fs/nilfs2/page.c

   1 /*
   2  * page.c - buffer/page management specific to NILFS
   3  *
   4  * Copyright (C) 2005-2008 Nippon Telegraph and Telephone Corporation.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License as published by
   8  * the Free Software Foundation; either version 2 of the License, or
   9  * (at your option) any later version.
  10  *
  11  * This program is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14  * GNU General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  19  *
  20  * Written by Ryusuke Konishi <ryusuke@osrg.net>,
  21  *            Seiji Kihara <kihara@osrg.net>.
  22  */
  23
  24 #include <linux/pagemap.h>
  25 #include <linux/writeback.h>
  26 #include <linux/swap.h>
  27 #include <linux/bitops.h>
  28 #include <linux/page-flags.h>
  29 #include <linux/list.h>
  30 #include <linux/highmem.h>
  31 #include <linux/pagevec.h>
  32 #include <linux/gfp.h>
  33 #include "nilfs.h"
  34 #include "page.h"
  35 #include "mdt.h"
  36
  37
  38 #define NILFS_BUFFER_INHERENT_BITS  \
  39         ((1UL << BH_Uptodate) | (1UL << BH_Mapped) | (1UL << BH_NILFS_Node) | \
  40          (1UL << BH_NILFS_Volatile) | (1UL << BH_NILFS_Checked))
  41
  42 static struct buffer_head *
  43 __nilfs_get_page_block(struct page *page, unsigned long block, pgoff_t index,
  44                        int blkbits, unsigned long b_state)
  45
  46 {
  47         unsigned long first_block;
  48         struct buffer_head *bh;
  49
  50         if (!page_has_buffers(page))
  51                 create_empty_buffers(page, 1 << blkbits, b_state);
  52
  53         first_block = (unsigned long)index << (PAGE_CACHE_SHIFT - blkbits);
  54         bh = nilfs_page_get_nth_block(page, block - first_block);
  55
  56         touch_buffer(bh);
  57         wait_on_buffer(bh);
  58         return bh;
  59 }
  60
  61 struct buffer_head *nilfs_grab_buffer(struct inode *inode,
  62                                       struct address_space *mapping,
  63                                       unsigned long blkoff,
  64                                       unsigned long b_state)
  65 {
  66         int blkbits = inode->i_blkbits;
  67         pgoff_t index = blkoff >> (PAGE_CACHE_SHIFT - blkbits);
  68         struct page *page;
  69         struct buffer_head *bh;
  70
  71         page = grab_cache_page(mapping, index);
  72         if (unlikely(!page))
  73                 return NULL;
  74
  75         bh = __nilfs_get_page_block(page, blkoff, index, blkbits, b_state);
  76         if (unlikely(!bh)) {
  77                 unlock_page(page);
  78                 page_cache_release(page);
  79                 return NULL;
  80         }
  81         return bh;
  82 }
  83
  84 /**
  85  * nilfs_forget_buffer - discard dirty state
  86  * @inode: owner inode of the buffer
  87  * @bh: buffer head of the buffer to be discarded
  88  */
  89 void nilfs_forget_buffer(struct buffer_head *bh)
  90 {
  91         struct page *page = bh->b_page;
  92
  93         lock_buffer(bh);
  94         clear_buffer_nilfs_volatile(bh);
  95         clear_buffer_nilfs_checked(bh);
  96         clear_buffer_nilfs_redirected(bh);
  97         clear_buffer_async_write(bh);
  98         clear_buffer_dirty(bh);
  99         if (nilfs_page_buffers_clean(page))
 100                 __nilfs_clear_page_dirty(page);
 101
 102         clear_buffer_uptodate(bh);
 103         clear_buffer_mapped(bh);
 104         bh->b_blocknr = -1;
 105         ClearPageUptodate(page);
 106         ClearPageMappedToDisk(page);
 107         unlock_buffer(bh);
 108         brelse(bh);
 109 }
 110
 111 /**
