lustre/kernel_patches/patches/jbd-static-wbuf-2.6.7.patch

   1 Index: linux-2.6.7/include/linux/jbd.h
   2 ===================================================================
   3 --- linux-2.6.7.orig/include/linux/jbd.h        2004-08-26 17:12:16.000000000 +0400
   4 +++ linux-2.6.7/include/linux/jbd.h     2005-01-19 17:08:33.144512008 +0300
   5 @@ -826,6 +826,12 @@
   6         struct jbd_revoke_table_s *j_revoke_table[2];
   7
   8         /*
   9 +        * array of bhs for journal_commit_transaction
  10 +        */
  11 +       struct buffer_head      **j_wbuf;
  12 +       int                     j_wbufsize;
  13 +
  14 +       /*
  15          * An opaque pointer to fs-private information.  ext3 puts its
  16          * superblock pointer here
  17          */
  18 Index: linux-2.6.7/include/linux/journal-head.h
  19 ===================================================================
  20 --- linux-2.6.7.orig/include/linux/journal-head.h       2003-06-24 18:05:26.000000000 +0400
  21 +++ linux-2.6.7/include/linux/journal-head.h    2005-01-19 14:09:59.000000000 +0300
  22 @@ -80,6 +80,11 @@
  23          * [j_list_lock]
  24          */
  25         struct journal_head *b_cpnext, *b_cpprev;
  26 +
  27 +       /*
  28 +        * counter to track users of the buffer in current transaction
  29 +        */
  30 +       int b_tcount;
  31  };
  32
  33  #endif         /* JOURNAL_HEAD_H_INCLUDED */
  34 Index: linux-2.6.7/fs/jbd/commit.c
  35 ===================================================================
  36 --- linux-2.6.7.orig/fs/jbd/commit.c    2004-08-26 17:12:40.000000000 +0400
  37 +++ linux-2.6.7/fs/jbd/commit.c 2005-01-19 17:28:32.965111552 +0300
  38 @@ -103,7 +103,7 @@
  39  {
  40         transaction_t *commit_transaction;
  41         struct journal_head *jh, *new_jh, *descriptor;
  42 -       struct buffer_head *wbuf[64];
  43 +       struct buffer_head **wbuf = journal->j_wbuf;
  44         int bufs;
  45         int flags;
  46         int err;
  47 @@ -271,7 +283,7 @@
  48                                 BUFFER_TRACE(bh, "start journal writeout");
  49                                 get_bh(bh);
  50                                 wbuf[bufs++] = bh;
  51 -                               if (bufs == ARRAY_SIZE(wbuf)) {
  52 +                               if (bufs == journal->j_wbufsize) {
  53                                         jbd_debug(2, "submit %d writes\n",
  54                                                         bufs);
  55                                         spin_unlock(&journal->j_list_lock);
  56 @@ -488,7 +500,7 @@
  57                 /* If there's no more to do, or if the descriptor is full,
  58                    let the IO rip! */
  59
  60 -               if (bufs == ARRAY_SIZE(wbuf) ||
  61 +               if (bufs == journal->j_wbufsize ||
  62                     commit_transaction->t_buffers == NULL ||
  63                     space_left < sizeof(journal_block_tag_t) + 16) {
  64
  65 Index: linux-2.6.7/fs/jbd/journal.c
  66 ===================================================================
  67 --- linux-2.6.7.orig/fs/jbd/journal.c   2005-01-19 12:07:59.000000000 +0300
  68 +++ linux-2.6.7/fs/jbd/journal.c        2005-01-19 17:11:08.589880720 +0300
  69 @@ -687,6 +687,7 @@
  70  {
  71         journal_t *journal = journal_init_common();
  72         struct buffer_head *bh;
  73 +       int n;
  74
  75         if (!journal)
  76                 return NULL;
  77 @@ -702,6 +703,17 @@
  78         journal->j_sb_buffer = bh;
  79         journal->j_superblock = (journal_superblock_t *)bh->b_data;
  80
  81 +       /* journal descriptor can store upto n blocks -bzzz */
  82 +       n = journal->j_blocksize / sizeof(journal_block_tag_t);
  83 +       journal->j_wbufsize = n;
