blob: 3b46c5433b7acd853e5f9ca9381f6f6b30db6884 [file] [log] [blame]
David Howells3cb98952013-09-24 10:35:17 +01001/* Generic associative array implementation.
2 *
3 * See Documentation/assoc_array.txt 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
Pranith Kumar990428b2014-12-30 00:46:21 -050014#include <linux/rcupdate.h>
David Howells3cb98952013-09-24 10:35:17 +010015#include <linux/slab.h>
David Howellsb2a4df22013-09-24 10:35:18 +010016#include <linux/err.h>
David Howells3cb98952013-09-24 10:35:17 +010017#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 smp_read_barrier_depends();
42 cursor = ACCESS_ONCE(shortcut->next_node);
43 }
44
45 node = assoc_array_ptr_to_node(cursor);
46 smp_read_barrier_depends();
47 slot = 0;
48
49 /* We perform two passes of each node.
50 *
51 * The first pass does all the leaves in this node. This means we
52 * don't miss any leaves if the node is split up by insertion whilst
53 * we're iterating over the branches rooted here (we may, however, see
54 * some leaves twice).
55 */
56 has_meta = 0;
57 for (; slot < ASSOC_ARRAY_FAN_OUT; slot++) {
58 ptr = ACCESS_ONCE(node->slots[slot]);
59 has_meta |= (unsigned long)ptr;
60 if (ptr && assoc_array_ptr_is_leaf(ptr)) {
61 /* We need a barrier between the read of the pointer
62 * and dereferencing the pointer - but only if we are
63 * actually going to dereference it.
64 */
65 smp_read_barrier_depends();
66
67 /* Invoke the callback */
68 ret = iterator(assoc_array_ptr_to_leaf(ptr),
69 iterator_data);
70 if (ret)
71 return ret;
72 }
73 }
74
75 /* The second pass attends to all the metadata pointers. If we follow
76 * one of these we may find that we don't come back here, but rather go
77 * back to a replacement node with the leaves in a different layout.
78 *
79 * We are guaranteed to make progress, however, as the slot number for
80 * a particular portion of the key space cannot change - and we
81 * continue at the back pointer + 1.
82 */
83 if (!(has_meta & ASSOC_ARRAY_PTR_META_TYPE))
84 goto finished_node;
85 slot = 0;
86
87continue_node:
88 node = assoc_array_ptr_to_node(cursor);
89 smp_read_barrier_depends();
90
91 for (; slot < ASSOC_ARRAY_FAN_OUT; slot++) {
92 ptr = ACCESS_ONCE(node->slots[slot]);
93 if (assoc_array_ptr_is_meta(ptr)) {
94 cursor = ptr;
95 goto begin_node;
96 }
97 }
98
99finished_node:
100 /* Move up to the parent (may need to skip back over a shortcut) */
101 parent = ACCESS_ONCE(node->back_pointer);
102 slot = node->parent_slot;
103 if (parent == stop)
104 return 0;
105
106 if (assoc_array_ptr_is_shortcut(parent)) {
107 shortcut = assoc_array_ptr_to_shortcut(parent);
108 smp_read_barrier_depends();
109 cursor = parent;
110 parent = ACCESS_ONCE(shortcut->back_pointer);
111 slot = shortcut->parent_slot;
112 if (parent == stop)
113 return 0;
114 }
115
116 /* Ascend to next slot in parent node */
117 cursor = parent;
118 slot++;
119 goto continue_node;
120}
121
122/**
123 * assoc_array_iterate - Pass all objects in the array to a callback
124 * @array: The array to iterate over.
125 * @iterator: The callback function.
126 * @iterator_data: Private data for the callback function.
127 *
128 * Iterate over all the objects in an associative array. Each one will be
129 * presented to the iterator function.
130 *
131 * If the array is being modified concurrently with the iteration then it is
132 * possible that some objects in the array will be passed to the iterator
133 * callback more than once - though every object should be passed at least
134 * once. If this is undesirable then the caller must lock against modification
135 * for the duration of this function.
