fs/nilfs2/btnode.c

   1 // SPDX-License-Identifier: GPL-2.0+
   2 /*
   3  * btnode.c - NILFS B-tree node cache
   4  *
   5  * Copyright (C) 2005-2008 Nippon Telegraph and Telephone Corporation.
   6  *
   7  * Originally written by Seiji Kihara.
   8  * Fully revised by Ryusuke Konishi for stabilization and simplification.
   9  *
  10  */
  11
  12 #include <linux/types.h>
  13 #include <linux/buffer_head.h>
  14 #include <linux/mm.h>
  15 #include <linux/backing-dev.h>
  16 #include <linux/gfp.h>
  17 #include "nilfs.h"
  18 #include "mdt.h"
  19 #include "dat.h"
  20 #include "page.h"
  21 #include "btnode.h"
  22
  23 void nilfs_btnode_cache_clear(struct address_space *btnc)
  24 {
  25         invalidate_mapping_pages(btnc, 0, -1);
  26         truncate_inode_pages(btnc, 0);
  27 }
  28
  29 struct buffer_head *
  30 nilfs_btnode_create_block(struct address_space *btnc, __u64 blocknr)
  31 {
  32         struct inode *inode = NILFS_BTNC_I(btnc);
  33         struct buffer_head *bh;
  34
  35         bh = nilfs_grab_buffer(inode, btnc, blocknr, BIT(BH_NILFS_Node));
  36         if (unlikely(!bh))
  37                 return NULL;
  38
  39         if (unlikely(buffer_mapped(bh) || buffer_uptodate(bh) ||
  40                      buffer_dirty(bh))) {
  41                 brelse(bh);
  42                 BUG();
  43         }
  44         memset(bh->b_data, 0, i_blocksize(inode));
  45         bh->b_bdev = inode->i_sb->s_bdev;
  46         bh->b_blocknr = blocknr;
  47         set_buffer_mapped(bh);
  48         set_buffer_uptodate(bh);
  49
  50         unlock_page(bh->b_page);
  51         put_page(bh->b_page);
  52         return bh;
  53 }
  54
  55 int nilfs_btnode_submit_block(struct address_space *btnc, __u64 blocknr,
  56                               sector_t pblocknr, int mode, int mode_flags,
  57                               struct buffer_head **pbh, sector_t *submit_ptr)
  58 {
  59         struct buffer_head *bh;
  60         struct inode *inode = NILFS_BTNC_I(btnc);
  61         struct page *page;
  62         int err;
  63
  64         bh = nilfs_grab_buffer(inode, btnc, blocknr, BIT(BH_NILFS_Node));
  65         if (unlikely(!bh))
  66                 return -ENOMEM;
  67
  68         err = -EEXIST; /* internal code */
  69         page = bh->b_page;
  70
  71         if (buffer_uptodate(bh) || buffer_dirty(bh))
  72                 goto found;
  73
  74         if (pblocknr == 0) {
  75                 pblocknr = blocknr;
  76                 if (inode->i_ino != NILFS_DAT_INO) {
  77                         struct the_nilfs *nilfs = inode->i_sb->s_fs_info;
  78
  79                         /* blocknr is a virtual block number */
  80                         err = nilfs_dat_translate(nilfs->ns_dat, blocknr,
  81                                                   &pblocknr);
  82                         if (unlikely(err)) {
  83                                 brelse(bh);
  84                                 goto out_locked;
  85                         }
  86                 }
  87         }
  88
  89         if (mode_flags & REQ_RAHEAD) {
  90                 if (pblocknr != *submit_ptr + 1 || !trylock_buffer(bh)) {
  91                         err = -EBUSY; /* internal code */
  92                         brelse(bh);
  93                         goto out_locked;
  94                 }
  95         } else { /* mode == READ */
  96                 lock_buffer(bh);
  97         }
  98         if (buffer_uptodate(bh)) {
  99                 unlock_buffer(bh);
 100                 err = -EEXIST; /* internal code */
 101                 goto found;
 102         }
 103         set_buffer_mapped(bh);
 104         bh->b_bdev = inode->i_sb->s_bdev;
 105         bh->b_blocknr = pblocknr; /* set block address for read */
 106         bh->b_end_io = end_buffer_read_sync;
 107         get_bh(bh);
 108         submit_bh(mode, mode_flags, bh);
 109         bh->b_blocknr = blocknr; /* set back to the given block address */
 110         *submit_ptr = pblocknr;
 111         err = 0;
 112 found:
 113         *pbh = bh;
 114
 115 out_locked:
 116         unlock_page(page);
 117         put_page(page);
 118         return err;
 119 }
 120
 121 /**
 122  * nilfs_btnode_delete - delete B-tree node buffer
 123  * @bh: buffer to be deleted
 124  *
 125  * nilfs_btnode_delete() invalidates the specified buffer and delete the page
 126  * including the buffer if the page gets unbusy.
 127  */
 128 void nilfs_btnode_delete(struct buffer_head *bh)
 129 {
 130         struct address_space *mapping;
 131         struct page *page = bh->b_page;
 132         pgoff_t index = page_index(page);
 133         int still_dirty;
 134
 135         get_page(page);
 136         lock_page(page);
 137         wait_on_page_writeback(page);
 138
 139         nilfs_forget_buffer(bh);
 140         still_dirty = PageDirty(page);
 141         mapping = page->mapping;
 142         unlock_page(page);
 143         put_page(page);
 144
 145         if (!still_dirty && mapping)
 146                 invalidate_inode_pages2_range(mapping, index, index);
 147 }
 148
 149 /**
 150  * nilfs_btnode_prepare_change_key
 151  *  prepare to move contents of the block for old key to one of new key.
