_to_map(page, mp, offset);
 622                 if (!tmp)
 623                         continue;
 624                 ASSERT(!(tmp->iomap_flags & IOMAP_HOLE));
 625                 ASSERT(!(tmp->iomap_flags & IOMAP_DELAY));
 626
 627                 /* If this is a new unwritten extent buffer (i.e. one
 628                  * that we haven't passed in private data for, we must
 629                  * now map this buffer too.
 630                  */
 631                 if (buffer_unwritten(bh) && !bh->b_end_io) {
 632                         ASSERT(tmp->iomap_flags & IOMAP_UNWRITTEN);
 633                         xfs_map_unwritten(inode, page, head, bh, offset,
 634                                         bbits, tmp, wbc, startio, all_bh);
 635                 } else if (! (buffer_unwritten(bh) && buffer_locked(bh))) {
 636                         xfs_map_at_offset(page, bh, offset, bbits, tmp);
 637                         if (buffer_unwritten(bh)) {
 638                                 set_buffer_unwritten_io(bh);
 639                                 bh->b_private = private;
 640                                 ASSERT(private);
 641                         }
 642                 }
 643                 if (startio && (offset < end)) {
 644                         bh_arr[index++] = bh;
 645                 } else {
 646                         set_buffer_dirty(bh);
 647                         unlock_buffer(bh);
 648                         mark_buffer_dirty(bh);
 649                 }
 650         } while (i++, (bh = bh->b_this_page) != head);
 651
 652         if (startio) {
 653                 wbc->nr_to_write--;
 654                 xfs_submit_page(page, bh_arr, index);
 655         } else {
 656                 unlock_page(page);
 657         }
 658 }
 659
 660 /*
 661  * Convert & write out a cluster of pages in the same extent as defined
 662  * by mp and following the start page.
 663  */
 664 STATIC void
 665 xfs_cluster_write(
 666         struct inode            *inode,
 667         pgoff_t                 tindex,
 668         xfs_iomap_t             *iomapp,
 669         struct writeback_control *wbc,
 670         int                     startio,
 671         int                     all_bh,
 672         pgoff_t                 tlast)
 673 {
 674         struct page             *page;
 675
 676         for (; tindex <= tlast; tindex++) {
 677                 page = xfs_probe_delalloc_page(inode, tindex);
 678                 if (!page)
 679                         break;
 680                 xfs_convert_page(inode, page, iomapp, wbc, NULL,
 681                                 startio, all_bh);
 682         }
 683 }
 684
 685 /*
 686  * Calling this without startio set means we are being asked to make a dirty
 687  * page ready for freeing it's buffers.  When called with startio set then
 688  * we are coming from writepage.
 689  *
 690  * When called with startio set it is important that we write the WHOLE
 691  * page if possible.
 692  * The bh->b_state's cannot know if any of the blocks or which block for
 693  * that matter are dirty due to mmap writes, and therefore bh uptodate is
 694  * only vaild if the page itself isn't completely uptodate.  Some layers
 695  * may clear the page dirty flag prior to calling write page, under the
 696  * assumption the entire page will be written out; by not writing out the
 697  * whole page the page can be reused before all valid dirty data is
 698  * written out.  Note: in the case of a page that has been dirty'd by
 699  * mapwrite and but partially setup by block_prepare_write the
 700  * bh->b_states's will not agree and only ones setup by BPW/BCW will have
 701  * valid state, thus the whole page must be written out thing.