  84 +       journal->j_wbuf = kmalloc(n * sizeof(struct buffer_head*), GFP_KERNEL);
  85 +       if (!journal->j_wbuf) {
  86 +               printk(KERN_ERR "%s: Cant allocate bhs for commit thread\n",
  87 +                       __FUNCTION__);
  88 +               kfree(journal);
  89 +               journal = NULL;
  90 +       }
  91 +
  92         return journal;
  93  }
  94
  95 @@ -717,7 +729,7 @@
  96  {
  97         struct buffer_head *bh;
  98         journal_t *journal = journal_init_common();
  99 -       int err;
 100 +       int err, n;
 101         unsigned long blocknr;
 102
 103         if (!journal)
 104 @@ -734,6 +746,17 @@
 105         journal->j_maxlen = inode->i_size >> inode->i_sb->s_blocksize_bits;
 106         journal->j_blocksize = inode->i_sb->s_blocksize;
 107
 108 +       /* journal descriptor can store upto n blocks -bzzz */
 109 +       n = journal->j_blocksize / sizeof(journal_block_tag_t);
 110 +       journal->j_wbufsize = n;
 111 +       journal->j_wbuf = kmalloc(n * sizeof(struct buffer_head*), GFP_KERNEL);
 112 +       if (!journal->j_wbuf) {
 113 +               printk(KERN_ERR "%s: Cant allocate bhs for commit thread\n",
 114 +                       __FUNCTION__);
 115 +               kfree(journal);
 116 +               return NULL;
 117 +       }
 118 +
 119         err = journal_bmap(journal, 0, &blocknr);
 120         /* If that failed, give up */
 121         if (err) {
 122 @@ -1107,6 +1130,10 @@
 123                 iput(journal->j_inode);
 124         if (journal->j_revoke)
 125                 journal_destroy_revoke(journal);
 126 +       if (journal->j_wbuf) {
 127 +               kfree(journal->j_wbuf);
 128 +               journal->j_wbuf = NULL;
 129 +       }
 130         kfree(journal);
 131  }
 132
 133 Index: linux-2.6.7/fs/jbd/transaction.c
 134 ===================================================================
 135 --- linux-2.6.7.orig/fs/jbd/transaction.c       2004-08-26 17:12:40.000000000 +0400
 136 +++ linux-2.6.7/fs/jbd/transaction.c    2005-01-19 17:23:30.058160408 +0300
 137 @@ -611,6 +611,10 @@
 138                 handle->h_buffer_credits--;
 139                 if (credits)
 140                         (*credits)++;
 141 +
 142 +               /* the block's becoming member of the trasaction -bzzz */
 143 +               jh->b_tcount = 0;
 144 +
 145                 goto done;
 146         }
 147
 148 @@ -694,6 +698,9 @@
 149         if (credits)
 150                 (*credits)++;
 151
 152 +       /* the block's becoming member of the trasaction -bzzz */
 153 +       jh->b_tcount = 0;
 154 +
 155         /*
 156          * Finally, if the buffer is not journaled right now, we need to make
 157          * sure it doesn't get written to disk before the caller actually
 158 @@ -723,6 +730,11 @@
 159                 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
 160                 kunmap_atomic(source, KM_USER0);
 161         }
 162 +
 163 +       /* track all references to the block to be able to recognize the
 164 +        * situation when the buffer is not part of transaction -bzzz */
 165 +       jh->b_tcount++;
 166 +
 167         jbd_unlock_bh_state(bh);
 168
 169         /*
 170 @@ -822,11 +834,20 @@
 171                 jh->b_transaction = transaction;
 172                 JBUFFER_TRACE(jh, "file as BJ_Reserved");
 173                 __journal_file_buffer(jh, transaction, BJ_Reserved);
 174 +               jh->b_tcount = 0;
 175         } else if (jh->b_transaction == journal->j_committing_transaction) {
 176                 JBUFFER_TRACE(jh, "set next transaction");
 177                 jh->b_next_transaction = transaction;
 178 +               jh->b_tcount = 0;
 179         }
 180         spin_unlock(&journal->j_list_lock);
