SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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 "nilfs.h"
  31 #include "mdt.h"
  32 #include "dat.h"
  33 #include "page.h"
  34 #include "btnode.h"
  35
  36
  37 void nilfs_btnode_cache_init_once(struct address_space *btnc)
  38 {
  39         memset(btnc, 0, sizeof(*btnc));
  40         INIT_RADIX_TREE(&btnc->page_tree, GFP_ATOMIC);
  41         spin_lock_init(&btnc->tree_lock);
  42         INIT_LIST_HEAD(&btnc->private_list);
  43         spin_lock_init(&btnc->private_lock);
  44
  45         spin_lock_init(&btnc->i_mmap_lock);
  46         INIT_RAW_PRIO_TREE_ROOT(&btnc->i_mmap);
  47         INIT_LIST_HEAD(&btnc->i_mmap_nonlinear);
  48 }
  49
  50 static const struct address_space_operations def_btnode_aops = {
  51         .sync_page              = block_sync_page,
  52 };
  53
  54 void nilfs_btnode_cache_init(struct address_space *btnc,
  55                              struct backing_dev_info *bdi)
  56 {
  57         btnc->host = NULL;  /* can safely set to host inode ? */
  58         btnc->flags = 0;
  59         mapping_set_gfp_mask(btnc, GFP_NOFS);
  60         btnc->assoc_mapping = NULL;
  61         btnc->backing_dev_info = bdi;
  62         btnc->a_ops = &def_btnode_aops;
  63 }
  64
  65 void nilfs_btnode_cache_clear(struct address_space *btnc)
  66 {
  67         invalidate_mapping_pages(btnc, 0, -1);
  68         truncate_inode_pages(btnc, 0);
  69 }
  70
  71 struct buffer_head *
  72 nilfs_btnode_create_block(struct address_space *btnc, __u64 blocknr)
  73 {
  74         struct inode *inode = NILFS_BTNC_I(btnc);
  75         struct buffer_head *bh;
  76
  77         bh = nilfs_grab_buffer(inode, btnc, blocknr, 1 << BH_NILFS_Node);
  78         if (unlikely(!bh))
  79                 return NULL;
  80
  81         if (unlikely(buffer_mapped(bh) || buffer_uptodate(bh) ||
  82                      buffer_dirty(bh))) {
  83                 brelse(bh);
  84                 BUG();
  85         }
  86         memset(bh->b_data, 0, 1 << inode->i_blkbits);
  87         bh->b_bdev = NILFS_I_NILFS(inode)->ns_bdev;
  88         bh->b_blocknr = blocknr;
  89         set_buffer_mapped(bh);
  90         set_buffer_uptodate(bh);
  91
  92         unlock_page(bh->b_page);
  93         page_cache_release(bh->b_page);
  94         return bh;
  95 }
  96
  97 int nilfs_btnode_submit_block(struct address_space *btnc, __u64 blocknr,
  98                               sector_t pblocknr, struct buffer_head **pbh,
  99                               int newblk)
 100 {
 101         struct buffer_head *bh;
 102         struct inode *inode = NILFS_BTNC_I(btnc);
 103         int err;
 104
 105         bh = nilfs_grab_buffer(inode, btnc, blocknr, 1 << BH_NILFS_Node);
 106         if (unlikely(!bh))
 107                 return -ENOMEM;
 108
 109         err = -EEXIST; /* internal code */
 110         if (newblk) {
 111                 if (unlikely(buffer_mapped(bh) || buffer_uptodate(bh) ||
 112                              buffer_dirty(bh))) {
 113                         brelse(bh);
 114                         BUG();
 115                 }
 116                 memset(bh->b_data, 0, 1 << inode->i_blkbits);
 117                 bh->b_bdev = NILFS_I_NILFS(inode)->ns_bdev;
 118                 bh->b_blocknr = blocknr;
 119                 set_buffer_mapped(bh);
 120                 set_buffer_uptodate(bh);
 121                 goto found;
 122         }
 123
 124         if (buffer_uptodate(bh) || buffer_dirty(bh))
 125                 goto found;
 126
 127         if (pblocknr == 0) {
 128                 pblocknr = blocknr;
 129                 if (inode->i_ino != NILFS_DAT_INO) {
 130                         struct inode *dat =
 131                                 nilfs_dat_inode(NILFS_I_NILFS(inode));
 132
