SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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 /*
 108  * Schedule IO completion handling on a xfsdatad if this was
 109  * the final hold on this ioend.
 110  */
 111 STATIC void
 112 xfs_finish_ioend(
 113         xfs_ioend_t             *ioend)
 114 {
 115         if (atomic_dec_and_test(&ioend->io_remaining))
 116                 queue_work(xfsdatad_workqueue, &ioend->io_work);
 117 }
 118
 119 STATIC void
 120 xfs_destroy_ioend(
 121         xfs_ioend_t             *ioend)
 122 {
 123         vn_iowake(ioend->io_vnode);
 124         mempool_free(ioend, xfs_ioend_pool);
 125 }
 126
 127 /*
 128  * Issue transactions to convert a buffer range from unwritten
 129  * to written extents.
 130  */
 131 STATIC void
 132 xfs_end_bio_unwritten(
 133         void                    *data)
 134 {
 135         xfs_ioend_t             *ioend = data;
 136         vnode_t                 *vp = ioend->io_vnode;
 137         xfs_off_t               offset = ioend->io_offset;
 138         size_t                  size = ioend->io_size;
 139         int                     error;
 140
 141         if (ioend->io_uptodate)
 142                 VOP_BMAP(vp, offset, size, BMAPI_UNWRITTEN, NULL, NULL, error);
 143         xfs_destroy_ioend(ioend);
 144 }
 145
 146 /*
 147  * Allocate and initialise an IO completion structure.
 148  * We need to track unwritten extent write completion here initially.
 149  * We'll need to extend this for updating the ondisk inode size later
 150  * (vs. incore size).
 151  */
 152 STATIC xfs_ioend_t *
 153 xfs_alloc_ioend(
 154         struct inode            *inode)
 155 {
 156         xfs_ioend_t             *ioend;
 157
 158         ioend = mempool_alloc(xfs_ioend_pool, GFP_NOFS);
 159
 160         /*
 161          * Set the count to 1 initially, which will prevent an I/O
 162          * completion callback from happening before we have started
 163          * all the I/O from calling the completion routine too early.
 164          */
 165         atomic_set(&ioend->io_remaining, 1);
 166         ioend->io_uptodate = 1; /* cleared if any I/O fails */
 167         ioend->io_vnode = LINVFS_GET_VP(inode);
 168         atomic_inc(&ioend->io_vnode->v_iocount);
 169         ioend->io_offset = 0;
 170         ioend->io_size = 0;
 171
 172         INIT_WORK(&ioend->io_work, xfs_end_bio_unwritten, ioend);
 173
 174         return ioend;
 175 }
 176
 177 void
 178 linvfs_unwritten_done(
 179         struct buffer_head      *bh,
 180         int                     uptodate)
 181 {
 182         xfs_ioend_t             *ioend = bh->b_private;
 183
 184         ASSERT(buffer_unwritten(bh));
 185         bh->b_end_io = NULL;
 186         clear_buffer_unwritten(bh);
 187         if (!uptodate)
 188                 ioend->io_uptodate = 0;
 189
 190         xfs_finish_ioend(ioend);
 191         end_buffer_async_write(bh, uptodate);
 192 }
 193
 194 STATIC int
 195 xfs_map_blocks(
 196         struct inode            *inode,
 197         loff_t                  offset,
 198         ssize_t                 count,
 199         xfs_iomap_t             *mapp,
 200         int                     flags)
 201 {
 202         vnode_t                 *vp = LINVFS_GET_VP(inode);
 203         int                     error, nmaps = 1;
 204
 205         VOP_BMAP(vp, offset, count, flags, mapp, &nmaps, error);
 206         if (!error && (flags & (BMAPI_WRITE|BMAPI_ALLOCATE)))
 207                 VMODIFY(vp);
 208         return -error;
 209 }
 210
 211 /*
 212  * Finds the corresponding mapping in block @map array of the
 213  * given @offset within a @page.
 214  */
 215 STATIC xfs_iomap_t *
 216 xfs_offset_to_map(
 217         struct page             *page,
 218         xfs_iomap_t             *iomapp,
 219         unsigned long           offset)
 220 {
 221         loff_t                  full_offset;    /* offset from start of file */
 222
 223         ASSERT(offset < PAGE_CACHE_SIZE);
 224
 225         full_offset = page->index;              /* NB: using 64bit number */
 226         full_offset <<= PAGE_CACHE_SHIFT;       /* offset from file start */
 227         full_offset += offset;                  /* offset from page start */
 228
 229         if (full_offset < iomapp->iomap_offset)
 230                 return NULL;
 231         if (iomapp->iomap_offset + (iomapp->iomap_bsize -1) >= full_offset)
 232                 return iomapp;
 233         return NULL;
 234 }
 235
 236 STATIC void
 237 xfs_map_at_offset(
 238         struct page             *page,
 239         struct buffer_head      *bh,
 240         unsigned long           offset,
 241         int                     block_bits,
 242         xfs_iomap_t             *iomapp)
 243 {
 244         xfs_daddr_t             bn;
 245         loff_t                  delta;
 246         int                     sector_shift;
 247
 248         ASSERT(!(iomapp->iomap_flags & IOMAP_HOLE));
 249         ASSERT(!(iomapp->iomap_flags & IOMAP_DELAY));
 250         ASSERT(iomapp->iomap_bn != IOMAP_DADDR_NULL);
 251
 252         delta = page->index;
 253         delta <<= PAGE_CACHE_SHIFT;
 254         delta += offset;
 255         delta -= iomapp->iomap_offset;
 256         delta >>= block_bits;
 257
 258         sector_shift = block_bits - BBSHIFT;
 259         bn = iomapp->iomap_bn >> sector_shift;
 260         bn += delta;
 261         BUG_ON(!bn && !(iomapp->iomap_flags & IOMAP_REALTIME));
 262         ASSERT((bn << sector_shift) >= iomapp->iomap_bn);
 263
 264         lock_buffer(bh);
 265         bh->b_blocknr = bn;
 266         bh->b_bdev = iomapp->iomap_target->pbr_bdev;
 267         set_buffer_mapped(bh);
 268         clear_buffer_delay(bh);
 269 }
 270
 271 /*
 272  * Look for a page at index which is unlocked and contains our
 273  * unwritten extent flagged buffers at its head.  Returns page
 274  * locked and with an extra reference count, and length of the
 275  * unwritten extent component on this page that we can write,
 276  * in units of filesystem blocks.
