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