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 * $FreeBSD$ 40 */ 41 42 #include <sys/queue.h> 43 44 /* 45 * Allocation dependencies are handled with undo/redo on the in-memory 46 * copy of the data. A particular data dependency is eliminated when 47 * it is ALLCOMPLETE: that is ATTACHED, DEPCOMPLETE, and COMPLETE. 48 * 49 * ATTACHED means that the data is not currently being written to 50 * disk. UNDONE means that the data has been rolled back to a safe 51 * state for writing to the disk. When the I/O completes, the data is 52 * restored to its current form and the state reverts to ATTACHED. 53 * The data must be locked throughout the rollback, I/O, and roll 54 * forward so that the rolled back information is never visible to 55 * user processes. The COMPLETE flag indicates that the item has been 56 * written. For example, a dependency that requires that an inode be 57 * written will be marked COMPLETE after the inode has been written 58 * to disk. The DEPCOMPLETE flag indicates the completion of any other 59 * dependencies such as the writing of a cylinder group map has been 60 * completed. A dependency structure may be freed only when both it 61 * and its dependencies have completed and any rollbacks that are in 62 * progress have finished as indicated by the set of ALLCOMPLETE flags 63 * all being set. The two MKDIR flags indicate additional dependencies 64 * that must be done when creating a new directory. MKDIR_BODY is 65 * cleared when the directory data block containing the "." and ".." 66 * entries has been written. MKDIR_PARENT is cleared when the parent 67 * inode with the increased link count for ".." has been written. When 68 * both MKDIR flags have been cleared, the DEPCOMPLETE flag is set to 69 * indicate that the directory dependencies have been completed. The 70 * writing of the directory inode itself sets the COMPLETE flag which 71 * then allows the directory entry for the new directory to be written 72 * to disk. The RMDIR flag marks a dirrem structure as representing 73 * the removal of a directory rather than a file. When the removal 74 * dependencies are completed, additional work needs to be done 75 * (truncation of the "." and ".." entries, an additional decrement 76 * of the associated inode, and a decrement of the parent inode). The 77 * DIRCHG flag marks a diradd structure as representing the changing 78 * of an existing entry rather than the addition of a new one. When 79 * the update is complete the dirrem associated with the inode for 80 * the old name must be added to the worklist to do the necessary 81 * reference count decrement. The GOINGAWAY flag indicates that the 82 * data structure is frozen from further change until its dependencies 83 * have been completed and its resources freed after which it will be 84 * discarded. The IOSTARTED flag prevents multiple calls to the I/O 85 * start routine from doing multiple rollbacks. The SPACECOUNTED flag 86 * says that the files space has been accounted to the pending free 87 * space count. The NEWBLOCK flag marks pagedep structures that have 88 * just been allocated, so must be claimed by the inode before all 89 * dependencies are complete. The INPROGRESS flag marks worklist 90 * structures that are still on the worklist, but are being considered 91 * for action by some process. The UFS1FMT flag indicates that the 92 * inode being processed is a ufs1 format. The ONWORKLIST flag shows 93 * whether the structure is currently linked onto a worklist. 94 */ 95 #define ATTACHED 0x0001 96 #define UNDONE 0x0002 97 #define COMPLETE 0x0004 98 #define DEPCOMPLETE 0x0008 99 #define MKDIR_PARENT 0x0010 /* diradd & mkdir only */ 100 #define MKDIR_BODY 0x0020 /* diradd & mkdir only */ 101 #define RMDIR 0x0040 /* dirrem only */ 102 #define DIRCHG 0x0080 /* diradd & dirrem only */ 103 #define GOINGAWAY 0x0100 /* indirdep only */ 104 #define IOSTARTED 0x0200 /* inodedep & pagedep only */ 105 #define SPACECOUNTED 0x0400 /* inodedep only */ 106 #define NEWBLOCK 0x0800 /* pagedep only */ 107 #define INPROGRESS 0x1000 /* dirrem, freeblks, freefrag, freefile only */ 108 #define UFS1FMT 0x2000 /* indirdep only */ 109 #define ONWORKLIST 0x8000 110 111 #define ALLCOMPLETE (ATTACHED | COMPLETE | DEPCOMPLETE) 112 113 /* 114 * The workitem queue. 