 702  */
 703
 704 STATIC int
 705 xfs_page_state_convert(
 706         struct inode    *inode,
 707         struct page     *page,
 708         struct writeback_control *wbc,
 709         int             startio,
 710         int             unmapped) /* also implies page uptodate */
 711 {
 712         struct buffer_head      *bh_arr[MAX_BUF_PER_PAGE], *bh, *head;
 713         xfs_iomap_t             *iomp, iomap;
 714         loff_t                  offset;
 715         unsigned long           p_offset = 0;
 716         __uint64_t              end_offset;
 717         pgoff_t                 end_index, last_index, tlast;
 718         int                     len, err, i, cnt = 0, uptodate = 1;
 719         int                     flags = startio ? 0 : BMAPI_TRYLOCK;
 720         int                     page_dirty = 1;
 721         int                     delalloc = 0;
 722
 723
 724         /* Are we off the end of the file ? */
 725         offset = i_size_read(inode);
 726         end_index = offset >> PAGE_CACHE_SHIFT;
 727         last_index = (offset - 1) >> PAGE_CACHE_SHIFT;
 728         if (page->index >= end_index) {
 729                 if ((page->index >= end_index + 1) ||
 730                     !(i_size_read(inode) & (PAGE_CACHE_SIZE - 1))) {
 731                         err = -EIO;
 732                         goto error;
 733                 }
 734         }
 735
 736         offset = (loff_t)page->index << PAGE_CACHE_SHIFT;
 737         end_offset = min_t(unsigned long long,
 738                         offset + PAGE_CACHE_SIZE, i_size_read(inode));
 739
 740         bh = head = page_buffers(page);
 741         iomp = NULL;
 742
 743         len = bh->b_size;
 744         do {
 745                 if (offset >= end_offset)
 746                         break;
 747                 if (!buffer_uptodate(bh))
 748                         uptodate = 0;
 749                 if (!(PageUptodate(page) || buffer_uptodate(bh)) && !startio)
 750                         continue;
 751
 752                 if (iomp) {
 753                         iomp = xfs_offset_to_map(page, &iomap, p_offset);
 754                 }
 755
 756                 /*
 757                  * First case, map an unwritten extent and prepare for
 758                  * extent state conversion transaction on completion.
 759                  */
 760                 if (buffer_unwritten(bh)) {
 761                         if (!startio)
 762                                 continue;
 763                         if (!iomp) {
 764                                 err = xfs_map_blocks(inode, offset, len, &iomap,
 765                                                 BMAPI_READ|BMAPI_IGNSTATE);
 766                                 if (err) {
 767                                         goto error;
 768                                 }
 769                                 iomp = xfs_offset_to_map(page, &iomap,
 770                                                                 p_offset);
 771                         }
 772                         if (iomp) {
 773                                 if (!bh->b_end_io) {
 774                                         err = xfs_map_unwritten(inode, page,
 775                                                         head, bh, p_offset,
 776                                                         inode->i_blkbits, iomp,
 777                                                         wbc, startio, unmapped);
 778                                         if (err) {
 779                                                 goto error;
 780                                         }
 781                                 } else {
 782                                         set_bit(BH_Lock, &bh->b_state);
 783                                 }
 784                                 BUG_ON(!buffer_locked(bh));
 785                                 bh_arr[cnt++] = bh;
 786                                 page_dirty = 0;
 787                         }
 788                 /*
 789                  * Second case, allocate space for a delalloc buffer.
 790                  * We can return EAGAIN here in the release page case.
 791                  */
 792                 } else if (buffer_delay(bh)) {
 793                         if (!iomp) {
 794                                 delalloc = 1;
 795                                 err = xfs_map_blocks(inode, offset, len, &iomap,
 796                                                 BMAPI_ALLOCATE | flags);
 797                                 if (err) {
 798                                         goto error;
 799                                 }
 800                                 iomp = xfs_offset_to_map(page, &iomap,
 801                                                                 p_offset);
 802                         }
 803                         if (iomp) {
 804                                 xfs_map_at_offset(page, bh, p_offset,
 805                                                 inode->i_blkbits, iomp);
 806                                 if (startio) {
 807                                         bh_arr[cnt++] = bh;
 808                                 } else {
 809                                         set_buffer_dirty(bh);
 810                                         unlock_buffer(bh);
 811                                         mark_buffer_dirty(bh);
 812                                 }
 813                                 page_dirty = 0;
 814                         }
 815                 } else if ((buffer_uptodate(bh) || PageUptodate(page)) &&
 816                            (unmapped || startio)) {
 817
 818                         if (!buffer_mapped(bh)) {
 819                                 int     size;
 820
 821                                 /*
 822                                  * Getting here implies an unmapped buffer
 823                                  * was found, and we are in a path where we
 824                                  * need to write the whole page out.
