sys/vfs/hammer2/hammer2.h

   1 /*
   2  * Copyright (c) 2011-2014 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@dragonflybsd.org>
   6  * by Venkatesh Srinivas <vsrinivas@dragonflybsd.org>
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  *
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in
  16  *    the documentation and/or other materials provided with the
  17  *    distribution.
  18  * 3. Neither the name of The DragonFly Project nor the names of its
  19  *    contributors may be used to endorse or promote products derived
  20  *    from this software without specific, prior written permission.
  21  *
  22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  */
  35
  36 /*
  37  * This header file contains structures used internally by the HAMMER2
  38  * implementation.  See hammer2_disk.h for on-disk structures.
  39  */
  40
  41 #ifndef _VFS_HAMMER2_HAMMER2_H_
  42 #define _VFS_HAMMER2_HAMMER2_H_
  43
  44 #include <sys/param.h>
  45 #include <sys/types.h>
  46 #include <sys/kernel.h>
  47 #include <sys/conf.h>
  48 #include <sys/systm.h>
  49 #include <sys/tree.h>
  50 #include <sys/malloc.h>
  51 #include <sys/mount.h>
  52 #include <sys/vnode.h>
  53 #include <sys/proc.h>
  54 #include <sys/mountctl.h>
  55 #include <sys/priv.h>
  56 #include <sys/stat.h>
  57 #include <sys/thread.h>
  58 #include <sys/globaldata.h>
  59 #include <sys/lockf.h>
  60 #include <sys/buf.h>
  61 #include <sys/queue.h>
  62 #include <sys/limits.h>
  63 #include <sys/buf2.h>
  64 #include <sys/signal2.h>
  65 #include <sys/dmsg.h>
  66 #include <sys/mutex.h>
  67 #include <sys/mutex2.h>
  68
  69 #include "hammer2_disk.h"
  70 #include "hammer2_mount.h"
  71 #include "hammer2_ioctl.h"
  72 #include "hammer2_ccms.h"
  73
  74 struct hammer2_chain;
  75 struct hammer2_inode;
  76 struct hammer2_mount;
  77 struct hammer2_pfsmount;
  78 struct hammer2_span;
  79 struct hammer2_state;
  80 struct hammer2_msg;
  81
  82 /*
  83  * The chain structure tracks a portion of the media topology from the
  84  * root (volume) down.  Chains represent volumes, inodes, indirect blocks,
  85  * data blocks, and freemap nodes and leafs.
  86  *
  87  * The chain structure can be multi-homed and its topological recursion
  88  * (chain->core) can be shared amongst several chains.  Chain structures
  89  * are topologically stable once placed in the in-memory topology (they
  90  * don't move around).  Modifications which cross flush synchronization
  91  * boundaries, renames, resizing, or any move of the chain to elsewhere
  92  * in the topology is accomplished via the DELETE-DUPLICATE mechanism.
  93  *
  94  * Deletions and delete-duplicates:
  95  *
  96  *      Any movement of chains within the topology utilize a delete-duplicate
  97  *      operation instead of a simple rename.  That is, the chain must be
  98  *      deleted from its original location and then duplicated to the new
  99  *      location.  A new chain structure is allocated while the old is
 100  *      deleted.  Deleted chains are removed from the above chain_core's
 101  *      rbtree but remain linked via the shadow topology for flush
 102  *      synchronization purposes.
 103  *
 104  *      delete_bmap is allocated and a bit set if the chain was originally
 105  *      loaded via the blockmap.
 106  *
 107  * Flush synchronization:
 108  *
 109  *      Flushes must synchronize chains up through the root.  To do this
 110  *      the in-memory topology would normally have to be frozen during the
 111  *      flush.  To avoid freezing the topology and to allow concurrent
 112  *      foreground / flush activity, any new modifications made while a
 113  *      flush is in progress retains the original chain in a shadow topology
 114  *      that is only visible to the flush code.  Only one flush can be
 115  *      running at a time so the shadow hierarchy can be implemented with
 116  *      just a few link fields in our in-memory data structures.
 117  *
 118  * Advantages:
 119  *
 120  *      (1) Fully coherent snapshots can be taken without requiring
 121  *          a pre-flush, resulting in extremely fast (sub-millisecond)
 122  *          snapshots.
 123  *
 124  *      (2) Multiple synchronization points can be in-flight at the same
 125  *          time, representing multiple snapshots or flushes.
 126  *
 127  *      (3) The algorithms needed to keep track of everything are actually
 128  *          not that complex.
 129  *
 130  * Special Considerations:
 131  *
 132  *      A chain is ref-counted on a per-chain basis, but the chain's lock
 133  *      is associated with the shared chain_core and is not per-chain.
 134  *
 135  *      The power-of-2 nature of the media radix tree ensures that there
 136  *      will be no overlaps which straddle edges.
 137  */
 138 RB_HEAD(hammer2_chain_tree, hammer2_chain);
 139 TAILQ_HEAD(h2_flush_deferral_list, hammer2_chain);
 140 TAILQ_HEAD(h2_core_list, hammer2_chain);
 141
 142 #define CHAIN_CORE_DELETE_BMAP_ENTRIES  \
 143         (HAMMER2_PBUFSIZE / sizeof(hammer2_blockref_t) / sizeof(uint32_t))
 144
 145 struct hammer2_chain_core {
 146         int             good;
 147         struct ccms_cst cst;
 148         struct h2_core_list ownerq;       /* all chains sharing this core */
 149         struct hammer2_chain_tree rbtree; /* live chains */
 150         struct hammer2_chain_tree dbtree; /* bmapped deletions */
 151         struct h2_core_list dbq;          /* other deletions */
 152         int             live_zero;      /* blockref array opt */
 153         u_int           sharecnt;
 154         u_int           flags;
 155         u_int           live_count;     /* live (not deleted) chains in tree */
 156         u_int           chain_count;    /* live + deleted chains under core */
 157         int             generation;     /* generation number (inserts only) */
 158 };
 159
 160 typedef struct hammer2_chain_core hammer2_chain_core_t;
 161
 162 #define HAMMER2_CORE_UNUSED0001         0x0001
 163 #define HAMMER2_CORE_COUNTEDBREFS       0x0002
 164
 165 /*
 166  * H2 is a copy-on-write filesystem.  In order to allow chains to allocate
 167  * smaller blocks (down to 64-bytes), but improve performance and make
 168  * clustered I/O possible using larger block sizes, the kernel buffer cache
 169  * is abstracted via the hammer2_io structure.