 277  */
 278 STATIC struct page *
 279 xfs_probe_unwritten_page(
 280         struct address_space    *mapping,
 281         pgoff_t                 index,
 282         xfs_iomap_t             *iomapp,
 283         xfs_ioend_t             *ioend,
 284         unsigned long           max_offset,
 285         unsigned long           *fsbs,
 286         unsigned int            bbits)
 287 {
 288         struct page             *page;
 289
 290         page = find_trylock_page(mapping, index);
 291         if (!page)
 292                 return NULL;
 293         if (PageWriteback(page))
 294                 goto out;
 295
 296         if (page->mapping && page_has_buffers(page)) {
 297                 struct buffer_head      *bh, *head;
 298                 unsigned long           p_offset = 0;
 299
 300                 *fsbs = 0;
 301                 bh = head = page_buffers(page);
 302                 do {
 303                         if (!buffer_unwritten(bh) || !buffer_uptodate(bh))
 304                                 break;
 305                         if (!xfs_offset_to_map(page, iomapp, p_offset))
 306                                 break;
 307                         if (p_offset >= max_offset)
 308                                 break;
 309                         xfs_map_at_offset(page, bh, p_offset, bbits, iomapp);
 310                         set_buffer_unwritten_io(bh);
 311                         bh->b_private = ioend;
 312                         p_offset += bh->b_size;
 313                         (*fsbs)++;
 314                 } while ((bh = bh->b_this_page) != head);
 315
 316                 if (p_offset)
 317                         return page;
 318         }
 319
 320 out:
 321         unlock_page(page);
 322         return NULL;
 323 }
 324
 325 /*
 326  * Look for a page at index which is unlocked and not mapped
 327  * yet - clustering for mmap write case.
 328  */
 329 STATIC unsigned int
 330 xfs_probe_unmapped_page(
 331         struct address_space    *mapping,
 332         pgoff_t                 index,
 333         unsigned int            pg_offset)
 334 {
 335         struct page             *page;
 336         int                     ret = 0;
 337
 338         page = find_trylock_page(mapping, index);
 339         if (!page)
 340                 return 0;
 341         if (PageWriteback(page))
 342                 goto out;
 343
 344         if (page->mapping && PageDirty(page)) {
 345                 if (page_has_buffers(page)) {
 346                         struct buffer_head      *bh, *head;
 347
 348                         bh = head = page_buffers(page);
 349                         do {
 350                                 if (buffer_mapped(bh) || !buffer_uptodate(bh))
 351                                         break;
 352                                 ret += bh->b_size;
 353                                 if (ret >= pg_offset)
 354                                         break;
 355                         } while ((bh = bh->b_this_page) != head);
 356                 } else
 357                         ret = PAGE_CACHE_SIZE;
 358         }
 359
 360 out:
 361         unlock_page(page);
 362         return ret;
 363 }
 364
 365 STATIC unsigned int
 366 xfs_probe_unmapped_cluster(
 367         struct inode            *inode,
 368         struct page             *startpage,
 369         struct buffer_head      *bh,
 370         struct buffer_head      *head)
 371 {
 372         pgoff_t                 tindex, tlast, tloff;
 373         unsigned int            pg_offset, len, total = 0;
 374         struct address_space    *mapping = inode->i_mapping;
 375
 376         /* First sum forwards in this page */
 377         do {
 378                 if (buffer_mapped(bh))
 379                         break;
 380                 total += bh->b_size;
 381         } while ((bh = bh->b_this_page) != head);
 382
 383         /* If we reached the end of the page, sum forwards in
 384          * following pages.
 385          */
 386         if (bh == head) {
 387                 tlast = i_size_read(inode) >> PAGE_CACHE_SHIFT;
 388                 /* Prune this back to avoid pathological behavior */
 389                 tloff = min(tlast, startpage->index + 64);
 390                 for (tindex = startpage->index + 1; tindex < tloff; tindex++) {
 391                         len = xfs_probe_unmapped_page(mapping, tindex,
 392                                                         PAGE_CACHE_SIZE);
 393                         if (!len)
 394                                 return total;
 395                         total += len;
 396                 }
 397                 if (tindex == tlast &&
 398                     (pg_offset = i_size_read(inode) & (PAGE_CACHE_SIZE - 1))) {
 399                         total += xfs_probe_unmapped_page(mapping,
 400                                                         tindex, pg_offset);
 401                 }
 402         }
 403         return total;
 404 }
 405
 406 /*
 407  * Probe for a given page (index) in the inode and test if it is delayed
 408  * and without unwritten buffers.  Returns page locked and with an extra
 409  * reference count.
