fs/gfs2/ops_address.c

   1 /*
   2  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
   3  * Copyright (C) 2004-2007 Red Hat, Inc.  All rights reserved.
   4  *
   5  * This copyrighted material is made available to anyone wishing to use,
   6  * modify, copy, or redistribute it subject to the terms and conditions
   7  * of the GNU General Public License version 2.
   8  */
   9
  10 #include <linux/sched.h>
  11 #include <linux/slab.h>
  12 #include <linux/spinlock.h>
  13 #include <linux/completion.h>
  14 #include <linux/buffer_head.h>
  15 #include <linux/pagemap.h>
  16 #include <linux/pagevec.h>
  17 #include <linux/mpage.h>
  18 #include <linux/fs.h>
  19 #include <linux/writeback.h>
  20 #include <linux/gfs2_ondisk.h>
  21 #include <linux/lm_interface.h>
  22
  23 #include "gfs2.h"
  24 #include "incore.h"
  25 #include "bmap.h"
  26 #include "glock.h"
  27 #include "inode.h"
  28 #include "log.h"
  29 #include "meta_io.h"
  30 #include "ops_address.h"
  31 #include "quota.h"
  32 #include "trans.h"
  33 #include "rgrp.h"
  34 #include "ops_file.h"
  35 #include "util.h"
  36 #include "glops.h"
  37
  38
  39 static void gfs2_page_add_databufs(struct gfs2_inode *ip, struct page *page,
  40                                    unsigned int from, unsigned int to)
  41 {
  42         struct buffer_head *head = page_buffers(page);
  43         unsigned int bsize = head->b_size;
  44         struct buffer_head *bh;
  45         unsigned int start, end;
  46
  47         for (bh = head, start = 0; bh != head || !start;
  48              bh = bh->b_this_page, start = end) {
  49                 end = start + bsize;
  50                 if (end <= from || start >= to)
  51                         continue;
  52                 gfs2_trans_add_bh(ip->i_gl, bh, 0);
  53         }
  54 }
  55
  56 /**
  57  * gfs2_get_block - Fills in a buffer head with details about a block
  58  * @inode: The inode
  59  * @lblock: The block number to look up
  60  * @bh_result: The buffer head to return the result in
  61  * @create: Non-zero if we may add block to the file
  62  *
  63  * Returns: errno
  64  */
  65
  66 int gfs2_get_block(struct inode *inode, sector_t lblock,
  67                    struct buffer_head *bh_result, int create)
  68 {
  69         return gfs2_block_map(inode, lblock, create, bh_result);
  70 }
  71
  72 /**
  73  * gfs2_get_block_noalloc - Fills in a buffer head with details about a block
  74  * @inode: The inode
  75  * @lblock: The block number to look up
  76  * @bh_result: The buffer head to return the result in
  77  * @create: Non-zero if we may add block to the file
  78  *
  79  * Returns: errno
  80  */
  81
  82 static int gfs2_get_block_noalloc(struct inode *inode, sector_t lblock,
  83                                   struct buffer_head *bh_result, int create)
  84 {
  85         int error;
  86
  87         error = gfs2_block_map(inode, lblock, 0, bh_result);
  88         if (error)
  89                 return error;
  90         if (bh_result->b_blocknr == 0)
  91                 return -EIO;
  92         return 0;
  93 }
  94
  95 static int gfs2_get_block_direct(struct inode *inode, sector_t lblock,
  96                                  struct buffer_head *bh_result, int create)
  97 {
  98         return gfs2_block_map(inode, lblock, 0, bh_result);
  99 }
 100
 101 /**
 102  * gfs2_writepage - Write complete page
 103  * @page: Page to write
 104  *
 105  * Returns: errno
 106  *
 107  * Some of this is copied from block_write_full_page() although we still
 108  * call it to do most of the work.