 170  */
 171 RB_HEAD(hammer2_io_tree, hammer2_io);
 172
 173 struct hammer2_io {
 174         RB_ENTRY(hammer2_io) rbnode;    /* indexed by device offset */
 175         struct spinlock spin;
 176         struct hammer2_mount *hmp;
 177         struct buf      *bp;
 178         struct bio      *bio;
 179         off_t           pbase;
 180         int             psize;
 181         void            (*callback)(struct hammer2_io *dio,
 182                                     struct hammer2_chain *chain,
 183                                     void *arg1, off_t arg2);
 184         struct hammer2_chain *arg_c;            /* INPROG I/O only */
 185         void            *arg_p;                 /* INPROG I/O only */
 186         off_t           arg_o;                  /* INPROG I/O only */
 187         int             refs;
 188         int             act;                    /* activity */
 189 };
 190
 191 typedef struct hammer2_io hammer2_io_t;
 192
 193 /*
 194  * Primary chain structure keeps track of the topology in-memory.
 195  */
 196 struct hammer2_chain {
 197         TAILQ_ENTRY(hammer2_chain) core_entry;  /* contemporary chains */
 198         RB_ENTRY(hammer2_chain) rbnode;         /* live chain(s) */
 199         TAILQ_ENTRY(hammer2_chain) db_entry;    /* non bmapped deletions */
 200         hammer2_blockref_t      bref;
 201         hammer2_chain_core_t    *core;
 202         hammer2_chain_core_t    *above;
 203         struct hammer2_state    *state;         /* if active cache msg */
 204         struct hammer2_mount    *hmp;
 205         struct hammer2_pfsmount *pmp;           /* can be NULL */
 206
 207         hammer2_blockref_t      dsrc;                   /* DEBUG */
 208         int                     ninserts;               /* DEBUG */
 209         int                     nremoves;               /* DEBUG */
 210         hammer2_tid_t           dsrc_dupfromat;         /* DEBUG */
 211         uint32_t                dsrc_dupfromflags;      /* DEBUG */
 212         int                     dsrc_reason;            /* DEBUG */
 213         int                     dsrc_ninserts;          /* DEBUG */
 214         uint32_t                dsrc_flags;             /* DEBUG */
 215         hammer2_tid_t           dsrc_modify;            /* DEBUG */
 216         hammer2_tid_t           dsrc_delete;            /* DEBUG */
 217         hammer2_tid_t           dsrc_update_lo;         /* DEBUG */
 218         struct hammer2_chain    *dsrc_original;         /* DEBUG */
 219
 220         hammer2_tid_t   modify_tid;             /* flush filter */
 221         hammer2_tid_t   delete_tid;             /* flush filter */
 222         hammer2_tid_t   update_lo;              /* flush propagation */
 223         hammer2_tid_t   update_hi;              /* setsubmod propagation */
 224         hammer2_key_t   data_count;             /* delta's to apply */
 225         hammer2_key_t   inode_count;            /* delta's to apply */
 226         hammer2_io_t    *dio;                   /* physical data buffer */
 227         u_int           bytes;                  /* physical data size */
 228         u_int           flags;
 229         u_int           refs;
 230         u_int           lockcnt;
 231         hammer2_media_data_t *data;             /* data pointer shortcut */
 232         TAILQ_ENTRY(hammer2_chain) flush_node;  /* flush deferral list */
 233
 234         int             inode_reason;
 235 };
 236
 237 typedef struct hammer2_chain hammer2_chain_t;
 238
 239 int hammer2_chain_cmp(hammer2_chain_t *chain1, hammer2_chain_t *chain2);
 240 RB_PROTOTYPE(hammer2_chain_tree, hammer2_chain, rbnode, hammer2_chain_cmp);
 241
 242 /*
 243  * Special notes on flags:
 244  *
 245  * INITIAL - This flag allows a chain to be created and for storage to
 246  *           be allocated without having to immediately instantiate the
 247  *           related buffer.  The data is assumed to be all-zeros.  It
 248  *           is primarily used for indirect blocks.
 249  *
 250  * MODIFIED- The chain's media data has been modified.
 251  */
 252 #define HAMMER2_CHAIN_MODIFIED          0x00000001      /* dirty chain data */
 253 #define HAMMER2_CHAIN_ALLOCATED         0x00000002      /* kmalloc'd chain */
 254 #define HAMMER2_CHAIN_FLUSH_TEMPORARY   0x00000004
 255 #define HAMMER2_CHAIN_FORCECOW          0x00000008      /* force copy-on-wr */
 256 #define HAMMER2_CHAIN_DELETED           0x00000010      /* deleted chain */
 257 #define HAMMER2_CHAIN_INITIAL           0x00000020      /* initial create */
 258 #define HAMMER2_CHAIN_FLUSH_CREATE      0x00000040      /* needs flush blkadd */
 259 #define HAMMER2_CHAIN_FLUSH_DELETE      0x00000080      /* needs flush blkdel */
 260 #define HAMMER2_CHAIN_IOFLUSH           0x00000100      /* bawrite on put */
 261 #define HAMMER2_CHAIN_DEFERRED          0x00000200      /* on a deferral list */
 262 #define HAMMER2_CHAIN_UNLINKED          0x00000400      /* delete on reclaim */
 263 #define HAMMER2_CHAIN_VOLUMESYNC        0x00000800      /* needs volume sync */
 264 #define HAMMER2_CHAIN_ONDBQ             0x00001000      /* !bmapped deletes */
 265 #define HAMMER2_CHAIN_MOUNTED           0x00002000      /* PFS is mounted */
 266 #define HAMMER2_CHAIN_ONRBTREE          0x00004000      /* on parent RB tree */
 267 #define HAMMER2_CHAIN_SNAPSHOT          0x00008000      /* snapshot special */
 268 #define HAMMER2_CHAIN_EMBEDDED          0x00010000      /* embedded data */
 269 #define HAMMER2_CHAIN_RELEASE           0x00020000      /* don't keep around */
 270 #define HAMMER2_CHAIN_BMAPPED           0x00040000      /* in parent blkmap */
 271 #define HAMMER2_CHAIN_ONDBTREE          0x00080000      /* bmapped deletes */
 272 #define HAMMER2_CHAIN_DUPLICATED        0x00100000      /* fwd delete-dup */
 273 #define HAMMER2_CHAIN_PFSROOT           0x00200000      /* in pfs->cluster */
 274
 275 /*
 276  * Flags passed to hammer2_chain_lookup() and hammer2_chain_next()
 277  *
 278  * NOTE: MATCHIND allows an indirect block / freemap node to be returned
 279  *       when the passed key range matches the radix.  Remember that key_end
 280  *       is inclusive (e.g. {0x000,0xFFF}, not {0x000,0x1000}).
