1/* Generic associative array implementation.
2 *
3 * See Documentation/core-api/assoc_array.rst for information.
4 *
5 * Copyright (C) 2013 Red Hat, Inc. All Rights Reserved.
6 * Written by David Howells (dhowells@redhat.com)
7 *
8 * This program is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU General Public Licence
10 * as published by the Free Software Foundation; either version
11 * 2 of the Licence, or (at your option) any later version.
12 */
13//#define DEBUG
14#include <linux/rcupdate.h>
15#include <linux/slab.h>
16#include <linux/err.h>
17#include <linux/assoc_array_priv.h>
18
19/*
20 * Iterate over an associative array. The caller must hold the RCU read lock
21 * or better.
22 */
23static int assoc_array_subtree_iterate(const struct assoc_array_ptr *root,
24 const struct assoc_array_ptr *stop,
25 int (*iterator)(const void *leaf,
26 void *iterator_data),
27 void *iterator_data)
28{
29 const struct assoc_array_shortcut *shortcut;
30 const struct assoc_array_node *node;
31 const struct assoc_array_ptr *cursor, *ptr, *parent;
32 unsigned long has_meta;
33 int slot, ret;
34
35 cursor = root;
36
37begin_node:
38 if (assoc_array_ptr_is_shortcut(cursor)) {
39 /* Descend through a shortcut */
40 shortcut = assoc_array_ptr_to_shortcut(cursor);
41 cursor = READ_ONCE(shortcut->next_node); /* Address dependency. */
42 }
43
44 node = assoc_array_ptr_to_node(cursor);
45 slot = 0;
46
47 /* We perform two passes of each node.
48 *
49 * The first pass does all the leaves in this node. This means we
50 * don't miss any leaves if the node is split up by insertion whilst
51 * we're iterating over the branches rooted here (we may, however, see
52 * some leaves twice).
53 */
54 has_meta = 0;
55 for (; slot < ASSOC_ARRAY_FAN_OUT; slot++) {
56 ptr = READ_ONCE(node->slots[slot]); /* Address dependency. */
57 has_meta |= (unsigned long)ptr;
58 if (ptr && assoc_array_ptr_is_leaf(ptr)) {
59 /* We need a barrier between the read of the pointer,
60 * which is supplied by the above READ_ONCE().
61 */
62 /* Invoke the callback */
63 ret = iterator(assoc_array_ptr_to_leaf(ptr),
64 iterator_data);
65 if (ret)
66 return ret;
67 }
68 }
69
70 /* The second pass attends to all the metadata pointers. If we follow
71 * one of these we may find that we don't come back here, but rather go
72 * back to a replacement node with the leaves in a different layout.
73 *
74 * We are guaranteed to make progress, however, as the slot number for
75 * a particular portion of the key space cannot change - and we
76 * continue at the back pointer + 1.
77 */
78 if (!(has_meta & ASSOC_ARRAY_PTR_META_TYPE))
79 goto finished_node;
80 slot = 0;
81
82continue_node:
83 node = assoc_array_ptr_to_node(cursor);
84 for (; slot < ASSOC_ARRAY_FAN_OUT; slot++) {
85 ptr = READ_ONCE(node->slots[slot]); /* Address dependency. */
86 if (assoc_array_ptr_is_meta(ptr)) {
87 cursor = ptr;
88 goto begin_node;
89 }
90 }
91
92finished_node:
93 /* Move up to the parent (may need to skip back over a shortcut) */
94 parent = READ_ONCE(node->back_pointer); /* Address dependency. */
95 slot = node->parent_slot;
96 if (parent == stop)
97 return 0;
98
99 if (assoc_array_ptr_is_shortcut(parent)) {
100 shortcut = assoc_array_ptr_to_shortcut(parent);
101 cursor = parent;
102 parent = READ_ONCE(shortcut->back_pointer); /* Address dependency. */
103 slot = shortcut->parent_slot;
104 if (parent == stop)
105 return 0;
106 }
107
108 /* Ascend to next slot in parent node */
109 cursor = parent;
110 slot++;
111 goto continue_node;
112}
113
114/**
115 * assoc_array_iterate - Pass all objects in the array to a callback
116 * @array: The array to iterate over.
117 * @iterator: The callback function.
118 * @iterator_data: Private data for the callback function.
119 *
120 * Iterate over all the objects in an associative array. Each one will be
121 * presented to the iterator function.
122 *
123 * If the array is being modified concurrently with the iteration then it is
124 * possible that some objects in the array will be passed to the iterator
125 * callback more than once - though every object should be passed at least
126 * once. If this is undesirable then the caller must lock against modification
127 * for the duration of this function.
128 *
129 * The function will return 0 if no objects were in the array or else it will
130 * return the result of the last iterator function called. Iteration stops
131 * immediately if any call to the iteration function results in a non-zero
132 * return.
133 *
134 * The caller should hold the RCU read lock or better if concurrent
135 * modification is possible.
136 */
137int assoc_array_iterate(const struct assoc_array *array,
138 int (*iterator)(const void *object,
139 void *iterator_data),
140 void *iterator_data)
141{
142 struct assoc_array_ptr *root = READ_ONCE(array->root); /* Address dependency. */
143
144 if (!root)
145 return 0;
146 return assoc_array_subtree_iterate(root, NULL, iterator, iterator_data);
147}
148
149enum assoc_array_walk_status {
150 assoc_array_walk_tree_empty,
151 assoc_array_walk_found_terminal_node,
152 assoc_array_walk_found_wrong_shortcut,
153};
154
155struct assoc_array_walk_result {
156 struct {
157 struct assoc_array_node *node; /* Node in which leaf might be found */
158 int level;
159 int slot;
160 } terminal_node;
161 struct {
162 struct assoc_array_shortcut *shortcut;
163 int level;
164 int sc_level;
165 unsigned long sc_segments;
166 unsigned long dissimilarity;
167 } wrong_shortcut;
168};
169
170/*
171 * Navigate through the internal tree looking for the closest node to the key.
172 */
173static enum assoc_array_walk_status
174assoc_array_walk(const struct assoc_array *array,
175 const struct assoc_array_ops *ops,
176 const void *index_key,
177 struct assoc_array_walk_result *result)
178{
179 struct assoc_array_shortcut *shortcut;
180 struct assoc_array_node *node;
181 struct assoc_array_ptr *cursor, *ptr;
182 unsigned long sc_segments, dissimilarity;
183 unsigned long segments;
184 int level, sc_level, next_sc_level;
185 int slot;
186
187 pr_devel("-->%s()\n", __func__);
188
189 cursor = READ_ONCE(array->root); /* Address dependency. */
190 if (!cursor)
191 return assoc_array_walk_tree_empty;
192
193 level = 0;
194
195 /* Use segments from the key for the new leaf to navigate through the
196 * internal tree, skipping through nodes and shortcuts that are on
197 * route to the destination. Eventually we'll come to a slot that is
198 * either empty or contains a leaf at which point we've found a node in
199 * which the leaf we're looking for might be found or into which it
200 * should be inserted.
201 */
202jumped:
203 segments = ops->get_key_chunk(index_key, level);
204 pr_devel("segments[%d]: %lx\n", level, segments);
205
206 if (assoc_array_ptr_is_shortcut(cursor))
207 goto follow_shortcut;
208
209consider_node:
210 node = assoc_array_ptr_to_node(cursor);
211 slot = segments >> (level & ASSOC_ARRAY_KEY_CHUNK_MASK);
212 slot &= ASSOC_ARRAY_FAN_MASK;
213 ptr = READ_ONCE(node->slots[slot]); /* Address dependency. */
214
215 pr_devel("consider slot %x [ix=%d type=%lu]\n",
216 slot, level, (unsigned long)ptr & 3);
217
218 if (!assoc_array_ptr_is_meta(ptr)) {
219 /* The node doesn't have a node/shortcut pointer in the slot
220 * corresponding to the index key that we have to follow.
221 */
222 result->terminal_node.node = node;
223 result->terminal_node.level = level;
224 result->terminal_node.slot = slot;
225 pr_devel("<--%s() = terminal_node\n", __func__);
226 return assoc_array_walk_found_terminal_node;
227 }
228
229 if (assoc_array_ptr_is_node(ptr)) {
230 /* There is a pointer to a node in the slot corresponding to
231 * this index key segment, so we need to follow it.
232 */
233 cursor = ptr;
234 level += ASSOC_ARRAY_LEVEL_STEP;
235 if ((level & ASSOC_ARRAY_KEY_CHUNK_MASK) != 0)
236 goto consider_node;
237 goto jumped;
238 }
239
240 /* There is a shortcut in the slot corresponding to the index key
241 * segment. We follow the shortcut if its partial index key matches
242 * this leaf's. Otherwise we need to split the shortcut.