 152  *  the old buffer will not be removed, but might be reused for new buffer.
 153  *  it might return -ENOMEM because of memory allocation errors,
 154  *  and might return -EIO because of disk read errors.
 155  */
 156 int nilfs_btnode_prepare_change_key(struct address_space *btnc,
 157                                     struct nilfs_btnode_chkey_ctxt *ctxt)
 158 {
 159         struct buffer_head *obh, *nbh;
 160         struct inode *inode = NILFS_BTNC_I(btnc);
 161         __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
 162         int err;
 163
 164         if (oldkey == newkey)
 165                 return 0;
 166
 167         obh = ctxt->bh;
 168         ctxt->newbh = NULL;
 169
 170         if (inode->i_blkbits == PAGE_SHIFT) {
 171                 lock_page(obh->b_page);
 172                 /*
 173                  * We cannot call radix_tree_preload for the kernels older
 174                  * than 2.6.23, because it is not exported for modules.
 175                  */
 176 retry:
 177                 err = radix_tree_preload(GFP_NOFS & ~__GFP_HIGHMEM);
 178                 if (err)
 179                         goto failed_unlock;
 180                 /* BUG_ON(oldkey != obh->b_page->index); */
 181                 if (unlikely(oldkey != obh->b_page->index))
 182                         NILFS_PAGE_BUG(obh->b_page,
 183                                        "invalid oldkey %lld (newkey=%lld)",
 184                                        (unsigned long long)oldkey,
 185                                        (unsigned long long)newkey);
 186
 187                 xa_lock_irq(&btnc->i_pages);
 188                 err = radix_tree_insert(&btnc->i_pages, newkey, obh->b_page);
 189                 xa_unlock_irq(&btnc->i_pages);
 190                 /*
 191                  * Note: page->index will not change to newkey until
 192                  * nilfs_btnode_commit_change_key() will be called.
 193                  * To protect the page in intermediate state, the page lock
 194                  * is held.
 195                  */
 196                 radix_tree_preload_end();
 197                 if (!err)
 198                         return 0;
 199                 else if (err != -EEXIST)
 200                         goto failed_unlock;
 201
 202                 err = invalidate_inode_pages2_range(btnc, newkey, newkey);
 203                 if (!err)
 204                         goto retry;
 205                 /* fallback to copy mode */
 206                 unlock_page(obh->b_page);
 207         }
 208
 209         nbh = nilfs_btnode_create_block(btnc, newkey);
 210         if (!nbh)
 211                 return -ENOMEM;
 212
 213         BUG_ON(nbh == obh);
 214         ctxt->newbh = nbh;
 215         return 0;
 216
 217  failed_unlock:
 218         unlock_page(obh->b_page);
 219         return err;
 220 }
 221
 222 /**
 223  * nilfs_btnode_commit_change_key
 224  *  commit the change_key operation prepared by prepare_change_key().
 225  */
 226 void nilfs_btnode_commit_change_key(struct address_space *btnc,
 227                                     struct nilfs_btnode_chkey_ctxt *ctxt)
 228 {
 229         struct buffer_head *obh = ctxt->bh, *nbh = ctxt->newbh;
 230         __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
 231         struct page *opage;
 232
 233         if (oldkey == newkey)
 234                 return;
 235
 236         if (nbh == NULL) {      /* blocksize == pagesize */
 237                 opage = obh->b_page;
 238                 if (unlikely(oldkey != opage->index))
 239                         NILFS_PAGE_BUG(opage,
 240                                        "invalid oldkey %lld (newkey=%lld)",
 241                                        (unsigned long long)oldkey,
 242                                        (unsigned long long)newkey);
 243                 mark_buffer_dirty(obh);
 244
 245                 xa_lock_irq(&btnc->i_pages);
 246                 radix_tree_delete(&btnc->i_pages, oldkey);
 247                 radix_tree_tag_set(&btnc->i_pages, newkey,
 248                                    PAGECACHE_TAG_DIRTY);
 249                 xa_unlock_irq(&btnc->i_pages);
 250
 251                 opage->index = obh->b_blocknr = newkey;
 252                 unlock_page(opage);
 253         } else {
 254                 nilfs_copy_buffer(nbh, obh);
 255                 mark_buffer_dirty(nbh);
 256
 257                 nbh->b_blocknr = newkey;
 258                 ctxt->bh = nbh;
 259                 nilfs_btnode_delete(obh); /* will decrement bh->b_count */
 260         }
 261 }
 262
 263 /**
 264  * nilfs_btnode_abort_change_key
 265  *  abort the change_key operation prepared by prepare_change_key().
 266  */
 267 void nilfs_btnode_abort_change_key(struct address_space *btnc,
 268                                    struct nilfs_btnode_chkey_ctxt *ctxt)
 269 {
 270         struct buffer_head *nbh = ctxt->newbh;
 271         __u64 oldkey = ctxt->oldkey, newkey = ctxt->newkey;
 272
 273         if (oldkey == newkey)
 274                 return;
 275
 276         if (nbh == NULL) {      /* blocksize == pagesize */
 277                 xa_lock_irq(&btnc->i_pages);
 278                 radix_tree_delete(&btnc->i_pages, newkey);
 279                 xa_unlock_irq(&btnc->i_pages);
 280                 unlock_page(ctxt->bh->b_page);
 281         } else
 282                 brelse(nbh);
 283 }