 281  */
 282 #define HAMMER2_LOOKUP_NOLOCK           0x00000001      /* ref only */
 283 #define HAMMER2_LOOKUP_NODATA           0x00000002      /* data left NULL */
 284 #define HAMMER2_LOOKUP_SHARED           0x00000100
 285 #define HAMMER2_LOOKUP_MATCHIND         0x00000200      /* return all chains */
 286 #define HAMMER2_LOOKUP_UNUSED0400       0x00000400
 287 #define HAMMER2_LOOKUP_ALWAYS           0x00000800      /* resolve data */
 288
 289 /*
 290  * Flags passed to hammer2_chain_modify() and hammer2_chain_resize()
 291  *
 292  * NOTE: OPTDATA allows us to avoid instantiating buffers for INDIRECT
 293  *       blocks in the INITIAL-create state.
 294  */
 295 #define HAMMER2_MODIFY_OPTDATA          0x00000002      /* data can be NULL */
 296 #define HAMMER2_MODIFY_NO_MODIFY_TID    0x00000004
 297 #define HAMMER2_MODIFY_ASSERTNOCOPY     0x00000008      /* assert no del-dup */
 298 #define HAMMER2_MODIFY_NOREALLOC        0x00000010
 299 #define HAMMER2_MODIFY_INPLACE          0x00000020      /* don't del-dup */
 300
 301 /*
 302  * Flags passed to hammer2_chain_lock()
 303  */
 304 #define HAMMER2_RESOLVE_NEVER           1
 305 #define HAMMER2_RESOLVE_MAYBE           2
 306 #define HAMMER2_RESOLVE_ALWAYS          3
 307 #define HAMMER2_RESOLVE_MASK            0x0F
 308
 309 #define HAMMER2_RESOLVE_SHARED          0x10    /* request shared lock */
 310 #define HAMMER2_RESOLVE_NOREF           0x20    /* already ref'd on lock */
 311
 312 /*
 313  * Flags passed to hammer2_chain_delete()
 314  */
 315 #define HAMMER2_DELETE_UNUSED0001       0x0001
 316
 317 /*
 318  * Flags passed to hammer2_chain_delete_duplicate()
 319  */
 320 #define HAMMER2_DELDUP_RECORE           0x0001
 321
 322 /*
 323  * Cluster different types of storage together for allocations
 324  */
 325 #define HAMMER2_FREECACHE_INODE         0
 326 #define HAMMER2_FREECACHE_INDIR         1
 327 #define HAMMER2_FREECACHE_DATA          2
 328 #define HAMMER2_FREECACHE_UNUSED3       3
 329 #define HAMMER2_FREECACHE_TYPES         4
 330
 331 /*
 332  * hammer2_freemap_alloc() block preference
 333  */
 334 #define HAMMER2_OFF_NOPREF              ((hammer2_off_t)-1)
 335
 336 /*
 337  * BMAP read-ahead maximum parameters
 338  */
 339 #define HAMMER2_BMAP_COUNT              16      /* max bmap read-ahead */
 340 #define HAMMER2_BMAP_BYTES              (HAMMER2_PBUFSIZE * HAMMER2_BMAP_COUNT)
 341
 342 /*
 343  * Misc
 344  */
 345 #define HAMMER2_FLUSH_DEPTH_LIMIT       10      /* stack recursion limit */
 346
 347 /*
 348  * hammer2_freemap_adjust()
 349  */
 350 #define HAMMER2_FREEMAP_DORECOVER       1
 351 #define HAMMER2_FREEMAP_DOMAYFREE       2
 352 #define HAMMER2_FREEMAP_DOREALFREE      3
 353
 354 /*
 355  * HAMMER2 IN-MEMORY CACHE OF MEDIA STRUCTURES
 356  *
 357  * There is an in-memory representation of all on-media data structure.
 358  * Basically everything is represented by a hammer2_chain structure
 359  * in-memory and other higher-level structures map to chains.
 360  *
 361  * A great deal of data is accessed simply via its buffer cache buffer,
 362  * which is mapped for the duration of the chain's lock.  However, because
 363  * chains may represent blocks smaller than the 16KB minimum we impose
 364  * on buffer cache buffers, we cannot hold related buffer cache buffers
 365  * locked for smaller blocks.  In these situations we kmalloc() a copy
 366  * of the block.
 367  *
 368  * When modifications are made to a chain a new filesystem block must be
 369  * allocated.  Multiple modifications do not necessarily allocate new
 370  * blocks.  However, when a flush occurs a flush synchronization point
 371  * is created and any new modifications made after this point will allocate
 372  * a new block even if the chain is already in a modified state.
 373  *
 374  * The in-memory representation may remain cached (for example in order to
 375  * placemark clustering locks) even after the related data has been
 376  * detached.
