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
  33 /**
  34  * nilfs_init_btnc_inode - initialize B-tree node cache inode
  35  * @btnc_inode: inode to be initialized
  36  *
  37  * nilfs_init_btnc_inode() sets up an inode for B-tree node cache.
  38  */
  39 void nilfs_init_btnc_inode(struct inode *btnc_inode)
  40 {
  41         struct nilfs_inode_info *ii = NILFS_I(btnc_inode);
  42
  43         btnc_inode->i_mode = S_IFREG;
  44         ii->i_flags = 0;
  45         memset(&ii->i_bmap_data, 0, sizeof(struct nilfs_bmap));
  46         mapping_set_gfp_mask(btnc_inode->i_mapping, GFP_NOFS);
  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 = btnc->host;
  59         struct buffer_head *bh;
  60
  61         bh = nilfs_grab_buffer(inode, btnc, blocknr, BIT(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, i_blocksize(inode));
  71         bh->b_bdev = inode->i_sb->s_bdev;
  72         bh->b_blocknr = blocknr;
  73         set_buffer_mapped(bh);
  74         set_buffer_uptodate(bh);
  75
  76         unlock_page(bh->b_page);
  77         put_page(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, int mode_flags,
  83                               struct buffer_head **pbh, sector_t *submit_ptr)
  84 {
  85         struct buffer_head *bh;
  86         struct inode *inode = btnc->host;
  87         struct page *page;
  88         int err;
  89
  90         bh = nilfs_grab_buffer(inode, btnc, blocknr, BIT(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 the_nilfs *nilfs = inode->i_sb->s_fs_info;
 104
 105                         /* blocknr is a virtual block number */
 106                         err = nilfs_dat_translate(nilfs->ns_dat, blocknr,
 107                                                   &pblocknr);
 108                         if (unlikely(err)) {
 109                                 brelse(bh);
 110                                 goto out_locked;
 111                         }
 112                 }
 113         }
 114
 115         if (mode_flags & REQ_RAHEAD) {
 116                 if (pblocknr != *submit_ptr + 1 || !trylock_buffer(bh)) {
 117                         err = -EBUSY; /* internal code */
 118                         brelse(bh);
 119                         goto out_locked;
 120                 }
 121         } else { /* mode == READ */
 122                 lock_buffer(bh);
 123         }
 124         if (buffer_uptodate(bh)) {
 125                 unlock_buffer(bh);
 126                 err = -EEXIST; /* internal code */
 127                 goto found;
 128         }
 129         set_buffer_mapped(bh);
 130         bh->b_bdev = inode->i_sb->s_bdev;
 131         bh->b_blocknr = pblocknr; /* set block address for read */
 132         bh->b_end_io = end_buffer_read_sync;
 133         get_bh(bh);
 134         submit_bh(mode, mode_flags, bh);
 135         bh->b_blocknr = blocknr; /* set back to the given block address */
 136         *submit_ptr = pblocknr;
 137         err = 0;
 138 found:
 139         *pbh = bh;
 140
 141 out_locked:
 142         unlock_page(page);
 143         put_page(page);
 144         return err;
 145 }
 146
 147 /**
 148  * nilfs_btnode_delete - delete B-tree node buffer
 149  * @bh: buffer to be deleted
 150  *
 151  * nilfs_btnode_delete() invalidates the specified buffer and delete the page
 152  * including the buffer if the page gets unbusy.
 153  */
 154 void nilfs_btnode_delete(struct buffer_head *bh)
 155 {
 156         struct address_space *mapping;
 157         struct page *page = bh->b_page;
 158         pgoff_t index = page_index(page);
 159         int still_dirty;
 160
 161         get_page(page);
 162         lock_page(page);
 163         wait_on_page_writeback(page);
 164
 165         nilfs_forget_buffer(bh);
 166         still_dirty = PageDirty(page);
 167         mapping = page->mapping;
 168         unlock_page(page);
 169         put_page(page);
 170
 171         if (!still_dirty && mapping)
 172                 invalidate_inode_pages2_range(mapping, index, index);
 173 }
 174
 175 /**
 176  * nilfs_btnode_prepare_change_key
 177  *  prepare to move contents of the block for old key to one of new key.
 178  *  the old buffer will not be removed, but might be reused for new buffer.
 179  *  it might return -ENOMEM because of memory allocation errors,
 180  *  and might return -EIO because of disk read errors.
 181  */
 182 int nilfs_btnode_prepare_change_key(struct address_space *btnc,
 183                                     struct nilfs_btnode_chkey_ctxt *ctxt)
 184 {
 185         struct buffer_head *obh, *nbh;
 186         struct inode *inode = btnc->host;
 187         __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
 188         int err;
 189
 190         if (oldkey == newkey)
 191                 return 0;
 192
 193         obh = ctxt->bh;
 194         ctxt->newbh = NULL;
 195
 196         if (inode->i_blkbits == PAGE_SHIFT) {
 197                 lock_page(obh->b_page);
 198                 /*
 199                  * We cannot call radix_tree_preload for the kernels older
 200                  * than 2.6.23, because it is not exported for modules.