 112  * nilfs_copy_buffer -- copy buffer data and flags
 113  * @dbh: destination buffer
 114  * @sbh: source buffer
 115  */
 116 void nilfs_copy_buffer(struct buffer_head *dbh, struct buffer_head *sbh)
 117 {
 118         void *kaddr0, *kaddr1;
 119         unsigned long bits;
 120         struct page *spage = sbh->b_page, *dpage = dbh->b_page;
 121         struct buffer_head *bh;
 122
 123         kaddr0 = kmap_atomic(spage, KM_USER0);
 124         kaddr1 = kmap_atomic(dpage, KM_USER1);
 125         memcpy(kaddr1 + bh_offset(dbh), kaddr0 + bh_offset(sbh), sbh->b_size);
 126         kunmap_atomic(kaddr1, KM_USER1);
 127         kunmap_atomic(kaddr0, KM_USER0);
 128
 129         dbh->b_state = sbh->b_state & NILFS_BUFFER_INHERENT_BITS;
 130         dbh->b_blocknr = sbh->b_blocknr;
 131         dbh->b_bdev = sbh->b_bdev;
 132
 133         bh = dbh;
 134         bits = sbh->b_state & ((1UL << BH_Uptodate) | (1UL << BH_Mapped));
 135         while ((bh = bh->b_this_page) != dbh) {
 136                 lock_buffer(bh);
 137                 bits &= bh->b_state;
 138                 unlock_buffer(bh);
 139         }
 140         if (bits & (1UL << BH_Uptodate))
 141                 SetPageUptodate(dpage);
 142         else
 143                 ClearPageUptodate(dpage);
 144         if (bits & (1UL << BH_Mapped))
 145                 SetPageMappedToDisk(dpage);
 146         else
 147                 ClearPageMappedToDisk(dpage);
 148 }
 149
 150 /**
 151  * nilfs_page_buffers_clean - check if a page has dirty buffers or not.
 152  * @page: page to be checked
 153  *
 154  * nilfs_page_buffers_clean() returns zero if the page has dirty buffers.
 155  * Otherwise, it returns non-zero value.
 156  */
 157 int nilfs_page_buffers_clean(struct page *page)
 158 {
 159         struct buffer_head *bh, *head;
 160
 161         bh = head = page_buffers(page);
 162         do {
 163                 if (buffer_dirty(bh))
 164                         return 0;
 165                 bh = bh->b_this_page;
 166         } while (bh != head);
 167         return 1;
 168 }
 169
 170 void nilfs_page_bug(struct page *page)
 171 {
 172         struct address_space *m;
 173         unsigned long ino;
 174
 175         if (unlikely(!page)) {
 176                 printk(KERN_CRIT "NILFS_PAGE_BUG(NULL)\n");
 177                 return;
 178         }
 179
 180         m = page->mapping;
 181         ino = m ? m->host->i_ino : 0;
 182
 183         printk(KERN_CRIT "NILFS_PAGE_BUG(%p): cnt=%d index#=%llu flags=0x%lx "
 184                "mapping=%p ino=%lu\n",
 185                page, atomic_read(&page->_count),
 186                (unsigned long long)page->index, page->flags, m, ino);
 187
 188         if (page_has_buffers(page)) {
 189                 struct buffer_head *bh, *head;
 190                 int i = 0;
 191
 192                 bh = head = page_buffers(page);
 193                 do {
 194                         printk(KERN_CRIT
 195                                " BH[%d] %p: cnt=%d block#=%llu state=0x%lx\n",
 196                                i++, bh, atomic_read(&bh->b_count),
 197                                (unsigned long long)bh->b_blocknr, bh->b_state);
 198                         bh = bh->b_this_page;
 199                 } while (bh != head);
 200         }
 201 }
 202
 203 /**
 204  * nilfs_copy_page -- copy the page with buffers
 205  * @dst: destination page
 206  * @src: source page
 207  * @copy_dirty: flag whether to copy dirty states on the page's buffer heads.
 208  *
 209  * This function is for both data pages and btnode pages.  The dirty flag
 210  * should be treated by caller.  The page must not be under i/o.
