1 /*- 2 * Copyright 1998, 2000 Marshall Kirk McKusick. All Rights Reserved. 3 * 4 * The soft updates code is derived from the appendix of a University 5 * of Michigan technical report (Gregory R. Ganger and Yale N. Patt, 6 * "Soft Updates: A Solution to the Metadata Update Problem in File 7 * Systems", CSE-TR-254-95, August 1995). 8 * 9 * Further information about soft updates can be obtained from: 10 * 11 * Marshall Kirk McKusick http://www.mckusick.com/softdep/ 12 * 1614 Oxford Street mckusick@mckusick.com 13 * Berkeley, CA 94709-1608 +1-510-843-9542 14 * USA 15 * 16 * Redistribution and use in source and binary forms, with or without 17 * modification, are permitted provided that the following conditions 18 * are met: 19 * 20 * 1. Redistributions of source code must retain the above copyright 21 * notice, this list of conditions and the following disclaimer. 22 * 2. Redistributions in binary form must reproduce the above copyright 23 * notice, this list of conditions and the following disclaimer in the 24 * documentation and/or other materials provided with the distribution. 25 * 26 * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY 27 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 28 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 29 * DISCLAIMED. IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR 30 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 * 38 * @(#)softdep.h 9.7 (McKusick) 6/21/00 39 * $FreeBSD: src/sys/ufs/ffs/softdep.h,v 1.7.2.1 2000/06/22 19:27:42 peter Exp $ 40 * $DragonFly: src/sys/vfs/ufs/softdep.h,v 1.5 2006/04/03 02:02:37 dillon Exp $ 41 */ 42 43 #include <sys/queue.h> 44 45 /* 46 * Allocation dependencies are handled with undo/redo on the in-memory 47 * copy of the data. A particular data dependency is eliminated when 48 * it is ALLCOMPLETE: that is ATTACHED, DEPCOMPLETE, and COMPLETE. 49 * 50 * ATTACHED means that the data is not currently being written to 51 * disk. UNDONE means that the data has been rolled back to a safe 52 * state for writing to the disk. When the I/O completes, the data is 53 * restored to its current form and the state reverts to ATTACHED. 54 * The data must be locked throughout the rollback, I/O, and roll 55 * forward so that the rolled back information is never visible to 56 * user processes. The COMPLETE flag indicates that the item has been 57 * written. For example, a dependency that requires that an inode be 58 * written will be marked COMPLETE after the inode has been written 59 * to disk. The DEPCOMPLETE flag indicates the completion of any other 60 * dependencies such as the writing of a cylinder group map has been 61 * completed. A dependency structure may be freed only when both it 62 * and its dependencies have completed and any rollbacks that are in 63 * progress have finished as indicated by the set of ALLCOMPLETE flags 64 * all being set. The two MKDIR flags indicate additional dependencies 65 * that must be done when creating a new directory. MKDIR_BODY is 66 * cleared when the directory data block containing the "." and ".." 67 * entries has been written. MKDIR_PARENT is cleared when the parent 68 * inode with the increased link count for ".." has been written. When 69 * both MKDIR flags have been cleared, the DEPCOMPLETE flag is set to 70 * indicate that the directory dependencies have been completed. The 71 * writing of the directory inode itself sets the COMPLETE flag which 72 * then allows the directory entry for the new directory to be written 73 * to disk. The RMDIR flag marks a dirrem structure as representing 74 * the removal of a directory rather than a file. When the removal 75 * dependencies are completed, additional work needs to be done 76 * (truncation of the "." and ".." entries, an additional decrement 77 * of the associated inode, and a decrement of the parent inode). The 78 * DIRCHG flag marks a diradd structure as representing the changing 79 * of an existing entry rather than the addition of a new one. When 80 * the update is complete the dirrem associated with the inode for 81 * the old name must be added to the worklist to do the necessary 82 * reference count decrement. The GOINGAWAY flag indicates that the 83 * data structure is frozen from further change until its dependencies 84 * have been completed and its resources freed after which it will be 85 * discarded. The IOSTARTED flag prevents multiple calls to the I/O 86 * start routine from doing multiple rollbacks. The ONWORKLIST flag 87 * shows whether the structure is currently linked onto a worklist. 