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  * Originally written by Seiji Kihara.
  17  * Fully revised by Ryusuke Konishi for stabilization and simplification.
  18  *
  19  */
  20
  21 #include <linux/types.h>
  22 #include <linux/buffer_head.h>
  23 #include <linux/mm.h>
  24 #include <linux/backing-dev.h>
  25 #include <linux/gfp.h>
  26 #include "nilfs.h"
  27 #include "mdt.h"
  28 #include "dat.h"
  29 #include "page.h"
  30 #include "btnode.h"
  31
  32 void nilfs_btnode_cache_clear(struct address_space *btnc)
  33 {
  34         invalidate_mapping_pages(btnc, 0, -1);
  35         truncate_inode_pages(btnc, 0);
  36 }
  37
  38 struct buffer_head *
  39 nilfs_btnode_create_block(struct address_space *btnc, __u64 blocknr)
  40 {
  41         struct inode *inode = NILFS_BTNC_I(btnc);
  42         struct buffer_head *bh;
  43
  44         bh = nilfs_grab_buffer(inode, btnc, blocknr, 1 << BH_NILFS_Node);
  45         if (unlikely(!bh))
  46                 return NULL;
  47
  48         if (unlikely(buffer_mapped(bh) || buffer_uptodate(bh) ||
  49                      buffer_dirty(bh))) {
  50                 brelse(bh);
  51                 BUG();
  52         }
  53         memset(bh->b_data, 0, 1 << inode->i_blkbits);
  54         bh->b_bdev = inode->i_sb->s_bdev;
  55         bh->b_blocknr = blocknr;
  56         set_buffer_mapped(bh);
  57         set_buffer_uptodate(bh);
  58
  59         unlock_page(bh->b_page);
  60         put_page(bh->b_page);
  61         return bh;
  62 }
  63
  64 int nilfs_btnode_submit_block(struct address_space *btnc, __u64 blocknr,
  65                               sector_t pblocknr, int mode, int mode_flags,
  66                               struct buffer_head **pbh, sector_t *submit_ptr)
  67 {
  68         struct buffer_head *bh;
  69         struct inode *inode = NILFS_BTNC_I(btnc);
  70         struct page *page;
  71         int err;
  72
  73         bh = nilfs_grab_buffer(inode, btnc, blocknr, 1 << BH_NILFS_Node);
  74         if (unlikely(!bh))
  75                 return -ENOMEM;
  76
  77         err = -EEXIST; /* internal code */
  78         page = bh->b_page;
  79
  80         if (buffer_uptodate(bh) || buffer_dirty(bh))
  81                 goto found;
  82
  83         if (pblocknr == 0) {
  84                 pblocknr = blocknr;
  85                 if (inode->i_ino != NILFS_DAT_INO) {
  86                         struct the_nilfs *nilfs = inode->i_sb->s_fs_info;
  87
  88                         /* blocknr is a virtual block number */
  89                         err = nilfs_dat_translate(nilfs->ns_dat, blocknr,
  90                                                   &pblocknr);
  91                         if (unlikely(err)) {
  92                                 brelse(bh);
  93                                 goto out_locked;
  94                         }
  95                 }
  96         }
  97
  98         if (mode_flags & REQ_RAHEAD) {
  99                 if (pblocknr != *submit_ptr + 1 || !trylock_buffer(bh)) {
 100                         err = -EBUSY; /* internal code */
 101                         brelse(bh);
 102                         goto out_locked;
 103                 }
 104         } else { /* mode == READ */
 105                 lock_buffer(bh);
 106         }
 107         if (buffer_uptodate(bh)) {
 108                 unlock_buffer(bh);
 109                 err = -EEXIST; /* internal code */
 110                 goto found;
 111         }
 112         set_buffer_mapped(bh);
 113         bh->b_bdev = inode->i_sb->s_bdev;
 114         bh->b_blocknr = pblocknr; /* set block address for read */
 115         bh->b_end_io = end_buffer_read_sync;
 116         get_bh(bh);
 117         submit_bh(mode, mode_flags, bh);
 118         bh->b_blocknr = blocknr; /* set back to the given block address */
 119         *submit_ptr = pblocknr;
 120         err = 0;
 121 found:
 122         *pbh = bh;
 123
 124 out_locked:
 125         unlock_page(page);
 126         put_page(page);
 127         return err;
 128 }
 129
 130 /**
 131  * nilfs_btnode_delete - delete B-tree node buffer
 132  * @bh: buffer to be deleted
 133  *
 134  * nilfs_btnode_delete() invalidates the specified buffer and delete the page
 135  * including the buffer if the page gets unbusy.
 136  */
 137 void nilfs_btnode_delete(struct buffer_head *bh)
 138 {
 139         struct address_space *mapping;
 140         struct page *page = bh->b_page;
 141         pgoff_t index = page_index(page);
 142         int still_dirty;
 143
 144         get_page(page);
 145         lock_page(page);
 146         wait_on_page_writeback(page);
 147
 148         nilfs_forget_buffer(bh);
 149         still_dirty = PageDirty(page);
 150         mapping = page->mapping;
 151         unlock_page(page);
 152         put_page(page);
 153
 154         if (!still_dirty && mapping)
 155                 invalidate_inode_pages2_range(mapping, index, index);
 156 }
 157
 158 /**
 159  * nilfs_btnode_prepare_change_key
 160  *  prepare to move contents of the block for old key to one of new key.