243 */
244 cursor = ptr;
245follow_shortcut:
246 shortcut = assoc_array_ptr_to_shortcut(cursor);
247 pr_devel("shortcut to %d\n", shortcut->skip_to_level);
248 sc_level = level + ASSOC_ARRAY_LEVEL_STEP;
249 BUG_ON(sc_level > shortcut->skip_to_level);
250
251 do {
252 /* Check the leaf against the shortcut's index key a word at a
253 * time, trimming the final word (the shortcut stores the index
254 * key completely from the root to the shortcut's target).
255 */
256 if ((sc_level & ASSOC_ARRAY_KEY_CHUNK_MASK) == 0)
257 segments = ops->get_key_chunk(index_key, sc_level);
258
259 sc_segments = shortcut->index_key[sc_level >> ASSOC_ARRAY_KEY_CHUNK_SHIFT];
260 dissimilarity = segments ^ sc_segments;
261
262 if (round_up(sc_level, ASSOC_ARRAY_KEY_CHUNK_SIZE) > shortcut->skip_to_level) {
263 /* Trim segments that are beyond the shortcut */
264 int shift = shortcut->skip_to_level & ASSOC_ARRAY_KEY_CHUNK_MASK;
265 dissimilarity &= ~(ULONG_MAX << shift);
266 next_sc_level = shortcut->skip_to_level;
267 } else {
268 next_sc_level = sc_level + ASSOC_ARRAY_KEY_CHUNK_SIZE;
269 next_sc_level = round_down(next_sc_level, ASSOC_ARRAY_KEY_CHUNK_SIZE);
270 }
271
272 if (dissimilarity != 0) {
273 /* This shortcut points elsewhere */
274 result->wrong_shortcut.shortcut = shortcut;
275 result->wrong_shortcut.level = level;
276 result->wrong_shortcut.sc_level = sc_level;
277 result->wrong_shortcut.sc_segments = sc_segments;
278 result->wrong_shortcut.dissimilarity = dissimilarity;
279 return assoc_array_walk_found_wrong_shortcut;
280 }
281
282 sc_level = next_sc_level;
283 } while (sc_level < shortcut->skip_to_level);
284
285 /* The shortcut matches the leaf's index to this point. */
286 cursor = READ_ONCE(shortcut->next_node); /* Address dependency. */
287 if (((level ^ sc_level) & ~ASSOC_ARRAY_KEY_CHUNK_MASK) != 0) {
288 level = sc_level;
289 goto jumped;
290 } else {
291 level = sc_level;
292 goto consider_node;
293 }
294}
295
296/**
297 * assoc_array_find - Find an object by index key
298 * @array: The associative array to search.
299 * @ops: The operations to use.
300 * @index_key: The key to the object.
301 *
302 * Find an object in an associative array by walking through the internal tree
303 * to the node that should contain the object and then searching the leaves
304 * there. NULL is returned if the requested object was not found in the array.
305 *
306 * The caller must hold the RCU read lock or better.
307 */
308void *assoc_array_find(const struct assoc_array *array,
309 const struct assoc_array_ops *ops,
310 const void *index_key)
311{
312 struct assoc_array_walk_result result;
313 const struct assoc_array_node *node;
314 const struct assoc_array_ptr *ptr;
315 const void *leaf;
316 int slot;
317
318 if (assoc_array_walk(array, ops, index_key, &result) !=
319 assoc_array_walk_found_terminal_node)
320 return NULL;
321
322 node = result.terminal_node.node;
323
324 /* If the target key is available to us, it's has to be pointed to by
325 * the terminal node.
326 */
327 for (slot = 0; slot < ASSOC_ARRAY_FAN_OUT; slot++) {
328 ptr = READ_ONCE(node->slots[slot]); /* Address dependency. */
329 if (ptr && assoc_array_ptr_is_leaf(ptr)) {
330 /* We need a barrier between the read of the pointer
331 * and dereferencing the pointer - but only if we are
332 * actually going to dereference it.
333 */
334 leaf = assoc_array_ptr_to_leaf(ptr);
335 if (ops->compare_object(leaf, index_key))
336 return (void *)leaf;
337 }
338 }
339
340 return NULL;
341}
342
343/*
344 * Destructively iterate over an associative array. The caller must prevent
345 * other simultaneous accesses.
346 */
347static void assoc_array_destroy_subtree(struct assoc_array_ptr *root,
348 const struct assoc_array_ops *ops)
349{
350 struct assoc_array_shortcut *shortcut;
351 struct assoc_array_node *node;
352 struct assoc_array_ptr *cursor, *parent = NULL;
353 int slot = -1;
354
355 pr_devel("-->%s()\n", __func__);
356
357 cursor = root;
358 if (!cursor) {
359 pr_devel("empty\n");
360 return;
361 }
362
363move_to_meta:
364 if (assoc_array_ptr_is_shortcut(cursor)) {
365 /* Descend through a shortcut */
366 pr_devel("[%d] shortcut\n", slot);
367 BUG_ON(!assoc_array_ptr_is_shortcut(cursor));
368 shortcut = assoc_array_ptr_to_shortcut(cursor);
369 BUG_ON(shortcut->back_pointer != parent);
370 BUG_ON(slot != -1 && shortcut->parent_slot != slot);
371 parent = cursor;
372 cursor = shortcut->next_node;
373 slot = -1;
374 BUG_ON(!assoc_array_ptr_is_node(cursor));
375 }
376
377 pr_devel("[%d] node\n", slot);
378 node = assoc_array_ptr_to_node(cursor);
379 BUG_ON(node->back_pointer != parent);
380 BUG_ON(slot != -1 && node->parent_slot != slot);
381 slot = 0;
382
383continue_node:
384 pr_devel("Node %p [back=%p]\n", node, node->back_pointer);
385 for (; slot < ASSOC_ARRAY_FAN_OUT; slot++) {
386 struct assoc_array_ptr *ptr = node->slots[slot];
387 if (!ptr)
388 continue;
389 if (assoc_array_ptr_is_meta(ptr)) {
390 parent = cursor;
391 cursor = ptr;
392 goto move_to_meta;
393 }
394
395 if (ops) {
396 pr_devel("[%d] free leaf\n", slot);
397 ops->free_object(assoc_array_ptr_to_leaf(ptr));
398 }
399 }
400
401 parent = node->back_pointer;
402 slot = node->parent_slot;
403 pr_devel("free node\n");
404 kfree(node);
405 if (!parent)
406 return; /* Done */
407
408 /* Move back up to the parent (may need to free a shortcut on
409 * the way up) */
410 if (assoc_array_ptr_is_shortcut(parent)) {
411 shortcut = assoc_array_ptr_to_shortcut(parent);
412 BUG_ON(shortcut->next_node != cursor);
413 cursor = parent;
414 parent = shortcut->back_pointer;
415 slot = shortcut->parent_slot;
416 pr_devel("free shortcut\n");
417 kfree(shortcut);
418 if (!parent)
419 return;
420
421 BUG_ON(!assoc_array_ptr_is_node(parent));
422 }
423
424 /* Ascend to next slot in parent node */
425 pr_devel("ascend to %p[%d]\n", parent, slot);
426 cursor = parent;
427 node = assoc_array_ptr_to_node(cursor);
428 slot++;
429 goto continue_node;
430}
431
432/**
433 * assoc_array_destroy - Destroy an associative array
434 * @array: The array to destroy.
435 * @ops: The operations to use.
436 *
437 * Discard all metadata and free all objects in an associative array. The
438 * array will be empty and ready to use again upon completion. This function
439 * cannot fail.
440 *
441 * The caller must prevent all other accesses whilst this takes place as no
442 * attempt is made to adjust pointers gracefully to permit RCU readlock-holding
443 * accesses to continue. On the other hand, no memory allocation is required.
444 */
445void assoc_array_destroy(struct assoc_array *array,
446 const struct assoc_array_ops *ops)
447{
448 assoc_array_destroy_subtree(array->root, ops);
449 array->root = NULL;
450}
451
452/*
453 * Handle insertion into an empty tree.
454 */
455static bool assoc_array_insert_in_empty_tree(struct assoc_array_edit *edit)
456{
457 struct assoc_array_node *new_n0;
458
459 pr_devel("-->%s()\n", __func__);
460
461 new_n0 = kzalloc(sizeof(struct assoc_array_node), GFP_KERNEL);
462 if (!new_n0)
463 return false;
464
465 edit->new_meta[0] = assoc_array_node_to_ptr(new_n0);
466 edit->leaf_p = &new_n0->slots[0];
467 edit->adjust_count_on = new_n0;
468 edit->set[0].ptr = &edit->array->root;
469 edit->set[0].to = assoc_array_node_to_ptr(new_n0);
470
471 pr_devel("<--%s() = ok [no root]\n", __func__);
472 return true;
473}
474
475/*
476 * Handle insertion into a terminal node.