 133                         /* blocknr is a virtual block number */
 134                         err = nilfs_dat_translate(dat, blocknr, &pblocknr);
 135                         if (unlikely(err)) {
 136                                 brelse(bh);
 137                                 goto out_locked;
 138                         }
 139                 }
 140         }
 141         lock_buffer(bh);
 142         if (buffer_uptodate(bh)) {
 143                 unlock_buffer(bh);
 144                 err = -EEXIST; /* internal code */
 145                 goto found;
 146         }
 147         set_buffer_mapped(bh);
 148         bh->b_bdev = NILFS_I_NILFS(inode)->ns_bdev;
 149         bh->b_blocknr = pblocknr; /* set block address for read */
 150         bh->b_end_io = end_buffer_read_sync;
 151         get_bh(bh);
 152         submit_bh(READ, bh);
 153         bh->b_blocknr = blocknr; /* set back to the given block address */
 154         err = 0;
 155 found:
 156         *pbh = bh;
 157
 158 out_locked:
 159         unlock_page(bh->b_page);
 160         page_cache_release(bh->b_page);
 161         return err;
 162 }
 163
 164 int nilfs_btnode_get(struct address_space *btnc, __u64 blocknr,
 165                      sector_t pblocknr, struct buffer_head **pbh, int newblk)
 166 {
 167         struct buffer_head *bh;
 168         int err;
 169
 170         err = nilfs_btnode_submit_block(btnc, blocknr, pblocknr, pbh, newblk);
 171         if (err == -EEXIST) /* internal code (cache hit) */
 172                 return 0;
 173         if (unlikely(err))
 174                 return err;
 175
 176         bh = *pbh;
 177         wait_on_buffer(bh);
 178         if (!buffer_uptodate(bh)) {
 179                 brelse(bh);
 180                 return -EIO;
 181         }
 182         return 0;
 183 }
 184
 185 /**
 186  * nilfs_btnode_delete - delete B-tree node buffer
 187  * @bh: buffer to be deleted
 188  *
 189  * nilfs_btnode_delete() invalidates the specified buffer and delete the page
 190  * including the buffer if the page gets unbusy.
 191  */
 192 void nilfs_btnode_delete(struct buffer_head *bh)
 193 {
 194         struct address_space *mapping;
 195         struct page *page = bh->b_page;
 196         pgoff_t index = page_index(page);
 197         int still_dirty;
 198
 199         page_cache_get(page);
 200         lock_page(page);
 201         wait_on_page_writeback(page);
 202
 203         nilfs_forget_buffer(bh);
 204         still_dirty = PageDirty(page);
 205         mapping = page->mapping;
 206         unlock_page(page);
 207         page_cache_release(page);
 208
 209         if (!still_dirty && mapping)
 210                 invalidate_inode_pages2_range(mapping, index, index);
 211 }
 212
 213 /**
 214  * nilfs_btnode_prepare_change_key
 215  *  prepare to move contents of the block for old key to one of new key.
 216  *  the old buffer will not be removed, but might be reused for new buffer.
 217  *  it might return -ENOMEM because of memory allocation errors,
 218  *  and might return -EIO because of disk read errors.
 219  */
 220 int nilfs_btnode_prepare_change_key(struct address_space *btnc,
 221                                     struct nilfs_btnode_chkey_ctxt *ctxt)
 222 {
 223         struct buffer_head *obh, *nbh;
 224         struct inode *inode = NILFS_BTNC_I(btnc);
 225         __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
 226         int err;
 227
 228         if (oldkey == newkey)
 229                 return 0;
 230
 231         obh = ctxt->bh;
 232         ctxt->newbh = NULL;
 233
 234         if (inode->i_blkbits == PAGE_CACHE_SHIFT) {
 235                 lock_page(obh->b_page);
 236                 /*
 237                  * We cannot call radix_tree_preload for the kernels older
 238                  * than 2.6.23, because it is not exported for modules.