 161  *  the old buffer will not be removed, but might be reused for new buffer.
 162  *  it might return -ENOMEM because of memory allocation errors,
 163  *  and might return -EIO because of disk read errors.
 164  */
 165 int nilfs_btnode_prepare_change_key(struct address_space *btnc,
 166                                     struct nilfs_btnode_chkey_ctxt *ctxt)
 167 {
 168         struct buffer_head *obh, *nbh;
 169         struct inode *inode = NILFS_BTNC_I(btnc);
 170         __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
 171         int err;
 172
 173         if (oldkey == newkey)
 174                 return 0;
 175
 176         obh = ctxt->bh;
 177         ctxt->newbh = NULL;
 178
 179         if (inode->i_blkbits == PAGE_SHIFT) {
 180                 lock_page(obh->b_page);
 181                 /*
 182                  * We cannot call radix_tree_preload for the kernels older
 183                  * than 2.6.23, because it is not exported for modules.
 184                  */
 185 retry:
 186                 err = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
 187                 if (err)
 188                         goto failed_unlock;
 189                 /* BUG_ON(oldkey != obh->b_page->index); */
 190                 if (unlikely(oldkey != obh->b_page->index))
 191                         NILFS_PAGE_BUG(obh->b_page,
 192                                        "invalid oldkey %lld (newkey=%lld)",
 193                                        (unsigned long long)oldkey,
 194                                        (unsigned long long)newkey);
 195
 196                 spin_lock_irq(&btnc->tree_lock);
 197                 err = radix_tree_insert(&btnc->page_tree, newkey, obh->b_page);
 198                 spin_unlock_irq(&btnc->tree_lock);
 199                 /*
 200                  * Note: page->index will not change to newkey until
 201                  * nilfs_btnode_commit_change_key() will be called.
 202                  * To protect the page in intermediate state, the page lock
 203                  * is held.
 204                  */
 205                 radix_tree_preload_end();
 206                 if (!err)
 207                         return 0;
 208                 else if (err != -EEXIST)
 209                         goto failed_unlock;
 210
 211                 err = invalidate_inode_pages2_range(btnc, newkey, newkey);
 212                 if (!err)
 213                         goto retry;
 214                 /* fallback to copy mode */
 215                 unlock_page(obh->b_page);
 216         }
 217
 218         nbh = nilfs_btnode_create_block(btnc, newkey);
 219         if (!nbh)
 220                 return -ENOMEM;
 221
 222         BUG_ON(nbh == obh);
 223         ctxt->newbh = nbh;
 224         return 0;
 225
 226  failed_unlock:
 227         unlock_page(obh->b_page);
 228         return err;
 229 }
 230
 231 /**
 232  * nilfs_btnode_commit_change_key
 233  *  commit the change_key operation prepared by prepare_change_key().
 234  */
 235 void nilfs_btnode_commit_change_key(struct address_space *btnc,
 236                                     struct nilfs_btnode_chkey_ctxt *ctxt)
 237 {
 238         struct buffer_head *obh = ctxt->bh, *nbh = ctxt->newbh;
 239         __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
 240         struct page *opage;
 241
 242         if (oldkey == newkey)
 243                 return;
 244
 245         if (nbh == NULL) {      /* blocksize == pagesize */
 246                 opage = obh->b_page;
 247                 if (unlikely(oldkey != opage->index))
 248                         NILFS_PAGE_BUG(opage,
 249                                        "invalid oldkey %lld (newkey=%lld)",
 250                                        (unsigned long long)oldkey,
 251                                        (unsigned long long)newkey);
 252                 mark_buffer_dirty(obh);
 253
 254                 spin_lock_irq(&btnc->tree_lock);
 255                 radix_tree_delete(&btnc->page_tree, oldkey);
 256                 radix_tree_tag_set(&btnc->page_tree, newkey,
 257                                    PAGECACHE_TAG_DIRTY);
 258                 spin_unlock_irq(&btnc->tree_lock);
 259
 260                 opage->index = obh->b_blocknr = newkey;
 261                 unlock_page(opage);
 262         } else {
 263                 nilfs_copy_buffer(nbh, obh);
 264                 mark_buffer_dirty(nbh);
 265
 266                 nbh->b_blocknr = newkey;
 267                 ctxt->bh = nbh;
 268                 nilfs_btnode_delete(obh); /* will decrement bh->b_count */
 269         }
 270 }
 271
 272 /**
 273  * nilfs_btnode_abort_change_key
 274  *  abort the change_key operation prepared by prepare_change_key().
 275  */
 276 void nilfs_btnode_abort_change_key(struct address_space *btnc,
 277                                    struct nilfs_btnode_chkey_ctxt *ctxt)
 278 {
 279         struct buffer_head *nbh = ctxt->newbh;
 280         __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
 281
 282         if (oldkey == newkey)
 283                 return;
 284
 285         if (nbh == NULL) {      /* blocksize == pagesize */
 286                 spin_lock_irq(&btnc->tree_lock);
 287                 radix_tree_delete(&btnc->page_tree, newkey);
 288                 spin_unlock_irq(&btnc->tree_lock);
 289                 unlock_page(ctxt->bh->b_page);
 290         } else
 291                 brelse(nbh);
 292 }