linux/Documentation/assoc_array.txt
<<
>>
Prefs
   1                   ========================================
   2                   GENERIC ASSOCIATIVE ARRAY IMPLEMENTATION
   3                   ========================================
   4
   5Contents:
   6
   7 - Overview.
   8
   9 - The public API.
  10   - Edit script.
  11   - Operations table.
  12   - Manipulation functions.
  13   - Access functions.
  14   - Index key form.
  15
  16 - Internal workings.
  17   - Basic internal tree layout.
  18   - Shortcuts.
  19   - Splitting and collapsing nodes.
  20   - Non-recursive iteration.
  21   - Simultaneous alteration and iteration.
  22
  23
  24========
  25OVERVIEW
  26========
  27
  28This associative array implementation is an object container with the following
  29properties:
  30
  31 (1) Objects are opaque pointers.  The implementation does not care where they
  32     point (if anywhere) or what they point to (if anything).
  33
  34     [!] NOTE: Pointers to objects _must_ be zero in the least significant bit.
  35
  36 (2) Objects do not need to contain linkage blocks for use by the array.  This
  37     permits an object to be located in multiple arrays simultaneously.
  38     Rather, the array is made up of metadata blocks that point to objects.
  39
  40 (3) Objects require index keys to locate them within the array.
  41
  42 (4) Index keys must be unique.  Inserting an object with the same key as one
  43     already in the array will replace the old object.
  44
  45 (5) Index keys can be of any length and can be of different lengths.
  46
  47 (6) Index keys should encode the length early on, before any variation due to
  48     length is seen.
  49
  50 (7) Index keys can include a hash to scatter objects throughout the array.
  51
  52 (8) The array can iterated over.  The objects will not necessarily come out in
  53     key order.
  54
  55 (9) The array can be iterated over whilst it is being modified, provided the
  56     RCU readlock is being held by the iterator.  Note, however, under these
  57     circumstances, some objects may be seen more than once.  If this is a
  58     problem, the iterator should lock against modification.  Objects will not
  59     be missed, however, unless deleted.
  60
  61(10) Objects in the array can be looked up by means of their index key.
  62
  63(11) Objects can be looked up whilst the array is being modified, provided the
  64     RCU readlock is being held by the thread doing the look up.
  65
  66The implementation uses a tree of 16-pointer nodes internally that are indexed
  67on each level by nibbles from the index key in the same manner as in a radix
  68tree.  To improve memory efficiency, shortcuts can be emplaced to skip over
  69what would otherwise be a series of single-occupancy nodes.  Further, nodes
  70pack leaf object pointers into spare space in the node rather than making an
  71extra branch until as such time an object needs to be added to a full node.
  72
  73
  74==============
  75THE PUBLIC API
  76==============
  77
  78The public API can be found in <linux/assoc_array.h>.  The associative array is
  79rooted on the following structure:
  80
  81        struct assoc_array {
  82                ...
  83        };
  84
  85The code is selected by enabling CONFIG_ASSOCIATIVE_ARRAY.
  86
  87
  88EDIT SCRIPT
  89-----------
  90
  91The insertion and deletion functions produce an 'edit script' that can later be
  92applied to effect the changes without risking ENOMEM.  This retains the
  93preallocated metadata blocks that will be installed in the internal tree and
  94keeps track of the metadata blocks that will be removed from the tree when the
  95script is applied.
  96
  97This is also used to keep track of dead blocks and dead objects after the
  98script has been applied so that they can be freed later.  The freeing is done
  99after an RCU grace period has passed - thus allowing access functions to
 100proceed under the RCU read lock.
 101
 102The script appears as outside of the API as a pointer of the type:
 103
 104        struct assoc_array_edit;
 105
 106There are two functions for dealing with the script:
 107
 108 (1) Apply an edit script.
 109
 110        void assoc_array_apply_edit(struct assoc_array_edit *edit);
 111
 112     This will perform the edit functions, interpolating various write barriers
 113     to permit accesses under the RCU read lock to continue.  The edit script
 114     will then be passed to call_rcu() to free it and any dead stuff it points
 115     to.
 116
 117 (2) Cancel an edit script.
 118
 119        void assoc_array_cancel_edit(struct assoc_array_edit *edit);
 120
 121     This frees the edit script and all preallocated memory immediately.  If
 122     this was for insertion, the new object is _not_ released by this function,
 123     but must rather be released by the caller.