 211  * Both src and dst page must be locked
 212  */
 213 static void nilfs_copy_page(struct page *dst, struct page *src, int copy_dirty)
 214 {
 215         struct buffer_head *dbh, *dbufs, *sbh, *sbufs;
 216         unsigned long mask = NILFS_BUFFER_INHERENT_BITS;
 217
 218         BUG_ON(PageWriteback(dst));
 219
 220         sbh = sbufs = page_buffers(src);
 221         if (!page_has_buffers(dst))
 222                 create_empty_buffers(dst, sbh->b_size, 0);
 223
 224         if (copy_dirty)
 225                 mask |= (1UL << BH_Dirty);
 226
 227         dbh = dbufs = page_buffers(dst);
 228         do {
 229                 lock_buffer(sbh);
 230                 lock_buffer(dbh);
 231                 dbh->b_state = sbh->b_state & mask;
 232                 dbh->b_blocknr = sbh->b_blocknr;
 233                 dbh->b_bdev = sbh->b_bdev;
 234                 sbh = sbh->b_this_page;
 235                 dbh = dbh->b_this_page;
 236         } while (dbh != dbufs);
 237
 238         copy_highpage(dst, src);
 239
 240         if (PageUptodate(src) && !PageUptodate(dst))
 241                 SetPageUptodate(dst);
 242         else if (!PageUptodate(src) && PageUptodate(dst))
 243                 ClearPageUptodate(dst);
 244         if (PageMappedToDisk(src) && !PageMappedToDisk(dst))
 245                 SetPageMappedToDisk(dst);
 246         else if (!PageMappedToDisk(src) && PageMappedToDisk(dst))
 247                 ClearPageMappedToDisk(dst);
 248
 249         do {
 250                 unlock_buffer(sbh);
 251                 unlock_buffer(dbh);
 252                 sbh = sbh->b_this_page;
 253                 dbh = dbh->b_this_page;
 254         } while (dbh != dbufs);
 255 }
 256
 257 int nilfs_copy_dirty_pages(struct address_space *dmap,
 258                            struct address_space *smap)
 259 {
 260         struct pagevec pvec;
 261         unsigned int i;
 262         pgoff_t index = 0;
 263         int err = 0;
 264
 265         pagevec_init(&pvec, 0);
 266 repeat:
 267         if (!pagevec_lookup_tag(&pvec, smap, &index, PAGECACHE_TAG_DIRTY,
 268                                 PAGEVEC_SIZE))
 269                 return 0;
 270
 271         for (i = 0; i < pagevec_count(&pvec); i++) {
 272                 struct page *page = pvec.pages[i], *dpage;
 273
 274                 lock_page(page);
 275                 if (unlikely(!PageDirty(page)))
 276                         NILFS_PAGE_BUG(page, "inconsistent dirty state");
 277
 278                 dpage = grab_cache_page(dmap, page->index);
 279                 if (unlikely(!dpage)) {
 280                         /* No empty page is added to the page cache */
 281                         err = -ENOMEM;
 282                         unlock_page(page);
 283                         break;
 284                 }
 285                 if (unlikely(!page_has_buffers(page)))
 286                         NILFS_PAGE_BUG(page,
 287                                        "found empty page in dat page cache");
 288
 289                 nilfs_copy_page(dpage, page, 1);
 290                 __set_page_dirty_nobuffers(dpage);
 291
 292                 unlock_page(dpage);
 293                 page_cache_release(dpage);
 294                 unlock_page(page);
 295         }
 296         pagevec_release(&pvec);
 297         cond_resched();
 298
 299         if (likely(!err))
 300                 goto repeat;
 301         return err;
 302 }
 303
 304 /**
 305  * nilfs_copy_back_pages -- copy back pages to original cache from shadow cache
 306  * @dmap: destination page cache
 307  * @smap: source page cache
 308  *
 309  * No pages must no be added to the cache during this process.
 310  * This must be ensured by the caller.