 109  */
 110
 111 static int gfs2_writepage(struct page *page, struct writeback_control *wbc)
 112 {
 113         struct inode *inode = page->mapping->host;
 114         struct gfs2_inode *ip = GFS2_I(inode);
 115         struct gfs2_sbd *sdp = GFS2_SB(inode);
 116         loff_t i_size = i_size_read(inode);
 117         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
 118         unsigned offset;
 119         int error;
 120         int done_trans = 0;
 121
 122         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl))) {
 123                 unlock_page(page);
 124                 return -EIO;
 125         }
 126         if (current->journal_info)
 127                 goto out_ignore;
 128
 129         /* Is the page fully outside i_size? (truncate in progress) */
 130         offset = i_size & (PAGE_CACHE_SIZE-1);
 131         if (page->index > end_index || (page->index == end_index && !offset)) {
 132                 page->mapping->a_ops->invalidatepage(page, 0);
 133                 unlock_page(page);
 134                 return 0; /* don't care */
 135         }
 136
 137         if (sdp->sd_args.ar_data == GFS2_DATA_ORDERED || gfs2_is_jdata(ip)) {
 138                 error = gfs2_trans_begin(sdp, RES_DINODE + 1, 0);
 139                 if (error)
 140                         goto out_ignore;
 141                 if (!page_has_buffers(page)) {
 142                         create_empty_buffers(page, inode->i_sb->s_blocksize,
 143                                              (1 << BH_Dirty)|(1 << BH_Uptodate));
 144                 }
 145                 gfs2_page_add_databufs(ip, page, 0, sdp->sd_vfs->s_blocksize-1);
 146                 done_trans = 1;
 147         }
 148         error = block_write_full_page(page, gfs2_get_block_noalloc, wbc);
 149         if (done_trans)
 150                 gfs2_trans_end(sdp);
 151         gfs2_meta_cache_flush(ip);
 152         return error;
 153
 154 out_ignore:
 155         redirty_page_for_writepage(wbc, page);
 156         unlock_page(page);
 157         return 0;
 158 }
 159
 160 /**
 161  * gfs2_writepages - Write a bunch of dirty pages back to disk
 162  * @mapping: The mapping to write
 163  * @wbc: Write-back control
 164  *
 165  * For journaled files and/or ordered writes this just falls back to the
 166  * kernel's default writepages path for now. We will probably want to change
 167  * that eventually (i.e. when we look at allocate on flush).
 168  *
 169  * For the data=writeback case though we can already ignore buffer heads
 170  * and write whole extents at once. This is a big reduction in the
 171  * number of I/O requests we send and the bmap calls we make in this case.
 172  */
 173 static int gfs2_writepages(struct address_space *mapping,
 174                            struct writeback_control *wbc)
 175 {
 176         struct inode *inode = mapping->host;
 177         struct gfs2_inode *ip = GFS2_I(inode);
 178         struct gfs2_sbd *sdp = GFS2_SB(inode);
 179
 180         if (sdp->sd_args.ar_data == GFS2_DATA_WRITEBACK && !gfs2_is_jdata(ip))
 181                 return mpage_writepages(mapping, wbc, gfs2_get_block_noalloc);
 182
 183         return generic_writepages(mapping, wbc);
 184 }
 185
 186 /**
 187  * stuffed_readpage - Fill in a Linux page with stuffed file data
 188  * @ip: the inode
 189  * @page: the page
 190  *
 191  * Returns: errno
 192  */
 193
 194 static int stuffed_readpage(struct gfs2_inode *ip, struct page *page)
 195 {
 196         struct buffer_head *dibh;
 197         void *kaddr;
 198         int error;
 199
 200         /*
 201          * Due to the order of unstuffing files and ->nopage(), we can be
 202          * asked for a zero page in the case of a stuffed file being extended,
 203          * so we need to supply one here. It doesn't happen often.