 124
 125These functions are guaranteed not to fail.
 126
 127
 128OPERATIONS TABLE
 129----------------
 130
 131Various functions take a table of operations:
 132
 133        struct assoc_array_ops {
 134                ...
 135        };
 136
 137This points to a number of methods, all of which need to be provided:
 138
 139 (1) Get a chunk of index key from caller data:
 140
 141        unsigned long (*get_key_chunk)(const void *index_key, int level);
 142
 143     This should return a chunk of caller-supplied index key starting at the
 144     *bit* position given by the level argument.  The level argument will be a
 145     multiple of ASSOC_ARRAY_KEY_CHUNK_SIZE and the function should return
 146     ASSOC_ARRAY_KEY_CHUNK_SIZE bits.  No error is possible.
 147
 148
 149 (2) Get a chunk of an object's index key.
 150
 151        unsigned long (*get_object_key_chunk)(const void *object, int level);
 152
 153     As the previous function, but gets its data from an object in the array
 154     rather than from a caller-supplied index key.
 155
 156
 157 (3) See if this is the object we're looking for.
 158
 159        bool (*compare_object)(const void *object, const void *index_key);
 160
 161     Compare the object against an index key and return true if it matches and
 162     false if it doesn't.
 163
 164
 165 (4) Diff the index keys of two objects.
 166
 167        int (*diff_objects)(const void *object, const void *index_key);
 168
 169     Return the bit position at which the index key of the specified object
 170     differs from the given index key or -1 if they are the same.
 171
 172
 173 (5) Free an object.
 174
 175        void (*free_object)(void *object);
 176
 177     Free the specified object.  Note that this may be called an RCU grace
 178     period after assoc_array_apply_edit() was called, so synchronize_rcu() may
 179     be necessary on module unloading.
 180
 181
 182MANIPULATION FUNCTIONS
 183----------------------
 184
 185There are a number of functions for manipulating an associative array:
 186
 187 (1) Initialise an associative array.
 188
 189        void assoc_array_init(struct assoc_array *array);
 190
 191     This initialises the base structure for an associative array.  It can't
 192     fail.
 193
 194
 195 (2) Insert/replace an object in an associative array.
 196
 197        struct assoc_array_edit *
 198        assoc_array_insert(struct assoc_array *array,
 199                           const struct assoc_array_ops *ops,
 200                           const void *index_key,
 201                           void *object);
 202
 203     This inserts the given object into the array.  Note that the least
 204     significant bit of the pointer must be zero as it's used to type-mark
 205     pointers internally.
 206
 207     If an object already exists for that key then it will be replaced with the
 208     new object and the old one will be freed automatically.
 209
 210     The index_key argument should hold index key information and is
 211     passed to the methods in the ops table when they are called.
 212
 213     This function makes no alteration to the array itself, but rather returns
 214     an edit script that must be applied.  -ENOMEM is returned in the case of
 215     an out-of-memory error.
 216
 217     The caller should lock exclusively against other modifiers of the array.
 218
 219
 220 (3) Delete an object from an associative array.
 221
 222        struct assoc_array_edit *
 223        assoc_array_delete(struct assoc_array *array,
 224                           const struct assoc_array_ops *ops,
 225                           const void *index_key);
 226
 227     This deletes an object that matches the specified data from the array.
 228
 229     The index_key argument should hold index key information and is
 230     passed to the methods in the ops table when they are called.
 231
 232     This function makes no alteration to the array itself, but rather returns
 233     an edit script that must be applied.  -ENOMEM is returned in the case of
 234     an out-of-memory error.  NULL will be returned if the specified object is
 235     not found within the array.
 236
 237     The caller should lock exclusively against other modifiers of the array.
 238
 239
 240 (4) Delete all objects from an associative array.
 241
 242        struct assoc_array_edit *
 243        assoc_array_clear(struct assoc_array *array,
 244                          const struct assoc_array_ops *ops);
 245
 246     This deletes all the objects from an associative array and leaves it
 247     completely empty.
 248
 249     This function makes no alteration to the array itself, but rather returns
 250     an edit script that must be applied.  -ENOMEM is returned in the case of
 251     an out-of-memory error.