 377  *
 378  *                              CORE SHARING
 379  *
 380  * In order to support concurrent flushes a flush synchronization point
 381  * is created represented by a transaction id.  Among other things,
 382  * operations may move filesystem objects from one part of the topology
 383  * to another (for example, if you rename a file or when indirect blocks
 384  * are created or destroyed, and a few other things).  When this occurs
 385  * across a flush synchronization point the flusher needs to be able to
 386  * recurse down BOTH the 'before' version of the topology and the 'after'
 387  * version.
 388  *
 389  * To facilitate this modifications to chains do what is called a
 390  * DELETE-DUPLICATE operation.  Chains are not actually moved in-memory.
 391  * Instead the chain we wish to move is deleted and a new chain is created
 392  * at the target location in the topology.  ANY SUBCHAINS PLACED UNDER THE
 393  * CHAIN BEING MOVED HAVE TO EXIST IN BOTH PLACES.  To make this work
 394  * all sub-chains are managed by the hammer2_chain_core structure.  This
 395  * structure can be multi-homed, meaning that it can have more than one
 396  * chain as its parent.  When a chain is delete-duplicated the chain's core
 397  * becomes shared under both the old and new chain.
 398  *
 399  *                              STALE CHAINS
 400  *
 401  * When a chain is delete-duplicated the old chain typically becomes stale.
 402  * This is detected via the HAMMER2_CHAIN_DUPLICATED flag in chain->flags.
 403  * To avoid executing live filesystem operations on stale chains, the inode
 404  * locking code will follow stale chains via core->ownerq until it finds
 405  * the live chain.  The lock prevents ripups by other threads.  Lookups
 406  * must properly order locking operations to prevent other threads from
 407  * racing the lookup operation and will also follow stale chains when
 408  * required.
 409  */
 410
 411 RB_HEAD(hammer2_inode_tree, hammer2_inode);
 412
 413 /*
 414  * A hammer2 inode.
 415  *
 416  * NOTE: The inode's attribute CST which is also used to lock the inode
 417  *       is embedded in the chain (chain.cst) and aliased w/ attr_cst.
 418  */
 419 struct hammer2_inode {
 420         RB_ENTRY(hammer2_inode) rbnode;         /* inumber lookup (HL) */
 421         ccms_cst_t              topo_cst;       /* directory topology cst */
 422         struct hammer2_pfsmount *pmp;           /* PFS mount */
 423         struct hammer2_inode    *pip;           /* parent inode */
 424         struct vnode            *vp;
 425         hammer2_chain_t         *chain;         /* NOTE: rehomed on rename */
 426         struct lockf            advlock;
 427         hammer2_tid_t           inum;
 428         u_int                   flags;
 429         u_int                   refs;           /* +vpref, +flushref */
 430         uint8_t                 comp_heuristic;
 431         hammer2_off_t           size;
 432         uint64_t                mtime;
 433 };
 434
 435 typedef struct hammer2_inode hammer2_inode_t;
 436
 437 #define HAMMER2_INODE_MODIFIED          0x0001
 438 #define HAMMER2_INODE_SROOT             0x0002  /* kmalloc special case */
 439 #define HAMMER2_INODE_RENAME_INPROG     0x0004
 440 #define HAMMER2_INODE_ONRBTREE          0x0008
 441 #define HAMMER2_INODE_RESIZED           0x0010
 442 #define HAMMER2_INODE_MTIME             0x0020
 443
 444 int hammer2_inode_cmp(hammer2_inode_t *ip1, hammer2_inode_t *ip2);
 445 RB_PROTOTYPE2(hammer2_inode_tree, hammer2_inode, rbnode, hammer2_inode_cmp,
 446                 hammer2_tid_t);
 447
 448 /*
 449  * A hammer2 transaction and flush sequencing structure.
 450  *
 451  * This global structure is tied into hammer2_mount and is used
 452  * to sequence modifying operations and flushes.
 453  *
 454  * (a) Any modifying operations with sync_tid >= flush_tid will stall until
 455  *     all modifying operating with sync_tid < flush_tid complete.
 456  *
 457  *     The flush related to flush_tid stalls until all modifying operations
 458  *     with sync_tid < flush_tid complete.
 459  *
 460  * (b) Once unstalled, modifying operations with sync_tid > flush_tid are
 461  *     allowed to run.  All modifications cause modify/duplicate operations
 462  *     to occur on the related chains.  Note that most INDIRECT blocks will
 463  *     be unaffected because the modifications just overload the RBTREE
 464  *     structurally instead of actually modifying the indirect blocks.
 465  *
 466  * (c) The actual flush unstalls and RUNS CONCURRENTLY with (b), but only
 467  *     utilizes the chain structures with sync_tid <= flush_tid.  The
 468  *     flush will modify related indirect blocks and inodes in-place
 469  *     (rather than duplicate) since the adjustments are compatible with
 470  *     (b)'s RBTREE overloading
 471  *
 472  *     SPECIAL NOTE:  Inode modifications have to also propagate along any
 473  *                    modify/duplicate chains.  File writes detect the flush
 474  *                    and force out the conflicting buffer cache buffer(s)
 475  *                    before reusing them.
 476  *
 477  * (d) Snapshots can be made instantly but must be flushed and disconnected
 478  *     from their duplicative source before they can be mounted.  This is
 479  *     because while H2's on-media structure supports forks, its in-memory
 480  *     structure only supports very simple forking for background flushing
 481  *     purposes.
 482  *
 483  * TODO: Flush merging.  When fsync() is called on multiple discrete files
 484  *       concurrently there is no reason to stall the second fsync.
 485  *       The final flush that reaches to root can cover both fsync()s.
 486  *
 487  *     The chains typically terminate as they fly onto the disk.  The flush
 488  *     ultimately reaches the volume header.