 204          */
 205         if (unlikely(page->index)) {
 206                 kaddr = kmap_atomic(page, KM_USER0);
 207                 memset(kaddr, 0, PAGE_CACHE_SIZE);
 208                 kunmap_atomic(kaddr, KM_USER0);
 209                 flush_dcache_page(page);
 210                 SetPageUptodate(page);
 211                 return 0;
 212         }
 213
 214         error = gfs2_meta_inode_buffer(ip, &dibh);
 215         if (error)
 216                 return error;
 217
 218         kaddr = kmap_atomic(page, KM_USER0);
 219         memcpy(kaddr, dibh->b_data + sizeof(struct gfs2_dinode),
 220                ip->i_di.di_size);
 221         memset(kaddr + ip->i_di.di_size, 0, PAGE_CACHE_SIZE - ip->i_di.di_size);
 222         kunmap_atomic(kaddr, KM_USER0);
 223         flush_dcache_page(page);
 224         brelse(dibh);
 225         SetPageUptodate(page);
 226
 227         return 0;
 228 }
 229
 230
 231 /**
 232  * gfs2_readpage - readpage with locking
 233  * @file: The file to read a page for. N.B. This may be NULL if we are
 234  * reading an internal file.
 235  * @page: The page to read
 236  *
 237  * Returns: errno
 238  */
 239
 240 static int gfs2_readpage(struct file *file, struct page *page)
 241 {
 242         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
 243         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 244         struct gfs2_file *gf = NULL;
 245         struct gfs2_holder gh;
 246         int error;
 247         int do_unlock = 0;
 248
 249         if (likely(file != &gfs2_internal_file_sentinel)) {
 250                 if (file) {
 251                         gf = file->private_data;
 252                         if (test_bit(GFF_EXLOCK, &gf->f_flags))
 253                                 /* gfs2_sharewrite_nopage has grabbed the ip->i_gl already */
 254                                 goto skip_lock;
 255                 }
 256                 gfs2_holder_init(ip->i_gl, LM_ST_SHARED, GL_ATIME|LM_FLAG_TRY_1CB, &gh);
 257                 do_unlock = 1;
 258                 error = gfs2_glock_nq_atime(&gh);
 259                 if (unlikely(error))
 260                         goto out_unlock;
 261         }
 262
 263 skip_lock:
 264         if (gfs2_is_stuffed(ip)) {
 265                 error = stuffed_readpage(ip, page);
 266                 unlock_page(page);
 267         } else
 268                 error = mpage_readpage(page, gfs2_get_block);
 269
 270         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 271                 error = -EIO;
 272
 273         if (do_unlock) {
 274                 gfs2_glock_dq_m(1, &gh);
 275                 gfs2_holder_uninit(&gh);
 276         }
 277 out:
 278         return error;
 279 out_unlock:
 280         unlock_page(page);
 281         if (error == GLR_TRYFAILED) {
 282                 error = AOP_TRUNCATED_PAGE;
 283                 yield();
 284         }
 285         if (do_unlock)
 286                 gfs2_holder_uninit(&gh);
 287         goto out;
 288 }
 289
 290 /**
 291  * gfs2_readpages - Read a bunch of pages at once
 292  *
 293  * Some notes:
 294  * 1. This is only for readahead, so we can simply ignore any things
 295  *    which are slightly inconvenient (such as locking conflicts between
 296  *    the page lock and the glock) and return having done no I/O. Its
 297  *    obviously not something we'd want to do on too regular a basis.
 298  *    Any I/O we ignore at this time will be done via readpage later.
 299  * 2. We don't handle stuffed files here we let readpage do the honours.
 300  * 3. mpage_readpages() does most of the heavy lifting in the common case.
 301  * 4. gfs2_get_block() is relied upon to set BH_Boundary in the right places.
 302  * 5. We use LM_FLAG_TRY_1CB here, effectively we then have lock-ahead as
 303  *    well as read-ahead.