 252
 253     The caller should lock exclusively against other modifiers of the array.
 254
 255
 256 (5) Destroy an associative array, deleting all objects.
 257
 258        void assoc_array_destroy(struct assoc_array *array,
 259                                 const struct assoc_array_ops *ops);
 260
 261     This destroys the contents of the associative array and leaves it
 262     completely empty.  It is not permitted for another thread to be traversing
 263     the array under the RCU read lock at the same time as this function is
 264     destroying it as no RCU deferral is performed on memory release -
 265     something that would require memory to be allocated.
 266
 267     The caller should lock exclusively against other modifiers and accessors
 268     of the array.
 269
 270
 271 (6) Garbage collect an associative array.
 272
 273        int assoc_array_gc(struct assoc_array *array,
 274                           const struct assoc_array_ops *ops,
 275                           bool (*iterator)(void *object, void *iterator_data),
 276                           void *iterator_data);
 277
 278     This iterates over the objects in an associative array and passes each one
 279     to iterator().  If iterator() returns true, the object is kept.  If it
 280     returns false, the object will be freed.  If the iterator() function
 281     returns true, it must perform any appropriate refcount incrementing on the
 282     object before returning.
 283
 284     The internal tree will be packed down if possible as part of the iteration
 285     to reduce the number of nodes in it.
 286
 287     The iterator_data is passed directly to iterator() and is otherwise
 288     ignored by the function.
 289
 290     The function will return 0 if successful and -ENOMEM if there wasn't
 291     enough memory.
 292
 293     It is possible for other threads to iterate over or search the array under
 294     the RCU read lock whilst this function is in progress.  The caller should
 295     lock exclusively against other modifiers of the array.
 296
 297
 298ACCESS FUNCTIONS
 299----------------
 300
 301There are two functions for accessing an associative array:
 302
 303 (1) Iterate over all the objects in an associative array.
 304
 305        int assoc_array_iterate(const struct assoc_array *array,
 306                                int (*iterator)(const void *object,
 307                                                void *iterator_data),
 308                                void *iterator_data);
 309
 310     This passes each object in the array to the iterator callback function.
 311     iterator_data is private data for that function.
 312
 313     This may be used on an array at the same time as the array is being
 314     modified, provided the RCU read lock is held.  Under such circumstances,
 315     it is possible for the iteration function to see some objects twice.  If
 316     this is a problem, then modification should be locked against.  The
 317     iteration algorithm should not, however, miss any objects.
 318
 319     The function will return 0 if no objects were in the array or else it will
 320     return the result of the last iterator function called.  Iteration stops
 321     immediately if any call to the iteration function results in a non-zero
 322     return.
 323
 324
 325 (2) Find an object in an associative array.
 326
 327        void *assoc_array_find(const struct assoc_array *array,
 328                               const struct assoc_array_ops *ops,
 329                               const void *index_key);
 330
 331     This walks through the array's internal tree directly to the object
 332     specified by the index key..
 333
 334     This may be used on an array at the same time as the array is being
 335     modified, provided the RCU read lock is held.
 336
 337     The function will return the object if found (and set *_type to the object
 338     type) or will return NULL if the object was not found.
 339
 340
 341INDEX KEY FORM
 342--------------
 343
 344The index key can be of any form, but since the algorithms aren't told how long
 345the key is, it is strongly recommended that the index key includes its length
 346very early on before any variation due to the length would have an effect on
 347comparisons.
 348
 349This will cause leaves with different length keys to scatter away from each
 350other - and those with the same length keys to cluster together.
 351
 352It is also recommended that the index key begin with a hash of the rest of the
 353key to maximise scattering throughout keyspace.
 354
 355The better the scattering, the wider and lower the internal tree will be.
 356
 357Poor scattering isn't too much of a problem as there are shortcuts and nodes
 358can contain mixtures of leaves and metadata pointers.
 359
 360The index key is read in chunks of machine word.  Each chunk is subdivided into
 361one nibble (4 bits) per level, so on a 32-bit CPU this is good for 8 levels and
 362on a 64-bit CPU, 16 levels.  Unless the scattering is really poor, it is
 363unlikely that more than one word of any particular index key will have to be
 364used.