 825                                  */
 826                                 if (!iomp) {
 827                                         size = xfs_probe_unmapped_cluster(
 828                                                         inode, page, bh, head);
 829                                         err = xfs_map_blocks(inode, offset,
 830                                                         size, &iomap,
 831                                                         BMAPI_WRITE|BMAPI_MMAP);
 832                                         if (err) {
 833                                                 goto error;
 834                                         }
 835                                         iomp = xfs_offset_to_map(page, &iomap,
 836                                                                      p_offset);
 837                                 }
 838                                 if (iomp) {
 839                                         xfs_map_at_offset(page,
 840                                                         bh, p_offset,
 841                                                         inode->i_blkbits, iomp);
 842                                         if (startio) {
 843                                                 bh_arr[cnt++] = bh;
 844                                         } else {
 845                                                 set_buffer_dirty(bh);
 846                                                 unlock_buffer(bh);
 847                                                 mark_buffer_dirty(bh);
 848                                         }
 849                                         page_dirty = 0;
 850                                 }
 851                         } else if (startio) {
 852                                 if (buffer_uptodate(bh) &&
 853                                     !test_and_set_bit(BH_Lock, &bh->b_state)) {
 854                                         bh_arr[cnt++] = bh;
 855                                         page_dirty = 0;
 856                                 }
 857                         }
 858                 }
 859         } while (offset += len, p_offset += len,
 860                 ((bh = bh->b_this_page) != head));
 861
 862         if (uptodate && bh == head)
 863                 SetPageUptodate(page);
 864
 865         if (startio)
 866                 xfs_submit_page(page, bh_arr, cnt);
 867
 868         if (iomp) {
 869                 tlast = (iomp->iomap_offset + iomp->iomap_bsize - 1) >>
 870                                         PAGE_CACHE_SHIFT;
 871                 if (delalloc && (tlast > last_index))
 872                         tlast = last_index;
 873                 xfs_cluster_write(inode, page->index + 1, iomp, wbc,
 874                                         startio, unmapped, tlast);
 875         }
 876
 877         return page_dirty;
 878
 879 error:
 880         for (i = 0; i < cnt; i++) {
 881                 unlock_buffer(bh_arr[i]);
 882         }
 883
 884         /*
 885          * If it's delalloc and we have nowhere to put it,
 886          * throw it away, unless the lower layers told
 887          * us to try again.
 888          */
 889         if (err != -EAGAIN) {
 890                 if (!unmapped) {
 891                         block_invalidatepage(page, 0);
 892                 }
 893                 ClearPageUptodate(page);
 894         }
 895         return err;
 896 }
 897
 898 STATIC int
 899 linvfs_get_block_core(
 900         struct inode            *inode,
 901         sector_t                iblock,
 902         unsigned long           blocks,
 903         struct buffer_head      *bh_result,
 904         int                     create,
 905         int                     direct,
 906         bmapi_flags_t           flags)
 907 {
 908         vnode_t                 *vp = LINVFS_GET_VP(inode);
 909         xfs_iomap_t             iomap;
 910         int                     retpbbm = 1;
 911         int                     error;
 912         ssize_t                 size;
 913         loff_t                  offset = (loff_t)iblock << inode->i_blkbits;
 914
 915         if (blocks)
 916                 size = blocks << inode->i_blkbits;
 917         else
 918                 size = 1 << inode->i_blkbits;
 919
 920         VOP_BMAP(vp, offset, size,
 921                 create ? flags : BMAPI_READ, &iomap, &retpbbm, error);
 922         if (error)
 923                 return -error;
 924
 925         if (retpbbm == 0)
 926                 return 0;
 927
 928         if (iomap.iomap_bn != IOMAP_DADDR_NULL) {
 929                 xfs_daddr_t             bn;
 930                 loff_t                  delta;
 931
 932                 /* For unwritten extents do not report a disk address on
 933                  * the read case (treat as if we're reading into a hole).