136 *
137 * The function will return 0 if no objects were in the array or else it will
138 * return the result of the last iterator function called. Iteration stops
139 * immediately if any call to the iteration function results in a non-zero
140 * return.
141 *
142 * The caller should hold the RCU read lock or better if concurrent
143 * modification is possible.
144 */
145int assoc_array_iterate(const struct assoc_array *array,
146 int (*iterator)(const void *object,
147 void *iterator_data),
148 void *iterator_data)
149{
150 struct assoc_array_ptr *root = ACCESS_ONCE(array->root);
151
152 if (!root)
153 return 0;
154 return assoc_array_subtree_iterate(root, NULL, iterator, iterator_data);
155}
156
157enum assoc_array_walk_status {
158 assoc_array_walk_tree_empty,
159 assoc_array_walk_found_terminal_node,
160 assoc_array_walk_found_wrong_shortcut,
Stephen Hemminger30b02c42014-01-23 13:24:09 +0000161};
David Howells3cb98952013-09-24 10:35:17 +0100162
163struct assoc_array_walk_result {
164 struct {
165 struct assoc_array_node *node; /* Node in which leaf might be found */
166 int level;
167 int slot;
168 } terminal_node;
169 struct {
170 struct assoc_array_shortcut *shortcut;
171 int level;
172 int sc_level;
173 unsigned long sc_segments;
174 unsigned long dissimilarity;
175 } wrong_shortcut;
176};
177
178/*
179 * Navigate through the internal tree looking for the closest node to the key.
180 */
181static enum assoc_array_walk_status
182assoc_array_walk(const struct assoc_array *array,
183 const struct assoc_array_ops *ops,
184 const void *index_key,
185 struct assoc_array_walk_result *result)
186{
187 struct assoc_array_shortcut *shortcut;
188 struct assoc_array_node *node;
189 struct assoc_array_ptr *cursor, *ptr;
190 unsigned long sc_segments, dissimilarity;
191 unsigned long segments;
192 int level, sc_level, next_sc_level;
193 int slot;
194
195 pr_devel("-->%s()\n", __func__);
196
197 cursor = ACCESS_ONCE(array->root);
198 if (!cursor)
199 return assoc_array_walk_tree_empty;
200
201 level = 0;
202
203 /* Use segments from the key for the new leaf to navigate through the
204 * internal tree, skipping through nodes and shortcuts that are on
205 * route to the destination. Eventually we'll come to a slot that is
206 * either empty or contains a leaf at which point we've found a node in
207 * which the leaf we're looking for might be found or into which it
208 * should be inserted.
209 */
210jumped:
211 segments = ops->get_key_chunk(index_key, level);
212 pr_devel("segments[%d]: %lx\n", level, segments);
213
214 if (assoc_array_ptr_is_shortcut(cursor))
215 goto follow_shortcut;
216
217consider_node:
218 node = assoc_array_ptr_to_node(cursor);
219 smp_read_barrier_depends();
220
221 slot = segments >> (level & ASSOC_ARRAY_KEY_CHUNK_MASK);
222 slot &= ASSOC_ARRAY_FAN_MASK;
223 ptr = ACCESS_ONCE(node->slots[slot]);
224
225 pr_devel("consider slot %x [ix=%d type=%lu]\n",
226 slot, level, (unsigned long)ptr & 3);
227
228 if (!assoc_array_ptr_is_meta(ptr)) {
229 /* The node doesn't have a node/shortcut pointer in the slot
230 * corresponding to the index key that we have to follow.
231 */
232 result->terminal_node.node = node;
233 result->terminal_node.level = level;
234 result->terminal_node.slot = slot;
235 pr_devel("<--%s() = terminal_node\n", __func__);
236 return assoc_array_walk_found_terminal_node;
237 }
238
239 if (assoc_array_ptr_is_node(ptr)) {
240 /* There is a pointer to a node in the slot corresponding to
241 * this index key segment, so we need to follow it.
242 */
243 cursor = ptr;
244 level += ASSOC_ARRAY_LEVEL_STEP;
245 if ((level & ASSOC_ARRAY_KEY_CHUNK_MASK) != 0)
246 goto consider_node;
247 goto jumped;
248 }
249
250 /* There is a shortcut in the slot corresponding to the index key
251 * segment. We follow the shortcut if its partial index key matches
252 * this leaf's. Otherwise we need to split the shortcut.
