1 /* -*- mode: c; c-basic-offset: 8; -*- 2 * vim: noexpandtab sw=8 ts=8 sts=0: 3 * 4 * journal.h 5 * 6 * Defines journalling api and structures. 7 * 8 * Copyright (C) 2003, 2005 Oracle. All rights reserved. 9 * 10 * This program is free software; you can redistribute it and/or 11 * modify it under the terms of the GNU General Public 12 * License as published by the Free Software Foundation; either 13 * version 2 of the License, or (at your option) any later version. 14 * 15 * This program is distributed in the hope that it will be useful, 16 * but WITHOUT ANY WARRANTY; without even the implied warranty of 17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 18 * General Public License for more details. 19 * 20 * You should have received a copy of the GNU General Public 21 * License along with this program; if not, write to the 22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330, 23 * Boston, MA 021110-1307, USA. 24 */ 25 26 #ifndef OCFS2_JOURNAL_H 27 #define OCFS2_JOURNAL_H 28 29 #include <linux/fs.h> 30 #include <linux/jbd.h> 31 32 enum ocfs2_journal_state { 33 OCFS2_JOURNAL_FREE = 0, 34 OCFS2_JOURNAL_LOADED, 35 OCFS2_JOURNAL_IN_SHUTDOWN, 36 }; 37 38 struct ocfs2_super; 39 struct ocfs2_dinode; 40 41 struct ocfs2_journal { 42 enum ocfs2_journal_state j_state; /* Journals current state */ 43 44 journal_t *j_journal; /* The kernels journal type */ 45 struct inode *j_inode; /* Kernel inode pointing to 46 * this journal */ 47 struct ocfs2_super *j_osb; /* pointer to the super 48 * block for the node 49 * we're currently 50 * running on -- not 51 * necessarily the super 52 * block from the node 53 * which we usually run 54 * from (recovery, 55 * etc) */ 56 struct buffer_head *j_bh; /* Journal disk inode block */ 57 atomic_t j_num_trans; /* Number of transactions 58 * currently in the system. */ 59 unsigned long j_trans_id; 60 struct rw_semaphore j_trans_barrier; 61 wait_queue_head_t j_checkpointed; 62 63 spinlock_t j_lock; 64 struct list_head j_la_cleanups; 65 struct work_struct j_recovery_work; 66 }; 67 68 extern spinlock_t trans_inc_lock; 69 70 /* wrap j_trans_id so we never have it equal to zero. */ 71 static inline unsigned long ocfs2_inc_trans_id(struct ocfs2_journal *j) 72 { 73 unsigned long old_id; 74 spin_lock(&trans_inc_lock); 75 old_id = j->j_trans_id++; 76 if (unlikely(!j->j_trans_id)) 77 j->j_trans_id = 1; 78 spin_unlock(&trans_inc_lock); 79 return old_id; 80 } 81 82 static inline void ocfs2_set_inode_lock_trans(struct ocfs2_journal *journal, 83 struct inode *inode) 84 { 85 spin_lock(&trans_inc_lock); 86 OCFS2_I(inode)->ip_last_trans = journal->j_trans_id; 87 spin_unlock(&trans_inc_lock); 88 } 89 90 /* Used to figure out whether it's safe to drop a metadata lock on an 91 * inode. Returns true if all the inodes changes have been 92 * checkpointed to disk. You should be holding the spinlock on the 93 * metadata lock while calling this to be sure that nobody can take 94 * the lock and put it on another transaction. */ 95 static inline int ocfs2_inode_fully_checkpointed(struct inode *inode) 96 { 97 int ret; 98 struct ocfs2_journal *journal = OCFS2_SB(inode->i_sb)->journal; 99 100 spin_lock(&trans_inc_lock); 101 ret = time_after(journal->j_trans_id, OCFS2_I(inode)->ip_last_trans); 102 spin_unlock(&trans_inc_lock); 103 return ret; 104 } 105 106 /* convenience function to check if an inode is still new (has never 107 * hit disk) Will do you a favor and set created_trans = 0 when you've 108 * been checkpointed. returns '1' if the inode is still new. */ 109 static inline int ocfs2_inode_is_new(struct inode *inode) 110 { 111 int ret; 112 113 /* System files are never "new" as they're written out by 114 * mkfs. This helps us early during mount, before we have the 115 * journal open and j_trans_id could be junk. */ 116 if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE) 117 return 0; 118 spin_lock(&trans_inc_lock); 119 ret = !