477 */
478static bool assoc_array_insert_into_terminal_node(struct assoc_array_edit *edit,
479 const struct assoc_array_ops *ops,
480 const void *index_key,
481 struct assoc_array_walk_result *result)
482{
483 struct assoc_array_shortcut *shortcut, *new_s0;
484 struct assoc_array_node *node, *new_n0, *new_n1, *side;
485 struct assoc_array_ptr *ptr;
486 unsigned long dissimilarity, base_seg, blank;
487 size_t keylen;
488 bool have_meta;
489 int level, diff;
490 int slot, next_slot, free_slot, i, j;
491
492 node = result->terminal_node.node;
493 level = result->terminal_node.level;
494 edit->segment_cache[ASSOC_ARRAY_FAN_OUT] = result->terminal_node.slot;
495
496 pr_devel("-->%s()\n", __func__);
497
498 /* We arrived at a node which doesn't have an onward node or shortcut
499 * pointer that we have to follow. This means that (a) the leaf we
500 * want must go here (either by insertion or replacement) or (b) we
501 * need to split this node and insert in one of the fragments.
502 */
503 free_slot = -1;
504
505 /* Firstly, we have to check the leaves in this node to see if there's
506 * a matching one we should replace in place.
507 */
508 for (i = 0; i < ASSOC_ARRAY_FAN_OUT; i++) {
509 ptr = node->slots[i];
510 if (!ptr) {
511 free_slot = i;
512 continue;
513 }
514 if (assoc_array_ptr_is_leaf(ptr) &&
515 ops->compare_object(assoc_array_ptr_to_leaf(ptr),
516 index_key)) {
517 pr_devel("replace in slot %d\n", i);
518 edit->leaf_p = &node->slots[i];
519 edit->dead_leaf = node->slots[i];
520 pr_devel("<--%s() = ok [replace]\n", __func__);
521 return true;
522 }
523 }
524
525 /* If there is a free slot in this node then we can just insert the
526 * leaf here.
527 */
528 if (free_slot >= 0) {
529 pr_devel("insert in free slot %d\n", free_slot);
530 edit->leaf_p = &node->slots[free_slot];
531 edit->adjust_count_on = node;
532 pr_devel("<--%s() = ok [insert]\n", __func__);
533 return true;
534 }
535
536 /* The node has no spare slots - so we're either going to have to split
537 * it or insert another node before it.
538 *
539 * Whatever, we're going to need at least two new nodes - so allocate
540 * those now. We may also need a new shortcut, but we deal with that
541 * when we need it.
542 */
543 new_n0 = kzalloc(sizeof(struct assoc_array_node), GFP_KERNEL);
544 if (!new_n0)
545 return false;
546 edit->new_meta[0] = assoc_array_node_to_ptr(new_n0);
547 new_n1 = kzalloc(sizeof(struct assoc_array_node), GFP_KERNEL);
548 if (!new_n1)
549 return false;
550 edit->new_meta[1] = assoc_array_node_to_ptr(new_n1);
551
552 /* We need to find out how similar the leaves are. */
553 pr_devel("no spare slots\n");
554 have_meta = false;
555 for (i = 0; i < ASSOC_ARRAY_FAN_OUT; i++) {
556 ptr = node->slots[i];
557 if (assoc_array_ptr_is_meta(ptr)) {
558 edit->segment_cache[i] = 0xff;
559 have_meta = true;
560 continue;
561 }
562 base_seg = ops->get_object_key_chunk(
563 assoc_array_ptr_to_leaf(ptr), level);
564 base_seg >>= level & ASSOC_ARRAY_KEY_CHUNK_MASK;
565 edit->segment_cache[i] = base_seg & ASSOC_ARRAY_FAN_MASK;
566 }
567
568 if (have_meta) {
569 pr_devel("have meta\n");
570 goto split_node;
571 }
572
573 /* The node contains only leaves */
574 dissimilarity = 0;
575 base_seg = edit->segment_cache[0];
576 for (i = 1; i < ASSOC_ARRAY_FAN_OUT; i++)
577 dissimilarity |= edit->segment_cache[i] ^ base_seg;
578
579 pr_devel("only leaves; dissimilarity=%lx\n", dissimilarity);
580
581 if ((dissimilarity & ASSOC_ARRAY_FAN_MASK) == 0) {
582 /* The old leaves all cluster in the same slot. We will need
583 * to insert a shortcut if the new node wants to cluster with them.
584 */
585 if ((edit->segment_cache[ASSOC_ARRAY_FAN_OUT] ^ base_seg) == 0)
586 goto all_leaves_cluster_together;
587
588 /* Otherwise all the old leaves cluster in the same slot, but
589 * the new leaf wants to go into a different slot - so we
590 * create a new node (n0) to hold the new leaf and a pointer to
591 * a new node (n1) holding all the old leaves.
592 *
593 * This can be done by falling through to the node splitting
594 * path.
595 */
596 pr_devel("present leaves cluster but not new leaf\n");
597 }
598
599split_node:
600 pr_devel("split node\n");
601
602 /* We need to split the current node. The node must contain anything
603 * from a single leaf (in the one leaf case, this leaf will cluster
604 * with the new leaf) and the rest meta-pointers, to all leaves, some
605 * of which may cluster.
606 *
607 * It won't contain the case in which all the current leaves plus the
608 * new leaves want to cluster in the same slot.
609 *
610 * We need to expel at least two leaves out of a set consisting of the
611 * leaves in the node and the new leaf. The current meta pointers can
612 * just be copied as they shouldn't cluster with any of the leaves.
613 *
614 * We need a new node (n0) to replace the current one and a new node to
615 * take the expelled nodes (n1).
616 */
617 edit->set[0].to = assoc_array_node_to_ptr(new_n0);
618 new_n0->back_pointer = node->back_pointer;
619 new_n0->parent_slot = node->parent_slot;
620 new_n1->back_pointer = assoc_array_node_to_ptr(new_n0);
621 new_n1->parent_slot = -1; /* Need to calculate this */
622
623do_split_node:
624 pr_devel("do_split_node\n");
625
626 new_n0->nr_leaves_on_branch = node->nr_leaves_on_branch;
627 new_n1->nr_leaves_on_branch = 0;
628
629 /* Begin by finding two matching leaves. There have to be at least two
630 * that match - even if there are meta pointers - because any leaf that
631 * would match a slot with a meta pointer in it must be somewhere
632 * behind that meta pointer and cannot be here. Further, given N
633 * remaining leaf slots, we now have N+1 leaves to go in them.
634 */
635 for (i = 0; i < ASSOC_ARRAY_FAN_OUT; i++) {
636 slot = edit->segment_cache[i];
637 if (slot != 0xff)
638 for (j = i + 1; j < ASSOC_ARRAY_FAN_OUT + 1; j++)
639 if (edit->segment_cache[j] == slot)
640 goto found_slot_for_multiple_occupancy;
641 }
642found_slot_for_multiple_occupancy:
643 pr_devel("same slot: %x %x [%02x]\n", i, j, slot);
644 BUG_ON(i >= ASSOC_ARRAY_FAN_OUT);
645 BUG_ON(j >= ASSOC_ARRAY_FAN_OUT + 1);
646 BUG_ON(slot >= ASSOC_ARRAY_FAN_OUT);
647
648 new_n1->parent_slot = slot;
649
650 /* Metadata pointers cannot change slot */
651 for (i = 0; i < ASSOC_ARRAY_FAN_OUT; i++)
652 if (assoc_array_ptr_is_meta(node->slots[i]))
653 new_n0->slots[i] = node->slots[i];
654 else
655 new_n0->slots[i] = NULL;
656 BUG_ON(new_n0->slots[slot] != NULL);
657 new_n0->slots[slot] = assoc_array_node_to_ptr(new_n1);
658
659 /* Filter the leaf pointers between the new nodes */
660 free_slot = -1;
661 next_slot = 0;
662 for (i = 0; i < ASSOC_ARRAY_FAN_OUT; i++) {
663 if (assoc_array_ptr_is_meta(node->slots[i]))
664 continue;
665 if (edit->segment_cache[i] == slot) {
666 new_n1->slots[next_slot++] = node->slots[i];
667 new_n1->nr_leaves_on_branch++;
668 } else {
669 do {
670 free_slot++;
671 } while (new_n0->slots[free_slot] != NULL);
672 new_n0->slots[free_slot] = node->slots[i];
673 }
674 }
675
676 pr_devel("filtered: f=%x n=%x\n", free_slot, next_slot);
677
678 if (edit->segment_cache[ASSOC_ARRAY_FAN_OUT] != slot) {
679 do {
680 free_slot++;
681 } while (new_n0->slots[free_slot] != NULL);
682 edit->leaf_p = &new_n0->slots[free_slot];
683 edit->adjust_count_on = new_n0;
684 } else {
685 edit->leaf_p = &new_n1->slots[next_slot++];
686 edit->adjust_count_on = new_n1;
687 }
688
689 BUG_ON(next_slot <= 1);
690
691 edit->set_backpointers_to = assoc_array_node_to_ptr(new_n0);
692 for (i = 0; i < ASSOC_ARRAY_FAN_OUT; i++) {
693 if (edit->segment_cache[i] == 0xff) {
694 ptr = node->slots[i];
695 BUG_ON(assoc_array_ptr_is_leaf(ptr));
696 if (assoc_array_ptr_is_node(ptr)) {
697 side = assoc_array_ptr_to_node(ptr);
698 edit->set_backpointers[i] = &side->back_pointer;
699 } else {
700 shortcut = assoc_array_ptr_to_shortcut(ptr);
701 edit->set_backpointers[i] = &shortcut->back_pointer;
702 }
703 }
704 }
705
706 ptr = node->back_pointer;
707 if (!ptr)
708 edit->set[0].ptr = &edit->array->root;
709 else if (assoc_array_ptr_is_node(ptr))
710 edit->set[0].ptr = &assoc_array_ptr_to_node(ptr)->slots[node->parent_slot];
711 else
712 edit->set[0].ptr = &assoc_array_ptr_to_shortcut(ptr)->next_node;
713 edit->excised_meta[0] = assoc_array_node_to_ptr(node);
714 pr_devel("<--%s() = ok [split node]\n", __func__);
715 return true;
716
717all_leaves_cluster_together:
718 /* All the leaves, new and old, want to cluster together in this node
719 * in the same slot, so we have to replace this node with a shortcut to
720 * skip over the identical parts of the key and then place a pair of
721 * nodes, one inside the other, at the end of the shortcut and
722 * distribute the keys between them.