 489  */
 490 struct hammer2_trans {
 491         TAILQ_ENTRY(hammer2_trans) entry;
 492         struct hammer2_pfsmount *pmp;           /* might be NULL */
 493         struct hammer2_mount    *hmp_single;    /* if single-targetted */
 494         hammer2_tid_t           orig_tid;
 495         hammer2_tid_t           sync_tid;       /* effective transaction id */
 496         hammer2_tid_t           inode_tid;
 497         thread_t                td;             /* pointer */
 498         int                     flags;
 499         int                     blocked;
 500         uint8_t                 inodes_created;
 501         uint8_t                 dummy[7];
 502 };
 503
 504 typedef struct hammer2_trans hammer2_trans_t;
 505
 506 #define HAMMER2_TRANS_ISFLUSH           0x0001  /* formal flush */
 507 #define HAMMER2_TRANS_CONCURRENT        0x0002  /* concurrent w/flush */
 508 #define HAMMER2_TRANS_BUFCACHE          0x0004  /* from bioq strategy write */
 509 #define HAMMER2_TRANS_NEWINODE          0x0008  /* caller allocating inode */
 510 #define HAMMER2_TRANS_ISALLOCATING      0x0010  /* in allocator */
 511
 512 #define HAMMER2_FREEMAP_HEUR_NRADIX     4       /* pwr 2 PBUFRADIX-MINIORADIX */
 513 #define HAMMER2_FREEMAP_HEUR_TYPES      8
 514 #define HAMMER2_FREEMAP_HEUR            (HAMMER2_FREEMAP_HEUR_NRADIX * \
 515                                          HAMMER2_FREEMAP_HEUR_TYPES)
 516
 517 /*
 518  * Global (per device) mount structure for device (aka vp->v_mount->hmp)
 519  */
 520 TAILQ_HEAD(hammer2_trans_queue, hammer2_trans);
 521
 522 struct hammer2_mount {
 523         struct vnode    *devvp;         /* device vnode */
 524         int             ronly;          /* read-only mount */
 525         int             pmp_count;      /* PFS mounts backed by us */
 526         TAILQ_ENTRY(hammer2_mount) mntentry; /* hammer2_mntlist */
 527
 528         struct malloc_type *mchain;
 529         int             nipstacks;
 530         int             maxipstacks;
 531         struct spinlock io_spin;        /* iotree access */
 532         struct hammer2_io_tree iotree;
 533         int             iofree_count;
 534         hammer2_chain_t vchain;         /* anchor chain (topology) */
 535         hammer2_chain_t fchain;         /* anchor chain (freemap) */
 536         hammer2_inode_t *sroot;         /* super-root localized to media */
 537         struct lock     alloclk;        /* lockmgr lock */
 538         struct lock     voldatalk;      /* lockmgr lock */
 539         struct hammer2_trans_queue transq; /* all in-progress transactions */
 540         hammer2_off_t   heur_freemap[HAMMER2_FREEMAP_HEUR];
 541         int             flushcnt;       /* #of flush trans on the list */
 542
 543         int             volhdrno;       /* last volhdrno written */
 544         hammer2_volume_data_t voldata;
 545         hammer2_volume_data_t volsync;  /* synchronized voldata */
 546 };
 547
 548 typedef struct hammer2_mount hammer2_mount_t;
 549
 550 /*
 551  * HAMMER2 cluster - a device/root associated with a PFS.
 552  *
 553  * A PFS may have several hammer2_cluster's associated with it.
 554  */
 555 #define HAMMER2_MAXCLUSTER      8
 556
 557 struct hammer2_cluster {
 558         int                     nchains;
 559         int                     status;
 560         hammer2_chain_t         *chains[HAMMER2_MAXCLUSTER];
 561 };
 562
 563 typedef struct hammer2_cluster hammer2_cluster_t;
 564
 565 /*
 566  * HAMMER2 PFS mount point structure (aka vp->v_mount->mnt_data).
 567  * This has a 1:1 correspondence to struct mount (note that the
 568  * hammer2_mount structure has a N:1 correspondence).
 569  *
 570  * This structure represents a cluster mount and not necessarily a
 571  * PFS under a specific device mount (HMP).  The distinction is important
 572  * because the elements backing a cluster mount can change on the fly.
 573  *
 574  * Usually the first element under the cluster represents the original
 575  * user-requested mount that bootstraps the whole mess.  In significant
 576  * setups the original is usually just a read-only media image (or
 577  * representitive file) that simply contains a bootstrap volume header
 578  * listing the configuration.
 579  */
 580 struct hammer2_pfsmount {
 581         struct mount            *mp;
 582         hammer2_cluster_t       cluster;
 583         hammer2_inode_t         *iroot;         /* PFS root inode */
 584         hammer2_inode_t         *ihidden;       /* PFS hidden directory */
 585         struct lock             lock;           /* PFS lock for certain ops */
 586         hammer2_off_t           inode_count;    /* copy of inode_count */
 587         ccms_domain_t           ccms_dom;
 588         struct netexport        export;         /* nfs export */
 589         int                     ronly;          /* read-only mount */
 590         struct malloc_type      *minode;
 591         struct malloc_type      *mmsg;
 592         kdmsg_iocom_t           iocom;
 593         struct spinlock         inum_spin;      /* inumber lookup */
 594         struct hammer2_inode_tree inum_tree;
 595         long                    inmem_inodes;
 596         long                    inmem_dirty_chains;
 597         int                     count_lwinprog; /* logical write in prog */
 598         thread_t                wthread_td;     /* write thread td */
 599         struct bio_queue_head   wthread_bioq;   /* logical buffer bioq */
 600         struct mtx              wthread_mtx;    /* interlock */
 601         int                     wthread_destroy;/* termination sequencing */
 602 };
 603
 604 typedef struct hammer2_pfsmount hammer2_pfsmount_t;
 605
 606 #define HAMMER2_DIRTYCHAIN_WAITING      0x80000000
 607 #define HAMMER2_DIRTYCHAIN_MASK         0x7FFFFFFF
 608
 609 #define HAMMER2_LWINPROG_WAITING        0x80000000
 610 #define HAMMER2_LWINPROG_MASK           0x7FFFFFFF
 611
 612 #if defined(_KERNEL)
 613
 614 MALLOC_DECLARE(M_HAMMER2);
 615
 616 #define VTOI(vp)        ((hammer2_inode_t *)(vp)->v_data)
 617 #define ITOV(ip)        ((ip)->vp)
 618
 619 /*
 620  * Currently locked chains retain the locked buffer cache buffer for
 621  * indirect blocks, and indirect blocks can be one of two sizes.  The
 622  * device buffer has to match the case to avoid deadlocking recursive
 623  * chains that might otherwise try to access different offsets within
 624  * the same device buffer.