88 */ 89 #define ATTACHED 0x0001 90 #define UNDONE 0x0002 91 #define COMPLETE 0x0004 92 #define DEPCOMPLETE 0x0008 93 #define MKDIR_PARENT 0x0010 94 #define MKDIR_BODY 0x0020 95 #define RMDIR 0x0040 96 #define DIRCHG 0x0080 97 #define GOINGAWAY 0x0100 98 #define IOSTARTED 0x0200 99 #define ONWORKLIST 0x8000 100 101 #define ALLCOMPLETE (ATTACHED | COMPLETE | DEPCOMPLETE) 102 103 /* 104 * The workitem queue. 105 * 106 * It is sometimes useful and/or necessary to clean up certain dependencies 107 * in the background rather than during execution of an application process 108 * or interrupt service routine. To realize this, we append dependency 109 * structures corresponding to such tasks to a "workitem" queue. In a soft 110 * updates implementation, most pending workitems should not wait for more 111 * than a couple of seconds, so the filesystem syncer process awakens once 112 * per second to process the items on the queue. 113 */ 114 115 /* LIST_HEAD(workhead, worklist); -- declared in buf.h */ 116 117 /* 118 * Each request can be linked onto a work queue through its worklist structure. 119 * To avoid the need for a pointer to the structure itself, this structure 120 * MUST be declared FIRST in each type in which it appears! If more than one 121 * worklist is needed in the structure, then a wk_data field must be added 122 * and the macros below changed to use it. 123 */ 124 struct worklist { 125 LIST_ENTRY(worklist) wk_list; /* list of work requests */ 126 unsigned short wk_type; /* type of request */ 127 unsigned short wk_state; /* state flags */ 128 }; 129 #define WK_DATA(wk) ((void *)(wk)) 130 #define WK_PAGEDEP(wk) ((struct pagedep *)(wk)) 131 #define WK_INODEDEP(wk) ((struct inodedep *)(wk)) 132 #define WK_NEWBLK(wk) ((struct newblk *)(wk)) 133 #define WK_BMSAFEMAP(wk) ((struct bmsafemap *)(wk)) 134 #define WK_ALLOCDIRECT(wk) ((struct allocdirect *)(wk)) 135 #define WK_INDIRDEP(wk) ((struct indirdep *)(wk)) 136 #define WK_ALLOCINDIR(wk) ((struct allocindir *)(wk)) 137 #define WK_FREEFRAG(wk) ((struct freefrag *)(wk)) 138 #define WK_FREEBLKS(wk) ((struct freeblks *)(wk)) 139 #define WK_FREEFILE(wk) ((struct freefile *)(wk)) 140 #define WK_DIRADD(wk) ((struct diradd *)(wk)) 141 #define WK_MKDIR(wk) ((struct mkdir *)(wk)) 142 #define WK_DIRREM(wk) ((struct dirrem *)(wk)) 143 144 /* 145 * Various types of lists 146 */ 147 LIST_HEAD(dirremhd, dirrem); 148 LIST_HEAD(diraddhd, diradd); 149 LIST_HEAD(newblkhd, newblk); 150 LIST_HEAD(inodedephd, inodedep); 151 LIST_HEAD(allocindirhd, allocindir); 152 LIST_HEAD(allocdirecthd, allocdirect); 153 TAILQ_HEAD(allocdirectlst, allocdirect); 154 155 /* 156 * The "pagedep" structure tracks the various dependencies related to 157 * a particular directory page. If a directory page has any dependencies, 158 * it will have a pagedep linked to its associated buffer. The 159 * pd_dirremhd list holds the list of dirrem requests which decrement 160 * inode reference counts. These requests are processed after the 161 * directory page with the corresponding zero'ed entries has been 162 * written. The pd_diraddhd list maintains the list of diradd requests 163 * which cannot be committed until their corresponding inode has been 164 * written to disk. Because a directory may have many new entries 165 * being created, several lists are maintained hashed on bits of the 166 * offset of the entry into the directory page to keep the lists from 167 * getting too long. Once a new directory entry has been cleared to 168 * be written, it is moved to the pd_pendinghd list. After the new 169 * entry has been written to disk it is removed from the pd_pendinghd 170 * list, any removed operations are done, and the dependency structure 171 * is freed. 