 934                  */
 935                 if (create || !(iomap.iomap_flags & IOMAP_UNWRITTEN)) {
 936                         delta = offset - iomap.iomap_offset;
 937                         delta >>= inode->i_blkbits;
 938
 939                         bn = iomap.iomap_bn >> (inode->i_blkbits - BBSHIFT);
 940                         bn += delta;
 941
 942                         bh_result->b_blocknr = bn;
 943                         bh_result->b_bdev = iomap.iomap_target->pbr_bdev;
 944                         set_buffer_mapped(bh_result);
 945                 }
 946                 if (create && (iomap.iomap_flags & IOMAP_UNWRITTEN)) {
 947                         if (direct)
 948                                 bh_result->b_private = inode;
 949                         set_buffer_unwritten(bh_result);
 950                         set_buffer_delay(bh_result);
 951                 }
 952         }
 953
 954         /* If this is a realtime file, data might be on a new device */
 955         bh_result->b_bdev = iomap.iomap_target->pbr_bdev;
 956
 957         /* If we previously allocated a block out beyond eof and
 958          * we are now coming back to use it then we will need to
 959          * flag it as new even if it has a disk address.
 960          */
 961         if (create &&
 962             ((!buffer_mapped(bh_result) && !buffer_uptodate(bh_result)) ||
 963              (offset >= i_size_read(inode)) || (iomap.iomap_flags & IOMAP_NEW))) {
 964                 set_buffer_new(bh_result);
 965         }
 966
 967         if (iomap.iomap_flags & IOMAP_DELAY) {
 968                 if (unlikely(direct))
 969                         BUG();
 970                 if (create) {
 971                         set_buffer_mapped(bh_result);
 972                         set_buffer_uptodate(bh_result);
 973                 }
 974                 bh_result->b_bdev = iomap.iomap_target->pbr_bdev;
 975                 set_buffer_delay(bh_result);
 976         }
 977
 978         if (blocks) {
 979                 loff_t iosize;
 980                 iosize = (iomap.iomap_bsize - iomap.iomap_delta);
 981                 bh_result->b_size =
 982                     (ssize_t)min(iosize, (loff_t)(blocks << inode->i_blkbits));
 983         }
 984
 985         return 0;
 986 }
 987
 988 int
 989 linvfs_get_block(
 990         struct inode            *inode,
 991         sector_t                iblock,
 992         struct buffer_head      *bh_result,
 993         int                     create)
 994 {
 995         return linvfs_get_block_core(inode, iblock, 0, bh_result,
 996                                         create, 0, BMAPI_WRITE);
 997 }
 998
 999 STATIC int
1000 linvfs_get_block_sync(
1001         struct inode            *inode,
1002         sector_t                iblock,
1003         struct buffer_head      *bh_result,
1004         int                     create)
1005 {
1006         return linvfs_get_block_core(inode, iblock, 0, bh_result,
1007                                         create, 0, BMAPI_SYNC|BMAPI_WRITE);
1008 }
1009
1010 STATIC int
1011 linvfs_get_blocks_direct(
1012         struct inode            *inode,
1013         sector_t                iblock,
1014         unsigned long           max_blocks,
1015         struct buffer_head      *bh_result,
1016         int                     create)
1017 {
1018         return linvfs_get_block_core(inode, iblock, max_blocks, bh_result,
1019                                         create, 1, BMAPI_WRITE|BMAPI_DIRECT);
1020 }
1021
1022 STATIC ssize_t
1023 linvfs_direct_IO(
1024         int                     rw,
1025         struct kiocb            *iocb,
1026         const struct iovec      *iov,
1027         loff_t                  offset,
1028         unsigned long           nr_segs)
1029 {
1030         struct file     *file = iocb->ki_filp;
1031         struct inode    *inode = file->f_mapping->host;
1032         vnode_t         *vp = LINVFS_GET_VP(inode);
1033         xfs_iomap_t     iomap;
1034         int             maps = 1;
1035         int             error;
1036
1037         VOP_BMAP(vp, offset, 0, BMAPI_DEVICE, &iomap, &maps, error);
1038         if (error)
1039                 return -error;
1040
1041         return blockdev_direct_IO_no_locking(rw, iocb, inode,
1042                 iomap.iomap_target->pbr_bdev,
1043                 iov, offset, nr_segs,
1044                 linvfs_get_blocks_direct,