 311  */
 312 void nilfs_copy_back_pages(struct address_space *dmap,
 313                            struct address_space *smap)
 314 {
 315         struct pagevec pvec;
 316         unsigned int i, n;
 317         pgoff_t index = 0;
 318         int err;
 319
 320         pagevec_init(&pvec, 0);
 321 repeat:
 322         n = pagevec_lookup(&pvec, smap, index, PAGEVEC_SIZE);
 323         if (!n)
 324                 return;
 325         index = pvec.pages[n - 1]->index + 1;
 326
 327         for (i = 0; i < pagevec_count(&pvec); i++) {
 328                 struct page *page = pvec.pages[i], *dpage;
 329                 pgoff_t offset = page->index;
 330
 331                 lock_page(page);
 332                 dpage = find_lock_page(dmap, offset);
 333                 if (dpage) {
 334                         /* override existing page on the destination cache */
 335                         WARN_ON(PageDirty(dpage));
 336                         nilfs_copy_page(dpage, page, 0);
 337                         unlock_page(dpage);
 338                         page_cache_release(dpage);
 339                 } else {
 340                         struct page *page2;
 341
 342                         /* move the page to the destination cache */
 343                         spin_lock_irq(&smap->tree_lock);
 344                         page2 = radix_tree_delete(&smap->page_tree, offset);
 345                         WARN_ON(page2 != page);
 346
 347                         smap->nrpages--;
 348                         spin_unlock_irq(&smap->tree_lock);
 349
 350                         spin_lock_irq(&dmap->tree_lock);
 351                         err = radix_tree_insert(&dmap->page_tree, offset, page);
 352                         if (unlikely(err < 0)) {
 353                                 WARN_ON(err == -EEXIST);
 354                                 page->mapping = NULL;
 355                                 page_cache_release(page); /* for cache */
 356                         } else {
 357                                 page->mapping = dmap;
 358                                 dmap->nrpages++;
 359                                 if (PageDirty(page))
 360                                         radix_tree_tag_set(&dmap->page_tree,
 361                                                            offset,
 362                                                            PAGECACHE_TAG_DIRTY);
 363                         }
 364                         spin_unlock_irq(&dmap->tree_lock);
 365                 }
 366                 unlock_page(page);
 367         }
 368         pagevec_release(&pvec);
 369         cond_resched();
 370
 371         goto repeat;
 372 }
 373
 374 void nilfs_clear_dirty_pages(struct address_space *mapping)
 375 {
 376         struct pagevec pvec;
 377         unsigned int i;
 378         pgoff_t index = 0;
 379
 380         pagevec_init(&pvec, 0);
 381
 382         while (pagevec_lookup_tag(&pvec, mapping, &index, PAGECACHE_TAG_DIRTY,
 383                                   PAGEVEC_SIZE)) {
 384                 for (i = 0; i < pagevec_count(&pvec); i++) {
 385                         struct page *page = pvec.pages[i];
 386                         struct buffer_head *bh, *head;
 387
 388                         lock_page(page);
 389                         ClearPageUptodate(page);
 390                         ClearPageMappedToDisk(page);
 391                         bh = head = page_buffers(page);
 392                         do {
 393                                 lock_buffer(bh);
 394                                 clear_buffer_async_write(bh);
 395                                 clear_buffer_dirty(bh);
 396                                 clear_buffer_nilfs_volatile(bh);
 397                                 clear_buffer_nilfs_checked(bh);
 398                                 clear_buffer_nilfs_redirected(bh);
 399                                 clear_buffer_uptodate(bh);
 400                                 clear_buffer_mapped(bh);
 401                                 unlock_buffer(bh);
 402                                 bh = bh->b_this_page;
 403                         } while (bh != head);
 404
 405                         __nilfs_clear_page_dirty(page);
 406                         unlock_page(page);
 407                 }
 408                 pagevec_release(&pvec);
 409                 cond_resched();
 410         }
 411 }
 412
 413 unsigned nilfs_page_count_clean_buffers(struct page *page,
 414                                         unsigned from, unsigned to)
 415 {
 416         unsigned block_start, block_end;
 417         struct buffer_head *bh, *head;
 418         unsigned nc = 0;
 419
 420         for (bh = head = page_buffers(page), block_start = 0;
 421              bh != head || !block_start;
 422              block_start = block_end, bh = bh->b_this_page) {
 423                 block_end = block_start + bh->b_size;
 424                 if (block_end > from && block_start < to && !buffer_dirty(bh))
 425                         nc++;
 426         }
 427         return nc;
 428 }
 429
 430 void nilfs_mapping_init(struct address_space *mapping, struct inode *inode,
 431                         struct backing_dev_info *bdi)
 432 {
 433         mapping->host = inode;
 434         mapping->flags = 0;
 435         mapping_set_gfp_mask(mapping, GFP_NOFS);
 436         mapping->assoc_mapping = NULL;
 437         mapping->backing_dev_info = bdi;
 438         mapping->a_ops = &empty_aops;
 439 }
 440
 441 /*
 442  * NILFS2 needs clear_page_dirty() in the following two cases:
 443  *
 444  * 1) For B-tree node pages and data pages of the dat/gcdat, NILFS2 clears
 445  *    page dirty flags when it copies back pages from the shadow cache
 446  *    (gcdat->{i_mapping,i_btnode_cache}) to its original cache
 447  *    (dat->{i_mapping,i_btnode_cache}).