 304  */
 305 static int gfs2_readpages(struct file *file, struct address_space *mapping,
 306                           struct list_head *pages, unsigned nr_pages)
 307 {
 308         struct inode *inode = mapping->host;
 309         struct gfs2_inode *ip = GFS2_I(inode);
 310         struct gfs2_sbd *sdp = GFS2_SB(inode);
 311         struct gfs2_holder gh;
 312         int ret = 0;
 313         int do_unlock = 0;
 314
 315         if (likely(file != &gfs2_internal_file_sentinel)) {
 316                 if (file) {
 317                         struct gfs2_file *gf = file->private_data;
 318                         if (test_bit(GFF_EXLOCK, &gf->f_flags))
 319                                 goto skip_lock;
 320                 }
 321                 gfs2_holder_init(ip->i_gl, LM_ST_SHARED,
 322                                  LM_FLAG_TRY_1CB|GL_ATIME, &gh);
 323                 do_unlock = 1;
 324                 ret = gfs2_glock_nq_atime(&gh);
 325                 if (ret == GLR_TRYFAILED)
 326                         goto out_noerror;
 327                 if (unlikely(ret))
 328                         goto out_unlock;
 329         }
 330 skip_lock:
 331         if (!gfs2_is_stuffed(ip))
 332                 ret = mpage_readpages(mapping, pages, nr_pages, gfs2_get_block);
 333
 334         if (do_unlock) {
 335                 gfs2_glock_dq_m(1, &gh);
 336                 gfs2_holder_uninit(&gh);
 337         }
 338 out:
 339         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
 340                 ret = -EIO;
 341         return ret;
 342 out_noerror:
 343         ret = 0;
 344 out_unlock:
 345         if (do_unlock)
 346                 gfs2_holder_uninit(&gh);
 347         goto out;
 348 }
 349
 350 /**
 351  * gfs2_prepare_write - Prepare to write a page to a file
 352  * @file: The file to write to
 353  * @page: The page which is to be prepared for writing
 354  * @from: From (byte range within page)
 355  * @to: To (byte range within page)
 356  *
 357  * Returns: errno
 358  */
 359
 360 static int gfs2_prepare_write(struct file *file, struct page *page,
 361                               unsigned from, unsigned to)
 362 {
 363         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
 364         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 365         unsigned int data_blocks, ind_blocks, rblocks;
 366         int alloc_required;
 367         int error = 0;
 368         loff_t pos = ((loff_t)page->index << PAGE_CACHE_SHIFT) + from;
 369         loff_t end = ((loff_t)page->index << PAGE_CACHE_SHIFT) + to;
 370         struct gfs2_alloc *al;
 371         unsigned int write_len = to - from;
 372
 373
 374         gfs2_holder_init(ip->i_gl, LM_ST_EXCLUSIVE, GL_ATIME|LM_FLAG_TRY_1CB, &ip->i_gh);
 375         error = gfs2_glock_nq_atime(&ip->i_gh);
 376         if (unlikely(error)) {
 377                 if (error == GLR_TRYFAILED) {
 378                         unlock_page(page);
 379                         error = AOP_TRUNCATED_PAGE;
 380                         yield();
 381                 }
 382                 goto out_uninit;
 383         }
 384
 385         gfs2_write_calc_reserv(ip, write_len, &data_blocks, &ind_blocks);
 386
 387         error = gfs2_write_alloc_required(ip, pos, write_len, &alloc_required);
 388         if (error)
 389                 goto out_unlock;
 390
 391
 392         ip->i_alloc.al_requested = 0;
 393         if (alloc_required) {
 394                 al = gfs2_alloc_get(ip);
 395
 396                 error = gfs2_quota_lock(ip, NO_QUOTA_CHANGE, NO_QUOTA_CHANGE);
 397                 if (error)
 398                         goto out_alloc_put;
 399
 400                 error = gfs2_quota_check(ip, ip->i_inode.i_uid, ip->i_inode.i_gid);
 401                 if (error)
 402                         goto out_qunlock;
 403
 404                 al->al_requested = data_blocks + ind_blocks;
 405                 error = gfs2_inplace_reserve(ip);
 406                 if (error)
 407                         goto out_qunlock;
 408         }
 409
 410         rblocks = RES_DINODE + ind_blocks;
 411         if (gfs2_is_jdata(ip))
 412                 rblocks += data_blocks ? data_blocks : 1;
 413         if (ind_blocks || data_blocks)
 414                 rblocks += RES_STATFS + RES_QUOTA;
 415