1045                 linvfs_unwritten_convert_direct);
1046 }
1047
1048
1049 STATIC sector_t
1050 linvfs_bmap(
1051         struct address_space    *mapping,
1052         sector_t                block)
1053 {
1054         struct inode            *inode = (struct inode *)mapping->host;
1055         vnode_t                 *vp = LINVFS_GET_VP(inode);
1056         int                     error;
1057
1058         vn_trace_entry(vp, "linvfs_bmap", (inst_t *)__return_address);
1059
1060         VOP_RWLOCK(vp, VRWLOCK_READ);
1061         VOP_FLUSH_PAGES(vp, (xfs_off_t)0, -1, 0, FI_REMAPF, error);
1062         VOP_RWUNLOCK(vp, VRWLOCK_READ);
1063         return generic_block_bmap(mapping, block, linvfs_get_block);
1064 }
1065
1066 STATIC int
1067 linvfs_readpage(
1068         struct file             *unused,
1069         struct page             *page)
1070 {
1071         return mpage_readpage(page, linvfs_get_block);
1072 }
1073
1074 STATIC int
1075 linvfs_readpages(
1076         struct file             *unused,
1077         struct address_space    *mapping,
1078         struct list_head        *pages,
1079         unsigned                nr_pages)
1080 {
1081         return mpage_readpages(mapping, pages, nr_pages, linvfs_get_block);
1082 }
1083
1084 STATIC void
1085 xfs_count_page_state(
1086         struct page             *page,
1087         int                     *delalloc,
1088         int                     *unmapped,
1089         int                     *unwritten)
1090 {
1091         struct buffer_head      *bh, *head;
1092
1093         *delalloc = *unmapped = *unwritten = 0;
1094
1095         bh = head = page_buffers(page);
1096         do {
1097                 if (buffer_uptodate(bh) && !buffer_mapped(bh))
1098                         (*unmapped) = 1;
1099                 else if (buffer_unwritten(bh) && !buffer_delay(bh))
1100                         clear_buffer_unwritten(bh);
1101                 else if (buffer_unwritten(bh))
1102                         (*unwritten) = 1;
1103                 else if (buffer_delay(bh))
1104                         (*delalloc) = 1;
1105         } while ((bh = bh->b_this_page) != head);
1106 }
1107
1108
1109 /*
1110  * writepage: Called from one of two places:
1111  *
1112  * 1. we are flushing a delalloc buffer head.
1113  *
1114  * 2. we are writing out a dirty page. Typically the page dirty
1115  *    state is cleared before we get here. In this case is it
1116  *    conceivable we have no buffer heads.
1117  *
1118  * For delalloc space on the page we need to allocate space and
1119  * flush it. For unmapped buffer heads on the page we should
1120  * allocate space if the page is uptodate. For any other dirty
1121  * buffer heads on the page we should flush them.
1122  *
1123  * If we detect that a transaction would be required to flush
1124  * the page, we have to check the process flags first, if we
1125  * are already in a transaction or disk I/O during allocations
1126  * is off, we need to fail the writepage and redirty the page.
1127  */
1128
1129 STATIC int
1130 linvfs_writepage(
1131         struct page             *page,
1132         struct writeback_control *wbc)
1133 {
1134         int                     error;
1135         int                     need_trans;
1136         int                     delalloc, unmapped, unwritten;
1137         struct inode            *inode = page->mapping->host;
1138
1139         xfs_page_trace(XFS_WRITEPAGE_ENTER, inode, page, 0);
1140
1141         /*
1142          * We need a transaction if:
1143          *  1. There are delalloc buffers on the page
1144          *  2. The page is uptodate and we have unmapped buffers
1145          *  3. The page is uptodate and we have no buffers
1146          *  4. There are unwritten buffers on the page
1147          */
1148
1149         if (!page_has_buffers(page)) {
1150                 unmapped = 1;
1151                 need_trans = 1;
1152         } else {
1153                 xfs_count_page_state(page, &delalloc, &unmapped, &unwritten);
1154                 if (!PageUptodate(page))
1155                         unmapped = 0;
1156                 need_trans = delalloc + unmapped + unwritten;
1157         }
1158
1159         /*
1160          * If we need a transaction and the process flags say
1161          * we are already in a transaction, or no IO is allowed
1162          * then mark the page dirty again and leave the page
1163          * as is.