115 * 116 * It is sometimes useful and/or necessary to clean up certain dependencies 117 * in the background rather than during execution of an application process 118 * or interrupt service routine. To realize this, we append dependency 119 * structures corresponding to such tasks to a "workitem" queue. In a soft 120 * updates implementation, most pending workitems should not wait for more 121 * than a couple of seconds, so the filesystem syncer process awakens once 122 * per second to process the items on the queue. 123 */ 124 125 /* LIST_HEAD(workhead, worklist); -- declared in buf.h */ 126 127 /* 128 * Each request can be linked onto a work queue through its worklist structure. 129 * To avoid the need for a pointer to the structure itself, this structure 130 * MUST be declared FIRST in each type in which it appears! If more than one 131 * worklist is needed in the structure, then a wk_data field must be added 132 * and the macros below changed to use it. 133 */ 134 struct worklist { 135 LIST_ENTRY(worklist) wk_list; /* list of work requests */ 136 unsigned short wk_type; /* type of request */ 137 unsigned short wk_state; /* state flags */ 138 }; 139 #define WK_DATA(wk) ((void *)(wk)) 140 #define WK_PAGEDEP(wk) ((struct pagedep *)(wk)) 141 #define WK_INODEDEP(wk) ((struct inodedep *)(wk)) 142 #define WK_NEWBLK(wk) ((struct newblk *)(wk)) 143 #define WK_BMSAFEMAP(wk) ((struct bmsafemap *)(wk)) 144 #define WK_ALLOCDIRECT(wk) ((struct allocdirect *)(wk)) 145 #define WK_INDIRDEP(wk) ((struct indirdep *)(wk)) 146 #define WK_ALLOCINDIR(wk) ((struct allocindir *)(wk)) 147 #define WK_FREEFRAG(wk) ((struct freefrag *)(wk)) 148 #define WK_FREEBLKS(wk) ((struct freeblks *)(wk)) 149 #define WK_FREEFILE(wk) ((struct freefile *)(wk)) 150 #define WK_DIRADD(wk) ((struct diradd *)(wk)) 151 #define WK_MKDIR(wk) ((struct mkdir *)(wk)) 152 #define WK_DIRREM(wk) ((struct dirrem *)(wk)) 153 #define WK_NEWDIRBLK(wk) ((struct newdirblk *)(wk)) 154 155 /* 156 * Various types of lists 157 */ 158 LIST_HEAD(dirremhd, dirrem); 159 LIST_HEAD(diraddhd, diradd); 160 LIST_HEAD(newblkhd, newblk); 161 LIST_HEAD(inodedephd, inodedep); 162 LIST_HEAD(allocindirhd, allocindir); 163 LIST_HEAD(allocdirecthd, allocdirect); 164 TAILQ_HEAD(allocdirectlst, allocdirect); 165 166 /* 167 * The "pagedep" structure tracks the various dependencies related to 168 * a particular directory page. If a directory page has any dependencies, 169 * it will have a pagedep linked to its associated buffer. The 170 * pd_dirremhd list holds the list of dirrem requests which decrement 171 * inode reference counts. These requests are processed after the 172 * directory page with the corresponding zero'ed entries has been 173 * written. The pd_diraddhd list maintains the list of diradd requests 174 * which cannot be committed until their corresponding inode has been 175 * written to disk. Because a directory may have many new entries 176 * being created, several lists are maintained hashed on bits of the 177 * offset of the entry into the directory page to keep the lists from 178 * getting too long. Once a new directory entry has been cleared to 179 * be written, it is moved to the pd_pendinghd list. After the new 180 * entry has been written to disk it is removed from the pd_pendinghd 181 * list, any removed operations are done, and the dependency structure 182 * is freed. 183 */ 184 #define DAHASHSZ 5 185 #define DIRADDHASH(offset) (((offset) >> 2) % DAHASHSZ) 186 struct pagedep { 187 struct worklist pd_list; /* page buffer */ 188 # define pd_state pd_list.wk_state /* check for multiple I/O starts */ 189 LIST_ENTRY(pagedep) pd_hash; /* hashed lookup */ 190 struct mount *pd_mnt; /* associated mount point */ 191 ino_t pd_ino; /* associated file */ 192 ufs_lbn_t pd_lbn; /* block within file */ 193 struct dirremhd pd_dirremhd; /* dirrem's waiting for page */ 194 struct diraddhd pd_diraddhd[DAHASHSZ]; /* diradd dir entry updates */ 195 struct diraddhd pd_pendinghd; /* directory entries awaiting write */ 196 }; 197 198 /* 199 * The "inodedep" structure tracks the set of dependencies associated 200 * with an inode. One task that it must manage is delayed operations 201 * (i.e., work requests that must be held until the inodedep's associated 202 * inode has been written to disk). Getting an inode from its incore 203 * state to the disk requires two steps to be taken by the filesystem 204 * in this order: first the inode