 416         error = gfs2_trans_begin(sdp, rblocks, 0);
 417         if (error)
 418                 goto out;
 419
 420         if (gfs2_is_stuffed(ip)) {
 421                 if (end > sdp->sd_sb.sb_bsize - sizeof(struct gfs2_dinode)) {
 422                         error = gfs2_unstuff_dinode(ip, page);
 423                         if (error == 0)
 424                                 goto prepare_write;
 425                 } else if (!PageUptodate(page))
 426                         error = stuffed_readpage(ip, page);
 427                 goto out;
 428         }
 429
 430 prepare_write:
 431         error = block_prepare_write(page, from, to, gfs2_get_block);
 432
 433 out:
 434         if (error) {
 435                 gfs2_trans_end(sdp);
 436                 if (alloc_required) {
 437                         gfs2_inplace_release(ip);
 438 out_qunlock:
 439                         gfs2_quota_unlock(ip);
 440 out_alloc_put:
 441                         gfs2_alloc_put(ip);
 442                 }
 443 out_unlock:
 444                 gfs2_glock_dq_m(1, &ip->i_gh);
 445 out_uninit:
 446                 gfs2_holder_uninit(&ip->i_gh);
 447         }
 448
 449         return error;
 450 }
 451
 452 /**
 453  * adjust_fs_space - Adjusts the free space available due to gfs2_grow
 454  * @inode: the rindex inode
 455  */
 456 static void adjust_fs_space(struct inode *inode)
 457 {
 458         struct gfs2_sbd *sdp = inode->i_sb->s_fs_info;
 459         struct gfs2_statfs_change_host *m_sc = &sdp->sd_statfs_master;
 460         struct gfs2_statfs_change_host *l_sc = &sdp->sd_statfs_local;
 461         u64 fs_total, new_free;
 462
 463         /* Total up the file system space, according to the latest rindex. */
 464         fs_total = gfs2_ri_total(sdp);
 465
 466         spin_lock(&sdp->sd_statfs_spin);
 467         if (fs_total > (m_sc->sc_total + l_sc->sc_total))
 468                 new_free = fs_total - (m_sc->sc_total + l_sc->sc_total);
 469         else
 470                 new_free = 0;
 471         spin_unlock(&sdp->sd_statfs_spin);
 472         fs_warn(sdp, "File system extended by %llu blocks.\n",
 473                 (unsigned long long)new_free);
 474         gfs2_statfs_change(sdp, new_free, new_free, 0);
 475 }
 476
 477 /**
 478  * gfs2_commit_write - Commit write to a file
 479  * @file: The file to write to
 480  * @page: The page containing the data
 481  * @from: From (byte range within page)
 482  * @to: To (byte range within page)
 483  *
 484  * Returns: errno
 485  */
 486
 487 static int gfs2_commit_write(struct file *file, struct page *page,
 488                              unsigned from, unsigned to)
 489 {
 490         struct inode *inode = page->mapping->host;
 491         struct gfs2_inode *ip = GFS2_I(inode);
 492         struct gfs2_sbd *sdp = GFS2_SB(inode);
 493         int error = -EOPNOTSUPP;
 494         struct buffer_head *dibh;
 495         struct gfs2_alloc *al = &ip->i_alloc;
 496         struct gfs2_dinode *di;
 497
 498         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_locked_by_me(ip->i_gl)))
 499                 goto fail_nounlock;
 500
 501         error = gfs2_meta_inode_buffer(ip, &dibh);
 502         if (error)
 503                 goto fail_endtrans;
 504
 505         gfs2_trans_add_bh(ip->i_gl, dibh, 1);
 506         di = (struct gfs2_dinode *)dibh->b_data;
 507
 508         if (gfs2_is_stuffed(ip)) {
 509                 u64 file_size;
 510                 void *kaddr;
 511
 512                 file_size = ((u64)page->index << PAGE_CACHE_SHIFT) + to;
 513
 514                 kaddr = kmap_atomic(page, KM_USER0);
 515                 memcpy(dibh->b_data + sizeof(struct gfs2_dinode) + from,
 516                        kaddr + from, to - from);
 517                 kunmap_atomic(kaddr, KM_USER0);
 518
 519                 SetPageUptodate(page);
 520
 521                 if (inode->i_size < file_size) {
 522                         i_size_write(inode, file_size);
 523                         mark_inode_dirty(inode);
 524                 }
 525         } else {
 526                 if (sdp->sd_args.ar_data == GFS2_DATA_ORDERED ||
 527                     gfs2_is_jdata(ip))
 528                         gfs2_page_add_databufs(ip, page, from, to);
 529                 error = generic_commit_write(file, page, from, to);
 530                 if (error)