172 */ 173 #define DAHASHSZ 6 174 #define DIRADDHASH(offset) (((offset) >> 2) % DAHASHSZ) 175 struct pagedep { 176 struct worklist pd_list; /* page buffer */ 177 # define pd_state pd_list.wk_state /* check for multiple I/O starts */ 178 LIST_ENTRY(pagedep) pd_hash; /* hashed lookup */ 179 struct mount *pd_mnt; /* associated mount point */ 180 ino_t pd_ino; /* associated file */ 181 ufs_lbn_t pd_lbn; /* block within file */ 182 struct dirremhd pd_dirremhd; /* dirrem's waiting for page */ 183 struct diraddhd pd_diraddhd[DAHASHSZ]; /* diradd dir entry updates */ 184 struct diraddhd pd_pendinghd; /* directory entries awaiting write */ 185 }; 186 187 /* 188 * The "inodedep" structure tracks the set of dependencies associated 189 * with an inode. One task that it must manage is delayed operations 190 * (i.e., work requests that must be held until the inodedep's associated 191 * inode has been written to disk). Getting an inode from its incore 192 * state to the disk requires two steps to be taken by the filesystem 193 * in this order: first the inode must be copied to its disk buffer by 194 * the VOP_UPDATE operation; second the inode's buffer must be written 195 * to disk. To ensure that both operations have happened in the required 196 * order, the inodedep maintains two lists. Delayed operations are 197 * placed on the id_inowait list. When the VOP_UPDATE is done, all 198 * operations on the id_inowait list are moved to the id_bufwait list. 199 * When the buffer is written, the items on the id_bufwait list can be 200 * safely moved to the work queue to be processed. A second task of the 201 * inodedep structure is to track the status of block allocation within 202 * the inode. Each block that is allocated is represented by an 203 * "allocdirect" structure (see below). It is linked onto the id_newinoupdt 204 * list until both its contents and its allocation in the cylinder 205 * group map have been written to disk. Once these dependencies have been 206 * satisfied, it is removed from the id_newinoupdt list and any followup 207 * actions such as releasing the previous block or fragment are placed 208 * on the id_inowait list. When an inode is updated (a VOP_UPDATE is 209 * done), the "inodedep" structure is linked onto the buffer through 210 * its worklist. Thus, it will be notified when the buffer is about 211 * to be written and when it is done. At the update time, all the 212 * elements on the id_newinoupdt list are moved to the id_inoupdt list 213 * since those changes are now relevant to the copy of the inode in the 214 * buffer. Also at update time, the tasks on the id_inowait list are 215 * moved to the id_bufwait list so that they will be executed when 216 * the updated inode has been written to disk. When the buffer containing 217 * the inode is written to disk, any updates listed on the id_inoupdt 218 * list are rolled back as they are not yet safe. Following the write, 219 * the changes are once again rolled forward and any actions on the 220 * id_bufwait list are processed (since those actions are now safe). 221 * The entries on the id_inoupdt and id_newinoupdt lists must be kept 222 * sorted by logical block number to speed the calculation of the size 223 * of the rolled back inode (see explanation in initiate_write_inodeblock). 224 * When a directory entry is created, it is represented by a diradd. 225 * The diradd is added to the id_inowait list as it cannot be safely 226 * written to disk until the inode that it represents is on disk. After 227 * the inode is written, the id_bufwait list is processed and the diradd 228 * entries are moved to the id_pendinghd list where they remain until 229 * the directory block containing the name has been written to disk. 230 * The purpose of keeping the entries on the id_pendinghd list is so that 231 * the softdep_fsync function can find and push the inode's directory 232 * name(s) as part of the fsync operation for that file. 233 */ 234 struct inodedep { 235 struct worklist id_list; /* buffer holding inode block */ 236 # define id_state