must be copied to its disk buffer by 205 * the VOP_UPDATE operation; second the inode's buffer must be written 206 * to disk. To ensure that both operations have happened in the required 207 * order, the inodedep maintains two lists. Delayed operations are 208 * placed on the id_inowait list. When the VOP_UPDATE is done, all 209 * operations on the id_inowait list are moved to the id_bufwait list. 210 * When the buffer is written, the items on the id_bufwait list can be 211 * safely moved to the work queue to be processed. A second task of the 212 * inodedep structure is to track the status of block allocation within 213 * the inode. Each block that is allocated is represented by an 214 * "allocdirect" structure (see below). It is linked onto the id_newinoupdt 215 * list until both its contents and its allocation in the cylinder 216 * group map have been written to disk. Once these dependencies have been 217 * satisfied, it is removed from the id_newinoupdt list and any followup 218 * actions such as releasing the previous block or fragment are placed 219 * on the id_inowait list. When an inode is updated (a VOP_UPDATE is 220 * done), the "inodedep" structure is linked onto the buffer through 221 * its worklist. Thus, it will be notified when the buffer is about 222 * to be written and when it is done. At the update time, all the 223 * elements on the id_newinoupdt list are moved to the id_inoupdt list 224 * since those changes are now relevant to the copy of the inode in the 225 * buffer. Also at update time, the tasks on the id_inowait list are 226 * moved to the id_bufwait list so that they will be executed when 227 * the updated inode has been written to disk. When the buffer containing 228 * the inode is written to disk, any updates listed on the id_inoupdt 229 * list are rolled back as they are not yet safe. Following the write, 230 * the changes are once again rolled forward and any actions on the 231 * id_bufwait list are processed (since those actions are now safe). 232 * The entries on the id_inoupdt and id_newinoupdt lists must be kept 233 * sorted by logical block number to speed the calculation of the size 234 * of the rolled back inode (see explanation in initiate_write_inodeblock). 235 * When a directory entry is created, it is represented by a diradd. 236 * The diradd is added to the id_inowait list as it cannot be safely 237 * written to disk until the inode that it represents is on disk. After 238 * the inode is written, the id_bufwait list is processed and the diradd 239 * entries are moved to the id_pendinghd list where they remain until 240 * the directory block containing the name has been written to disk. 241 * The purpose of keeping the entries on the id_pendinghd list is so that 242 * the softdep_fsync function can find and push the inode's directory 243 * name(s) as part of the fsync operation for that file. 244 */ 245 struct inodedep { 246 struct worklist id_list; /* buffer holding inode block */ 247 # define id_state id_list.wk_state /* inode dependency state */ 248 LIST_ENTRY(inodedep) id_hash; /* hashed lookup */ 249 struct fs *id_fs; /* associated filesystem */ 250 ino_t id_ino; /* dependent inode */ 251 nlink_t id_nlinkdelta; /* saved effective link count */ 252 LIST_ENTRY(inodedep) id_deps; /* bmsafemap's list of inodedep's */ 253 struct buf *id_buf; /* related bmsafemap (if pending) */ 254 off_t id_savedsize; /* file size saved during rollback */ 255 struct workhead id_pendinghd; /* entries awaiting directory write */ 256 struct workhead id_bufwait; /* operations after inode written */ 257 struct workhead id_inowait; /* operations waiting inode update */ 258 struct allocdirectlst id_inoupdt; /* updates before inode written */ 259 struct allocdirectlst id_newinoupdt; /* updates when inode written */ 260 union { 261 struct ufs1_dinode *idu_savedino1; /* saved ufs1_dinode contents */ 262 struct ufs2_dinode *idu_savedino2; /* saved ufs2_dinode contents */ 263 } id_un; 264 }; 265 #define id_savedino1 id_un.idu_savedino1 266 #define id_savedino2 id_un.idu_savedino2 267 268 /* 269 * A "newblk" structure is attached to a bmsafemap structure when a block 270 * or fragment is allocated from a cylinder group. Its state is set to 271 * DEPCOMPLETE when its cylinder group map is written. It is consumed by 272 * an associated allocdirect or allocindir allocation which will attach 273 * themselves to the bmsafemap structure if the newblk's DEPCOMPLETE flag 274 * is not set (i.e., its cylinder group map has not been written). 