(time_after(OCFS2_SB(inode->i_sb)->journal->j_trans_id, 120 OCFS2_I(inode)->ip_created_trans)); 121 if (!ret) 122 OCFS2_I(inode)->ip_created_trans = 0; 123 spin_unlock(&trans_inc_lock); 124 return ret; 125 } 126 127 static inline void ocfs2_inode_set_new(struct ocfs2_super *osb, 128 struct inode *inode) 129 { 130 spin_lock(&trans_inc_lock); 131 OCFS2_I(inode)->ip_created_trans = osb->journal->j_trans_id; 132 spin_unlock(&trans_inc_lock); 133 } 134 135 /* Exported only for the journal struct init code in super.c. Do not call. */ 136 void ocfs2_complete_recovery(struct work_struct *work); 137 void ocfs2_wait_for_recovery(struct ocfs2_super *osb); 138 139 int ocfs2_recovery_init(struct ocfs2_super *osb); 140 void ocfs2_recovery_exit(struct ocfs2_super *osb); 141 142 /* 143 * Journal Control: 144 * Initialize, Load, Shutdown, Wipe a journal. 145 * 146 * ocfs2_journal_init - Initialize journal structures in the OSB. 147 * ocfs2_journal_load - Load the given journal off disk. Replay it if 148 * there's transactions still in there. 149 * ocfs2_journal_shutdown - Shutdown a journal, this will flush all 150 * uncommitted, uncheckpointed transactions. 151 * ocfs2_journal_wipe - Wipe transactions from a journal. Optionally 152 * zero out each block. 153 * ocfs2_recovery_thread - Perform recovery on a node. osb is our own osb. 154 * ocfs2_mark_dead_nodes - Start recovery on nodes we won't get a heartbeat 155 * event on. 156 * ocfs2_start_checkpoint - Kick the commit thread to do a checkpoint. 157 */ 158 void ocfs2_set_journal_params(struct ocfs2_super *osb); 159 int ocfs2_journal_init(struct ocfs2_journal *journal, 160 int *dirty); 161 void ocfs2_journal_shutdown(struct ocfs2_super *osb); 162 int ocfs2_journal_wipe(struct ocfs2_journal *journal, 163 int full); 164 int ocfs2_journal_load(struct ocfs2_journal *journal, int local, 165 int replayed); 166 int ocfs2_check_journals_nolocks(struct ocfs2_super *osb); 167 void ocfs2_recovery_thread(struct ocfs2_super *osb, 168 int node_num); 169 int ocfs2_mark_dead_nodes(struct ocfs2_super *osb); 170 void ocfs2_complete_mount_recovery(struct ocfs2_super *osb); 171 172 static inline void ocfs2_start_checkpoint(struct ocfs2_super *osb) 173 { 174 atomic_set(&osb->needs_checkpoint, 1); 175 wake_up(&osb->checkpoint_event); 176 } 177 178 static inline void ocfs2_checkpoint_inode(struct inode *inode) 179 { 180 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 181 182 if (ocfs2_mount_local(osb)) 183 return; 184 185 if (!ocfs2_inode_fully_checkpointed(inode)) { 186 /* WARNING: This only kicks off a single 187 * checkpoint. If someone races you and adds more 188 * metadata to the journal, you won't know, and will 189 * wind up waiting *alot* longer than necessary. Right 190 * now we only use this in clear_inode so that's 191 * OK. */ 192 ocfs2_start_checkpoint(osb); 193 194 wait_event(osb->journal->j_checkpointed, 195 ocfs2_inode_fully_checkpointed(inode)); 196 } 197 } 198 199 /* 200 * Transaction Handling: 201 * Manage the lifetime of a transaction handle. 202 * 203 * ocfs2_start_trans - Begin a transaction. Give it an upper estimate of 204 * the number of blocks that will be changed during 205 * this handle. 206 * ocfs2_commit_trans - Complete a handle. It might return -EIO if 207 * the journal was aborted. The majority of paths don't 208 * check the return value as an error there comes too 209 * late to do anything (and will be picked up in a 210 * later transaction). 211 * ocfs2_extend_trans - Extend a handle by nblocks credits. This may 212 * commit the handle to disk in the process, but will 213 * not release any locks taken during the transaction. 214 * ocfs2_journal_access - Notify the handle that we want to journal this 215 * buffer. Will have to call ocfs2_journal_dirty once 216 * we've actually dirtied it. Type is one of . or . 