 410  */
 411 STATIC struct page *
 412 xfs_probe_delalloc_page(
 413         struct inode            *inode,
 414         pgoff_t                 index)
 415 {
 416         struct page             *page;
 417
 418         page = find_trylock_page(inode->i_mapping, index);
 419         if (!page)
 420                 return NULL;
 421         if (PageWriteback(page))
 422                 goto out;
 423
 424         if (page->mapping && page_has_buffers(page)) {
 425                 struct buffer_head      *bh, *head;
 426                 int                     acceptable = 0;
 427
 428                 bh = head = page_buffers(page);
 429                 do {
 430                         if (buffer_unwritten(bh)) {
 431                                 acceptable = 0;
 432                                 break;
 433                         } else if (buffer_delay(bh)) {
 434                                 acceptable = 1;
 435                         }
 436                 } while ((bh = bh->b_this_page) != head);
 437
 438                 if (acceptable)
 439                         return page;
 440         }
 441
 442 out:
 443         unlock_page(page);
 444         return NULL;
 445 }
 446
 447 STATIC int
 448 xfs_map_unwritten(
 449         struct inode            *inode,
 450         struct page             *start_page,
 451         struct buffer_head      *head,
 452         struct buffer_head      *curr,
 453         unsigned long           p_offset,
 454         int                     block_bits,
 455         xfs_iomap_t             *iomapp,
 456         struct writeback_control *wbc,
 457         int                     startio,
 458         int                     all_bh)
 459 {
 460         struct buffer_head      *bh = curr;
 461         xfs_iomap_t             *tmp;
 462         xfs_ioend_t             *ioend;
 463         loff_t                  offset;
 464         unsigned long           nblocks = 0;
 465
 466         offset = start_page->index;
 467         offset <<= PAGE_CACHE_SHIFT;
 468         offset += p_offset;
 469
 470         ioend = xfs_alloc_ioend(inode);
 471
 472         /* First map forwards in the page consecutive buffers
 473          * covering this unwritten extent
 474          */
 475         do {
 476                 if (!buffer_unwritten(bh))
 477                         break;
 478                 tmp = xfs_offset_to_map(start_page, iomapp, p_offset);
 479                 if (!tmp)
 480                         break;
 481                 xfs_map_at_offset(start_page, bh, p_offset, block_bits, iomapp);
 482                 set_buffer_unwritten_io(bh);
 483                 bh->b_private = ioend;
 484                 p_offset += bh->b_size;
 485                 nblocks++;
 486         } while ((bh = bh->b_this_page) != head);
 487
 488         atomic_add(nblocks, &ioend->io_remaining);
 489
 490         /* If we reached the end of the page, map forwards in any
 491          * following pages which are also covered by this extent.
 492          */
 493         if (bh == head) {
 494                 struct address_space    *mapping = inode->i_mapping;
 495                 pgoff_t                 tindex, tloff, tlast;
 496                 unsigned long           bs;
 497                 unsigned int            pg_offset, bbits = inode->i_blkbits;
 498                 struct page             *page;
 499
 500                 tlast = i_size_read(inode) >> PAGE_CACHE_SHIFT;
 501                 tloff = (iomapp->iomap_offset + iomapp->iomap_bsize) >> PAGE_CACHE_SHIFT;
 502                 tloff = min(tlast, tloff);
 503                 for (tindex = start_page->index + 1; tindex < tloff; tindex++) {
 504                         page = xfs_probe_unwritten_page(mapping,
 505                                                 tindex, iomapp, ioend,
 506                                                 PAGE_CACHE_SIZE, &bs, bbits);
 507                         if (!page)
 508                                 break;
 509                         nblocks += bs;
 510                         atomic_add(bs, &ioend->io_remaining);
 511                         xfs_convert_page(inode, page, iomapp, wbc, ioend,
 512                                                         startio, all_bh);
 513                         /* stop if converting the next page might add
 514                          * enough blocks that the corresponding byte
 515                          * count won't fit in our ulong page buf length */
 516                         if (nblocks >= ((ULONG_MAX - PAGE_SIZE) >> block_bits))
 517                                 goto enough;
 518                 }
 519
 520                 if (tindex == tlast &&
 521                     (pg_offset = (i_size_read(inode) & (PAGE_CACHE_SIZE - 1)))) {
 522                         page = xfs_probe_unwritten_page(mapping,
 523                                                         tindex, iomapp, ioend,
 524                                                         pg_offset, &bs, bbits);
 525                         if (page) {
 526                                 nblocks += bs;
 527                                 atomic_add(bs, &ioend->io_remaining);
 528                                 xfs_convert_page(inode, page, iomapp, wbc, ioend,
 529                                                         startio, all_bh);
 530                                 if (nblocks >= ((ULONG_MAX - PAGE_SIZE) >> block_bits))
 531                                         goto enough;
 532                         }
 533                 }
 534         }
 535
 536 enough:
 537         ioend->io_size = (xfs_off_t)nblocks << block_bits;
 538         ioend->io_offset = offset;
 539         xfs_finish_ioend(ioend);
 540         return 0;
 541 }
 542
 543 STATIC void
 544 xfs_submit_page(
 545         struct page             *page,
 546         struct writeback_control *wbc,
 547         struct buffer_head      *bh_arr[],
 548         int                     bh_count,
 549         int                     probed_page,
 550         int                     clear_dirty)
 551 {
 552         struct buffer_head      *bh;
 553         int                     i;
 554
 555         BUG_ON(PageWriteback(page));
 556         if (bh_count)
 557                 set_page_writeback(page);
 558         if (clear_dirty)
 559                 clear_page_dirty(page);
 560         unlock_page(page);
 561
 562         if (bh_count) {
 563                 for (i = 0; i < bh_count; i++) {
 564                         bh = bh_arr[i];
 565                         mark_buffer_async_write(bh);
 566                         if (buffer_unwritten(bh))
 567                                 set_buffer_unwritten_io(bh);
 568                         set_buffer_uptodate(bh);
 569                         clear_buffer_dirty(bh);
 570                 }
 571
 572                 for (i = 0; i < bh_count; i++)
 573                         submit_bh(WRITE, bh_arr[i]);
 574
 575                 if (probed_page && clear_dirty)
 576                         wbc->nr_to_write--;     /* Wrote an "extra" page */
 577         }
 578 }
 579
 580 /*
 581  * Allocate & map buffers for page given the extent map. Write it out.