723 *
724 * Firstly we need to work out where the leaves start diverging as a
725 * bit position into their keys so that we know how big the shortcut
726 * needs to be.
727 *
728 * We only need to make a single pass of N of the N+1 leaves because if
729 * any keys differ between themselves at bit X then at least one of
730 * them must also differ with the base key at bit X or before.
731 */
732 pr_devel("all leaves cluster together\n");
733 diff = INT_MAX;
734 for (i = 0; i < ASSOC_ARRAY_FAN_OUT; i++) {
735 int x = ops->diff_objects(assoc_array_ptr_to_leaf(node->slots[i]),
736 index_key);
737 if (x < diff) {
738 BUG_ON(x < 0);
739 diff = x;
740 }
741 }
742 BUG_ON(diff == INT_MAX);
743 BUG_ON(diff < level + ASSOC_ARRAY_LEVEL_STEP);
744
745 keylen = round_up(diff, ASSOC_ARRAY_KEY_CHUNK_SIZE);
746 keylen >>= ASSOC_ARRAY_KEY_CHUNK_SHIFT;
747
748 new_s0 = kzalloc(sizeof(struct assoc_array_shortcut) +
749 keylen * sizeof(unsigned long), GFP_KERNEL);
750 if (!new_s0)
751 return false;
752 edit->new_meta[2] = assoc_array_shortcut_to_ptr(new_s0);
753
754 edit->set[0].to = assoc_array_shortcut_to_ptr(new_s0);
755 new_s0->back_pointer = node->back_pointer;
756 new_s0->parent_slot = node->parent_slot;
757 new_s0->next_node = assoc_array_node_to_ptr(new_n0);
758 new_n0->back_pointer = assoc_array_shortcut_to_ptr(new_s0);
759 new_n0->parent_slot = 0;
760 new_n1->back_pointer = assoc_array_node_to_ptr(new_n0);
761 new_n1->parent_slot = -1; /* Need to calculate this */
762
763 new_s0->skip_to_level = level = diff & ~ASSOC_ARRAY_LEVEL_STEP_MASK;
764 pr_devel("skip_to_level = %d [diff %d]\n", level, diff);
765 BUG_ON(level <= 0);
766
767 for (i = 0; i < keylen; i++)
768 new_s0->index_key[i] =
769 ops->get_key_chunk(index_key, i * ASSOC_ARRAY_KEY_CHUNK_SIZE);
770
771 if (level & ASSOC_ARRAY_KEY_CHUNK_MASK) {
772 blank = ULONG_MAX << (level & ASSOC_ARRAY_KEY_CHUNK_MASK);
773 pr_devel("blank off [%zu] %d: %lx\n", keylen - 1, level, blank);
774 new_s0->index_key[keylen - 1] &= ~blank;
775 }
776
777 /* This now reduces to a node splitting exercise for which we'll need
778 * to regenerate the disparity table.
779 */
780 for (i = 0; i < ASSOC_ARRAY_FAN_OUT; i++) {
781 ptr = node->slots[i];
782 base_seg = ops->get_object_key_chunk(assoc_array_ptr_to_leaf(ptr),
783 level);
784 base_seg >>= level & ASSOC_ARRAY_KEY_CHUNK_MASK;
785 edit->segment_cache[i] = base_seg & ASSOC_ARRAY_FAN_MASK;
786 }
787
788 base_seg = ops->get_key_chunk(index_key, level);
789 base_seg >>= level & ASSOC_ARRAY_KEY_CHUNK_MASK;
790 edit->segment_cache[ASSOC_ARRAY_FAN_OUT] = base_seg & ASSOC_ARRAY_FAN_MASK;
791 goto do_split_node;
792}
793
794/*
795 * Handle insertion into the middle of a shortcut.
796 */
797static bool assoc_array_insert_mid_shortcut(struct assoc_array_edit *edit,
798 const struct assoc_array_ops *ops,
799 struct assoc_array_walk_result *result)
800{
801 struct assoc_array_shortcut *shortcut, *new_s0, *new_s1;
802 struct assoc_array_node *node, *new_n0, *side;
803 unsigned long sc_segments, dissimilarity, blank;
804 size_t keylen;
805 int level, sc_level, diff;
806 int sc_slot;
807
808 shortcut = result->wrong_shortcut.shortcut;
809 level = result->wrong_shortcut.level;
810 sc_level = result->wrong_shortcut.sc_level;
811 sc_segments = result->wrong_shortcut.sc_segments;
812 dissimilarity = result->wrong_shortcut.dissimilarity;
813
814 pr_devel("-->%s(ix=%d dis=%lx scix=%d)\n",
815 __func__, level, dissimilarity, sc_level);
816
817 /* We need to split a shortcut and insert a node between the two
818 * pieces. Zero-length pieces will be dispensed with entirely.
819 *
820 * First of all, we need to find out in which level the first
821 * difference was.
822 */
823 diff = __ffs(dissimilarity);
824 diff &= ~ASSOC_ARRAY_LEVEL_STEP_MASK;
825 diff += sc_level & ~ASSOC_ARRAY_KEY_CHUNK_MASK;
826 pr_devel("diff=%d\n", diff);
827
828 if (!shortcut->back_pointer) {
829 edit->set[0].ptr = &edit->array->root;
830 } else if (assoc_array_ptr_is_node(shortcut->back_pointer)) {
831 node = assoc_array_ptr_to_node(shortcut->back_pointer);
832 edit->set[0].ptr = &node->slots[shortcut->parent_slot];
833 } else {
834 BUG();
835 }
836
837 edit->excised_meta[0] = assoc_array_shortcut_to_ptr(shortcut);
838
839 /* Create a new node now since we're going to need it anyway */
840 new_n0 = kzalloc(sizeof(struct assoc_array_node), GFP_KERNEL);
841 if (!new_n0)
842 return false;
843 edit->new_meta[0] = assoc_array_node_to_ptr(new_n0);
844 edit->adjust_count_on = new_n0;
845
846 /* Insert a new shortcut before the new node if this segment isn't of
847 * zero length - otherwise we just connect the new node directly to the
848 * parent.