253 */
254 cursor = ptr;
255follow_shortcut:
256 shortcut = assoc_array_ptr_to_shortcut(cursor);
257 smp_read_barrier_depends();
258 pr_devel("shortcut to %d\n", shortcut->skip_to_level);
259 sc_level = level + ASSOC_ARRAY_LEVEL_STEP;
260 BUG_ON(sc_level > shortcut->skip_to_level);
261
262 do {
263 /* Check the leaf against the shortcut's index key a word at a
264 * time, trimming the final word (the shortcut stores the index
265 * key completely from the root to the shortcut's target).
266 */
267 if ((sc_level & ASSOC_ARRAY_KEY_CHUNK_MASK) == 0)
268 segments = ops->get_key_chunk(index_key, sc_level);
269
270 sc_segments = shortcut->index_key[sc_level >> ASSOC_ARRAY_KEY_CHUNK_SHIFT];
271 dissimilarity = segments ^ sc_segments;
272
273 if (round_up(sc_level, ASSOC_ARRAY_KEY_CHUNK_SIZE) > shortcut->skip_to_level) {
274 /* Trim segments that are beyond the shortcut */
275 int shift = shortcut->skip_to_level & ASSOC_ARRAY_KEY_CHUNK_MASK;
276 dissimilarity &= ~(ULONG_MAX << shift);
277 next_sc_level = shortcut->skip_to_level;
278 } else {
279 next_sc_level = sc_level + ASSOC_ARRAY_KEY_CHUNK_SIZE;
280 next_sc_level = round_down(next_sc_level, ASSOC_ARRAY_KEY_CHUNK_SIZE);
281 }
282
283 if (dissimilarity != 0) {
284 /* This shortcut points elsewhere */
285 result->wrong_shortcut.shortcut = shortcut;
286 result->wrong_shortcut.level = level;
287 result->wrong_shortcut.sc_level = sc_level;
288 result->wrong_shortcut.sc_segments = sc_segments;
289 result->wrong_shortcut.dissimilarity = dissimilarity;
290 return assoc_array_walk_found_wrong_shortcut;
291 }
292
293 sc_level = next_sc_level;
294 } while (sc_level < shortcut->skip_to_level);
295
296 /* The shortcut matches the leaf's index to this point. */
297 cursor = ACCESS_ONCE(shortcut->next_node);
298 if (((level ^ sc_level) & ~ASSOC_ARRAY_KEY_CHUNK_MASK) != 0) {
299 level = sc_level;
300 goto jumped;
301 } else {
302 level = sc_level;
303 goto consider_node;
304 }
305}
306
307/**
308 * assoc_array_find - Find an object by index key
309 * @array: The associative array to search.
310 * @ops: The operations to use.
311 * @index_key: The key to the object.
312 *
313 * Find an object in an associative array by walking through the internal tree
314 * to the node that should contain the object and then searching the leaves
315 * there. NULL is returned if the requested object was not found in the array.
316 *
317 * The caller must hold the RCU read lock or better.
318 */
319void *assoc_array_find(const struct assoc_array *array,
320 const struct assoc_array_ops *ops,
321 const void *index_key)
322{
323 struct assoc_array_walk_result result;
324 const struct assoc_array_node *node;
325 const struct assoc_array_ptr *ptr;
326 const void *leaf;
327 int slot;
328
329 if (assoc_array_walk(array, ops, index_key, &result) !=
330 assoc_array_walk_found_terminal_node)
331 return NULL;
332
333 node = result.terminal_node.node;
334 smp_read_barrier_depends();
335
336 /* If the target key is available to us, it's has to be pointed to by
337 * the terminal node.
338 */
339 for (slot = 0; slot < ASSOC_ARRAY_FAN_OUT; slot++) {
340 ptr = ACCESS_ONCE(node->slots[slot]);
341 if (ptr && assoc_array_ptr_is_leaf(ptr)) {
342 /* We need a barrier between the read of the pointer
343 * and dereferencing the pointer - but only if we are
344 * actually going to dereference it.