 531                         goto fail;
 532         }
 533
 534         if (ip->i_di.di_size < inode->i_size) {
 535                 ip->i_di.di_size = inode->i_size;
 536                 di->di_size = cpu_to_be64(inode->i_size);
 537         }
 538
 539         if (inode == sdp->sd_rindex)
 540                 adjust_fs_space(inode);
 541
 542         brelse(dibh);
 543         gfs2_trans_end(sdp);
 544         if (al->al_requested) {
 545                 gfs2_inplace_release(ip);
 546                 gfs2_quota_unlock(ip);
 547                 gfs2_alloc_put(ip);
 548         }
 549         unlock_page(page);
 550         gfs2_glock_dq_m(1, &ip->i_gh);
 551         lock_page(page);
 552         gfs2_holder_uninit(&ip->i_gh);
 553         return 0;
 554
 555 fail:
 556         brelse(dibh);
 557 fail_endtrans:
 558         gfs2_trans_end(sdp);
 559         if (al->al_requested) {
 560                 gfs2_inplace_release(ip);
 561                 gfs2_quota_unlock(ip);
 562                 gfs2_alloc_put(ip);
 563         }
 564         unlock_page(page);
 565         gfs2_glock_dq_m(1, &ip->i_gh);
 566         lock_page(page);
 567         gfs2_holder_uninit(&ip->i_gh);
 568 fail_nounlock:
 569         ClearPageUptodate(page);
 570         return error;
 571 }
 572
 573 /**
 574  * gfs2_bmap - Block map function
 575  * @mapping: Address space info
 576  * @lblock: The block to map
 577  *
 578  * Returns: The disk address for the block or 0 on hole or error
 579  */
 580
 581 static sector_t gfs2_bmap(struct address_space *mapping, sector_t lblock)
 582 {
 583         struct gfs2_inode *ip = GFS2_I(mapping->host);
 584         struct gfs2_holder i_gh;
 585         sector_t dblock = 0;
 586         int error;
 587
 588         error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
 589         if (error)
 590                 return 0;
 591
 592         if (!gfs2_is_stuffed(ip))
 593                 dblock = generic_block_bmap(mapping, lblock, gfs2_get_block);
 594
 595         gfs2_glock_dq_uninit(&i_gh);
 596
 597         return dblock;
 598 }
 599
 600 static void discard_buffer(struct gfs2_sbd *sdp, struct buffer_head *bh)
 601 {
 602         struct gfs2_bufdata *bd;
 603
 604         gfs2_log_lock(sdp);
 605         bd = bh->b_private;
 606         if (bd) {
 607                 bd->bd_bh = NULL;
 608                 bh->b_private = NULL;
 609         }
 610         gfs2_log_unlock(sdp);
 611
 612         lock_buffer(bh);
 613         clear_buffer_dirty(bh);
 614         bh->b_bdev = NULL;
 615         clear_buffer_mapped(bh);
 616         clear_buffer_req(bh);
 617         clear_buffer_new(bh);
 618         clear_buffer_delay(bh);
 619         unlock_buffer(bh);
 620 }
 621
 622 static void gfs2_invalidatepage(struct page *page, unsigned long offset)
 623 {
 624         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
 625         struct buffer_head *head, *bh, *next;
 626         unsigned int curr_off = 0;
 627
 628         BUG_ON(!PageLocked(page));
 629         if (!page_has_buffers(page))
 630                 return;
 631
 632         bh = head = page_buffers(page);
 633         do {
 634                 unsigned int next_off = curr_off + bh->b_size;
 635                 next = bh->b_this_page;
 636
 637                 if (offset <= curr_off)
 638                         discard_buffer(sdp, bh);
 639
 640                 curr_off = next_off;
 641                 bh = next;
 642         } while (bh != head);
 643
 644         if (!offset)
 645                 try_to_release_page(page, 0);
 646
 647         return;
 648 }
 649
 650 /**
 651  * gfs2_ok_for_dio - check that dio is valid on this file
 652  * @ip: The inode
 653  * @rw: READ or WRITE
 654  * @offset: The offset at which we are reading or writing
 655  *
 656  * Returns: 0 (to ignore the i/o request and thus fall back to buffered i/o)
 657  *          1 (to accept the i/o request)
 658  */
 659 static int gfs2_ok_for_dio(struct gfs2_inode *ip, int rw, loff_t offset)
 660 {
 661         /*
 662          * Should we return an error here? I can't see that O_DIRECT for
 663          * a journaled file makes any sense. For now we'll silently fall
 664          * back to buffered I/O, likewise we do the same for stuffed
 665          * files since they are (a) small and (b) unaligned.