849 */
850 level += ASSOC_ARRAY_LEVEL_STEP;
851 if (diff > level) {
852 pr_devel("pre-shortcut %d...%d\n", level, diff);
853 keylen = round_up(diff, ASSOC_ARRAY_KEY_CHUNK_SIZE);
854 keylen >>= ASSOC_ARRAY_KEY_CHUNK_SHIFT;
855
856 new_s0 = kzalloc(sizeof(struct assoc_array_shortcut) +
857 keylen * sizeof(unsigned long), GFP_KERNEL);
858 if (!new_s0)
859 return false;
860 edit->new_meta[1] = assoc_array_shortcut_to_ptr(new_s0);
861 edit->set[0].to = assoc_array_shortcut_to_ptr(new_s0);
862 new_s0->back_pointer = shortcut->back_pointer;
863 new_s0->parent_slot = shortcut->parent_slot;
864 new_s0->next_node = assoc_array_node_to_ptr(new_n0);
865 new_s0->skip_to_level = diff;
866
867 new_n0->back_pointer = assoc_array_shortcut_to_ptr(new_s0);
868 new_n0->parent_slot = 0;
869
870 memcpy(new_s0->index_key, shortcut->index_key,
871 keylen * sizeof(unsigned long));
872
873 blank = ULONG_MAX << (diff & ASSOC_ARRAY_KEY_CHUNK_MASK);
874 pr_devel("blank off [%zu] %d: %lx\n", keylen - 1, diff, blank);
875 new_s0->index_key[keylen - 1] &= ~blank;
876 } else {
877 pr_devel("no pre-shortcut\n");
878 edit->set[0].to = assoc_array_node_to_ptr(new_n0);
879 new_n0->back_pointer = shortcut->back_pointer;
880 new_n0->parent_slot = shortcut->parent_slot;
881 }
882
883 side = assoc_array_ptr_to_node(shortcut->next_node);
884 new_n0->nr_leaves_on_branch = side->nr_leaves_on_branch;
885
886 /* We need to know which slot in the new node is going to take a
887 * metadata pointer.
888 */
889 sc_slot = sc_segments >> (diff & ASSOC_ARRAY_KEY_CHUNK_MASK);
890 sc_slot &= ASSOC_ARRAY_FAN_MASK;
891
892 pr_devel("new slot %lx >> %d -> %d\n",
893 sc_segments, diff & ASSOC_ARRAY_KEY_CHUNK_MASK, sc_slot);
894
895 /* Determine whether we need to follow the new node with a replacement
896 * for the current shortcut. We could in theory reuse the current
897 * shortcut if its parent slot number doesn't change - but that's a
898 * 1-in-16 chance so not worth expending the code upon.
899 */
900 level = diff + ASSOC_ARRAY_LEVEL_STEP;
901 if (level < shortcut->skip_to_level) {
902 pr_devel("post-shortcut %d...%d\n", level, shortcut->skip_to_level);
903 keylen = round_up(shortcut->skip_to_level, ASSOC_ARRAY_KEY_CHUNK_SIZE);
904 keylen >>= ASSOC_ARRAY_KEY_CHUNK_SHIFT;
905
906 new_s1 = kzalloc(sizeof(struct assoc_array_shortcut) +
907 keylen * sizeof(unsigned long), GFP_KERNEL);
908 if (!new_s1)
909 return false;
910 edit->new_meta[2] = assoc_array_shortcut_to_ptr(new_s1);
911
912 new_s1->back_pointer = assoc_array_node_to_ptr(new_n0);
913 new_s1->parent_slot = sc_slot;
914 new_s1->next_node = shortcut->next_node;
915 new_s1->skip_to_level = shortcut->skip_to_level;
916
917 new_n0->slots[sc_slot] = assoc_array_shortcut_to_ptr(new_s1);
918
919 memcpy(new_s1->index_key, shortcut->index_key,
920 keylen * sizeof(unsigned long));
921
922 edit->set[1].ptr = &side->back_pointer;
923 edit->set[1].to = assoc_array_shortcut_to_ptr(new_s1);
924 } else {
925 pr_devel("no post-shortcut\n");
926
927 /* We don't have to replace the pointed-to node as long as we
928 * use memory barriers to make sure the parent slot number is
929 * changed before the back pointer (the parent slot number is
930 * irrelevant to the old parent shortcut).
931 */
932 new_n0->slots[sc_slot] = shortcut->next_node;
933 edit->set_parent_slot[0].p = &side->parent_slot;
934 edit->set_parent_slot[0].to = sc_slot;
935 edit->set[1].ptr = &side->back_pointer;
936 edit->set[1].to = assoc_array_node_to_ptr(new_n0);
937 }
938
939 /* Install the new leaf in a spare slot in the new node. */
940 if (sc_slot == 0)
941 edit->leaf_p = &new_n0->slots[1];
942 else
943 edit->leaf_p = &new_n0->slots[0];
944
945 pr_devel("<--%s() = ok [split shortcut]\n", __func__);
946 return edit;
947}
948
949/**
950 * assoc_array_insert - Script insertion of an object into an associative array
951 * @array: The array to insert into.
952 * @ops: The operations to use.
953 * @index_key: The key to insert at.
954 * @object: The object to insert.
955 *
956 * Precalculate and preallocate a script for the insertion or replacement of an
957 * object in an associative array. This results in an edit script that can
958 * either be applied or cancelled.
959 *
960 * The function returns a pointer to an edit script or -ENOMEM.
961 *
962 * The caller should lock against other modifications and must continue to hold
963 * the lock until assoc_array_apply_edit() has been called.
964 *
965 * Accesses to the tree may take place concurrently with this function,
966 * provided they hold the RCU read lock.
967 */
968struct assoc_array_edit *assoc_array_insert(struct assoc_array *array,
969 const struct assoc_array_ops *ops,
970 const void *index_key,
971 void *object)
972{
973 struct assoc_array_walk_result result;
974 struct assoc_array_edit *edit;
975
976 pr_devel("-->%s()\n", __func__);
977
978 /* The leaf pointer we're given must not have the bottom bit set as we
979 * use those for type-marking the pointer. NULL pointers are also not
980 * allowed as they indicate an empty slot but we have to allow them
981 * here as they can be updated later.
982 */
983 BUG_ON(assoc_array_ptr_is_meta(object));
984
985 edit = kzalloc(sizeof(struct assoc_array_edit), GFP_KERNEL);
986 if (!edit)
987 return ERR_PTR(-ENOMEM);
988 edit->array = array;
989 edit->ops = ops;
990 edit->leaf = assoc_array_leaf_to_ptr(object);
991 edit->adjust_count_by = 1;
992
993 switch (assoc_array_walk(array, ops, index_key, &result)) {
994 case assoc_array_walk_tree_empty:
995 /* Allocate a root node if there isn't one yet */
996 if (!assoc_array_insert_in_empty_tree(edit))
997 goto enomem;
998 return edit;
999
1000 case assoc_array_walk_found_terminal_node:
1001 /* We found a node that doesn't have a node/shortcut pointer in
1002 * the slot corresponding to the index key that we have to
1003 * follow.
1004 */
1005 if (!assoc_array_insert_into_terminal_node(edit, ops, index_key,
1006 &result))
1007 goto enomem;
1008 return edit;
1009
1010 case assoc_array_walk_found_wrong_shortcut:
1011 /* We found a shortcut that didn't match our key in a slot we
1012 * needed to follow.
1013 */
1014 if (!assoc_array_insert_mid_shortcut(edit, ops, &result))
1015 goto enomem;
1016 return edit;
1017 }
1018
1019enomem:
1020 /* Clean up after an out of memory error */
1021 pr_devel("enomem\n");
1022 assoc_array_cancel_edit(edit);
1023 return ERR_PTR(-ENOMEM);
1024}
1025
1026/**
1027 * assoc_array_insert_set_object - Set the new object pointer in an edit script
1028 * @edit: The edit script to modify.
1029 * @object: The object pointer to set.
1030 *
1031 * Change the object to be inserted in an edit script. The object pointed to
1032 * by the old object is not freed. This must be done prior to applying the
1033 * script.
1034 */
1035void assoc_array_insert_set_object(struct assoc_array_edit *edit, void *object)
1036{
1037 BUG_ON(!object);
1038 edit->leaf = assoc_array_leaf_to_ptr(object);
1039}
1040
1041struct assoc_array_delete_collapse_context {
1042 struct assoc_array_node *node;
1043 const void *skip_leaf;
1044 int slot;
1045};
1046
1047/*
1048 * Subtree collapse to node iterator.
1049 */
1050static int assoc_array_delete_collapse_iterator(const void *leaf,
1051 void *iterator_data)
1052{
1053 struct assoc_array_delete_collapse_context *collapse = iterator_data;
1054
1055 if (leaf == collapse->skip_leaf)
1056 return 0;
1057
1058 BUG_ON(collapse->slot >= ASSOC_ARRAY_FAN_OUT);
1059
1060 collapse->node->slots[collapse->slot++] = assoc_array_leaf_to_ptr(leaf);
1061 return 0;
1062}
1063
1064/**
1065 * assoc_array_delete - Script deletion of an object from an associative array
1066 * @array: The array to search.
1067 * @ops: The operations to use.
1068 * @index_key: The key to the object.
1069 *
1070 * Precalculate and preallocate a script for the deletion of an object from an
1071 * associative array. This results in an edit script that can either be
1072 * applied or cancelled.
1073 *
1074 * The function returns a pointer to an edit script if the object was found,
1075 * NULL if the object was not found or -ENOMEM.