345 */
346 leaf = assoc_array_ptr_to_leaf(ptr);
347 smp_read_barrier_depends();
348 if (ops->compare_object(leaf, index_key))
349 return (void *)leaf;
350 }
351 }
352
353 return NULL;
354}
355
356/*
357 * Destructively iterate over an associative array. The caller must prevent
358 * other simultaneous accesses.
359 */
360static void assoc_array_destroy_subtree(struct assoc_array_ptr *root,
361 const struct assoc_array_ops *ops)
362{
363 struct assoc_array_shortcut *shortcut;
364 struct assoc_array_node *node;
365 struct assoc_array_ptr *cursor, *parent = NULL;
366 int slot = -1;
367
368 pr_devel("-->%s()\n", __func__);
369
370 cursor = root;
371 if (!cursor) {
372 pr_devel("empty\n");
373 return;
374 }
375
376move_to_meta:
377 if (assoc_array_ptr_is_shortcut(cursor)) {
378 /* Descend through a shortcut */
379 pr_devel("[%d] shortcut\n", slot);
380 BUG_ON(!assoc_array_ptr_is_shortcut(cursor));
381 shortcut = assoc_array_ptr_to_shortcut(cursor);
382 BUG_ON(shortcut->back_pointer != parent);
383 BUG_ON(slot != -1 && shortcut->parent_slot != slot);
384 parent = cursor;
385 cursor = shortcut->next_node;
386 slot = -1;
387 BUG_ON(!assoc_array_ptr_is_node(cursor));
388 }
389
390 pr_devel("[%d] node\n", slot);
391 node = assoc_array_ptr_to_node(cursor);
392 BUG_ON(node->back_pointer != parent);
393 BUG_ON(slot != -1 && node->parent_slot != slot);
394 slot = 0;
395
396continue_node:
397 pr_devel("Node %p [back=%p]\n", node, node->back_pointer);
398 for (; slot < ASSOC_ARRAY_FAN_OUT; slot++) {
399 struct assoc_array_ptr *ptr = node->slots[slot];
400 if (!ptr)
401 continue;
402 if (assoc_array_ptr_is_meta(ptr)) {
403 parent = cursor;
404 cursor = ptr;
405 goto move_to_meta;
406 }
407
408 if (ops) {
409 pr_devel("[%d] free leaf\n", slot);
410 ops->free_object(assoc_array_ptr_to_leaf(ptr));
411 }
412 }
413
414 parent = node->back_pointer;
415 slot = node->parent_slot;
416 pr_devel("free node\n");
417 kfree(node);
418 if (!parent)
419 return; /* Done */
420
421 /* Move back up to the parent (may need to free a shortcut on
422 * the way up) */
423 if (assoc_array_ptr_is_shortcut(parent)) {
424 shortcut = assoc_array_ptr_to_shortcut(parent);
425 BUG_ON(shortcut->next_node != cursor);
426 cursor = parent;
427 parent = shortcut->back_pointer;
428 slot = shortcut->parent_slot;
429 pr_devel("free shortcut\n");
430 kfree(shortcut);
431 if (!parent)
432 return;
433
434 BUG_ON(!assoc_array_ptr_is_node(parent));
435 }
436
437 /* Ascend to next slot in parent node */
438 pr_devel("ascend to %p[%d]\n", parent, slot);
439 cursor = parent;
440 node = assoc_array_ptr_to_node(cursor);
441 slot++;
442 goto continue_node;
443}
444
445/**
446 * assoc_array_destroy - Destroy an associative array
447 * @array: The array to destroy.
448 * @ops: The operations to use.
449 *
450 * Discard all metadata and free all objects in an associative array. The
451 * array will be empty and ready to use again upon completion. This function
452 * cannot fail.
453 *
454 * The caller must prevent all other accesses whilst this takes place as no
455 * attempt is made to adjust pointers gracefully to permit RCU readlock-holding
456 * accesses to continue. On the other hand, no memory allocation is required.