 625  */
 626 static __inline
 627 int
 628 hammer2_devblkradix(int radix)
 629 {
 630         if (radix <= HAMMER2_LBUFRADIX) {
 631                 return (HAMMER2_LBUFRADIX);
 632         } else {
 633                 return (HAMMER2_PBUFRADIX);
 634         }
 635 }
 636
 637 static __inline
 638 size_t
 639 hammer2_devblksize(size_t bytes)
 640 {
 641         if (bytes <= HAMMER2_LBUFSIZE) {
 642                 return(HAMMER2_LBUFSIZE);
 643         } else {
 644                 KKASSERT(bytes <= HAMMER2_PBUFSIZE &&
 645                          (bytes ^ (bytes - 1)) == ((bytes << 1) - 1));
 646                 return (HAMMER2_PBUFSIZE);
 647         }
 648 }
 649
 650
 651 static __inline
 652 hammer2_pfsmount_t *
 653 MPTOPMP(struct mount *mp)
 654 {
 655         return ((hammer2_pfsmount_t *)mp->mnt_data);
 656 }
 657
 658 extern struct vop_ops hammer2_vnode_vops;
 659 extern struct vop_ops hammer2_spec_vops;
 660 extern struct vop_ops hammer2_fifo_vops;
 661
 662 extern int hammer2_debug;
 663 extern int hammer2_cluster_enable;
 664 extern int hammer2_hardlink_enable;
 665 extern int hammer2_flush_pipe;
 666 extern int hammer2_synchronous_flush;
 667 extern long hammer2_limit_dirty_chains;
 668 extern long hammer2_iod_file_read;
 669 extern long hammer2_iod_meta_read;
 670 extern long hammer2_iod_indr_read;
 671 extern long hammer2_iod_fmap_read;
 672 extern long hammer2_iod_volu_read;
 673 extern long hammer2_iod_file_write;
 674 extern long hammer2_iod_meta_write;
 675 extern long hammer2_iod_indr_write;
 676 extern long hammer2_iod_fmap_write;
 677 extern long hammer2_iod_volu_write;
 678 extern long hammer2_ioa_file_read;
 679 extern long hammer2_ioa_meta_read;
 680 extern long hammer2_ioa_indr_read;
 681 extern long hammer2_ioa_fmap_read;
 682 extern long hammer2_ioa_volu_read;
 683 extern long hammer2_ioa_file_write;
 684 extern long hammer2_ioa_meta_write;
 685 extern long hammer2_ioa_indr_write;
 686 extern long hammer2_ioa_fmap_write;
 687 extern long hammer2_ioa_volu_write;
 688
 689 extern struct objcache *cache_buffer_read;
 690 extern struct objcache *cache_buffer_write;
 691
 692 extern int destroy;
 693 extern int write_thread_wakeup;
 694
 695 extern mtx_t thread_protect;
 696
 697 /*
 698  * hammer2_subr.c
 699  */
 700 #define hammer2_icrc32(buf, size)       iscsi_crc32((buf), (size))
 701 #define hammer2_icrc32c(buf, size, crc) iscsi_crc32_ext((buf), (size), (crc))
 702
 703 hammer2_chain_t *hammer2_inode_lock_ex(hammer2_inode_t *ip);
 704 hammer2_chain_t *hammer2_inode_lock_sh(hammer2_inode_t *ip);
 705 void hammer2_inode_unlock_ex(hammer2_inode_t *ip, hammer2_chain_t *chain);
 706 void hammer2_inode_unlock_sh(hammer2_inode_t *ip, hammer2_chain_t *chain);
 707 void hammer2_chain_refactor(hammer2_chain_t **chainp);
 708 void hammer2_voldata_lock(hammer2_mount_t *hmp);
 709 void hammer2_voldata_unlock(hammer2_mount_t *hmp, int modify);
 710 ccms_state_t hammer2_inode_lock_temp_release(hammer2_inode_t *ip);
 711 void hammer2_inode_lock_temp_restore(hammer2_inode_t *ip, ccms_state_t ostate);
 712 ccms_state_t hammer2_inode_lock_upgrade(hammer2_inode_t *ip);
 713 void hammer2_inode_lock_downgrade(hammer2_inode_t *ip, ccms_state_t ostate);
 714
 715 void hammer2_mount_exlock(hammer2_mount_t *hmp);
 716 void hammer2_mount_shlock(hammer2_mount_t *hmp);
 717 void hammer2_mount_unlock(hammer2_mount_t *hmp);
 718
 719 int hammer2_get_dtype(hammer2_chain_t *chain);
 720 int hammer2_get_vtype(hammer2_chain_t *chain);
 721 u_int8_t hammer2_get_obj_type(enum vtype vtype);
 722 void hammer2_time_to_timespec(u_int64_t xtime, struct timespec *ts);
 723 u_int64_t hammer2_timespec_to_time(struct timespec *ts);
 724 u_int32_t hammer2_to_unix_xid(uuid_t *uuid);
 725 void hammer2_guid_to_uuid(uuid_t *uuid, u_int32_t guid);
 726
 727 hammer2_key_t hammer2_dirhash(const unsigned char *name, size_t len);
 728 int hammer2_getradix(size_t bytes);
 729
 730 int hammer2_calc_logical(hammer2_inode_t *ip, hammer2_off_t uoff,
 731                         hammer2_key_t *lbasep, hammer2_key_t *leofp);
 732 int hammer2_calc_physical(hammer2_inode_t *ip, hammer2_key_t lbase);
 733 void hammer2_update_time(uint64_t *timep);
 734
 735 /*
 736  * hammer2_inode.c
 737  */
 738 struct vnode *hammer2_igetv(hammer2_inode_t *ip, int *errorp);
 739
 740 void hammer2_inode_lock_nlinks(hammer2_inode_t *ip);
 741 void hammer2_inode_unlock_nlinks(hammer2_inode_t *ip);
 742 hammer2_inode_t *hammer2_inode_lookup(hammer2_pfsmount_t *pmp,
 743                         hammer2_tid_t inum);
 744 hammer2_inode_t *hammer2_inode_get(hammer2_pfsmount_t *pmp,
 745                         hammer2_inode_t *dip, hammer2_chain_t *chain);
 746 void hammer2_inode_free(hammer2_inode_t *ip);
 747 void hammer2_inode_ref(hammer2_inode_t *ip);
 748 void hammer2_inode_drop(hammer2_inode_t *ip);
 749 void hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
 750                         hammer2_chain_t *chain);
 751
 752 hammer2_inode_t *hammer2_inode_create(hammer2_trans_t *trans,
 753                         hammer2_inode_t *dip,
 754                         struct vattr *vap, struct ucred *cred,
