fs/nilfs2/btnode.c

   1 /*
   2  * btnode.c - NILFS B-tree node cache
   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  * This file was originally written by Seiji Kihara <kihara@osrg.net>
  21  * and fully revised by Ryusuke Konishi <ryusuke@osrg.net> for
  22  * stabilization and simplification.
  23  *
  24  */
  25
  26 #include <linux/types.h>
  27 #include <linux/buffer_head.h>
  28 #include <linux/mm.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/gfp.h>
  31 #include "nilfs.h"
  32 #include "mdt.h"
  33 #include "dat.h"
  34 #include "page.h"
  35 #include "btnode.h"
  36
  37
  38 void nilfs_btnode_cache_init_once(struct address_space *btnc)
  39 {
  40         nilfs_mapping_init_once(btnc);
  41 }
  42
  43 void nilfs_btnode_cache_init(struct address_space *btnc,
  44                              struct backing_dev_info *bdi)
  45 {
  46         nilfs_mapping_init(btnc, bdi);
  47 }
  48
  49 void nilfs_btnode_cache_clear(struct address_space *btnc)
  50 {
  51         invalidate_mapping_pages(btnc, 0, -1);
  52         truncate_inode_pages(btnc, 0);
  53 }
  54
  55 struct buffer_head *
  56 nilfs_btnode_create_block(struct address_space *btnc, __u64 blocknr)
  57 {
  58         struct inode *inode = NILFS_BTNC_I(btnc);
  59         struct buffer_head *bh;
  60
  61         bh = nilfs_grab_buffer(inode, btnc, blocknr, 1 << BH_NILFS_Node);
  62         if (unlikely(!bh))
  63                 return NULL;
  64
  65         if (unlikely(buffer_mapped(bh) || buffer_uptodate(bh) ||
  66                      buffer_dirty(bh))) {
  67                 brelse(bh);
  68                 BUG();
  69         }
  70         memset(bh->b_data, 0, 1 << inode->i_blkbits);
  71         bh->b_bdev = NILFS_I_NILFS(inode)->ns_bdev;
  72         bh->b_blocknr = blocknr;
  73         set_buffer_mapped(bh);
  74         set_buffer_uptodate(bh);
  75
  76         unlock_page(bh->b_page);
  77         page_cache_release(bh->b_page);
  78         return bh;
  79 }
  80
  81 int nilfs_btnode_submit_block(struct address_space *btnc, __u64 blocknr,
  82                               sector_t pblocknr, int mode,
  83                               struct buffer_head **pbh, sector_t *submit_ptr)
  84 {
  85         struct buffer_head *bh;
  86         struct inode *inode = NILFS_BTNC_I(btnc);
  87         struct page *page;
  88         int err;
  89
  90         bh = nilfs_grab_buffer(inode, btnc, blocknr, 1 << BH_NILFS_Node);
  91         if (unlikely(!bh))
  92                 return -ENOMEM;
  93
  94         err = -EEXIST; /* internal code */
  95         page = bh->b_page;
  96
  97         if (buffer_uptodate(bh) || buffer_dirty(bh))
  98                 goto found;
  99
 100         if (pblocknr == 0) {
 101                 pblocknr = blocknr;
 102                 if (inode->i_ino != NILFS_DAT_INO) {
 103                         struct inode *dat = NILFS_I_NILFS(inode)->ns_dat;
 104
 105                         /* blocknr is a virtual block number */
 106                         err = nilfs_dat_translate(dat, blocknr, &pblocknr);
 107                         if (unlikely(err)) {
 108                                 brelse(bh);
 109                                 goto out_locked;
 110                         }
 111                 }
 112         }
 113
 114         if (mode == READA) {
 115                 if (pblocknr != *submit_ptr + 1 || !trylock_buffer(bh)) {
 116                         err = -EBUSY; /* internal code */
 117                         brelse(bh);
 118                         goto out_locked;
 119                 }
 120         } else { /* mode == READ */
 121                 lock_buffer(bh);
 122         }
 123         if (buffer_uptodate(bh)) {
 124                 unlock_buffer(bh);
 125                 err = -EEXIST; /* internal code */
 126                 goto found;
 127         }
 128         set_buffer_mapped(bh);
 129         bh->b_bdev = NILFS_I_NILFS(inode)->ns_bdev;
 130         bh->b_blocknr = pblocknr; /* set block address for read */
 131         bh->b_end_io = end_buffer_read_sync;
 132         get_bh(bh);
 133         submit_bh(mode, bh);
 134         bh->b_blocknr = blocknr; /* set back to the given block address */
 135         *submit_ptr = pblocknr;
 136         err = 0;
 137 found:
 138         *pbh = bh;
 139
 140 out_locked:
 141         unlock_page(page);
 142         page_cache_release(page);
 143         return err;
 144 }
 145
 146 /**
 147  * nilfs_btnode_delete - delete B-tree node buffer
 148  * @bh: buffer to be deleted
 149  *
 150  * nilfs_btnode_delete() invalidates the specified buffer and delete the page
 151  * including the buffer if the page gets unbusy.
 152  */
 153 void nilfs_btnode_delete(struct buffer_head *bh)
 154 {
 155         struct address_space *mapping;
 156         struct page *page = bh->b_page;
 157         pgoff_t index = page_index(page);
 158         int still_dirty;
 159
 160         page_cache_get(page);
 161         lock_page(page);
 162         wait_on_page_writeback(page);
 163
 164         nilfs_forget_buffer(bh);
 165         still_dirty = PageDirty(page);
 166         mapping = page->mapping;
 167         unlock_page(page);
 168         page_cache_release(page);
 169
 170         if (!still_dirty && mapping)
 171                 invalidate_inode_pages2_range(mapping, index, index);
 172 }
 173
 174 /**
 175  * nilfs_btnode_prepare_change_key
 176  *  prepare to move contents of the block for old key to one of new key.