457 */
458void assoc_array_destroy(struct assoc_array *array,
459 const struct assoc_array_ops *ops)
460{
461 assoc_array_destroy_subtree(array->root, ops);
462 array->root = NULL;
463}
464
465/*
466 * Handle insertion into an empty tree.
467 */
468static bool assoc_array_insert_in_empty_tree(struct assoc_array_edit *edit)
469{
470 struct assoc_array_node *new_n0;
471
472 pr_devel("-->%s()\n", __func__);
473
474 new_n0 = kzalloc(sizeof(struct assoc_array_node), GFP_KERNEL);
475 if (!new_n0)
476 return false;
477
478 edit->new_meta[0] = assoc_array_node_to_ptr(new_n0);
479 edit->leaf_p = &new_n0->slots[0];
480 edit->adjust_count_on = new_n0;
481 edit->set[0].ptr = &edit->array->root;
482 edit->set[0].to = assoc_array_node_to_ptr(new_n0);
483
484 pr_devel("<--%s() = ok [no root]\n", __func__);
485 return true;
486}
487
488/*
489 * Handle insertion into a terminal node.
490 */
491static bool assoc_array_insert_into_terminal_node(struct assoc_array_edit *edit,
492 const struct assoc_array_ops *ops,
493 const void *index_key,
494 struct assoc_array_walk_result *result)
495{
496 struct assoc_array_shortcut *shortcut, *new_s0;
497 struct assoc_array_node *node, *new_n0, *new_n1, *side;
498 struct assoc_array_ptr *ptr;
499 unsigned long dissimilarity, base_seg, blank;
500 size_t keylen;
501 bool have_meta;
502 int level, diff;
503 int slot, next_slot, free_slot, i, j;
504
505 node = result->terminal_node.node;
506 level = result->terminal_node.level;
507 edit->segment_cache[ASSOC_ARRAY_FAN_OUT] = result->terminal_node.slot;
508
509 pr_devel("-->%s()\n", __func__);
510
511 /* We arrived at a node which doesn't have an onward node or shortcut
512 * pointer that we have to follow. This means that (a) the leaf we
513 * want must go here (either by insertion or replacement) or (b) we
514 * need to split this node and insert in one of the fragments.
515 */
516 free_slot = -1;
517
518 /* Firstly, we have to check the leaves in this node to see if there's
519 * a matching one we should replace in place.
520 */
521 for (i = 0; i < ASSOC_ARRAY_FAN_OUT; i++) {
522 ptr = node->slots[i];
523 if (!ptr) {
524 free_slot = i;
525 continue;
526 }
Jerome Marchand8d4a2ec2016-04-06 14:06:48 +0100527 if (assoc_array_ptr_is_leaf(ptr) &&
528 ops->compare_object(assoc_array_ptr_to_leaf(ptr),
529 index_key)) {
David Howells3cb98952013-09-24 10:35:17 +0100530 pr_devel("replace in slot %d\n", i);
531 edit->leaf_p = &node->slots[i];
532 edit->dead_leaf = node->slots[i];
533 pr_devel("<--%s() = ok [replace]\n", __func__);
534 return true;
535 }
536 }
537
538 /* If there is a free slot in this node then we can just insert the
539 * leaf here.
540 */
541 if (free_slot >= 0) {
542 pr_devel("insert in free slot %d\n", free_slot);
543 edit->leaf_p = &node->slots[free_slot];
544 edit->adjust_count_on = node;
545 pr_devel("<--%s() = ok [insert]\n", __func__);
546 return true;
547 }
548
549 /* The node has no spare slots - so we're either going to have to split
550 * it or insert another node before it.
551 *
552 * Whatever, we're going to need at least two new nodes - so allocate
553 * those now. We may also need a new shortcut, but we deal with that
554 * when we need it.