 582  * except for the original page of a writepage, this is called on
 583  * delalloc/unwritten pages only, for the original page it is possible
 584  * that the page has no mapping at all.
 585  */
 586 STATIC void
 587 xfs_convert_page(
 588         struct inode            *inode,
 589         struct page             *page,
 590         xfs_iomap_t             *iomapp,
 591         struct writeback_control *wbc,
 592         void                    *private,
 593         int                     startio,
 594         int                     all_bh)
 595 {
 596         struct buffer_head      *bh_arr[MAX_BUF_PER_PAGE], *bh, *head;
 597         xfs_iomap_t             *mp = iomapp, *tmp;
 598         unsigned long           offset, end_offset;
 599         int                     index = 0;
 600         int                     bbits = inode->i_blkbits;
 601         int                     len, page_dirty;
 602
 603         end_offset = (i_size_read(inode) & (PAGE_CACHE_SIZE - 1));
 604
 605         /*
 606          * page_dirty is initially a count of buffers on the page before
 607          * EOF and is decrememted as we move each into a cleanable state.
 608          */
 609         len = 1 << inode->i_blkbits;
 610         end_offset = max(end_offset, PAGE_CACHE_SIZE);
 611         end_offset = roundup(end_offset, len);
 612         page_dirty = end_offset / len;
 613
 614         offset = 0;
 615         bh = head = page_buffers(page);
 616         do {
 617                 if (offset >= end_offset)
 618                         break;
 619                 if (!(PageUptodate(page) || buffer_uptodate(bh)))
 620                         continue;
 621                 if (buffer_mapped(bh) && all_bh &&
 622                     !(buffer_unwritten(bh) || buffer_delay(bh))) {
 623                         if (startio) {
 624                                 lock_buffer(bh);
 625                                 bh_arr[index++] = bh;
 626                                 page_dirty--;
 627                         }
 628                         continue;
 629                 }
 630                 tmp = xfs_offset_to_map(page, mp, offset);
 631                 if (!tmp)
 632                         continue;
 633                 ASSERT(!(tmp->iomap_flags & IOMAP_HOLE));
 634                 ASSERT(!(tmp->iomap_flags & IOMAP_DELAY));
 635
 636                 /* If this is a new unwritten extent buffer (i.e. one
 637                  * that we haven't passed in private data for, we must
 638                  * now map this buffer too.
 639                  */
 640                 if (buffer_unwritten(bh) && !bh->b_end_io) {
 641                         ASSERT(tmp->iomap_flags & IOMAP_UNWRITTEN);
 642                         xfs_map_unwritten(inode, page, head, bh, offset,
 643                                         bbits, tmp, wbc, startio, all_bh);
 644                 } else if (! (buffer_unwritten(bh) && buffer_locked(bh))) {
 645                         xfs_map_at_offset(page, bh, offset, bbits, tmp);
 646                         if (buffer_unwritten(bh)) {
 647                                 set_buffer_unwritten_io(bh);
 648                                 bh->b_private = private;
 649                                 ASSERT(private);
 650                         }
 651                 }
 652                 if (startio) {
 653                         bh_arr[index++] = bh;
 654                 } else {
 655                         set_buffer_dirty(bh);
 656                         unlock_buffer(bh);
 657                         mark_buffer_dirty(bh);
 658                 }
 659                 page_dirty--;
 660         } while (offset += len, (bh = bh->b_this_page) != head);
 661
 662         if (startio && index) {
 663                 xfs_submit_page(page, wbc, bh_arr, index, 1, !page_dirty);
 664         } else {
 665                 unlock_page(page);
 666         }
 667 }
 668
 669 /*
 670  * Convert & write out a cluster of pages in the same extent as defined
 671  * by mp and following the start page.
 672  */
 673 STATIC void
 674 xfs_cluster_write(
 675         struct inode            *inode,
 676         pgoff_t                 tindex,
 677         xfs_iomap_t             *iomapp,
 678         struct writeback_control *wbc,
 679         int                     startio,
 680         int                     all_bh,
 681         pgoff_t                 tlast)
 682 {
 683         struct page             *page;
 684
 685         for (; tindex <= tlast; tindex++) {
 686                 page = xfs_probe_delalloc_page(inode, tindex);
 687                 if (!page)
 688                         break;
 689                 xfs_convert_page(inode, page, iomapp, wbc, NULL,
 690                                 startio, all_bh);
 691         }
 692 }
 693
 694 /*
 695  * Calling this without startio set means we are being asked to make a dirty
 696  * page ready for freeing it's buffers.  When called with startio set then
 697  * we are coming from writepage.
 698  *
 699  * When called with startio set it is important that we write the WHOLE
 700  * page if possible.
 701  * The bh->b_state's cannot know if any of the blocks or which block for
 702  * that matter are dirty due to mmap writes, and therefore bh uptodate is
 703  * only vaild if the page itself isn't completely uptodate.  Some layers
 704  * may clear the page dirty flag prior to calling write page, under the
 705  * assumption the entire page will be written out; by not writing out the
 706  * whole page the page can be reused before all valid dirty data is
 707  * written out.  Note: in the case of a page that has been dirty'd by
 708  * mapwrite and but partially setup by block_prepare_write the
 709  * bh->b_states's will not agree and only ones setup by BPW/BCW will have
 710  * valid state, thus the whole page must be written out thing.