 239                  */
 240 retry:
 241                 err = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
 242                 if (err)
 243                         goto failed_unlock;
 244                 /* BUG_ON(oldkey != obh->b_page->index); */
 245                 if (unlikely(oldkey != obh->b_page->index))
 246                         NILFS_PAGE_BUG(obh->b_page,
 247                                        "invalid oldkey %lld (newkey=%lld)",
 248                                        (unsigned long long)oldkey,
 249                                        (unsigned long long)newkey);
 250
 251                 spin_lock_irq(&btnc->tree_lock);
 252                 err = radix_tree_insert(&btnc->page_tree, newkey, obh->b_page);
 253                 spin_unlock_irq(&btnc->tree_lock);
 254                 /*
 255                  * Note: page->index will not change to newkey until
 256                  * nilfs_btnode_commit_change_key() will be called.
 257                  * To protect the page in intermediate state, the page lock
 258                  * is held.
 259                  */
 260                 radix_tree_preload_end();
 261                 if (!err)
 262                         return 0;
 263                 else if (err != -EEXIST)
 264                         goto failed_unlock;
 265
 266                 err = invalidate_inode_pages2_range(btnc, newkey, newkey);
 267                 if (!err)
 268                         goto retry;
 269                 /* fallback to copy mode */
 270                 unlock_page(obh->b_page);
 271         }
 272
 273         nbh = nilfs_btnode_create_block(btnc, newkey);
 274         if (!nbh)
 275                 return -ENOMEM;
 276
 277         BUG_ON(nbh == obh);
 278         ctxt->newbh = nbh;
 279         return 0;
 280
 281  failed_unlock:
 282         unlock_page(obh->b_page);
 283         return err;
 284 }
 285
 286 /**
 287  * nilfs_btnode_commit_change_key
 288  *  commit the change_key operation prepared by prepare_change_key().
 289  */
 290 void nilfs_btnode_commit_change_key(struct address_space *btnc,
 291                                     struct nilfs_btnode_chkey_ctxt *ctxt)
 292 {
 293         struct buffer_head *obh = ctxt->bh, *nbh = ctxt->newbh;
 294         __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
 295         struct page *opage;
 296
 297         if (oldkey == newkey)
 298                 return;
 299
 300         if (nbh == NULL) {      /* blocksize == pagesize */
 301                 opage = obh->b_page;
 302                 if (unlikely(oldkey != opage->index))
 303                         NILFS_PAGE_BUG(opage,
 304                                        "invalid oldkey %lld (newkey=%lld)",
 305                                        (unsigned long long)oldkey,
 306                                        (unsigned long long)newkey);
 307                 nilfs_btnode_mark_dirty(obh);
 308
 309                 spin_lock_irq(&btnc->tree_lock);
 310                 radix_tree_delete(&btnc->page_tree, oldkey);
 311                 radix_tree_tag_set(&btnc->page_tree, newkey,
 312                                    PAGECACHE_TAG_DIRTY);
 313                 spin_unlock_irq(&btnc->tree_lock);
 314
 315                 opage->index = obh->b_blocknr = newkey;
 316                 unlock_page(opage);
 317         } else {
 318                 nilfs_copy_buffer(nbh, obh);
 319                 nilfs_btnode_mark_dirty(nbh);
 320
 321                 nbh->b_blocknr = newkey;
 322                 ctxt->bh = nbh;
 323                 nilfs_btnode_delete(obh); /* will decrement bh->b_count */
 324         }
 325 }
 326
 327 /**
 328  * nilfs_btnode_abort_change_key
 329  *  abort the change_key operation prepared by prepare_change_key().
 330  */
 331 void nilfs_btnode_abort_change_key(struct address_space *btnc,
 332                                    struct nilfs_btnode_chkey_ctxt *ctxt)
 333 {
 334         struct buffer_head *nbh = ctxt->newbh;
 335         __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
 336
 337         if (oldkey == newkey)
 338                 return;
 339
 340         if (nbh == NULL) {      /* blocksize == pagesize */
 341                 spin_lock_irq(&btnc->tree_lock);
 342                 radix_tree_delete(&btnc->page_tree, newkey);
 343                 spin_unlock_irq(&btnc->tree_lock);
 344                 unlock_page(ctxt->bh->b_page);
 345         } else
 346                 brelse(nbh);
 347 }