555 */
556 new_n0 = kzalloc(sizeof(struct assoc_array_node), GFP_KERNEL);
557 if (!new_n0)
558 return false;
559 edit->new_meta[0] = assoc_array_node_to_ptr(new_n0);
560 new_n1 = kzalloc(sizeof(struct assoc_array_node), GFP_KERNEL);
561 if (!new_n1)
562 return false;
563 edit->new_meta[1] = assoc_array_node_to_ptr(new_n1);
564
565 /* We need to find out how similar the leaves are. */
566 pr_devel("no spare slots\n");
567 have_meta = false;
568 for (i = 0; i < ASSOC_ARRAY_FAN_OUT; i++) {
569 ptr = node->slots[i];
570 if (assoc_array_ptr_is_meta(ptr)) {
571 edit->segment_cache[i] = 0xff;
572 have_meta = true;
573 continue;
574 }
575 base_seg = ops->get_object_key_chunk(
576 assoc_array_ptr_to_leaf(ptr), level);
577 base_seg >>= level & ASSOC_ARRAY_KEY_CHUNK_MASK;
578 edit->segment_cache[i] = base_seg & ASSOC_ARRAY_FAN_MASK;
579 }
580
581 if (have_meta) {
582 pr_devel("have meta\n");
583 goto split_node;
584 }
585
586 /* The node contains only leaves */
587 dissimilarity = 0;
588 base_seg = edit->segment_cache[0];
589 for (i = 1; i < ASSOC_ARRAY_FAN_OUT; i++)
590 dissimilarity |= edit->segment_cache[i] ^ base_seg;
591
592 pr_devel("only leaves; dissimilarity=%lx\n", dissimilarity);
593
594 if ((dissimilarity & ASSOC_ARRAY_FAN_MASK) == 0) {
595 /* The old leaves all cluster in the same slot. We will need
596 * to insert a shortcut if the new node wants to cluster with them.
597 */
598 if ((edit->segment_cache[ASSOC_ARRAY_FAN_OUT] ^ base_seg) == 0)
599 goto all_leaves_cluster_together;
600
David Howells67bcc5e2017-10-11 23:32:27 +0100601 /* Otherwise all the old leaves cluster in the same slot, but
602 * the new leaf wants to go into a different slot - so we
603 * create a new node (n0) to hold the new leaf and a pointer to
604 * a new node (n1) holding all the old leaves.
605 *
606 * This can be done by falling through to the node splitting
607 * path.
David Howells3cb98952013-09-24 10:35:17 +0100608 */
David Howells67bcc5e2017-10-11 23:32:27 +0100609 pr_devel("present leaves cluster but not new leaf\n");
David Howells3cb98952013-09-24 10:35:17 +0100610 }
611
612split_node:
613 pr_devel("split node\n");
614
David Howells67bcc5e2017-10-11 23:32:27 +0100615 /* We need to split the current node. The node must contain anything
616 * from a single leaf (in the one leaf case, this leaf will cluster
617 * with the new leaf) and the rest meta-pointers, to all leaves, some
618 * of which may cluster.
619 *
620 * It won't contain the case in which all the current leaves plus the
621 * new leaves want to cluster in the same slot.
David Howells3cb98952013-09-24 10:35:17 +0100622 *
623 * We need to expel at least two leaves out of a set consisting of the
David Howells67bcc5e2017-10-11 23:32:27 +0100624 * leaves in the node and the new leaf. The current meta pointers can
625 * just be copied as they shouldn't cluster with any of the leaves.
David Howells3cb98952013-09-24 10:35:17 +0100626 *
627 * We need a new node (n0) to replace the current one and a new node to
628 * take the expelled nodes (n1).
629 */
630 edit->set[0].to = assoc_array_node_to_ptr(new_n0);
631 new_n0->back_pointer = node->back_pointer;
632 new_n0->parent_slot = node->parent_slot;
633 new_n1->back_pointer = assoc_array_node_to_ptr(new_n0);
634 new_n1->parent_slot = -1; /* Need to calculate this */
635
636do_split_node:
637 pr_devel("do_split_node\n");
638
639 new_n0->nr_leaves_on_branch = node->nr_leaves_on_branch;
640 new_n1->nr_leaves_on_branch = 0;
641
642 /* Begin by finding two matching leaves. There have to be at least two
643 * that match - even if there are meta pointers - because any leaf that
644 * would match a slot with a meta pointer in it must be somewhere
645 * behind that meta pointer and cannot be here. Further, given N
646 * remaining leaf slots, we now have N+1 leaves to go in them.