id_list.wk_state /* inode dependency state */ 237 LIST_ENTRY(inodedep) id_hash; /* hashed lookup */ 238 struct fs *id_fs; /* associated filesystem */ 239 ino_t id_ino; /* dependent inode */ 240 nlink_t id_nlinkdelta; /* saved effective link count */ 241 struct ufs1_dinode *id_savedino; /* saved dinode contents */ 242 LIST_ENTRY(inodedep) id_deps; /* bmsafemap's list of inodedep's */ 243 struct buf *id_buf; /* related bmsafemap (if pending) */ 244 off_t id_savedsize; /* file size saved during rollback */ 245 struct workhead id_pendinghd; /* entries awaiting directory write */ 246 struct workhead id_bufwait; /* operations after inode written */ 247 struct workhead id_inowait; /* operations waiting inode update */ 248 struct allocdirectlst id_inoupdt; /* updates before inode written */ 249 struct allocdirectlst id_newinoupdt; /* updates when inode written */ 250 }; 251 252 /* 253 * A "newblk" structure is attached to a bmsafemap structure when a block 254 * or fragment is allocated from a cylinder group. Its state is set to 255 * DEPCOMPLETE when its cylinder group map is written. It is consumed by 256 * an associated allocdirect or allocindir allocation which will attach 257 * themselves to the bmsafemap structure if the newblk's DEPCOMPLETE flag 258 * is not set (i.e., its cylinder group map has not been written). 259 */ 260 struct newblk { 261 LIST_ENTRY(newblk) nb_hash; /* hashed lookup */ 262 struct fs *nb_fs; /* associated filesystem */ 263 ufs_daddr_t nb_newblkno; /* allocated block number */ 264 int nb_state; /* state of bitmap dependency */ 265 LIST_ENTRY(newblk) nb_deps; /* bmsafemap's list of newblk's */ 266 struct bmsafemap *nb_bmsafemap; /* associated bmsafemap */ 267 }; 268 269 /* 270 * A "bmsafemap" structure maintains a list of dependency structures 271 * that depend on the update of a particular cylinder group map. 272 * It has lists for newblks, allocdirects, allocindirs, and inodedeps. 273 * It is attached to the buffer of a cylinder group block when any of 274 * these things are allocated from the cylinder group. It is freed 275 * after the cylinder group map is written and the state of its 276 * dependencies are updated with DEPCOMPLETE to indicate that it has 277 * been processed. 278 */ 279 struct bmsafemap { 280 struct worklist sm_list; /* cylgrp buffer */ 281 struct buf *sm_buf; /* associated buffer */ 282 struct allocdirecthd sm_allocdirecthd; /* allocdirect deps */ 283 struct allocindirhd sm_allocindirhd; /* allocindir deps */ 284 struct inodedephd sm_inodedephd; /* inodedep deps */ 285 struct newblkhd sm_newblkhd; /* newblk deps */ 286 }; 287 288 /* 289 * An "allocdirect" structure is attached to an "inodedep" when a new block 290 * or fragment is allocated and pointed to by the inode described by 291 * "inodedep". The worklist is linked to the buffer that holds the block. 292 * When the block is first allocated, it is linked to the bmsafemap 293 * structure associated with the buffer holding the cylinder group map 294 * from which it was allocated. When the cylinder group map is written 295 * to disk, ad_state has the DEPCOMPLETE flag set. When the block itself 296 * is written, the COMPLETE flag is set. Once both the cylinder group map 297 * and the data itself have been written, it is safe to write the inode 298 * that claims the block. If there was a previous fragment that had been 299 * allocated before the file was increased in size, the old fragment may 300 * be freed once the inode claiming the new block is written to disk. 301 * This ad_fragfree request is attached to the id_inowait list of the 302 * associated inodedep (pointed to by ad_inodedep) for processing after 303 * the inode is written. 