 365
 366
 367=================
 368INTERNAL WORKINGS
 369=================
 370
 371The associative array data structure has an internal tree.  This tree is
 372constructed of two types of metadata blocks: nodes and shortcuts.
 373
 374A node is an array of slots.  Each slot can contain one of four things:
 375
 376 (*) A NULL pointer, indicating that the slot is empty.
 377
 378 (*) A pointer to an object (a leaf).
 379
 380 (*) A pointer to a node at the next level.
 381
 382 (*) A pointer to a shortcut.
 383
 384
 385BASIC INTERNAL TREE LAYOUT
 386--------------------------
 387
 388Ignoring shortcuts for the moment, the nodes form a multilevel tree.  The index
 389key space is strictly subdivided by the nodes in the tree and nodes occur on
 390fixed levels.  For example:
 391
 392 Level: 0               1               2               3
 393        =============== =============== =============== ===============
 394                                                        NODE D
 395                        NODE B          NODE C  +------>+---+
 396                +------>+---+   +------>+---+   |       | 0 |
 397        NODE A  |       | 0 |   |       | 0 |   |       +---+
 398        +---+   |       +---+   |       +---+   |       :   :
 399        | 0 |   |       :   :   |       :   :   |       +---+
 400        +---+   |       +---+   |       +---+   |       | f |
 401        | 1 |---+       | 3 |---+       | 7 |---+       +---+
 402        +---+           +---+           +---+
 403        :   :           :   :           | 8 |---+
 404        +---+           +---+           +---+   |       NODE E
 405        | e |---+       | f |           :   :   +------>+---+
 406        +---+   |       +---+           +---+           | 0 |
 407        | f |   |                       | f |           +---+
 408        +---+   |                       +---+           :   :
 409                |       NODE F                          +---+
 410                +------>+---+                           | f |
 411                        | 0 |           NODE G          +---+
 412                        +---+   +------>+---+
 413                        :   :   |       | 0 |
 414                        +---+   |       +---+
 415                        | 6 |---+       :   :
 416                        +---+           +---+
 417                        :   :           | f |
 418                        +---+           +---+
 419                        | f |
 420                        +---+
 421
 422In the above example, there are 7 nodes (A-G), each with 16 slots (0-f).
 423Assuming no other meta data nodes in the tree, the key space is divided thusly:
 424
 425        KEY PREFIX      NODE
 426        ==========      ====
 427        137*            D
 428        138*            E
 429        13[0-69-f]*     C
 430        1[0-24-f]*      B
 431        e6*             G
 432        e[0-57-f]*      F
 433        [02-df]*        A
 434
 435So, for instance, keys with the following example index keys will be found in
 436the appropriate nodes:
 437
 438        INDEX KEY       PREFIX  NODE
 439        =============== ======= ====
 440        13694892892489  13      C
 441        13795289025897  137     D
 442        13889dde88793   138     E
 443        138bbb89003093  138     E
 444        1394879524789   12      C
 445        1458952489      1       B
 446        9431809de993ba  -       A
 447        b4542910809cd   -       A
 448        e5284310def98   e       F
 449        e68428974237    e6      G
 450        e7fffcbd443     e       F
 451        f3842239082     -       A
 452
 453To save memory, if a node can hold all the leaves in its portion of keyspace,
 454then the node will have all those leaves in it and will not have any metadata
 455pointers - even if some of those leaves would like to be in the same slot.
 456
 457A node can contain a heterogeneous mix of leaves and metadata pointers.
 458Metadata pointers must be in the slots that match their subdivisions of key
 459space.  The leaves can be in any slot not occupied by a metadata pointer.  It
 460is guaranteed that none of the leaves in a node will match a slot occupied by a
 461metadata pointer.  If the metadata pointer is there, any leaf whose key matches
 462the metadata key prefix must be in the subtree that the metadata pointer points
 463to.
 464
 465In the above example list of index keys, node A will contain:
 466
 467        SLOT    CONTENT         INDEX KEY (PREFIX)
 468        ====    =============== ==================
 469        1       PTR TO NODE B   1*
 470        any     LEAF            9431809de993ba
 471        any     LEAF            b4542910809cd
 472        e       PTR TO NODE F   e*
 473        any     LEAF            f3842239082
 474
 475and node B:
 476
 477        3       PTR TO NODE C   13*
 478        any     LEAF            1458952489
 479
 480
 481SHORTCUTS
 482---------
 483
 484Shortcuts are metadata records that jump over a piece of keyspace.  A shortcut
 485is a replacement for a series of single-occupancy nodes ascending through the
 486levels.  Shortcuts exist to save memory and to speed up traversal.