 711  */
 712
 713 STATIC int
 714 xfs_page_state_convert(
 715         struct inode    *inode,
 716         struct page     *page,
 717         struct writeback_control *wbc,
 718         int             startio,
 719         int             unmapped) /* also implies page uptodate */
 720 {
 721         struct buffer_head      *bh_arr[MAX_BUF_PER_PAGE], *bh, *head;
 722         xfs_iomap_t             *iomp, iomap;
 723         loff_t                  offset;
 724         unsigned long           p_offset = 0;
 725         __uint64_t              end_offset;
 726         pgoff_t                 end_index, last_index, tlast;
 727         int                     len, err, i, cnt = 0, uptodate = 1;
 728         int                     flags;
 729         int                     page_dirty;
 730
 731         /* wait for other IO threads? */
 732         flags = (startio && wbc->sync_mode != WB_SYNC_NONE) ? 0 : BMAPI_TRYLOCK;
 733
 734         /* Is this page beyond the end of the file? */
 735         offset = i_size_read(inode);
 736         end_index = offset >> PAGE_CACHE_SHIFT;
 737         last_index = (offset - 1) >> PAGE_CACHE_SHIFT;
 738         if (page->index >= end_index) {
 739                 if ((page->index >= end_index + 1) ||
 740                     !(i_size_read(inode) & (PAGE_CACHE_SIZE - 1))) {
 741                         err = -EIO;
 742                         goto error;
 743                 }
 744         }
 745
 746         end_offset = min_t(unsigned long long,
 747                         (loff_t)(page->index + 1) << PAGE_CACHE_SHIFT, offset);
 748         offset = (loff_t)page->index << PAGE_CACHE_SHIFT;
 749
 750         /*
 751          * page_dirty is initially a count of buffers on the page before
 752          * EOF and is decrememted as we move each into a cleanable state.
 753          */
 754         len = 1 << inode->i_blkbits;
 755         p_offset = max(p_offset, PAGE_CACHE_SIZE);
 756         p_offset = roundup(p_offset, len);
 757         page_dirty = p_offset / len;
 758
 759         iomp = NULL;
 760         p_offset = 0;
 761         bh = head = page_buffers(page);
 762
 763         do {
 764                 if (offset >= end_offset)
 765                         break;
 766                 if (!buffer_uptodate(bh))
 767                         uptodate = 0;
 768                 if (!(PageUptodate(page) || buffer_uptodate(bh)) && !startio)
 769                         continue;
 770
 771                 if (iomp) {
 772                         iomp = xfs_offset_to_map(page, &iomap, p_offset);
 773                 }
 774
 775                 /*
 776                  * First case, map an unwritten extent and prepare for
 777                  * extent state conversion transaction on completion.
 778                  */
 779                 if (buffer_unwritten(bh)) {
 780                         if (!startio)
 781                                 continue;
 782                         if (!iomp) {
 783                                 err = xfs_map_blocks(inode, offset, len, &iomap,
 784                                                 BMAPI_READ|BMAPI_IGNSTATE);
 785                                 if (err) {
 786                                         goto error;
 787                                 }
 788                                 iomp = xfs_offset_to_map(page, &iomap,
 789                                                                 p_offset);
 790                         }
 791                         if (iomp) {
 792                                 if (!bh->b_end_io) {
 793                                         err = xfs_map_unwritten(inode, page,
 794                                                         head, bh, p_offset,
 795                                                         inode->i_blkbits, iomp,
 796                                                         wbc, startio, unmapped);
 797                                         if (err) {
 798                                                 goto error;
 799                                         }
 800                                 } else {
 801                                         set_bit(BH_Lock, &bh->b_state);
 802                                 }
 803                                 BUG_ON(!buffer_locked(bh));
 804                                 bh_arr[cnt++] = bh;
 805                                 page_dirty--;
 806                         }
 807                 /*
 808                  * Second case, allocate space for a delalloc buffer.
 809                  * We can return EAGAIN here in the release page case.
 810                  */
 811                 } else if (buffer_delay(bh)) {
 812                         if (!iomp) {
 813                                 err = xfs_map_blocks(inode, offset, len, &iomap,
 814                                                 BMAPI_ALLOCATE | flags);
 815                                 if (err) {
 816                                         goto error;
 817                                 }
 818                                 iomp = xfs_offset_to_map(page, &iomap,
 819                                                                 p_offset);
 820                         }
 821                         if (iomp) {
 822                                 xfs_map_at_offset(page, bh, p_offset,
 823                                                 inode->i_blkbits, iomp);
 824                                 if (startio) {
 825                                         bh_arr[cnt++] = bh;
 826                                 } else {
 827                                         set_buffer_dirty(bh);
 828                                         unlock_buffer(bh);
 829                                         mark_buffer_dirty(bh);
 830                                 }
 831                                 page_dirty--;
 832                         }
 833                 } else if ((buffer_uptodate(bh) || PageUptodate(page)) &&
 834                            (unmapped || startio)) {
 835
 836                         if (!buffer_mapped(bh)) {
 837                                 int     size;
 838
 839                                 /*
 840                                  * Getting here implies an unmapped buffer
 841                                  * was found, and we are in a path where we
 842                                  * need to write the whole page out.