217 * ocfs2_journal_dirty - Mark a journalled buffer as having dirty data. 218 * ocfs2_journal_dirty_data - Indicate that a data buffer should go out before 219 * the current handle commits. 220 */ 221 222 /* You must always start_trans with a number of buffs > 0, but it's 223 * perfectly legal to go through an entire transaction without having 224 * dirtied any buffers. */ 225 handle_t *ocfs2_start_trans(struct ocfs2_super *osb, 226 int max_buffs); 227 int ocfs2_commit_trans(struct ocfs2_super *osb, 228 handle_t *handle); 229 int ocfs2_extend_trans(handle_t *handle, int nblocks); 230 231 /* 232 * Create access is for when we get a newly created buffer and we're 233 * not gonna read it off disk, but rather fill it ourselves. Right 234 * now, we don't do anything special with this (it turns into a write 235 * request), but this is a good placeholder in case we do... 236 * 237 * Write access is for when we read a block off disk and are going to 238 * modify it. This way the journalling layer knows it may need to make 239 * a copy of that block (if it's part of another, uncommitted 240 * transaction) before we do so. 241 */ 242 #define OCFS2_JOURNAL_ACCESS_CREATE 0 243 #define OCFS2_JOURNAL_ACCESS_WRITE 1 244 #define OCFS2_JOURNAL_ACCESS_UNDO 2 245 246 int ocfs2_journal_access(handle_t *handle, 247 struct inode *inode, 248 struct buffer_head *bh, 249 int type); 250 /* 251 * A word about the journal_access/journal_dirty "dance". It is 252 * entirely legal to journal_access a buffer more than once (as long 253 * as the access type is the same -- I'm not sure what will happen if 254 * access type is different but this should never happen anyway) It is 255 * also legal to journal_dirty a buffer more than once. In fact, you 256 * can even journal_access a buffer after you've done a 257 * journal_access/journal_dirty pair. The only thing you cannot do 258 * however, is journal_dirty a buffer which you haven't yet passed to 259 * journal_access at least once. 260 * 261 * That said, 99% of the time this doesn't matter and this is what the 262 * path looks like: 263 * 264 * <read a bh> 265 * ocfs2_journal_access(handle, bh, OCFS2_JOURNAL_ACCESS_WRITE); 266 * <modify the bh> 267 * ocfs2_journal_dirty(handle, bh); 268 */ 269 int ocfs2_journal_dirty(handle_t *handle, 270 struct buffer_head *bh); 271 int ocfs2_journal_dirty_data(handle_t *handle, 272 struct buffer_head *bh); 273 274 /* 275 * Credit Macros: 276 * Convenience macros to calculate number of credits needed. 277 * 278 * For convenience sake, I have a set of macros here which calculate 279 * the *maximum* number of sectors which will be changed for various 280 * metadata updates. 281 */ 282 283 /* simple file updates like chmod, etc. */ 284 #define OCFS2_INODE_UPDATE_CREDITS 1 285 286 /* group extend. inode update and last group update. */ 287 #define OCFS2_GROUP_EXTEND_CREDITS (OCFS2_INODE_UPDATE_CREDITS + 1) 288 289 /* group add. inode update and the new group update. */ 290 #define OCFS2_GROUP_ADD_CREDITS (OCFS2_INODE_UPDATE_CREDITS + 1) 291 292 /* get one bit out of a suballocator: dinode + group descriptor + 293 * prev. group desc. if we relink. */ 294 #define OCFS2_SUBALLOC_ALLOC (3) 295 296 #define OCFS2_INLINE_TO_EXTENTS_CREDITS (OCFS2_SUBALLOC_ALLOC \ 297 + OCFS2_INODE_UPDATE_CREDITS) 298 299 /* dinode + group descriptor update. We don't relink on free yet. */ 300 #define OCFS2_SUBALLOC_FREE (2) 301 302 #define OCFS2_TRUNCATE_LOG_UPDATE OCFS2_INODE_UPDATE_CREDITS 303 #define OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC (OCFS2_SUBALLOC_FREE \ 304 + OCFS2_TRUNCATE_LOG_UPDATE) 305 306 #define OCFS2_REMOVE_EXTENT_CREDITS (OCFS2_TRUNCATE_LOG_UPDATE + OCFS2_INODE_UPDATE_CREDITS) 307 308 /* data block for new dir/symlink, 2 for bitmap updates (bitmap fe + 309 * bitmap block for the new bit) */ 310 #define OCFS2_DIR_LINK_ADDITIONAL_CREDITS (1 + 2) 311 312 /* parent fe, parent block, new file entry, inode alloc fe, inode alloc 313 * group descriptor + mkdir/symlink