 181 +
 182 +       /*
 183 +        * track all reference to the block to be able to recognize
 184 +        * the situation when the buffer is not part of transaction -bzzz
 185 +        */
 186 +       jh->b_tcount++;
 187 +
 188         jbd_unlock_bh_state(bh);
 189
 190         /*
 191 @@ -1178,8 +1199,40 @@
 192  void
 193  journal_release_buffer(handle_t *handle, struct buffer_head *bh, int credits)
 194  {
 195 +       journal_t *journal = handle->h_transaction->t_journal;
 196 +       struct journal_head *jh = bh2jh(bh);
 197 +
 198         BUFFER_TRACE(bh, "entry");
 199 -       handle->h_buffer_credits += credits;
 200 +
 201 +       /* return credit back to the handle if it was really spent */
 202 +       if (credits)
 203 +               handle->h_buffer_credits++;
 204 +
 205 +       jbd_lock_bh_state(bh);
 206 +       J_ASSERT(jh->b_tcount > 0);
 207 +
 208 +       jh->b_tcount--;
 209 +       if (jh->b_tcount == 0) {
 210 +               /* we can drop it from the transaction -bzzz */
 211 +               J_ASSERT(jh->b_transaction == handle->h_transaction ||
 212 +                               jh->b_next_transaction == handle->h_transaction);
 213 +               if (jh->b_transaction == handle->h_transaction) {
 214 +                       spin_lock(&journal->j_list_lock);
 215 +                       __journal_unfile_buffer(jh);
 216 +                       spin_unlock(&journal->j_list_lock);
 217 +               } else if(jh->b_next_transaction) {
 218 +                       jh->b_next_transaction = NULL;
 219 +               }
 220 +
 221 +               /*
 222 +                * this was last reference to the block from the current
 223 +                * transaction and we'd like to return credit to the
 224 +                * whole transaction -bzzz
 225 +                */
 226 +               if (!credits)
 227 +                       handle->h_buffer_credits++;
 228 +       }
 229 +       jbd_unlock_bh_state(bh);
 230  }
 231
 232  /**
 233 @@ -1204,6 +1257,7 @@
 234         transaction_t *transaction = handle->h_transaction;
 235         journal_t *journal = transaction->t_journal;
 236         struct journal_head *jh;
 237 +       int drop_reserve = 0;
 238
 239         BUFFER_TRACE(bh, "entry");
 240
 241 @@ -1227,6 +1281,7 @@
 242                 J_ASSERT_JH(jh, !jh->b_committed_data);
 243
 244                 __journal_unfile_buffer(jh);
 245 +               drop_reserve = 1;
 246
 247                 /*
 248                  * We are no longer going to journal this buffer.
 249 @@ -1249,7 +1304,7 @@
 250                                 spin_unlock(&journal->j_list_lock);
 251                                 jbd_unlock_bh_state(bh);
 252                                 __bforget(bh);
 253 -                               return;
 254 +                               goto drop;
 255                         }
 256                 }
 257         } else if (jh->b_transaction) {
 258 @@ -1264,6 +1319,7 @@
 259                 if (jh->b_next_transaction) {
 260                         J_ASSERT(jh->b_next_transaction == transaction);
 261                         jh->b_next_transaction = NULL;
 262 +                       drop_reserve = 1;
 263                 }
 264         }
 265
 266 @@ -1271,6 +1327,15 @@
 267         spin_unlock(&journal->j_list_lock);
 268         jbd_unlock_bh_state(bh);
 269         __brelse(bh);
 270 +
 271 +drop:
 272 +       if (drop_reserve) {
 273 +               /* no need to reserve log space for this block -bzzz */
 274 +               spin_lock(&transaction->t_handle_lock);
 275 +               transaction->t_outstanding_credits--;
 276 +               spin_unlock(&transaction->t_handle_lock);
 277 +       }
 278 +
 279         return;
 280  }
 281