275 */ 276 struct newblk { 277 LIST_ENTRY(newblk) nb_hash; /* hashed lookup */ 278 struct fs *nb_fs; /* associated filesystem */ 279 int nb_state; /* state of bitmap dependency */ 280 ufs2_daddr_t nb_newblkno; /* allocated block number */ 281 LIST_ENTRY(newblk) nb_deps; /* bmsafemap's list of newblk's */ 282 struct bmsafemap *nb_bmsafemap; /* associated bmsafemap */ 283 }; 284 285 /* 286 * A "bmsafemap" structure maintains a list of dependency structures 287 * that depend on the update of a particular cylinder group map. 288 * It has lists for newblks, allocdirects, allocindirs, and inodedeps. 289 * It is attached to the buffer of a cylinder group block when any of 290 * these things are allocated from the cylinder group. It is freed 291 * after the cylinder group map is written and the state of its 292 * dependencies are updated with DEPCOMPLETE to indicate that it has 293 * been processed. 294 */ 295 struct bmsafemap { 296 struct worklist sm_list; /* cylgrp buffer */ 297 struct buf *sm_buf; /* associated buffer */ 298 struct allocdirecthd sm_allocdirecthd; /* allocdirect deps */ 299 struct allocindirhd sm_allocindirhd; /* allocindir deps */ 300 struct inodedephd sm_inodedephd; /* inodedep deps */ 301 struct newblkhd sm_newblkhd; /* newblk deps */ 302 }; 303 304 /* 305 * An "allocdirect" structure is attached to an "inodedep" when a new block 306 * or fragment is allocated and pointed to by the inode described by 307 * "inodedep". The worklist is linked to the buffer that holds the block. 308 * When the block is first allocated, it is linked to the bmsafemap 309 * structure associated with the buffer holding the cylinder group map 310 * from which it was allocated. When the cylinder group map is written 311 * to disk, ad_state has the DEPCOMPLETE flag set. When the block itself 312 * is written, the COMPLETE flag is set. Once both the cylinder group map 313 * and the data itself have been written, it is safe to write the inode 314 * that claims the block. If there was a previous fragment that had been 315 * allocated before the file was increased in size, the old fragment may 316 * be freed once the inode claiming the new block is written to disk. 317 * This ad_fragfree request is attached to the id_inowait list of the 318 * associated inodedep (pointed to by ad_inodedep) for processing after 319 * the inode is written. When a block is allocated to a directory, an 320 * fsync of a file whose name is within that block must ensure not only 321 * that the block containing the file name has been written, but also 322 * that the on-disk inode references that block. When a new directory 323 * block is created, we allocate a newdirblk structure which is linked 324 * to the associated allocdirect (on its ad_newdirblk list). When the 325 * allocdirect has been satisfied, the newdirblk structure is moved to 326 * the inodedep id_bufwait list of its directory to await the inode 327 * being written. When the inode is written, the directory entries are 328 * fully committed and can be deleted from their pagedep->id_pendinghd 329 * and inodedep->id_pendinghd lists. 330 */ 331 struct allocdirect { 332 struct worklist ad_list; /* buffer holding block */ 333 # define ad_state ad_list.wk_state /* block pointer state */ 334 TAILQ_ENTRY(allocdirect) ad_next; /* inodedep's list of allocdirect's */ 335 ufs_lbn_t ad_lbn; /* block within file */ 336 ufs2_daddr_t ad_newblkno; /* new value of block pointer */ 337 ufs2_daddr_t ad_oldblkno; /* old value of block pointer */ 338 long ad_newsize; /* size of new block */ 339 long ad_oldsize; /* size of old block */ 340 LIST_ENTRY(allocdirect) ad_deps; /* bmsafemap's list of allocdirect's */ 341 struct buf *ad_buf; /* cylgrp buffer (if pending) */ 342 struct inodedep *ad_inodedep; /* associated inodedep */ 343 struct freefrag *ad_freefrag; /* fragment to be freed (if any) */ 344 struct workhead ad_newdirblk; /* dir block to notify when written */ 345 }; 346 347 /* 348 * A single "indirdep" structure manages all allocation dependencies for 349 * pointers in an indirect block. The up-to-date state of the indirect 350 * block is stored in ir_savedata. The set of pointers that may be safely 