 666          */
 667         if (gfs2_is_jdata(ip))
 668                 return 0;
 669
 670         if (gfs2_is_stuffed(ip))
 671                 return 0;
 672
 673         if (offset > i_size_read(&ip->i_inode))
 674                 return 0;
 675         return 1;
 676 }
 677
 678
 679
 680 static ssize_t gfs2_direct_IO(int rw, struct kiocb *iocb,
 681                               const struct iovec *iov, loff_t offset,
 682                               unsigned long nr_segs)
 683 {
 684         struct file *file = iocb->ki_filp;
 685         struct inode *inode = file->f_mapping->host;
 686         struct gfs2_inode *ip = GFS2_I(inode);
 687         struct gfs2_holder gh;
 688         int rv;
 689
 690         /*
 691          * Deferred lock, even if its a write, since we do no allocation
 692          * on this path. All we need change is atime, and this lock mode
 693          * ensures that other nodes have flushed their buffered read caches
 694          * (i.e. their page cache entries for this inode). We do not,
 695          * unfortunately have the option of only flushing a range like
 696          * the VFS does.
 697          */
 698         gfs2_holder_init(ip->i_gl, LM_ST_DEFERRED, GL_ATIME, &gh);
 699         rv = gfs2_glock_nq_atime(&gh);
 700         if (rv)
 701                 return rv;
 702         rv = gfs2_ok_for_dio(ip, rw, offset);
 703         if (rv != 1)
 704                 goto out; /* dio not valid, fall back to buffered i/o */
 705
 706         rv = blockdev_direct_IO_no_locking(rw, iocb, inode, inode->i_sb->s_bdev,
 707                                            iov, offset, nr_segs,
 708                                            gfs2_get_block_direct, NULL);
 709 out:
 710         gfs2_glock_dq_m(1, &gh);
 711         gfs2_holder_uninit(&gh);
 712         return rv;
 713 }
 714
 715 /**
 716  * stuck_releasepage - We're stuck in gfs2_releasepage().  Print stuff out.
 717  * @bh: the buffer we're stuck on
 718  *
 719  */
 720
 721 static void stuck_releasepage(struct buffer_head *bh)
 722 {
 723         struct inode *inode = bh->b_page->mapping->host;
 724         struct gfs2_sbd *sdp = inode->i_sb->s_fs_info;
 725         struct gfs2_bufdata *bd = bh->b_private;
 726         struct gfs2_glock *gl;
 727 static unsigned limit = 0;
 728
 729         if (limit > 3)
 730                 return;
 731         limit++;
 732
 733         fs_warn(sdp, "stuck in gfs2_releasepage() %p\n", inode);
 734         fs_warn(sdp, "blkno = %llu, bh->b_count = %d\n",
 735                 (unsigned long long)bh->b_blocknr, atomic_read(&bh->b_count));
 736         fs_warn(sdp, "pinned = %u\n", buffer_pinned(bh));
 737         fs_warn(sdp, "bh->b_private = %s\n", (bd) ? "!NULL" : "NULL");
 738
 739         if (!bd)
 740                 return;
 741
 742         gl = bd->bd_gl;
 743
 744         fs_warn(sdp, "gl = (%u, %llu)\n",
 745                 gl->gl_name.ln_type, (unsigned long long)gl->gl_name.ln_number);
 746
 747         fs_warn(sdp, "bd_list_tr = %s, bd_le.le_list = %s\n",
 748                 (list_empty(&bd->bd_list_tr)) ? "no" : "yes",
 749                 (list_empty(&bd->bd_le.le_list)) ? "no" : "yes");
 750
 751         if (gl->gl_ops == &gfs2_inode_glops) {
 752                 struct gfs2_inode *ip = gl->gl_object;
 753                 unsigned int x;
 754
 755                 if (!ip)
 756                         return;
 757