1076 *
1077 * The caller should lock against other modifications and must continue to hold
1078 * the lock until assoc_array_apply_edit() has been called.
1079 *
1080 * Accesses to the tree may take place concurrently with this function,
1081 * provided they hold the RCU read lock.
1082 */
1083struct assoc_array_edit *assoc_array_delete(struct assoc_array *array,
1084 const struct assoc_array_ops *ops,
1085 const void *index_key)
1086{
1087 struct assoc_array_delete_collapse_context collapse;
1088 struct assoc_array_walk_result result;
1089 struct assoc_array_node *node, *new_n0;
1090 struct assoc_array_edit *edit;
1091 struct assoc_array_ptr *ptr;
1092 bool has_meta;
1093 int slot, i;
1094
1095 pr_devel("-->%s()\n", __func__);
1096
1097 edit = kzalloc(sizeof(struct assoc_array_edit), GFP_KERNEL);
1098 if (!edit)
1099 return ERR_PTR(-ENOMEM);
1100 edit->array = array;
1101 edit->ops = ops;
1102 edit->adjust_count_by = -1;
1103
1104 switch (assoc_array_walk(array, ops, index_key, &result)) {
1105 case assoc_array_walk_found_terminal_node:
1106 /* We found a node that should contain the leaf we've been
1107 * asked to remove - *if* it's in the tree.
1108 */
1109 pr_devel("terminal_node\n");
1110 node = result.terminal_node.node;
1111
1112 for (slot = 0; slot < ASSOC_ARRAY_FAN_OUT; slot++) {
1113 ptr = node->slots[slot];
1114 if (ptr &&
1115 assoc_array_ptr_is_leaf(ptr) &&
1116 ops->compare_object(assoc_array_ptr_to_leaf(ptr),
1117 index_key))
1118 goto found_leaf;
1119 }
1120 /* fall through */
1121 case assoc_array_walk_tree_empty:
1122 case assoc_array_walk_found_wrong_shortcut:
1123 default:
1124 assoc_array_cancel_edit(edit);
1125 pr_devel("not found\n");
1126 return NULL;
1127 }
1128
1129found_leaf:
1130 BUG_ON(array->nr_leaves_on_tree <= 0);
1131
1132 /* In the simplest form of deletion we just clear the slot and release
1133 * the leaf after a suitable interval.
1134 */
1135 edit->dead_leaf = node->slots[slot];
1136 edit->set[0].ptr = &node->slots[slot];
1137 edit->set[0].to = NULL;
1138 edit->adjust_count_on = node;
1139
1140 /* If that concludes erasure of the last leaf, then delete the entire
1141 * internal array.
1142 */
1143 if (array->nr_leaves_on_tree == 1) {
1144 edit->set[1].ptr = &array->root;
1145 edit->set[1].to = NULL;
1146 edit->adjust_count_on = NULL;
1147 edit->excised_subtree = array->root;
1148 pr_devel("all gone\n");
1149 return edit;
1150 }
1151
1152 /* However, we'd also like to clear up some metadata blocks if we
1153 * possibly can.
1154 *
1155 * We go for a simple algorithm of: if this node has FAN_OUT or fewer
1156 * leaves in it, then attempt to collapse it - and attempt to
1157 * recursively collapse up the tree.
1158 *
1159 * We could also try and collapse in partially filled subtrees to take
1160 * up space in this node.
1161 */
1162 if (node->nr_leaves_on_branch <= ASSOC_ARRAY_FAN_OUT + 1) {
1163 struct assoc_array_node *parent, *grandparent;
1164 struct assoc_array_ptr *ptr;
1165
1166 /* First of all, we need to know if this node has metadata so
1167 * that we don't try collapsing if all the leaves are already
1168 * here.
1169 */
1170 has_meta = false;
1171 for (i = 0; i < ASSOC_ARRAY_FAN_OUT; i++) {
1172 ptr = node->slots[i];
1173 if (assoc_array_ptr_is_meta(ptr)) {
1174 has_meta = true;
1175 break;
1176 }
1177 }
1178
1179 pr_devel("leaves: %ld [m=%d]\n",
1180 node->nr_leaves_on_branch - 1, has_meta);
1181
1182 /* Look further up the tree to see if we can collapse this node
1183 * into a more proximal node too.
1184 */
1185 parent = node;
1186 collapse_up:
1187 pr_devel("collapse subtree: %ld\n", parent->nr_leaves_on_branch);
1188
1189 ptr = parent->back_pointer;
1190 if (!ptr)
1191 goto do_collapse;
1192 if (assoc_array_ptr_is_shortcut(ptr)) {
1193 struct assoc_array_shortcut *s = assoc_array_ptr_to_shortcut(ptr);
1194 ptr = s->back_pointer;
1195 if (!ptr)
1196 goto do_collapse;
1197 }
1198
1199 grandparent = assoc_array_ptr_to_node(ptr);
1200 if (grandparent->nr_leaves_on_branch <= ASSOC_ARRAY_FAN_OUT + 1) {
1201 parent = grandparent;
1202 goto collapse_up;
1203 }
1204
1205 do_collapse:
1206 /* There's no point collapsing if the original node has no meta
1207 * pointers to discard and if we didn't merge into one of that
1208 * node's ancestry.
1209 */
1210 if (has_meta || parent != node) {
1211 node = parent;
1212
1213 /* Create a new node to collapse into */
1214 new_n0 = kzalloc(sizeof(struct assoc_array_node), GFP_KERNEL);
1215 if (!new_n0)
1216 goto enomem;
1217 edit->new_meta[0] = assoc_array_node_to_ptr(new_n0);
1218
1219 new_n0->back_pointer = node->back_pointer;
1220 new_n0->parent_slot = node->parent_slot;
1221 new_n0->nr_leaves_on_branch = node->nr_leaves_on_branch;
1222 edit->adjust_count_on = new_n0;
1223
1224 collapse.node = new_n0;
1225 collapse.skip_leaf = assoc_array_ptr_to_leaf(edit->dead_leaf);
1226 collapse.slot = 0;
1227 assoc_array_subtree_iterate(assoc_array_node_to_ptr(node),
1228 node->back_pointer,
1229 assoc_array_delete_collapse_iterator,
1230 &collapse);
1231 pr_devel("collapsed %d,%lu\n", collapse.slot, new_n0->nr_leaves_on_branch);
1232 BUG_ON(collapse.slot != new_n0->nr_leaves_on_branch - 1);
1233
1234 if (!node->back_pointer) {
1235 edit->set[1].ptr = &array->root;
1236 } else if (assoc_array_ptr_is_leaf(node->back_pointer)) {
1237 BUG();
1238 } else if (assoc_array_ptr_is_node(node->back_pointer)) {
1239 struct assoc_array_node *p =
1240 assoc_array_ptr_to_node(node->back_pointer);
1241 edit->set[1].ptr = &p->slots[node->parent_slot];
1242 } else if (assoc_array_ptr_is_shortcut(node->back_pointer)) {
1243 struct assoc_array_shortcut *s =
1244 assoc_array_ptr_to_shortcut(node->back_pointer);
1245 edit->set[1].ptr = &s->next_node;
1246 }
1247 edit->set[1].to = assoc_array_node_to_ptr(new_n0);
1248 edit->excised_subtree = assoc_array_node_to_ptr(node);
1249 }
1250 }
1251
1252 return edit;
1253
1254enomem:
1255 /* Clean up after an out of memory error */
1256 pr_devel("enomem\n");
1257 assoc_array_cancel_edit(edit);
1258 return ERR_PTR(-ENOMEM);
1259}
1260
1261/**
1262 * assoc_array_clear - Script deletion of all objects from an associative array
1263 * @array: The array to clear.
1264 * @ops: The operations to use.
1265 *
1266 * Precalculate and preallocate a script for the deletion of all the objects
1267 * from an associative array. This results in an edit script that can either
1268 * be applied or cancelled.
1269 *
1270 * The function returns a pointer to an edit script if there are objects to be
1271 * deleted, NULL if there are no objects in the array or -ENOMEM.
1272 *
1273 * The caller should lock against other modifications and must continue to hold
1274 * the lock until assoc_array_apply_edit() has been called.
1275 *
1276 * Accesses to the tree may take place concurrently with this function,
1277 * provided they hold the RCU read lock.
1278 */
1279struct assoc_array_edit *assoc_array_clear(struct assoc_array *array,
1280 const struct assoc_array_ops *ops)
1281{
1282 struct assoc_array_edit *edit;
1283
1284 pr_devel("-->%s()\n", __func__);
1285
1286 if (!array->root)
1287 return NULL;
1288
1289 edit = kzalloc(sizeof(struct assoc_array_edit), GFP_KERNEL);
1290 if (!edit)
1291 return ERR_PTR(-ENOMEM);
1292 edit->array = array;
1293 edit->ops = ops;
1294 edit->set[1].ptr = &array->root;
1295 edit->set[1].to = NULL;
1296 edit->excised_subtree = array->root;
1297 edit->ops_for_excised_subtree = ops;
1298 pr_devel("all gone\n");
1299 return edit;
1300}
1301
1302/*
1303 * Handle the deferred destruction after an applied edit.