304 */ 305 struct allocdirect { 306 struct worklist ad_list; /* buffer holding block */ 307 # define ad_state ad_list.wk_state /* block pointer state */ 308 TAILQ_ENTRY(allocdirect) ad_next; /* inodedep's list of allocdirect's */ 309 ufs_lbn_t ad_lbn; /* block within file */ 310 ufs_daddr_t ad_newblkno; /* new value of block pointer */ 311 ufs_daddr_t ad_oldblkno; /* old value of block pointer */ 312 long ad_newsize; /* size of new block */ 313 long ad_oldsize; /* size of old block */ 314 LIST_ENTRY(allocdirect) ad_deps; /* bmsafemap's list of allocdirect's */ 315 struct buf *ad_buf; /* cylgrp buffer (if pending) */ 316 struct inodedep *ad_inodedep; /* associated inodedep */ 317 struct freefrag *ad_freefrag; /* fragment to be freed (if any) */ 318 }; 319 320 /* 321 * A single "indirdep" structure manages all allocation dependencies for 322 * pointers in an indirect block. The up-to-date state of the indirect 323 * block is stored in ir_savedata. The set of pointers that may be safely 324 * written to the disk is stored in ir_safecopy. The state field is used 325 * only to track whether the buffer is currently being written (in which 326 * case it is not safe to update ir_safecopy). Ir_deplisthd contains the 327 * list of allocindir structures, one for each block that needs to be 328 * written to disk. Once the block and its bitmap allocation have been 329 * written the safecopy can be updated to reflect the allocation and the 330 * allocindir structure freed. If ir_state indicates that an I/O on the 331 * indirect block is in progress when ir_safecopy is to be updated, the 332 * update is deferred by placing the allocindir on the ir_donehd list. 333 * When the I/O on the indirect block completes, the entries on the 334 * ir_donehd list are processed by updating their corresponding ir_safecopy 335 * pointers and then freeing the allocindir structure. 336 */ 337 struct indirdep { 338 struct worklist ir_list; /* buffer holding indirect block */ 339 # define ir_state ir_list.wk_state /* indirect block pointer state */ 340 caddr_t ir_saveddata; /* buffer cache contents */ 341 struct buf *ir_savebp; /* buffer holding safe copy */ 342 struct allocindirhd ir_donehd; /* done waiting to update safecopy */ 343 struct allocindirhd ir_deplisthd; /* allocindir deps for this block */ 344 }; 345 346 /* 347 * An "allocindir" structure is attached to an "indirdep" when a new block 348 * is allocated and pointed to by the indirect block described by the 349 * "indirdep". The worklist is linked to the buffer that holds the new block. 350 * When the block is first allocated, it is linked to the bmsafemap 351 * structure associated with the buffer holding the cylinder group map 352 * from which it was allocated. When the cylinder group map is written 353 * to disk, ai_state has the DEPCOMPLETE flag set. When the block itself 354 * is written, the COMPLETE flag is set. Once both the cylinder group map 355 * and the data itself have been written, it is safe to write the entry in 356 * the indirect block that claims the block; the "allocindir" dependency 357 * can then be freed as it is no longer applicable. 358 */ 359 struct allocindir { 360 struct worklist ai_list; /* buffer holding indirect block */ 361 # define ai_state ai_list.wk_state /* indirect block pointer state */ 362 LIST_ENTRY(allocindir) ai_next; /* indirdep's list of allocindir's */ 363 int ai_offset; /* pointer offset in indirect block */ 364 ufs_daddr_t ai_newblkno; /* new block pointer value */ 365 ufs_daddr_t ai_oldblkno; /* old block pointer value */ 366 struct freefrag *ai_freefrag; /* block to be freed when complete */ 367 struct indirdep *ai_indirdep; /* address of associated indirdep */ 368 LIST_ENTRY(allocindir) ai_deps; /* bmsafemap's list of allocindir's */ 369 struct buf *ai_buf; /* cylgrp buffer (if pending) */ 370 }; 371 372 /* 373 * A "freefrag" structure is attached to an "inodedep" when a previously 374 * allocated fragment is replaced with a larger fragment, rather than extended. 375 * The "freefrag" structure is constructed and attached when the replacement 376 * block is first allocated. It is processed after the inode claiming the 377 * bigger block that replaces it has been written to disk. Note that the 378 * ff_state field is is used to store the uid, so may lose data. However, 379 * the uid is used only in printing an error message, so is not critical. 380 * Keeping it in a short keeps the data structure down to 32 bytes. 