 487
 488It is possible for the root of the tree to be a shortcut - say, for example,
 489the tree contains at least 17 nodes all with key prefix '1111'.  The insertion
 490algorithm will insert a shortcut to skip over the '1111' keyspace in a single
 491bound and get to the fourth level where these actually become different.
 492
 493
 494SPLITTING AND COLLAPSING NODES
 495------------------------------
 496
 497Each node has a maximum capacity of 16 leaves and metadata pointers.  If the
 498insertion algorithm finds that it is trying to insert a 17th object into a
 499node, that node will be split such that at least two leaves that have a common
 500key segment at that level end up in a separate node rooted on that slot for
 501that common key segment.
 502
 503If the leaves in a full node and the leaf that is being inserted are
 504sufficiently similar, then a shortcut will be inserted into the tree.
 505
 506When the number of objects in the subtree rooted at a node falls to 16 or
 507fewer, then the subtree will be collapsed down to a single node - and this will
 508ripple towards the root if possible.
 509
 510
 511NON-RECURSIVE ITERATION
 512-----------------------
 513
 514Each node and shortcut contains a back pointer to its parent and the number of
 515slot in that parent that points to it.  None-recursive iteration uses these to
 516proceed rootwards through the tree, going to the parent node, slot N + 1 to
 517make sure progress is made without the need for a stack.
 518
 519The backpointers, however, make simultaneous alteration and iteration tricky.
 520
 521
 522SIMULTANEOUS ALTERATION AND ITERATION
 523-------------------------------------
 524
 525There are a number of cases to consider:
 526
 527 (1) Simple insert/replace.  This involves simply replacing a NULL or old
 528     matching leaf pointer with the pointer to the new leaf after a barrier.
 529     The metadata blocks don't change otherwise.  An old leaf won't be freed
 530     until after the RCU grace period.
 531
 532 (2) Simple delete.  This involves just clearing an old matching leaf.  The
 533     metadata blocks don't change otherwise.  The old leaf won't be freed until
 534     after the RCU grace period.
 535
 536 (3) Insertion replacing part of a subtree that we haven't yet entered.  This
 537     may involve replacement of part of that subtree - but that won't affect
 538     the iteration as we won't have reached the pointer to it yet and the
 539     ancestry blocks are not replaced (the layout of those does not change).
 540
 541 (4) Insertion replacing nodes that we're actively processing.  This isn't a
 542     problem as we've passed the anchoring pointer and won't switch onto the
 543     new layout until we follow the back pointers - at which point we've
 544     already examined the leaves in the replaced node (we iterate over all the
 545     leaves in a node before following any of its metadata pointers).
 546
 547     We might, however, re-see some leaves that have been split out into a new
 548     branch that's in a slot further along than we were at.
 549
 550 (5) Insertion replacing nodes that we're processing a dependent branch of.
 551     This won't affect us until we follow the back pointers.  Similar to (4).
 552
 553 (6) Deletion collapsing a branch under us.  This doesn't affect us because the
 554     back pointers will get us back to the parent of the new node before we
 555     could see the new node.  The entire collapsed subtree is thrown away
 556     unchanged - and will still be rooted on the same slot, so we shouldn't
 557     process it a second time as we'll go back to slot + 1.
 558
 559Note:
 560
 561 (*) Under some circumstances, we need to simultaneously change the parent
 562     pointer and the parent slot pointer on a node (say, for example, we
 563     inserted another node before it and moved it up a level).  We cannot do
 564     this without locking against a read - so we have to replace that node too.
 565
 566     However, when we're changing a shortcut into a node this isn't a problem
 567     as shortcuts only have one slot and so the parent slot number isn't used
 568     when traversing backwards over one.  This means that it's okay to change
 569     the slot number first - provided suitable barriers are used to make sure
 570     the parent slot number is read after the back pointer.
 571
 572Obsolete blocks and leaves are freed up after an RCU grace period has passed,
 573so as long as anyone doing walking or iteration holds the RCU read lock, the
 574old superstructure should not go away on them.
 575