 843                                  */
 844                                 if (!iomp) {
 845                                         size = xfs_probe_unmapped_cluster(
 846                                                         inode, page, bh, head);
 847                                         err = xfs_map_blocks(inode, offset,
 848                                                         size, &iomap,
 849                                                         BMAPI_WRITE|BMAPI_MMAP);
 850                                         if (err) {
 851                                                 goto error;
 852                                         }
 853                                         iomp = xfs_offset_to_map(page, &iomap,
 854                                                                      p_offset);
 855                                 }
 856                                 if (iomp) {
 857                                         xfs_map_at_offset(page,
 858                                                         bh, p_offset,
 859                                                         inode->i_blkbits, iomp);
 860                                         if (startio) {
 861                                                 bh_arr[cnt++] = bh;
 862                                         } else {
 863                                                 set_buffer_dirty(bh);
 864                                                 unlock_buffer(bh);
 865                                                 mark_buffer_dirty(bh);
 866                                         }
 867                                         page_dirty--;
 868                                 }
 869                         } else if (startio) {
 870                                 if (buffer_uptodate(bh) &&
 871                                     !test_and_set_bit(BH_Lock, &bh->b_state)) {
 872                                         bh_arr[cnt++] = bh;
 873                                         page_dirty--;
 874                                 }
 875                         }
 876                 }
 877         } while (offset += len, p_offset += len,
 878                 ((bh = bh->b_this_page) != head));
 879
 880         if (uptodate && bh == head)
 881                 SetPageUptodate(page);
 882
 883         if (startio) {
 884                 xfs_submit_page(page, wbc, bh_arr, cnt, 0, !page_dirty);
 885         }
 886
 887         if (iomp) {
 888                 offset = (iomp->iomap_offset + iomp->iomap_bsize - 1) >>
 889                                         PAGE_CACHE_SHIFT;
 890                 tlast = min_t(pgoff_t, offset, last_index);
 891                 xfs_cluster_write(inode, page->index + 1, iomp, wbc,
 892                                         startio, unmapped, tlast);
 893         }
 894
 895         return page_dirty;
 896
 897 error:
 898         for (i = 0; i < cnt; i++) {
 899                 unlock_buffer(bh_arr[i]);
 900         }
 901
 902         /*
 903          * If it's delalloc and we have nowhere to put it,
 904          * throw it away, unless the lower layers told
 905          * us to try again.
 906          */
 907         if (err != -EAGAIN) {
 908                 if (!unmapped) {
 909                         block_invalidatepage(page, 0);
 910                 }
 911                 ClearPageUptodate(page);
 912         }
 913         return err;
 914 }
 915
 916 STATIC int
 917 __linvfs_get_block(
 918         struct inode            *inode,
 919         sector_t                iblock,
 920         unsigned long           blocks,
 921         struct buffer_head      *bh_result,
 922         int                     create,
 923         int                     direct,
 924         bmapi_flags_t           flags)
 925 {
 926         vnode_t                 *vp = LINVFS_GET_VP(inode);
 927         xfs_iomap_t             iomap;
 928         int                     retpbbm = 1;
 929         int                     error;
 930         ssize_t                 size;
 931         loff_t                  offset = (loff_t)iblock << inode->i_blkbits;
 932
 933         if (blocks)
 934                 size = blocks << inode->i_blkbits;
 935         else
 936                 size = 1 << inode->i_blkbits;
 937
 938         VOP_BMAP(vp, offset, size,
 939                 create ? flags : BMAPI_READ, &iomap, &retpbbm, error);
 940         if (error)
 941                 return -error;
 942
 943         if (retpbbm == 0)
 944                 return 0;
 945
 946         if (iomap.iomap_bn != IOMAP_DADDR_NULL) {
 947                 xfs_daddr_t             bn;
 948                 loff_t                  delta;
 949
 950                 /* For unwritten extents do not report a disk address on
 951                  * the read case (treat as if we're reading into a hole).
 952                  */
 953                 if (create || !(iomap.iomap_flags & IOMAP_UNWRITTEN)) {
 954                         delta = offset - iomap.iomap_offset;
 955                         delta >>= inode->i_blkbits;
 956
 957                         bn = iomap.iomap_bn >> (inode->i_blkbits - BBSHIFT);
 958                         bn += delta;
 959                         BUG_ON(!bn && !(iomap.iomap_flags & IOMAP_REALTIME));
 960                         bh_result->b_blocknr = bn;
 961                         set_buffer_mapped(bh_result);
 962                 }
 963                 if (create && (iomap.iomap_flags & IOMAP_UNWRITTEN)) {
 964                         if (direct)
 965                                 bh_result->b_private = inode;
 966                         set_buffer_unwritten(bh_result);
 967                         set_buffer_delay(bh_result);
 968                 }
 969         }
 970
 971         /* If this is a realtime file, data might be on a new device */
 972         bh_result->b_bdev = iomap.iomap_target->pbr_bdev;
 973
 974         /* If we previously allocated a block out beyond eof and
 975          * we are now coming back to use it then we will need to
 976          * flag it as new even if it has a disk address.
 977          */
 978         if (create &&
 979             ((!buffer_mapped(bh_result) && !buffer_uptodate(bh_result)) ||
 980              (offset >= i_size_read(inode)) || (iomap.iomap_flags & IOMAP_NEW))) {
 981                 set_buffer_new(bh_result);
 982         }
 983
 984         if (iomap.iomap_flags & IOMAP_DELAY) {
 985                 BUG_ON(direct);
 986                 if (create) {
 987                         set_buffer_uptodate(bh_result);
 988                         set_buffer_mapped(bh_result);
 989                         set_buffer_delay(bh_result);
 990                 }
 991         }
 992
 993         if (blocks) {
 994                 bh_result->b_size = (ssize_t)min(
 995                         (loff_t)(iomap.iomap_bsize - iomap.iomap_delta),
 996                         (loff_t)(blocks << inode->i_blkbits));
 997         }
 998
 999         return 0;
1000 }
1001
1002 int
1003 linvfs_get_block(
1004         struct inode            *inode,
1005         sector_t                iblock,
1006         struct buffer_head      *bh_result,
1007         int                     create)
1008 {
1009         return __linvfs_get_block(inode, iblock, 0, bh_result,
1010                                         create, 0, BMAPI_WRITE);
1011 }
1012
1013 STATIC int
1014 linvfs_get_blocks_direct(
1015         struct inode            *inode,
1016         sector_t                iblock,
1017         unsigned long           max_blocks,
1018         struct buffer_head      *bh_result,
1019         int                     create)
1020 {
1021         return __linvfs_get_block(inode, iblock, max_blocks, bh_result,
1022                                         create, 1, BMAPI_WRITE|BMAPI_DIRECT);
1023 }
1024
1025 STATIC void
1026 linvfs_end_io_direct(
1027         struct kiocb    *iocb,
1028         loff_t          offset,
1029         ssize_t         size,
1030         void            *private)
1031 {
1032         xfs_ioend_t     *ioend = iocb->private;
1033
1034         /*
1035          * Non-NULL private data means we need to issue a transaction to
1036          * convert a range from unwritten to written extents.  This needs
1037          * to happen from process contect but aio+dio I/O completion
1038          * happens from irq context so we need to defer it to a workqueue.