381 */ 382 struct freefrag { 383 struct worklist ff_list; /* id_inowait or delayed worklist */ 384 # define ff_state ff_list.wk_state /* owning user; should be uid_t */ 385 struct vnode *ff_devvp; /* filesystem device vnode */ 386 struct fs *ff_fs; /* addr of superblock */ 387 ufs_daddr_t ff_blkno; /* fragment physical block number */ 388 long ff_fragsize; /* size of fragment being deleted */ 389 ino_t ff_inum; /* owning inode number */ 390 }; 391 392 /* 393 * A "freeblks" structure is attached to an "inodedep" when the 394 * corresponding file's length is reduced to zero. It records all 395 * the information needed to free the blocks of a file after its 396 * zero'ed inode has been written to disk. 397 */ 398 struct freeblks { 399 struct worklist fb_list; /* id_inowait or delayed worklist */ 400 # define fb_state fb_list.wk_state /* inode and dirty block state */ 401 ino_t fb_previousinum; /* inode of previous owner of blocks */ 402 struct vnode *fb_devvp; /* filesystem device vnode */ 403 struct fs *fb_fs; /* addr of superblock */ 404 off_t fb_oldsize; /* previous file size */ 405 off_t fb_newsize; /* new file size */ 406 int fb_chkcnt; /* used to check cnt of blks released */ 407 uid_t fb_uid; /* uid of previous owner of blocks */ 408 ufs_daddr_t fb_dblks[NDADDR]; /* direct blk ptrs to deallocate */ 409 ufs_daddr_t fb_iblks[NIADDR]; /* indirect blk ptrs to deallocate */ 410 }; 411 412 /* 413 * A "freefile" structure is attached to an inode when its 414 * link count is reduced to zero. It marks the inode as free in 415 * the cylinder group map after the zero'ed inode has been written 416 * to disk and any associated blocks and fragments have been freed. 417 */ 418 struct freefile { 419 struct worklist fx_list; /* id_inowait or delayed worklist */ 420 mode_t fx_mode; /* mode of inode */ 421 ino_t fx_oldinum; /* inum of the unlinked file */ 422 struct vnode *fx_devvp; /* filesystem device vnode */ 423 struct fs *fx_fs; /* addr of superblock */ 424 }; 425 426 /* 427 * A "diradd" structure is linked to an "inodedep" id_inowait list when a 428 * new directory entry is allocated that references the inode described 429 * by "inodedep". When the inode itself is written (either the initial 430 * allocation for new inodes or with the increased link count for 431 * existing inodes), the COMPLETE flag is set in da_state. If the entry 432 * is for a newly allocated inode, the "inodedep" structure is associated 433 * with a bmsafemap which prevents the inode from being written to disk 434 * until the cylinder group has been updated. Thus the da_state COMPLETE 435 * flag cannot be set until the inode bitmap dependency has been removed. 436 * When creating a new file, it is safe to write the directory entry that 437 * claims the inode once the referenced inode has been written. Since 438 * writing the inode clears the bitmap dependencies, the DEPCOMPLETE flag 439 * in the diradd can be set unconditionally when creating a file. When 440 * creating a directory, there are two additional dependencies described by 441 * mkdir structures (see their description below). When these dependencies 442 * are resolved the DEPCOMPLETE flag is set in the diradd structure. 443 * If there are multiple links created to the same inode, there will be 444 * a separate diradd structure created for each link. The diradd is 445 * linked onto the pg_diraddhd list of the pagedep for the directory 446 * page that contains the entry. When a directory page is written, 447 * the pg_diraddhd list is traversed to rollback any entries that are 448 * not yet ready to be written to disk. If a directory entry is being 449 * changed (by rename) rather than added, the DIRCHG flag is set and 450 * the da_previous entry points to the entry that will be "removed" 451 * once the new entry has been committed. During rollback, entries 452 * with da_previous are replaced with the previous inode number rather 453 * than zero. 454 * 455 * The overlaying of da_pagedep and da_previous is done to keep the 456 * structure down to 32 bytes in size on a 32-bit machine. If a 457 * da_previous entry is present, the pointer to its pagedep is available 458 * in the associated dirrem entry. If the DIRCHG flag is set, the 459 * da_previous entry is valid; if not set the da_pagedep entry is valid. 