351 * written to the disk is stored in ir_safecopy. The state field is used 352 * only to track whether the buffer is currently being written (in which 353 * case it is not safe to update ir_safecopy). Ir_deplisthd contains the 354 * list of allocindir structures, one for each block that needs to be 355 * written to disk. Once the block and its bitmap allocation have been 356 * written the safecopy can be updated to reflect the allocation and the 357 * allocindir structure freed. If ir_state indicates that an I/O on the 358 * indirect block is in progress when ir_safecopy is to be updated, the 359 * update is deferred by placing the allocindir on the ir_donehd list. 360 * When the I/O on the indirect block completes, the entries on the 361 * ir_donehd list are processed by updating their corresponding ir_safecopy 362 * pointers and then freeing the allocindir structure. 363 */ 364 struct indirdep { 365 struct worklist ir_list; /* buffer holding indirect block */ 366 # define ir_state ir_list.wk_state /* indirect block pointer state */ 367 caddr_t ir_saveddata; /* buffer cache contents */ 368 struct buf *ir_savebp; /* buffer holding safe copy */ 369 struct allocindirhd ir_donehd; /* done waiting to update safecopy */ 370 struct allocindirhd ir_deplisthd; /* allocindir deps for this block */ 371 }; 372 373 /* 374 * An "allocindir" structure is attached to an "indirdep" when a new block 375 * is allocated and pointed to by the indirect block described by the 376 * "indirdep". The worklist is linked to the buffer that holds the new block. 377 * When the block is first allocated, it is linked to the bmsafemap 378 * structure associated with the buffer holding the cylinder group map 379 * from which it was allocated. When the cylinder group map is written 380 * to disk, ai_state has the DEPCOMPLETE flag set. When the block itself 381 * is written, the COMPLETE flag is set. Once both the cylinder group map 382 * and the data itself have been written, it is safe to write the entry in 383 * the indirect block that claims the block; the "allocindir" dependency 384 * can then be freed as it is no longer applicable. 385 */ 386 struct allocindir { 387 struct worklist ai_list; /* buffer holding indirect block */ 388 # define ai_state ai_list.wk_state /* indirect block pointer state */ 389 LIST_ENTRY(allocindir) ai_next; /* indirdep's list of allocindir's */ 390 int ai_offset; /* pointer offset in indirect block */ 391 ufs2_daddr_t ai_newblkno; /* new block pointer value */ 392 ufs2_daddr_t ai_oldblkno; /* old block pointer value */ 393 struct freefrag *ai_freefrag; /* block to be freed when complete */ 394 struct indirdep *ai_indirdep; /* address of associated indirdep */ 395 LIST_ENTRY(allocindir) ai_deps; /* bmsafemap's list of allocindir's */ 396 struct buf *ai_buf; /* cylgrp buffer (if pending) */ 397 }; 398 399 /* 400 * A "freefrag" structure is attached to an "inodedep" when a previously 401 * allocated fragment is replaced with a larger fragment, rather than extended. 402 * The "freefrag" structure is constructed and attached when the replacement 403 * block is first allocated. It is processed after the inode claiming the 404 * bigger block that replaces it has been written to disk. Note that the 405 * ff_state field is is used to store the uid, so may lose data. However, 406 * the uid is used only in printing an error message, so is not critical. 407 * Keeping it in a short keeps the data structure down to 32 bytes. 408 */ 409 struct freefrag { 410 struct worklist ff_list; /* id_inowait or delayed worklist */ 411 # define ff_state ff_list.wk_state /* owning user; should be uid_t */ 412 struct mount *ff_mnt; /* associated mount point */ 413 ufs2_daddr_t ff_blkno; /* fragment physical block number */ 414 long ff_fragsize; /* size of fragment being deleted */ 415 ino_t ff_inum; /* owning inode number */ 416 }; 417 418 /* 419 * A "freeblks" structure is attached to an "inodedep" when the 420 * corresponding file's length is reduced to zero. It records all 421 * the information needed to free the blocks of a file after its 422 * zero'ed inode has been written to disk. 423 */ 424 struct freeblks { 425 struct worklist fb_list; /* id_inowait or delayed worklist */ 426 ino_t fb_previousinum; /* inode of previous owner of blocks */ 427 uid_t fb_uid; /* uid of previous owner of blocks */ 428 