blocks */ 314 #define OCFS2_MKNOD_CREDITS (3 + OCFS2_SUBALLOC_ALLOC \ 315 + OCFS2_DIR_LINK_ADDITIONAL_CREDITS) 316 317 /* local alloc metadata change + main bitmap updates */ 318 #define OCFS2_WINDOW_MOVE_CREDITS (OCFS2_INODE_UPDATE_CREDITS \ 319 + OCFS2_SUBALLOC_ALLOC + OCFS2_SUBALLOC_FREE) 320 321 /* used when we don't need an allocation change for a dir extend. One 322 * for the dinode, one for the new block. */ 323 #define OCFS2_SIMPLE_DIR_EXTEND_CREDITS (2) 324 325 /* file update (nlink, etc) + directory mtime/ctime + dir entry block */ 326 #define OCFS2_LINK_CREDITS (2*OCFS2_INODE_UPDATE_CREDITS + 1) 327 328 /* inode + dir inode (if we unlink a dir), + dir entry block + orphan 329 * dir inode link */ 330 #define OCFS2_UNLINK_CREDITS (2 * OCFS2_INODE_UPDATE_CREDITS + 1 \ 331 + OCFS2_LINK_CREDITS) 332 333 /* dinode + orphan dir dinode + inode alloc dinode + orphan dir entry + 334 * inode alloc group descriptor */ 335 #define OCFS2_DELETE_INODE_CREDITS (3 * OCFS2_INODE_UPDATE_CREDITS + 1 + 1) 336 337 /* dinode update, old dir dinode update, new dir dinode update, old 338 * dir dir entry, new dir dir entry, dir entry update for renaming 339 * directory + target unlink */ 340 #define OCFS2_RENAME_CREDITS (3 * OCFS2_INODE_UPDATE_CREDITS + 3 \ 341 + OCFS2_UNLINK_CREDITS) 342 343 static inline int ocfs2_calc_extend_credits(struct super_block *sb, 344 struct ocfs2_dinode *fe, 345 u32 bits_wanted) 346 { 347 int bitmap_blocks, sysfile_bitmap_blocks, dinode_blocks; 348 349 /* bitmap dinode, group desc. + relinked group. */ 350 bitmap_blocks = OCFS2_SUBALLOC_ALLOC; 351 352 /* we might need to shift tree depth so lets assume an 353 * absolute worst case of complete fragmentation. Even with 354 * that, we only need one update for the dinode, and then 355 * however many metadata chunks needed * a remaining suballoc 356 * alloc. */ 357 sysfile_bitmap_blocks = 1 + 358 (OCFS2_SUBALLOC_ALLOC - 1) * ocfs2_extend_meta_needed(fe); 359 360 /* this does not include *new* metadata blocks, which are 361 * accounted for in sysfile_bitmap_blocks. fe + 362 * prev. last_eb_blk + blocks along edge of tree. 363 * calc_symlink_credits passes because we just need 1 364 * credit for the dinode there. */ 365 dinode_blocks = 1 + 1 + le16_to_cpu(fe->id2.i_list.l_tree_depth); 366 367 return bitmap_blocks + sysfile_bitmap_blocks + dinode_blocks; 368 } 369 370 static inline int ocfs2_calc_symlink_credits(struct super_block *sb) 371 { 372 int blocks = OCFS2_MKNOD_CREDITS; 373 374 /* links can be longer than one block so we may update many 375 * within our single allocated extent. */ 376 blocks += ocfs2_clusters_to_blocks(sb, 1); 377 378 return blocks; 379 } 380 381 static inline int ocfs2_calc_group_alloc_credits(struct super_block *sb, 382 unsigned int cpg) 383 { 384 int blocks; 385 int bitmap_blocks = OCFS2_SUBALLOC_ALLOC + 1; 386 /* parent inode update + new block group header + bitmap inode update 387 + bitmap blocks affected */ 388 blocks = 1 + 1 + 1 + bitmap_blocks; 389 return blocks; 390 } 391 392 static inline int ocfs2_calc_tree_trunc_credits(struct super_block *sb, 393 unsigned int clusters_to_del, 394 struct ocfs2_dinode *fe, 395 struct ocfs2_extent_list *last_el) 396 { 397 /* for dinode + all headers in this pass + update to next leaf */ 398 u16 next_free = le16_to_cpu(last_el->l_next_free_rec); 399 u16 tree_depth = le16_to_cpu(fe->id2.i_list.l_tree_depth); 400 int credits = 1 + tree_depth + 1; 401 int i; 402 403 i = next_free - 1; 404 BUG_ON(i < 0); 405 406 /* We may be deleting metadata blocks, so metadata alloc dinode + 407 one desc. block for each possible delete. */ 408 if (tree_depth && next_free == 1 && 409 ocfs2_rec_clusters(last_el, &last_el->l_recs[i]) == clusters_to_del) 410 credits += 1 + tree_depth; 411 412 /* update to the truncate log. */ 413 credits += OCFS2_TRUNCATE_LOG_UPDATE; 414 415 return credits; 416 } 417 418 #endif /* OCFS2_JOURNAL_H */ 419