1304 */
1305static void assoc_array_rcu_cleanup(struct rcu_head *head)
1306{
1307 struct assoc_array_edit *edit =
1308 container_of(head, struct assoc_array_edit, rcu);
1309 int i;
1310
1311 pr_devel("-->%s()\n", __func__);
1312
1313 if (edit->dead_leaf)
1314 edit->ops->free_object(assoc_array_ptr_to_leaf(edit->dead_leaf));
1315 for (i = 0; i < ARRAY_SIZE(edit->excised_meta); i++)
1316 if (edit->excised_meta[i])
1317 kfree(assoc_array_ptr_to_node(edit->excised_meta[i]));
1318
1319 if (edit->excised_subtree) {
1320 BUG_ON(assoc_array_ptr_is_leaf(edit->excised_subtree));
1321 if (assoc_array_ptr_is_node(edit->excised_subtree)) {
1322 struct assoc_array_node *n =
1323 assoc_array_ptr_to_node(edit->excised_subtree);
1324 n->back_pointer = NULL;
1325 } else {
1326 struct assoc_array_shortcut *s =
1327 assoc_array_ptr_to_shortcut(edit->excised_subtree);
1328 s->back_pointer = NULL;
1329 }
1330 assoc_array_destroy_subtree(edit->excised_subtree,
1331 edit->ops_for_excised_subtree);
1332 }
1333
1334 kfree(edit);
1335}
1336
1337/**
1338 * assoc_array_apply_edit - Apply an edit script to an associative array
1339 * @edit: The script to apply.
1340 *
1341 * Apply an edit script to an associative array to effect an insertion,
1342 * deletion or clearance. As the edit script includes preallocated memory,
1343 * this is guaranteed not to fail.
1344 *
1345 * The edit script, dead objects and dead metadata will be scheduled for
1346 * destruction after an RCU grace period to permit those doing read-only
1347 * accesses on the array to continue to do so under the RCU read lock whilst
1348 * the edit is taking place.
1349 */
1350void assoc_array_apply_edit(struct assoc_array_edit *edit)
1351{
1352 struct assoc_array_shortcut *shortcut;
1353 struct assoc_array_node *node;
1354 struct assoc_array_ptr *ptr;
1355 int i;
1356
1357 pr_devel("-->%s()\n", __func__);
1358
1359 smp_wmb();
1360 if (edit->leaf_p)
1361 *edit->leaf_p = edit->leaf;
1362
1363 smp_wmb();
1364 for (i = 0; i < ARRAY_SIZE(edit->set_parent_slot); i++)
1365 if (edit->set_parent_slot[i].p)
1366 *edit->set_parent_slot[i].p = edit->set_parent_slot[i].to;
1367
1368 smp_wmb();
1369 for (i = 0; i < ARRAY_SIZE(edit->set_backpointers); i++)
1370 if (edit->set_backpointers[i])
1371 *edit->set_backpointers[i] = edit->set_backpointers_to;
1372
1373 smp_wmb();
1374 for (i = 0; i < ARRAY_SIZE(edit->set); i++)
1375 if (edit->set[i].ptr)
1376 *edit->set[i].ptr = edit->set[i].to;
1377
1378 if (edit->array->root == NULL) {
1379 edit->array->nr_leaves_on_tree = 0;
1380 } else if (edit->adjust_count_on) {
1381 node = edit->adjust_count_on;
1382 for (;;) {
1383 node->nr_leaves_on_branch += edit->adjust_count_by;
1384
1385 ptr = node->back_pointer;
1386 if (!ptr)
1387 break;
1388 if (assoc_array_ptr_is_shortcut(ptr)) {
1389 shortcut = assoc_array_ptr_to_shortcut(ptr);
1390 ptr = shortcut->back_pointer;
1391 if (!ptr)
1392 break;
1393 }
1394 BUG_ON(!assoc_array_ptr_is_node(ptr));
1395 node = assoc_array_ptr_to_node(ptr);
1396 }
1397
1398 edit->array->nr_leaves_on_tree += edit->adjust_count_by;
1399 }
1400
1401 call_rcu(&edit->rcu, assoc_array_rcu_cleanup);
1402}
1403
1404/**
1405 * assoc_array_cancel_edit - Discard an edit script.
1406 * @edit: The script to discard.
1407 *
1408 * Free an edit script and all the preallocated data it holds without making
1409 * any changes to the associative array it was intended for.
1410 *
1411 * NOTE! In the case of an insertion script, this does _not_ release the leaf
1412 * that was to be inserted. That is left to the caller.
1413 */
1414void assoc_array_cancel_edit(struct assoc_array_edit *edit)
1415{
1416 struct assoc_array_ptr *ptr;
1417 int i;
1418
1419 pr_devel("-->%s()\n", __func__);
1420
1421 /* Clean up after an out of memory error */
1422 for (i = 0; i < ARRAY_SIZE(edit->new_meta); i++) {
1423 ptr = edit->new_meta[i];
1424 if (ptr) {
1425 if (assoc_array_ptr_is_node(ptr))
1426 kfree(assoc_array_ptr_to_node(ptr));
1427 else
1428 kfree(assoc_array_ptr_to_shortcut(ptr));
1429 }
1430 }
1431 kfree(edit);
1432}
1433
1434/**
1435 * assoc_array_gc - Garbage collect an associative array.
1436 * @array: The array to clean.
1437 * @ops: The operations to use.
1438 * @iterator: A callback function to pass judgement on each object.
1439 * @iterator_data: Private data for the callback function.
1440 *
1441 * Collect garbage from an associative array and pack down the internal tree to
1442 * save memory.
1443 *
1444 * The iterator function is asked to pass judgement upon each object in the
1445 * array. If it returns false, the object is discard and if it returns true,
1446 * the object is kept. If it returns true, it must increment the object's
1447 * usage count (or whatever it needs to do to retain it) before returning.
1448 *
1449 * This function returns 0 if successful or -ENOMEM if out of memory. In the
1450 * latter case, the array is not changed.
1451 *
1452 * The caller should lock against other modifications and must continue to hold
1453 * the lock until assoc_array_apply_edit() has been called.
1454 *
1455 * Accesses to the tree may take place concurrently with this function,
1456 * provided they hold the RCU read lock.
1457 */
1458int assoc_array_gc(struct assoc_array *array,
1459 const struct assoc_array_ops *ops,
1460 bool (*iterator)(void *object, void *iterator_data),
1461 void *iterator_data)
1462{
1463 struct assoc_array_shortcut *shortcut, *new_s;
1464 struct assoc_array_node *node, *new_n;
1465 struct assoc_array_edit *edit;
1466 struct assoc_array_ptr *cursor, *ptr;
1467 struct assoc_array_ptr *new_root, *new_parent, **new_ptr_pp;
1468 unsigned long nr_leaves_on_tree;
1469 int keylen, slot, nr_free, next_slot, i;
1470
1471 pr_devel("-->%s()\n", __func__);
1472
1473 if (!array->root)
1474 return 0;
1475
1476 edit = kzalloc(sizeof(struct assoc_array_edit), GFP_KERNEL);
1477 if (!edit)
1478 return -ENOMEM;
1479 edit->array = array;
1480 edit->ops = ops;
1481 edit->ops_for_excised_subtree = ops;
1482 edit->set[0].ptr = &array->root;
1483 edit->excised_subtree = array->root;
1484
1485 new_root = new_parent = NULL;
1486 new_ptr_pp = &new_root;
1487 cursor = array->root;
1488
1489descend:
1490 /* If this point is a shortcut, then we need to duplicate it and
1491 * advance the target cursor.