460 * The DIRCHG flag never changes; it is set when the structure is created 461 * if appropriate and is never cleared. 462 */ 463 struct diradd { 464 struct worklist da_list; /* id_inowait or id_pendinghd list */ 465 # define da_state da_list.wk_state /* state of the new directory entry */ 466 LIST_ENTRY(diradd) da_pdlist; /* pagedep holding directory block */ 467 doff_t da_offset; /* offset of new dir entry in dir blk */ 468 ino_t da_newinum; /* inode number for the new dir entry */ 469 union { 470 struct dirrem *dau_previous; /* entry being replaced in dir change */ 471 struct pagedep *dau_pagedep; /* pagedep dependency for addition */ 472 } da_un; 473 }; 474 #define da_previous da_un.dau_previous 475 #define da_pagedep da_un.dau_pagedep 476 477 /* 478 * Two "mkdir" structures are needed to track the additional dependencies 479 * associated with creating a new directory entry. Normally a directory 480 * addition can be committed as soon as the newly referenced inode has been 481 * written to disk with its increased link count. When a directory is 482 * created there are two additional dependencies: writing the directory 483 * data block containing the "." and ".." entries (MKDIR_BODY) and writing 484 * the parent inode with the increased link count for ".." (MKDIR_PARENT). 485 * These additional dependencies are tracked by two mkdir structures that 486 * reference the associated "diradd" structure. When they have completed, 487 * they set the DEPCOMPLETE flag on the diradd so that it knows that its 488 * extra dependencies have been completed. The md_state field is used only 489 * to identify which type of dependency the mkdir structure is tracking. 490 * It is not used in the mainline code for any purpose other than consistency 491 * checking. All the mkdir structures in the system are linked together on 492 * a list. This list is needed so that a diradd can find its associated 493 * mkdir structures and deallocate them if it is prematurely freed (as for 494 * example if a mkdir is immediately followed by a rmdir of the same directory). 495 * Here, the free of the diradd must traverse the list to find the associated 496 * mkdir structures that reference it. The deletion would be faster if the 497 * diradd structure were simply augmented to have two pointers that referenced 498 * the associated mkdir's. However, this would increase the size of the diradd 499 * structure from 32 to 64-bits to speed a very infrequent operation. 500 */ 501 struct mkdir { 502 struct worklist md_list; /* id_inowait or buffer holding dir */ 503 # define md_state md_list.wk_state /* type: MKDIR_PARENT or MKDIR_BODY */ 504 struct diradd *md_diradd; /* associated diradd */ 505 struct buf *md_buf; /* MKDIR_BODY: buffer holding dir */ 506 LIST_ENTRY(mkdir) md_mkdirs; /* list of all mkdirs */ 507 }; 508 LIST_HEAD(mkdirlist, mkdir) mkdirlisthd; 509 510 /* 511 * A "dirrem" structure describes an operation to decrement the link 512 * count on an inode. The dirrem structure is attached to the pg_dirremhd 513 * list of the pagedep for the directory page that contains the entry. 514 * It is processed after the directory page with the deleted entry has 515 * been written to disk. 516 * 517 * The overlaying of dm_pagedep and dm_dirinum is done to keep the 518 * structure down to 32 bytes in size on a 32-bit machine. It works 519 * because they are never used concurrently. 520 */ 521 struct dirrem { 522 struct worklist dm_list; /* delayed worklist */ 523 # define dm_state dm_list.wk_state /* state of the old directory entry */ 524 LIST_ENTRY(dirrem) dm_next; /* pagedep's list of dirrem's */ 525 struct mount *dm_mnt; /* associated mount point */ 526 ino_t dm_oldinum; /* inum of the removed dir entry */ 527 union { 528 struct pagedep *dmu_pagedep; /* pagedep dependency for remove */ 529 ino_t dmu_dirinum; /* parent inode number (for rmdir) */ 530 } dm_un; 531 }; 532 #define dm_pagedep dm_un.dmu_pagedep 533 #define dm_dirinum dm_un.dmu_dirinum 534