 448  *
 449  * 2) Some B-tree operations like insertion or deletion may dispose buffers
 450  *    in dirty state, and this needs to cancel the dirty state of their pages.
 451  */
 452 int __nilfs_clear_page_dirty(struct page *page)
 453 {
 454         struct address_space *mapping = page->mapping;
 455
 456         if (mapping) {
 457                 spin_lock_irq(&mapping->tree_lock);
 458                 if (test_bit(PG_dirty, &page->flags)) {
 459                         radix_tree_tag_clear(&mapping->page_tree,
 460                                              page_index(page),
 461                                              PAGECACHE_TAG_DIRTY);
 462                         spin_unlock_irq(&mapping->tree_lock);
 463                         return clear_page_dirty_for_io(page);
 464                 }
 465                 spin_unlock_irq(&mapping->tree_lock);
 466                 return 0;
 467         }
 468         return TestClearPageDirty(page);
 469 }
 470
 471 /**
 472  * nilfs_find_uncommitted_extent - find extent of uncommitted data
 473  * @inode: inode
 474  * @start_blk: start block offset (in)
 475  * @blkoff: start offset of the found extent (out)
 476  *
 477  * This function searches an extent of buffers marked "delayed" which
 478  * starts from a block offset equal to or larger than @start_blk.  If
 479  * such an extent was found, this will store the start offset in
 480  * @blkoff and return its length in blocks.  Otherwise, zero is
 481  * returned.
 482  */
 483 unsigned long nilfs_find_uncommitted_extent(struct inode *inode,
 484                                             sector_t start_blk,
 485                                             sector_t *blkoff)
 486 {
 487         unsigned int i;
 488         pgoff_t index;
 489         unsigned int nblocks_in_page;
 490         unsigned long length = 0;
 491         sector_t b;
 492         struct pagevec pvec;
 493         struct page *page;
 494
 495         if (inode->i_mapping->nrpages == 0)
 496                 return 0;
 497
 498         index = start_blk >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
 499         nblocks_in_page = 1U << (PAGE_CACHE_SHIFT - inode->i_blkbits);
 500
 501         pagevec_init(&pvec, 0);
 502
 503 repeat:
 504         pvec.nr = find_get_pages_contig(inode->i_mapping, index, PAGEVEC_SIZE,
 505                                         pvec.pages);
 506         if (pvec.nr == 0)
 507                 return length;
 508
 509         if (length > 0 && pvec.pages[0]->index > index)
 510                 goto out;
 511
 512         b = pvec.pages[0]->index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
 513         i = 0;
 514         do {
 515                 page = pvec.pages[i];
 516
 517                 lock_page(page);
 518                 if (page_has_buffers(page)) {
 519                         struct buffer_head *bh, *head;
 520
 521                         bh = head = page_buffers(page);
 522                         do {
 523                                 if (b < start_blk)
 524                                         continue;
 525                                 if (buffer_delay(bh)) {
 526                                         if (length == 0)
 527                                                 *blkoff = b;
 528                                         length++;
 529                                 } else if (length > 0) {
 530                                         goto out_locked;
 531                                 }
 532                         } while (++b, bh = bh->b_this_page, bh != head);
 533                 } else {
 534                         if (length > 0)
 535                                 goto out_locked;
 536
 537                         b += nblocks_in_page;
 538                 }
 539                 unlock_page(page);
 540
 541         } while (++i < pagevec_count(&pvec));
 542
 543         index = page->index + 1;
 544         pagevec_release(&pvec);
 545         cond_resched();
 546         goto repeat;
 547
 548 out_locked:
 549         unlock_page(page);
 550 out:
 551         pagevec_release(&pvec);
 552         return length;
 553 }