 177  *  the old buffer will not be removed, but might be reused for new buffer.
 178  *  it might return -ENOMEM because of memory allocation errors,
 179  *  and might return -EIO because of disk read errors.
 180  */
 181 int nilfs_btnode_prepare_change_key(struct address_space *btnc,
 182                                     struct nilfs_btnode_chkey_ctxt *ctxt)
 183 {
 184         struct buffer_head *obh, *nbh;
 185         struct inode *inode = NILFS_BTNC_I(btnc);
 186         __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
 187         int err;
 188
 189         if (oldkey == newkey)
 190                 return 0;
 191
 192         obh = ctxt->bh;
 193         ctxt->newbh = NULL;
 194
 195         if (inode->i_blkbits == PAGE_CACHE_SHIFT) {
 196                 lock_page(obh->b_page);
 197                 /*
 198                  * We cannot call radix_tree_preload for the kernels older
 199                  * than 2.6.23, because it is not exported for modules.
 200                  */
 201 retry:
 202                 err = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
 203                 if (err)
 204                         goto failed_unlock;
 205                 /* BUG_ON(oldkey != obh->b_page->index); */
 206                 if (unlikely(oldkey != obh->b_page->index))
 207                         NILFS_PAGE_BUG(obh->b_page,
 208                                        "invalid oldkey %lld (newkey=%lld)",
 209                                        (unsigned long long)oldkey,
 210                                        (unsigned long long)newkey);
 211
 212                 spin_lock_irq(&btnc->tree_lock);
 213                 err = radix_tree_insert(&btnc->page_tree, newkey, obh->b_page);
 214                 spin_unlock_irq(&btnc->tree_lock);
 215                 /*
 216                  * Note: page->index will not change to newkey until
 217                  * nilfs_btnode_commit_change_key() will be called.
 218                  * To protect the page in intermediate state, the page lock
 219                  * is held.
 220                  */
 221                 radix_tree_preload_end();
 222                 if (!err)
 223                         return 0;
 224                 else if (err != -EEXIST)
 225                         goto failed_unlock;
 226
 227                 err = invalidate_inode_pages2_range(btnc, newkey, newkey);
 228                 if (!err)
 229                         goto retry;
 230                 /* fallback to copy mode */
 231                 unlock_page(obh->b_page);
 232         }
 233
 234         nbh = nilfs_btnode_create_block(btnc, newkey);
 235         if (!nbh)
 236                 return -ENOMEM;
 237
 238         BUG_ON(nbh == obh);
 239         ctxt->newbh = nbh;
 240         return 0;
 241
 242  failed_unlock:
 243         unlock_page(obh->b_page);
 244         return err;
 245 }
 246
 247 /**
 248  * nilfs_btnode_commit_change_key
 249  *  commit the change_key operation prepared by prepare_change_key().
 250  */
 251 void nilfs_btnode_commit_change_key(struct address_space *btnc,
 252                                     struct nilfs_btnode_chkey_ctxt *ctxt)
 253 {
 254         struct buffer_head *obh = ctxt->bh, *nbh = ctxt->newbh;
 255         __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
 256         struct page *opage;
 257
 258         if (oldkey == newkey)
 259                 return;
 260
 261         if (nbh == NULL) {      /* blocksize == pagesize */
 262                 opage = obh->b_page;
 263                 if (unlikely(oldkey != opage->index))
 264                         NILFS_PAGE_BUG(opage,
 265                                        "invalid oldkey %lld (newkey=%lld)",
 266                                        (unsigned long long)oldkey,
 267                                        (unsigned long long)newkey);
 268                 nilfs_btnode_mark_dirty(obh);
 269
 270                 spin_lock_irq(&btnc->tree_lock);
 271                 radix_tree_delete(&btnc->page_tree, oldkey);
 272                 radix_tree_tag_set(&btnc->page_tree, newkey,
 273                                    PAGECACHE_TAG_DIRTY);
 274                 spin_unlock_irq(&btnc->tree_lock);
 275
 276                 opage->index = obh->b_blocknr = newkey;
 277                 unlock_page(opage);
 278         } else {
 279                 nilfs_copy_buffer(nbh, obh);
 280                 nilfs_btnode_mark_dirty(nbh);
 281
 282                 nbh->b_blocknr = newkey;
 283                 ctxt->bh = nbh;
 284                 nilfs_btnode_delete(obh); /* will decrement bh->b_count */
 285         }
 286 }
 287
 288 /**
 289  * nilfs_btnode_abort_change_key
 290  *  abort the change_key operation prepared by prepare_change_key().
 291  */
 292 void nilfs_btnode_abort_change_key(struct address_space *btnc,
 293                                    struct nilfs_btnode_chkey_ctxt *ctxt)
 294 {
 295         struct buffer_head *nbh = ctxt->newbh;
 296         __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
 297
 298         if (oldkey == newkey)
 299                 return;
 300
 301         if (nbh == NULL) {      /* blocksize == pagesize */
 302                 spin_lock_irq(&btnc->tree_lock);
 303                 radix_tree_delete(&btnc->page_tree, newkey);
 304                 spin_unlock_irq(&btnc->tree_lock);
 305                 unlock_page(ctxt->bh->b_page);
 306         } else
 307                 brelse(nbh);
 308 }