 758                 fs_warn(sdp, "ip = %llu %llu\n",
 759                         (unsigned long long)ip->i_num.no_formal_ino,
 760                         (unsigned long long)ip->i_num.no_addr);
 761
 762                 for (x = 0; x < GFS2_MAX_META_HEIGHT; x++)
 763                         fs_warn(sdp, "ip->i_cache[%u] = %s\n",
 764                                 x, (ip->i_cache[x]) ? "!NULL" : "NULL");
 765         }
 766 }
 767
 768 /**
 769  * gfs2_releasepage - free the metadata associated with a page
 770  * @page: the page that's being released
 771  * @gfp_mask: passed from Linux VFS, ignored by us
 772  *
 773  * Call try_to_free_buffers() if the buffers in this page can be
 774  * released.
 775  *
 776  * Returns: 0
 777  */
 778
 779 int gfs2_releasepage(struct page *page, gfp_t gfp_mask)
 780 {
 781         struct inode *aspace = page->mapping->host;
 782         struct gfs2_sbd *sdp = aspace->i_sb->s_fs_info;
 783         struct buffer_head *bh, *head;
 784         struct gfs2_bufdata *bd;
 785         unsigned long t = jiffies + gfs2_tune_get(sdp, gt_stall_secs) * HZ;
 786
 787         if (!page_has_buffers(page))
 788                 goto out;
 789
 790         head = bh = page_buffers(page);
 791         do {
 792                 while (atomic_read(&bh->b_count)) {
 793                         if (!atomic_read(&aspace->i_writecount))
 794                                 return 0;
 795
 796                         if (!(gfp_mask & __GFP_WAIT))
 797                                 return 0;
 798
 799                         if (time_after_eq(jiffies, t)) {
 800                                 stuck_releasepage(bh);
 801                                 /* should we withdraw here? */
 802                                 return 0;
 803                         }
 804
 805                         yield();
 806                 }
 807
 808                 gfs2_assert_warn(sdp, !buffer_pinned(bh));
 809                 gfs2_assert_warn(sdp, !buffer_dirty(bh));
 810
 811                 gfs2_log_lock(sdp);
 812                 bd = bh->b_private;
 813                 if (bd) {
 814                         gfs2_assert_warn(sdp, bd->bd_bh == bh);
 815                         gfs2_assert_warn(sdp, list_empty(&bd->bd_list_tr));
 816                         gfs2_assert_warn(sdp, !bd->bd_ail);
 817                         bd->bd_bh = NULL;
 818                         if (!list_empty(&bd->bd_le.le_list))
 819                                 bd = NULL;
 820                         bh->b_private = NULL;
 821                 }
 822                 gfs2_log_unlock(sdp);
 823                 if (bd)
 824                         kmem_cache_free(gfs2_bufdata_cachep, bd);
 825
 826                 bh = bh->b_this_page;
 827         } while (bh != head);
 828
 829 out:
 830         return try_to_free_buffers(page);
 831 }
 832
 833 const struct address_space_operations gfs2_file_aops = {
 834         .writepage = gfs2_writepage,
 835         .writepages = gfs2_writepages,
 836         .readpage = gfs2_readpage,
 837         .readpages = gfs2_readpages,
 838         .sync_page = block_sync_page,
 839         .prepare_write = gfs2_prepare_write,
 840         .commit_write = gfs2_commit_write,
 841         .bmap = gfs2_bmap,
 842         .invalidatepage = gfs2_invalidatepage,
 843         .releasepage = gfs2_releasepage,
 844         .direct_IO = gfs2_direct_IO,
 845 };
 846