 755                         const uint8_t *name, size_t name_len,
 756                         hammer2_chain_t **chainp, int *errorp);
 757 int hammer2_inode_connect(hammer2_trans_t *trans,
 758                         hammer2_chain_t **chainp, int hlink,
 759                         hammer2_inode_t *dip, hammer2_chain_t **dchainp,
 760                         const uint8_t *name, size_t name_len,
 761                         hammer2_key_t key);
 762 hammer2_inode_t *hammer2_inode_common_parent(hammer2_inode_t *fdip,
 763                         hammer2_inode_t *tdip);
 764 void hammer2_inode_fsync(hammer2_trans_t *trans, hammer2_inode_t *ip,
 765                         hammer2_chain_t **parentp);
 766 int hammer2_unlink_file(hammer2_trans_t *trans, hammer2_inode_t *dip,
 767                         const uint8_t *name, size_t name_len, int isdir,
 768                         int *hlinkp, struct nchandle *nch);
 769 int hammer2_hardlink_consolidate(hammer2_trans_t *trans,
 770                         hammer2_inode_t *ip, hammer2_chain_t **chainp,
 771                         hammer2_inode_t *cdip, hammer2_chain_t **cdchainp,
 772                         int nlinks);
 773 int hammer2_hardlink_deconsolidate(hammer2_trans_t *trans, hammer2_inode_t *dip,
 774                         hammer2_chain_t **chainp, hammer2_chain_t **ochainp);
 775 int hammer2_hardlink_find(hammer2_inode_t *dip,
 776                         hammer2_chain_t **chainp, hammer2_chain_t **ochainp);
 777 void hammer2_inode_install_hidden(hammer2_pfsmount_t *pmp);
 778
 779 /*
 780  * hammer2_chain.c
 781  */
 782 void hammer2_modify_volume(hammer2_mount_t *hmp);
 783 hammer2_chain_t *hammer2_chain_alloc(hammer2_mount_t *hmp,
 784                                 hammer2_pfsmount_t *pmp,
 785                                 hammer2_trans_t *trans,
 786                                 hammer2_blockref_t *bref);
 787 void hammer2_chain_core_alloc(hammer2_trans_t *trans, hammer2_chain_t *nchain,
 788                                 hammer2_chain_t *ochain);
 789 void hammer2_chain_ref(hammer2_chain_t *chain);
 790 void hammer2_chain_drop(hammer2_chain_t *chain);
 791 int hammer2_chain_lock(hammer2_chain_t *chain, int how);
 792 void hammer2_chain_load_async(hammer2_chain_t *chain,
 793                                 void (*func)(hammer2_io_t *dio,
 794                                              hammer2_chain_t *chain,
 795                                              void *arg_p, off_t arg_o),
 796                                 void *arg_p, off_t arg_o);
 797 void hammer2_chain_moved(hammer2_chain_t *chain);
 798 void hammer2_chain_modify(hammer2_trans_t *trans,
 799                                 hammer2_chain_t **chainp, int flags);
 800 hammer2_inode_data_t *hammer2_chain_modify_ip(hammer2_trans_t *trans,
 801                                 hammer2_inode_t *ip, hammer2_chain_t **chainp,
 802                                 int flags);
 803 void hammer2_chain_resize(hammer2_trans_t *trans, hammer2_inode_t *ip,
 804                                 hammer2_chain_t *parent,
 805                                 hammer2_chain_t **chainp,
 806                                 int nradix, int flags);
 807 void hammer2_chain_unlock(hammer2_chain_t *chain);
 808 void hammer2_chain_wait(hammer2_chain_t *chain);
 809 hammer2_chain_t *hammer2_chain_get(hammer2_chain_t *parent, int generation,
 810                                 hammer2_blockref_t *bref);
 811 hammer2_chain_t *hammer2_chain_lookup_init(hammer2_chain_t *parent, int flags);
 812 void hammer2_chain_lookup_done(hammer2_chain_t *parent);
 813 hammer2_chain_t *hammer2_chain_lookup(hammer2_chain_t **parentp,
 814                                 hammer2_key_t *key_nextp,
 815                                 hammer2_key_t key_beg, hammer2_key_t key_end,
 816                                 int *cache_indexp, int flags);
 817 hammer2_chain_t *hammer2_chain_next(hammer2_chain_t **parentp,
 818                                 hammer2_chain_t *chain,
 819                                 hammer2_key_t *key_nextp,
 820                                 hammer2_key_t key_beg, hammer2_key_t key_end,
 821                                 int *cache_indexp, int flags);
 822 hammer2_chain_t *hammer2_chain_scan(hammer2_chain_t *parent,
 823                                 hammer2_chain_t *chain,
 824                                 int *cache_indexp, int flags);
 825
 826 int hammer2_chain_create(hammer2_trans_t *trans,
 827                                 hammer2_chain_t **parentp,
 828                                 hammer2_chain_t **chainp,
 829                                 hammer2_key_t key, int keybits,
 830                                 int type, size_t bytes);
 831 void hammer2_chain_duplicate(hammer2_trans_t *trans, hammer2_chain_t **parentp,
 832                                 hammer2_chain_t **chainp,
 833                                 hammer2_blockref_t *bref, int snapshot,
 834                                 int duplicate_reason);
 835 int hammer2_chain_snapshot(hammer2_trans_t *trans, hammer2_chain_t **chainp,