1039          * This is not nessecary for synchronous direct I/O, but we do
1040          * it anyway to keep the code uniform and simpler.
1041          *
1042          * The core direct I/O code might be changed to always call the
1043          * completion handler in the future, in which case all this can
1044          * go away.
1045          */
1046         if (private && size > 0) {
1047                 ioend->io_offset = offset;
1048                 ioend->io_size = size;
1049                 xfs_finish_ioend(ioend);
1050         } else {
1051                 ASSERT(size >= 0);
1052                 xfs_destroy_ioend(ioend);
1053         }
1054
1055         /*
1056          * blockdev_direct_IO can return an error even afer the I/O
1057          * completion handler was called.  Thus we need to protect
1058          * against double-freeing.
1059          */
1060         iocb->private = NULL;
1061 }
1062
1063 STATIC ssize_t
1064 linvfs_direct_IO(
1065         int                     rw,
1066         struct kiocb            *iocb,
1067         const struct iovec      *iov,
1068         loff_t                  offset,
1069         unsigned long           nr_segs)
1070 {
1071         struct file     *file = iocb->ki_filp;
1072         struct inode    *inode = file->f_mapping->host;
1073         vnode_t         *vp = LINVFS_GET_VP(inode);
1074         xfs_iomap_t     iomap;
1075         int             maps = 1;
1076         int             error;
1077         ssize_t         ret;
1078
1079         VOP_BMAP(vp, offset, 0, BMAPI_DEVICE, &iomap, &maps, error);
1080         if (error)
1081                 return -error;
1082
1083         iocb->private = xfs_alloc_ioend(inode);
1084
1085         ret = blockdev_direct_IO_own_locking(rw, iocb, inode,
1086                 iomap.iomap_target->pbr_bdev,
1087                 iov, offset, nr_segs,
1088                 linvfs_get_blocks_direct,
1089                 linvfs_end_io_direct);
1090
1091         if (unlikely(ret <= 0 && iocb->private))
1092                 xfs_destroy_ioend(iocb->private);
1093         return ret;
1094 }
1095
1096
1097 STATIC sector_t
1098 linvfs_bmap(
1099         struct address_space    *mapping,
1100         sector_t                block)
1101 {
1102         struct inode            *inode = (struct inode *)mapping->host;
1103         vnode_t                 *vp = LINVFS_GET_VP(inode);
1104         int                     error;
1105
1106         vn_trace_entry(vp, "linvfs_bmap", (inst_t *)__return_address);
1107
1108         VOP_RWLOCK(vp, VRWLOCK_READ);
1109         VOP_FLUSH_PAGES(vp, (xfs_off_t)0, -1, 0, FI_REMAPF, error);
1110         VOP_RWUNLOCK(vp, VRWLOCK_READ);
1111         return generic_block_bmap(mapping, block, linvfs_get_block);
1112 }
1113
1114 STATIC int
1115 linvfs_readpage(
1116         struct file             *unused,
1117         struct page             *page)
1118 {
1119         return mpage_readpage(page, linvfs_get_block);
1120 }
1121
1122 STATIC int
1123 linvfs_readpages(
1124         struct file             *unused,
1125         struct address_space    *mapping,
1126         struct list_head        *pages,
1127         unsigned                nr_pages)
1128 {
1129         return mpage_readpages(mapping, pages, nr_pages, linvfs_get_block);
1130 }
1131
1132 STATIC void
1133 xfs_count_page_state(
1134         struct page             *page,
1135         int                     *delalloc,
1136         int                     *unmapped,
1137         int                     *unwritten)
1138 {
1139         struct buffer_head      *bh, *head;
1140
1141         *delalloc = *unmapped = *unwritten = 0;
1142
1143         bh = head = page_buffers(page);
1144         do {
1145                 if (buffer_uptodate(bh) && !buffer_mapped(bh))
1146                         (*unmapped) = 1;
1147                 else if (buffer_unwritten(bh) && !buffer_delay(bh))
1148                         clear_buffer_unwritten(bh);
1149                 else if (buffer_unwritten(bh))
1150                         (*unwritten) = 1;
1151                 else if (buffer_delay(bh))
1152                         (*delalloc) = 1;
1153         } while ((bh = bh->b_this_page) != head);
1154 }
1155
1156
1157 /*
1158  * writepage: Called from one of two places:
1159  *
1160  * 1. we are flushing a delalloc buffer head.
1161  *
1162  * 2. we are writing out a dirty page. Typically the page dirty
1163  *    state is cleared before we get here. In this case is it
1164  *    conceivable we have no buffer heads.
1165  *
1166  * For delalloc space on the page we need to allocate space and
1167  * flush it. For unmapped buffer heads on the page we should
1168  * allocate space if the page is uptodate. For any other dirty
1169  * buffer heads on the page we should flush them.
1170  *
1171  * If we detect that a transaction would be required to flush
1172  * the page, we have to check the process flags first, if we
1173  * are already in a transaction or disk I/O during allocations
1174  * is off, we need to fail the writepage and redirty the page.