struct vnode *fb_devvp; /* filesystem device vnode */ 429 struct mount *fb_mnt; /* associated mount point */ 430 off_t fb_oldsize; /* previous file size */ 431 off_t fb_newsize; /* new file size */ 432 ufs2_daddr_t fb_chkcnt; /* used to check cnt of blks released */ 433 ufs2_daddr_t fb_dblks[NDADDR]; /* direct blk ptrs to deallocate */ 434 ufs2_daddr_t fb_iblks[NIADDR]; /* indirect blk ptrs to deallocate */ 435 }; 436 437 /* 438 * A "freefile" structure is attached to an inode when its 439 * link count is reduced to zero. It marks the inode as free in 440 * the cylinder group map after the zero'ed inode has been written 441 * to disk and any associated blocks and fragments have been freed. 442 */ 443 struct freefile { 444 struct worklist fx_list; /* id_inowait or delayed worklist */ 445 mode_t fx_mode; /* mode of inode */ 446 ino_t fx_oldinum; /* inum of the unlinked file */ 447 struct vnode *fx_devvp; /* filesystem device vnode */ 448 struct mount *fx_mnt; /* associated mount point */ 449 }; 450 451 /* 452 * A "diradd" structure is linked to an "inodedep" id_inowait list when a 453 * new directory entry is allocated that references the inode described 454 * by "inodedep". When the inode itself is written (either the initial 455 * allocation for new inodes or with the increased link count for 456 * existing inodes), the COMPLETE flag is set in da_state. If the entry 457 * is for a newly allocated inode, the "inodedep" structure is associated 458 * with a bmsafemap which prevents the inode from being written to disk 459 * until the cylinder group has been updated. Thus the da_state COMPLETE 460 * flag cannot be set until the inode bitmap dependency has been removed. 461 * When creating a new file, it is safe to write the directory entry that 462 * claims the inode once the referenced inode has been written. Since 463 * writing the inode clears the bitmap dependencies, the DEPCOMPLETE flag 464 * in the diradd can be set unconditionally when creating a file. When 465 * creating a directory, there are two additional dependencies described by 466 * mkdir structures (see their description below). When these dependencies 467 * are resolved the DEPCOMPLETE flag is set in the diradd structure. 468 * If there are multiple links created to the same inode, there will be 469 * a separate diradd structure created for each link. The diradd is 470 * linked onto the pg_diraddhd list of the pagedep for the directory 471 * page that contains the entry. When a directory page is written, 472 * the pg_diraddhd list is traversed to rollback any entries that are 473 * not yet ready to be written to disk. If a directory entry is being 474 * changed (by rename) rather than added, the DIRCHG flag is set and 475 * the da_previous entry points to the entry that will be "removed" 476 * once the new entry has been committed. During rollback, entries 477 * with da_previous are replaced with the previous inode number rather 478 * than zero. 479 * 480 * The overlaying of da_pagedep and da_previous is done to keep the 481 * structure down to 32 bytes in size on a 32-bit machine. If a 482 * da_previous entry is present, the pointer to its pagedep is available 483 * in the associated dirrem entry. If the DIRCHG flag is set, the 484 * da_previous entry is valid; if not set the da_pagedep entry is valid. 485 * The DIRCHG flag never changes; it is set when the structure is created 486 * if appropriate and is never cleared. 487 */ 488 struct diradd { 489 struct worklist da_list; /* id_inowait or id_pendinghd list */ 490 # define da_state da_list.wk_state /* state of the new directory entry */ 491 LIST_ENTRY(diradd) da_pdlist; /* pagedep holding directory block */ 492 doff_t da_offset; /* offset of new dir entry in dir blk */ 493 ino_t da_newinum; /* inode number for the new dir entry */ 494 union { 495 struct dirrem *dau_previous; /* entry being replaced in dir change */ 496 struct pagedep *dau_pagedep; /* pagedep dependency for addition */ 497 } da_un; 498 }; 499 #define da_previous da_un.dau_previous 500 #define da_pagedep da_un.dau_pagedep 501 502 /* 503 * Two "mkdir" structures are needed to track the additional dependencies 504 * associated with creating a new directory entry. Normally a directory 505 * addition can be committed as soon as the newly referenced inode has been 506 * written to disk with its increased link count. When a directory is 507 * created there are two additional dependencies: writing the directory 508 * data block containing the "." and ".." entries (MKDIR_BODY) and writing 509 * the parent inode with the increased link count for ".." (MKDIR_PARENT). 