1164          */
1165         if (PFLAGS_TEST_FSTRANS() && need_trans)
1166                 goto out_fail;
1167
1168         /*
1169          * Delay hooking up buffer heads until we have
1170          * made our go/no-go decision.
1171          */
1172         if (!page_has_buffers(page))
1173                 create_empty_buffers(page, 1 << inode->i_blkbits, 0);
1174
1175         /*
1176          * Convert delayed allocate, unwritten or unmapped space
1177          * to real space and flush out to disk.
1178          */
1179         error = xfs_page_state_convert(inode, page, wbc, 1, unmapped);
1180         if (error == -EAGAIN)
1181                 goto out_fail;
1182         if (unlikely(error < 0))
1183                 goto out_unlock;
1184
1185         return 0;
1186
1187 out_fail:
1188         set_page_dirty(page);
1189         unlock_page(page);
1190         return 0;
1191 out_unlock:
1192         unlock_page(page);
1193         return error;
1194 }
1195
1196 /*
1197  * Called to move a page into cleanable state - and from there
1198  * to be released. Possibly the page is already clean. We always
1199  * have buffer heads in this call.
1200  *
1201  * Returns 0 if the page is ok to release, 1 otherwise.
1202  *
1203  * Possible scenarios are:
1204  *
1205  * 1. We are being called to release a page which has been written
1206  *    to via regular I/O. buffer heads will be dirty and possibly
1207  *    delalloc. If no delalloc buffer heads in this case then we
1208  *    can just return zero.
1209  *
1210  * 2. We are called to release a page which has been written via
1211  *    mmap, all we need to do is ensure there is no delalloc
1212  *    state in the buffer heads, if not we can let the caller
1213  *    free them and we should come back later via writepage.
1214  */
1215 STATIC int
1216 linvfs_release_page(
1217         struct page             *page,
1218         int                     gfp_mask)
1219 {
1220         struct inode            *inode = page->mapping->host;
1221         int                     dirty, delalloc, unmapped, unwritten;
1222         struct writeback_control wbc = {
1223                 .sync_mode = WB_SYNC_ALL,
1224                 .nr_to_write = 1,
1225         };
1226
1227         xfs_page_trace(XFS_RELEASEPAGE_ENTER, inode, page, gfp_mask);
1228
1229         xfs_count_page_state(page, &delalloc, &unmapped, &unwritten);
1230         if (!delalloc && !unwritten)
1231                 goto free_buffers;
1232
1233         if (!(gfp_mask & __GFP_FS))
1234                 return 0;
1235
1236         /* If we are already inside a transaction or the thread cannot
1237          * do I/O, we cannot release this page.
1238          */
1239         if (PFLAGS_TEST_FSTRANS())
1240                 return 0;
1241
1242         /*
1243          * Convert delalloc space to real space, do not flush the
1244          * data out to disk, that will be done by the caller.
1245          * Never need to allocate space here - we will always
1246          * come back to writepage in that case.
1247          */
1248         dirty = xfs_page_state_convert(inode, page, &wbc, 0, 0);
1249         if (dirty == 0 && !unwritten)
1250                 goto free_buffers;
1251         return 0;
1252
1253 free_buffers:
1254         return try_to_free_buffers(page);
1255 }
1256
1257 STATIC int
1258 linvfs_prepare_write(
1259         struct file             *file,
1260         struct page             *page,
1261         unsigned int            from,
1262         unsigned int            to)
1263 {
1264         if (file && (file->f_flags & O_SYNC)) {
1265                 return block_prepare_write(page, from, to,
1266                                                 linvfs_get_block_sync);
1267         } else {
1268                 return block_prepare_write(page, from, to,
1269                                                 linvfs_get_block);
1270         }
1271 }
1272
1273 struct address_space_operations linvfs_aops = {
1274         .readpage               = linvfs_readpage,
1275         .readpages              = linvfs_readpages,
1276         .writepage              = linvfs_writepage,
1277         .sync_page              = block_sync_page,
1278         .releasepage            = linvfs_release_page,
1279         .prepare_write          = linvfs_prepare_write,
1280         .commit_write           = generic_commit_write,
1281         .bmap                   = linvfs_bmap,
1282         .direct_IO              = linvfs_direct_IO,
1283 };