1175  */
1176
1177 STATIC int
1178 linvfs_writepage(
1179         struct page             *page,
1180         struct writeback_control *wbc)
1181 {
1182         int                     error;
1183         int                     need_trans;
1184         int                     delalloc, unmapped, unwritten;
1185         struct inode            *inode = page->mapping->host;
1186
1187         xfs_page_trace(XFS_WRITEPAGE_ENTER, inode, page, 0);
1188
1189         /*
1190          * We need a transaction if:
1191          *  1. There are delalloc buffers on the page
1192          *  2. The page is uptodate and we have unmapped buffers
1193          *  3. The page is uptodate and we have no buffers
1194          *  4. There are unwritten buffers on the page
1195          */
1196
1197         if (!page_has_buffers(page)) {
1198                 unmapped = 1;
1199                 need_trans = 1;
1200         } else {
1201                 xfs_count_page_state(page, &delalloc, &unmapped, &unwritten);
1202                 if (!PageUptodate(page))
1203                         unmapped = 0;
1204                 need_trans = delalloc + unmapped + unwritten;
1205         }
1206
1207         /*
1208          * If we need a transaction and the process flags say
1209          * we are already in a transaction, or no IO is allowed
1210          * then mark the page dirty again and leave the page
1211          * as is.
1212          */
1213         if (PFLAGS_TEST_FSTRANS() && need_trans)
1214                 goto out_fail;
1215
1216         /*
1217          * Delay hooking up buffer heads until we have
1218          * made our go/no-go decision.
1219          */
1220         if (!page_has_buffers(page))
1221                 create_empty_buffers(page, 1 << inode->i_blkbits, 0);
1222
1223         /*
1224          * Convert delayed allocate, unwritten or unmapped space
1225          * to real space and flush out to disk.
1226          */
1227         error = xfs_page_state_convert(inode, page, wbc, 1, unmapped);
1228         if (error == -EAGAIN)
1229                 goto out_fail;
1230         if (unlikely(error < 0))
1231                 goto out_unlock;
1232
1233         return 0;
1234
1235 out_fail:
1236         redirty_page_for_writepage(wbc, page);
1237         unlock_page(page);
1238         return 0;
1239 out_unlock:
1240         unlock_page(page);
1241         return error;
1242 }
1243
1244 STATIC int
1245 linvfs_invalidate_page(
1246         struct page             *page,
1247         unsigned long           offset)
1248 {
1249         xfs_page_trace(XFS_INVALIDPAGE_ENTER,
1250                         page->mapping->host, page, offset);
1251         return block_invalidatepage(page, offset);
1252 }
1253
1254 /*
1255  * Called to move a page into cleanable state - and from there
1256  * to be released. Possibly the page is already clean. We always
1257  * have buffer heads in this call.
1258  *
1259  * Returns 0 if the page is ok to release, 1 otherwise.
1260  *
1261  * Possible scenarios are:
1262  *
1263  * 1. We are being called to release a page which has been written
1264  *    to via regular I/O. buffer heads will be dirty and possibly
1265  *    delalloc. If no delalloc buffer heads in this case then we
1266  *    can just return zero.
1267  *
1268  * 2. We are called to release a page which has been written via
1269  *    mmap, all we need to do is ensure there is no delalloc
1270  *    state in the buffer heads, if not we can let the caller
1271  *    free them and we should come back later via writepage.
1272  */
1273 STATIC int
1274 linvfs_release_page(
1275         struct page             *page,
1276         int                     gfp_mask)
1277 {
1278         struct inode            *inode = page->mapping->host;
1279         int                     dirty, delalloc, unmapped, unwritten;
1280         struct writeback_control wbc = {
1281                 .sync_mode = WB_SYNC_ALL,
1282                 .nr_to_write = 1,
1283         };
1284
1285         xfs_page_trace(XFS_RELEASEPAGE_ENTER, inode, page, gfp_mask);
1286
1287         xfs_count_page_state(page, &delalloc, &unmapped, &unwritten);
1288         if (!delalloc && !unwritten)
1289                 goto free_buffers;
1290
1291         if (!(gfp_mask & __GFP_FS))
1292                 return 0;
1293
1294         /* If we are already inside a transaction or the thread cannot
1295          * do I/O, we cannot release this page.
1296          */
1297         if (PFLAGS_TEST_FSTRANS())
1298                 return 0;
1299
1300         /*
1301          * Convert delalloc space to real space, do not flush the
1302          * data out to disk, that will be done by the caller.
1303          * Never need to allocate space here - we will always
1304          * come back to writepage in that case.
1305          */
1306         dirty = xfs_page_state_convert(inode, page, &wbc, 0, 0);
1307         if (dirty == 0 && !unwritten)
1308                 goto free_buffers;
1309         return 0;
1310
1311 free_buffers:
1312         return try_to_free_buffers(page);
1313 }
1314
1315 STATIC int
1316 linvfs_prepare_write(
1317         struct file             *file,
1318         struct page             *page,
1319         unsigned int            from,
1320         unsigned int            to)
1321 {
1322         return block_prepare_write(page, from, to, linvfs_get_block);
1323 }
1324
1325 struct address_space_operations linvfs_aops = {
1326         .readpage               = linvfs_readpage,
1327         .readpages              = linvfs_readpages,
1328         .writepage              = linvfs_writepage,
1329         .sync_page              = block_sync_page,
1330         .releasepage            = linvfs_release_page,
1331         .invalidatepage         = linvfs_invalidate_page,
1332         .prepare_write          = linvfs_prepare_write,
1333         .commit_write           = generic_commit_write,
1334         .bmap                   = linvfs_bmap,
1335         .direct_IO              = linvfs_direct_IO,
1336 };