510 * These additional dependencies are tracked by two mkdir structures that 511 * reference the associated "diradd" structure. When they have completed, 512 * they set the DEPCOMPLETE flag on the diradd so that it knows that its 513 * extra dependencies have been completed. The md_state field is used only 514 * to identify which type of dependency the mkdir structure is tracking. 515 * It is not used in the mainline code for any purpose other than consistency 516 * checking. All the mkdir structures in the system are linked together on 517 * a list. This list is needed so that a diradd can find its associated 518 * mkdir structures and deallocate them if it is prematurely freed (as for 519 * example if a mkdir is immediately followed by a rmdir of the same directory). 520 * Here, the free of the diradd must traverse the list to find the associated 521 * mkdir structures that reference it. The deletion would be faster if the 522 * diradd structure were simply augmented to have two pointers that referenced 523 * the associated mkdir's. However, this would increase the size of the diradd 524 * structure from 32 to 64-bits to speed a very infrequent operation. 525 */ 526 struct mkdir { 527 struct worklist md_list; /* id_inowait or buffer holding dir */ 528 # define md_state md_list.wk_state /* type: MKDIR_PARENT or MKDIR_BODY */ 529 struct diradd *md_diradd; /* associated diradd */ 530 struct buf *md_buf; /* MKDIR_BODY: buffer holding dir */ 531 LIST_ENTRY(mkdir) md_mkdirs; /* list of all mkdirs */ 532 }; 533 LIST_HEAD(mkdirlist, mkdir) mkdirlisthd; 534 535 /* 536 * A "dirrem" structure describes an operation to decrement the link 537 * count on an inode. The dirrem structure is attached to the pg_dirremhd 538 * list of the pagedep for the directory page that contains the entry. 539 * It is processed after the directory page with the deleted entry has 540 * been written to disk. 541 * 542 * The overlaying of dm_pagedep and dm_dirinum is done to keep the 543 * structure down to 32 bytes in size on a 32-bit machine. It works 544 * because they are never used concurrently. 545 */ 546 struct dirrem { 547 struct worklist dm_list; /* delayed worklist */ 548 # define dm_state dm_list.wk_state /* state of the old directory entry */ 549 LIST_ENTRY(dirrem) dm_next; /* pagedep's list of dirrem's */ 550 struct mount *dm_mnt; /* associated mount point */ 551 ino_t dm_oldinum; /* inum of the removed dir entry */ 552 union { 553 struct pagedep *dmu_pagedep; /* pagedep dependency for remove */ 554 ino_t dmu_dirinum; /* parent inode number (for rmdir) */ 555 } dm_un; 556 }; 557 #define dm_pagedep dm_un.dmu_pagedep 558 #define dm_dirinum dm_un.dmu_dirinum 559 560 /* 561 * A "newdirblk" structure tracks the progress of a newly allocated 562 * directory block from its creation until it is claimed by its on-disk 563 * inode. When a block is allocated to a directory, an fsync of a file 564 * whose name is within that block must ensure not only that the block 565 * containing the file name has been written, but also that the on-disk 566 * inode references that block. When a new directory block is created, 567 * we allocate a newdirblk structure which is linked to the associated 568 * allocdirect (on its ad_newdirblk list). When the allocdirect has been 569 * satisfied, the newdirblk structure is moved to the inodedep id_bufwait 570 * list of its directory to await the inode being written. When the inode 571 * is written, the directory entries are fully committed and can be 572 * deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd 573 * lists. Note that we could track directory blocks allocated to indirect 574 * blocks using a similar scheme with the allocindir structures. Rather 575 * than adding this level of complexity, we simply write those newly 576 * allocated indirect blocks synchronously as such allocations are rare. 577 */ 578 struct newdirblk { 579 struct worklist db_list; /* id_inowait or pg_newdirblk */ 580 # define db_state db_list.wk_state /* unused */ 581 struct pagedep *db_pagedep; /* associated pagedep */ 582 }; 583