1492 */
1493 if (assoc_array_ptr_is_shortcut(cursor)) {
1494 shortcut = assoc_array_ptr_to_shortcut(cursor);
1495 keylen = round_up(shortcut->skip_to_level, ASSOC_ARRAY_KEY_CHUNK_SIZE);
1496 keylen >>= ASSOC_ARRAY_KEY_CHUNK_SHIFT;
1497 new_s = kmalloc(sizeof(struct assoc_array_shortcut) +
1498 keylen * sizeof(unsigned long), GFP_KERNEL);
1499 if (!new_s)
1500 goto enomem;
1501 pr_devel("dup shortcut %p -> %p\n", shortcut, new_s);
1502 memcpy(new_s, shortcut, (sizeof(struct assoc_array_shortcut) +
1503 keylen * sizeof(unsigned long)));
1504 new_s->back_pointer = new_parent;
1505 new_s->parent_slot = shortcut->parent_slot;
1506 *new_ptr_pp = new_parent = assoc_array_shortcut_to_ptr(new_s);
1507 new_ptr_pp = &new_s->next_node;
1508 cursor = shortcut->next_node;
1509 }
1510
1511 /* Duplicate the node at this position */
1512 node = assoc_array_ptr_to_node(cursor);
1513 new_n = kzalloc(sizeof(struct assoc_array_node), GFP_KERNEL);
1514 if (!new_n)
1515 goto enomem;
1516 pr_devel("dup node %p -> %p\n", node, new_n);
1517 new_n->back_pointer = new_parent;
1518 new_n->parent_slot = node->parent_slot;
1519 *new_ptr_pp = new_parent = assoc_array_node_to_ptr(new_n);
1520 new_ptr_pp = NULL;
1521 slot = 0;
1522
1523continue_node:
1524 /* Filter across any leaves and gc any subtrees */
1525 for (; slot < ASSOC_ARRAY_FAN_OUT; slot++) {
1526 ptr = node->slots[slot];
1527 if (!ptr)
1528 continue;
1529
1530 if (assoc_array_ptr_is_leaf(ptr)) {
1531 if (iterator(assoc_array_ptr_to_leaf(ptr),
1532 iterator_data))
1533 /* The iterator will have done any reference
1534 * counting on the object for us.
1535 */
1536 new_n->slots[slot] = ptr;
1537 continue;
1538 }
1539
1540 new_ptr_pp = &new_n->slots[slot];
1541 cursor = ptr;
1542 goto descend;
1543 }
1544
1545 pr_devel("-- compress node %p --\n", new_n);
1546
1547 /* Count up the number of empty slots in this node and work out the
1548 * subtree leaf count.
1549 */
1550 new_n->nr_leaves_on_branch = 0;
1551 nr_free = 0;
1552 for (slot = 0; slot < ASSOC_ARRAY_FAN_OUT; slot++) {
1553 ptr = new_n->slots[slot];
1554 if (!ptr)
1555 nr_free++;
1556 else if (assoc_array_ptr_is_leaf(ptr))
1557 new_n->nr_leaves_on_branch++;
1558 }
1559 pr_devel("free=%d, leaves=%lu\n", nr_free, new_n->nr_leaves_on_branch);
1560
1561 /* See what we can fold in */
1562 next_slot = 0;
1563 for (slot = 0; slot < ASSOC_ARRAY_FAN_OUT; slot++) {
1564 struct assoc_array_shortcut *s;
1565 struct assoc_array_node *child;
1566
1567 ptr = new_n->slots[slot];
1568 if (!ptr || assoc_array_ptr_is_leaf(ptr))
1569 continue;
1570
1571 s = NULL;
1572 if (assoc_array_ptr_is_shortcut(ptr)) {
1573 s = assoc_array_ptr_to_shortcut(ptr);
1574 ptr = s->next_node;
1575 }
1576
1577 child = assoc_array_ptr_to_node(ptr);
1578 new_n->nr_leaves_on_branch += child->nr_leaves_on_branch;
1579
1580 if (child->nr_leaves_on_branch <= nr_free + 1) {
1581 /* Fold the child node into this one */
1582 pr_devel("[%d] fold node %lu/%d [nx %d]\n",
1583 slot, child->nr_leaves_on_branch, nr_free + 1,
1584 next_slot);
1585
1586 /* We would already have reaped an intervening shortcut
1587 * on the way back up the tree.
1588 */
1589 BUG_ON(s);
1590
1591 new_n->slots[slot] = NULL;
1592 nr_free++;
1593 if (slot < next_slot)
1594 next_slot = slot;
1595 for (i = 0; i < ASSOC_ARRAY_FAN_OUT; i++) {
1596 struct assoc_array_ptr *p = child->slots[i];
1597 if (!p)
1598 continue;
1599 BUG_ON(assoc_array_ptr_is_meta(p));
1600 while (new_n->slots[next_slot])
1601 next_slot++;
1602 BUG_ON(next_slot >= ASSOC_ARRAY_FAN_OUT);
1603 new_n->slots[next_slot++] = p;
1604 nr_free--;
1605 }
1606 kfree(child);
1607 } else {
1608 pr_devel("[%d] retain node %lu/%d [nx %d]\n",
1609 slot, child->nr_leaves_on_branch, nr_free + 1,
1610 next_slot);
1611 }
1612 }
1613
1614 pr_devel("after: %lu\n", new_n->nr_leaves_on_branch);
1615
1616 nr_leaves_on_tree = new_n->nr_leaves_on_branch;
1617
1618 /* Excise this node if it is singly occupied by a shortcut */
1619 if (nr_free == ASSOC_ARRAY_FAN_OUT - 1) {
1620 for (slot = 0; slot < ASSOC_ARRAY_FAN_OUT; slot++)
1621 if ((ptr = new_n->slots[slot]))
1622 break;
1623
1624 if (assoc_array_ptr_is_meta(ptr) &&
1625 assoc_array_ptr_is_shortcut(ptr)) {
1626 pr_devel("excise node %p with 1 shortcut\n", new_n);
1627 new_s = assoc_array_ptr_to_shortcut(ptr);
1628 new_parent = new_n->back_pointer;
1629 slot = new_n->parent_slot;
1630 kfree(new_n);
1631 if (!new_parent) {
1632 new_s->back_pointer = NULL;
1633 new_s->parent_slot = 0;
1634 new_root = ptr;
1635 goto gc_complete;
1636 }
1637
1638 if (assoc_array_ptr_is_shortcut(new_parent)) {
1639 /* We can discard any preceding shortcut also */
1640 struct assoc_array_shortcut *s =
1641 assoc_array_ptr_to_shortcut(new_parent);
1642
1643 pr_devel("excise preceding shortcut\n");
1644
1645 new_parent = new_s->back_pointer = s->back_pointer;
1646 slot = new_s->parent_slot = s->parent_slot;
1647 kfree(s);
1648 if (!new_parent) {
1649 new_s->back_pointer = NULL;
1650 new_s->parent_slot = 0;
1651 new_root = ptr;
1652 goto gc_complete;
1653 }
1654 }
1655
1656 new_s->back_pointer = new_parent;
1657 new_s->parent_slot = slot;
1658 new_n = assoc_array_ptr_to_node(new_parent);
1659 new_n->slots[slot] = ptr;
1660 goto ascend_old_tree;
1661 }
1662 }
1663
1664 /* Excise any shortcuts we might encounter that point to nodes that
1665 * only contain leaves.
1666 */
1667 ptr = new_n->back_pointer;
1668 if (!ptr)
1669 goto gc_complete;
1670
1671 if (assoc_array_ptr_is_shortcut(ptr)) {
1672 new_s = assoc_array_ptr_to_shortcut(ptr);
1673 new_parent = new_s->back_pointer;
1674 slot = new_s->parent_slot;
1675
1676 if (new_n->nr_leaves_on_branch <= ASSOC_ARRAY_FAN_OUT) {
1677 struct assoc_array_node *n;
1678
1679 pr_devel("excise shortcut\n");
1680 new_n->back_pointer = new_parent;
1681 new_n->parent_slot = slot;
1682 kfree(new_s);
1683 if (!new_parent) {
1684 new_root = assoc_array_node_to_ptr(new_n);
1685 goto gc_complete;
1686 }
1687
1688 n = assoc_array_ptr_to_node(new_parent);
1689 n->slots[slot] = assoc_array_node_to_ptr(new_n);
1690 }
1691 } else {
1692 new_parent = ptr;
1693 }
1694 new_n = assoc_array_ptr_to_node(new_parent);
1695
1696ascend_old_tree:
1697 ptr = node->back_pointer;
1698 if (assoc_array_ptr_is_shortcut(ptr)) {
1699 shortcut = assoc_array_ptr_to_shortcut(ptr);
1700 slot = shortcut->parent_slot;
1701 cursor = shortcut->back_pointer;
1702 if (!cursor)
1703 goto gc_complete;
1704 } else {
1705 slot = node->parent_slot;
1706 cursor = ptr;
1707 }
1708 BUG_ON(!cursor);
1709 node = assoc_array_ptr_to_node(cursor);
1710 slot++;
1711 goto continue_node;
1712
1713gc_complete:
1714 edit->set[0].to = new_root;
1715 assoc_array_apply_edit(edit);
1716 array->nr_leaves_on_tree = nr_leaves_on_tree;
1717 return 0;
1718
1719enomem:
1720 pr_devel("enomem\n");
1721 assoc_array_destroy_subtree(new_root, edit->ops);
1722 kfree(edit);
1723 return -ENOMEM;
1724}
1725