 836                                 hammer2_ioc_pfs_t *pfs);
 837 void hammer2_chain_delete(hammer2_trans_t *trans, hammer2_chain_t *chain,
 838                                 int flags);
 839 void hammer2_chain_delete_duplicate(hammer2_trans_t *trans,
 840                                 hammer2_chain_t **chainp, int flags);
 841 void hammer2_flush(hammer2_trans_t *trans, hammer2_chain_t **chainp);
 842 void hammer2_chain_commit(hammer2_trans_t *trans, hammer2_chain_t *chain);
 843 void hammer2_chain_setsubmod(hammer2_trans_t *trans, hammer2_chain_t *chain);
 844
 845 void hammer2_chain_memory_wait(hammer2_pfsmount_t *pmp);
 846 void hammer2_chain_memory_inc(hammer2_pfsmount_t *pmp);
 847 void hammer2_chain_memory_wakeup(hammer2_pfsmount_t *pmp);
 848 void hammer2_chain_countbrefs(hammer2_chain_t *chain,
 849                                 hammer2_blockref_t *base, int count);
 850
 851 int hammer2_base_find(hammer2_chain_t *chain,
 852                                 hammer2_blockref_t *base, int count,
 853                                 int *cache_indexp, hammer2_key_t *key_nextp,
 854                                 hammer2_key_t key_beg, hammer2_key_t key_end,
 855                                 int delete_filter);
 856 void hammer2_base_delete(hammer2_trans_t *trans, hammer2_chain_t *chain,
 857                                 hammer2_blockref_t *base, int count,
 858                                 int *cache_indexp, hammer2_chain_t *child);
 859 void hammer2_base_insert(hammer2_trans_t *trans, hammer2_chain_t *chain,
 860                                 hammer2_blockref_t *base, int count,
 861                                 int *cache_indexp, hammer2_chain_t *child);
 862
 863 /*
 864  * hammer2_trans.c
 865  */
 866 void hammer2_trans_init(hammer2_trans_t *trans, hammer2_pfsmount_t *pmp,
 867                                 hammer2_mount_t *hmp, int flags);
 868 void hammer2_trans_clear_invfsync(hammer2_trans_t *trans);
 869 void hammer2_trans_done(hammer2_trans_t *trans);
 870
 871 /*
 872  * hammer2_ioctl.c
 873  */
 874 int hammer2_ioctl(hammer2_inode_t *ip, u_long com, void *data,
 875                                 int fflag, struct ucred *cred);
 876
 877 /*
 878  * hammer2_io.c
 879  */
 880 hammer2_io_t *hammer2_io_getblk(hammer2_mount_t *hmp, off_t lbase,
 881                                 int lsize, int *ownerp);
 882 void hammer2_io_putblk(hammer2_io_t **diop);
 883 void hammer2_io_cleanup(hammer2_mount_t *hmp, struct hammer2_io_tree *tree);
 884 char *hammer2_io_data(hammer2_io_t *dio, off_t lbase);
 885 int hammer2_io_new(hammer2_mount_t *hmp, off_t lbase, int lsize,
 886                                 hammer2_io_t **diop);
 887 int hammer2_io_newnz(hammer2_mount_t *hmp, off_t lbase, int lsize,
 888                                 hammer2_io_t **diop);
 889 int hammer2_io_newq(hammer2_mount_t *hmp, off_t lbase, int lsize,
 890                                 hammer2_io_t **diop);
 891 int hammer2_io_bread(hammer2_mount_t *hmp, off_t lbase, int lsize,
 892                                 hammer2_io_t **diop);
 893 void hammer2_io_breadcb(hammer2_mount_t *hmp, off_t lbase, int lsize,
 894                                 void (*callback)(hammer2_io_t *dio,
 895                                                  hammer2_chain_t *arg_c,
 896                                                  void *arg_p, off_t arg_o),
 897                                 hammer2_chain_t *arg_c,
 898                                 void *arg_p, off_t arg_o);
 899 void hammer2_io_bawrite(hammer2_io_t **diop);
 900 void hammer2_io_bdwrite(hammer2_io_t **diop);
 901 int hammer2_io_bwrite(hammer2_io_t **diop);
 902 int hammer2_io_isdirty(hammer2_io_t *dio);
 903 void hammer2_io_setdirty(hammer2_io_t *dio);
 904 void hammer2_io_setinval(hammer2_io_t *dio, u_int bytes);
 905 void hammer2_io_brelse(hammer2_io_t **diop);
 906 void hammer2_io_bqrelse(hammer2_io_t **diop);
 907
 908 /*
 909  * hammer2_msgops.c
 910  */
 911 int hammer2_msg_dbg_rcvmsg(kdmsg_msg_t *msg);
 912 int hammer2_msg_adhoc_input(kdmsg_msg_t *msg);
 913
 914 /*
 915  * hammer2_vfsops.c
 916  */
 917 void hammer2_clusterctl_wakeup(kdmsg_iocom_t *iocom);
 918 void hammer2_volconf_update(hammer2_pfsmount_t *pmp, int index);
 919 void hammer2_cluster_reconnect(hammer2_pfsmount_t *pmp, struct file *fp);
 920 void hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx);
 921 void hammer2_bioq_sync(hammer2_pfsmount_t *pmp);
 922 int hammer2_vfs_sync(struct mount *mp, int waitflags);
 923 void hammer2_lwinprog_ref(hammer2_pfsmount_t *pmp);
 924 void hammer2_lwinprog_drop(hammer2_pfsmount_t *pmp);
 925 void hammer2_lwinprog_wait(hammer2_pfsmount_t *pmp);
 926
 927 /*
 928  * hammer2_freemap.c
 929  */
 930 int hammer2_freemap_alloc(hammer2_trans_t *trans, hammer2_chain_t *chain,
 931                                 size_t bytes);
 932 void hammer2_freemap_adjust(hammer2_trans_t *trans, hammer2_mount_t *hmp,
 933                                 hammer2_blockref_t *bref, int how);
 934
 935
 936 #endif /* !_KERNEL */
 937 #endif /* !_VFS_HAMMER2_HAMMER2_H_ */