 201                  */
 202 retry:
 203                 err = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
 204                 if (err)
 205                         goto failed_unlock;
 206                 /* BUG_ON(oldkey != obh->b_page->index); */
 207                 if (unlikely(oldkey != obh->b_page->index))
 208                         NILFS_PAGE_BUG(obh->b_page,
 209                                        "invalid oldkey %lld (newkey=%lld)",
 210                                        (unsigned long long)oldkey,
 211                                        (unsigned long long)newkey);
 212
 213                 spin_lock_irq(&btnc->tree_lock);
 214                 err = radix_tree_insert(&btnc->page_tree, newkey, obh->b_page);
 215                 spin_unlock_irq(&btnc->tree_lock);
 216                 /*
 217                  * Note: page->index will not change to newkey until
 218                  * nilfs_btnode_commit_change_key() will be called.
 219                  * To protect the page in intermediate state, the page lock
 220                  * is held.
 221                  */
 222                 radix_tree_preload_end();
 223                 if (!err)
 224                         return 0;
 225                 else if (err != -EEXIST)
 226                         goto failed_unlock;
 227
 228                 err = invalidate_inode_pages2_range(btnc, newkey, newkey);
 229                 if (!err)
 230                         goto retry;
 231                 /* fallback to copy mode */
 232                 unlock_page(obh->b_page);
 233         }
 234
 235         nbh = nilfs_btnode_create_block(btnc, newkey);
 236         if (!nbh)
 237                 return -ENOMEM;
 238
 239         BUG_ON(nbh == obh);
 240         ctxt->newbh = nbh;
 241         return 0;
 242
 243  failed_unlock:
 244         unlock_page(obh->b_page);
 245         return err;
 246 }
 247
 248 /**
 249  * nilfs_btnode_commit_change_key
 250  *  commit the change_key operation prepared by prepare_change_key().
 251  */
 252 void nilfs_btnode_commit_change_key(struct address_space *btnc,
 253                                     struct nilfs_btnode_chkey_ctxt *ctxt)
 254 {
 255         struct buffer_head *obh = ctxt->bh, *nbh = ctxt->newbh;
 256         __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
 257         struct page *opage;
 258
 259         if (oldkey == newkey)
 260                 return;
 261
 262         if (nbh == NULL) {      /* blocksize == pagesize */
 263                 opage = obh->b_page;
 264                 if (unlikely(oldkey != opage->index))
 265                         NILFS_PAGE_BUG(opage,
 266                                        "invalid oldkey %lld (newkey=%lld)",
 267                                        (unsigned long long)oldkey,
 268                                        (unsigned long long)newkey);
 269                 mark_buffer_dirty(obh);
 270
 271                 spin_lock_irq(&btnc->tree_lock);
 272                 radix_tree_delete(&btnc->page_tree, oldkey);
 273                 radix_tree_tag_set(&btnc->page_tree, newkey,
 274                                    PAGECACHE_TAG_DIRTY);
 275                 spin_unlock_irq(&btnc->tree_lock);
 276
 277                 opage->index = obh->b_blocknr = newkey;
 278                 unlock_page(opage);
 279         } else {
 280                 nilfs_copy_buffer(nbh, obh);
 281                 mark_buffer_dirty(nbh);
 282
 283                 nbh->b_blocknr = newkey;
 284                 ctxt->bh = nbh;
 285                 nilfs_btnode_delete(obh); /* will decrement bh->b_count */
 286         }
 287 }
 288
 289 /**
 290  * nilfs_btnode_abort_change_key
 291  *  abort the change_key operation prepared by prepare_change_key().
 292  */
 293 void nilfs_btnode_abort_change_key(struct address_space *btnc,
 294                                    struct nilfs_btnode_chkey_ctxt *ctxt)
 295 {
 296         struct buffer_head *nbh = ctxt->newbh;
 297         __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
 298
 299         if (oldkey == newkey)
 300                 return;
 301
 302         if (nbh == NULL) {      /* blocksize == pagesize */
 303                 spin_lock_irq(&btnc->tree_lock);
 304                 radix_tree_delete(&btnc->page_tree, newkey);
 305                 spin_unlock_irq(&btnc->tree_lock);
 306                 unlock_page(ctxt->bh->b_page);
 307         } else {
 308                 /*
 309                  * When canceling a buffer that a prepare operation has
 310                  * allocated to copy a node block to another location, use
 311                  * nilfs_btnode_delete() to initialize and release the buffer
 312                  * so that the buffer flags will not be in an inconsistent
 313                  * state when it is reallocated.
 314                  */
 315                 nilfs_btnode_delete(nbh);
 316         }
 317 }