647 */
648 for (i = 0; i < ASSOC_ARRAY_FAN_OUT; i++) {
649 slot = edit->segment_cache[i];
650 if (slot != 0xff)
651 for (j = i + 1; j < ASSOC_ARRAY_FAN_OUT + 1; j++)
652 if (edit->segment_cache[j] == slot)
653 goto found_slot_for_multiple_occupancy;
654 }
655found_slot_for_multiple_occupancy:
656 pr_devel("same slot: %x %x [%02x]\n", i, j, slot);
657 BUG_ON(i >= ASSOC_ARRAY_FAN_OUT);
658 BUG_ON(j >= ASSOC_ARRAY_FAN_OUT + 1);
659 BUG_ON(slot >= ASSOC_ARRAY_FAN_OUT);
660
661 new_n1->parent_slot = slot;
662
663 /* Metadata pointers cannot change slot */
664 for (i = 0; i < ASSOC_ARRAY_FAN_OUT; i++)
665 if (assoc_array_ptr_is_meta(node->slots[i]))
666 new_n0->slots[i] = node->slots[i];
667 else
668 new_n0->slots[i] = NULL;
669 BUG_ON(new_n0->slots[slot] != NULL);
670 new_n0->slots[slot] = assoc_array_node_to_ptr(new_n1);
671
672 /* Filter the leaf pointers between the new nodes */
673 free_slot = -1;
674 next_slot = 0;
675 for (i = 0; i < ASSOC_ARRAY_FAN_OUT; i++) {
676 if (assoc_array_ptr_is_meta(node->slots[i]))
677 continue;
678 if (edit->segment_cache[i] == slot) {
679 new_n1->slots[next_slot++] = node->slots[i];
680 new_n1->nr_leaves_on_branch++;
681 } else {
682 do {
683 free_slot++;
684 } while (new_n0->slots[free_slot] != NULL);
685 new_n0->slots[free_slot] = node->slots[i];
686 }
687 }
688
689 pr_devel("filtered: f=%x n=%x\n", free_slot, next_slot);
690
691 if (edit->segment_cache[ASSOC_ARRAY_FAN_OUT] != slot) {
692 do {
693 free_slot++;
694 } while (new_n0->slots[free_slot] != NULL);
695 edit->leaf_p = &new_n0->slots[free_slot];
696 edit->adjust_count_on = new_n0;
697 } else {
698 edit->leaf_p = &new_n1->slots[next_slot++];
699 edit->adjust_count_on = new_n1;
700 }
701
702 BUG_ON(next_slot <= 1);
703
704 edit->set_backpointers_to = assoc_array_node_to_ptr(new_n0);
705 for (i = 0; i < ASSOC_ARRAY_FAN_OUT; i++) {
706 if (edit->segment_cache[i] == 0xff) {
707 ptr = node->slots[i];
708 BUG_ON(assoc_array_ptr_is_leaf(ptr));
709 if (assoc_array_ptr_is_node(ptr)) {
710 side = assoc_array_ptr_to_node(ptr);
711 edit->set_backpointers[i] = &side->back_pointer;
712 } else {
713 shortcut = assoc_array_ptr_to_shortcut(ptr);
714 edit->set_backpointers[i] = &shortcut->back_pointer;
715 }
716 }
717 }
718
719 ptr = node->back_pointer;
720 if (!ptr)
721 edit->set[0].ptr = &edit->array->root;
722 else if (assoc_array_ptr_is_node(ptr))
723 edit->set[0].ptr = &assoc_array_ptr_to_node(ptr)->slots[node->parent_slot];
724 else
725 edit->set[0].ptr = &assoc_array_ptr_to_shortcut(ptr)->next_node;
726 edit->excised_meta[0] = assoc_array_node_to_ptr(node);
727 pr_devel("<--%s() = ok [split node]\n", __func__);
728 return true;
729
David Howells3cb98952013-09-24 10:35:17 +0100730all_leaves_cluster_together:
731 /* All the leaves, new and old, want to cluster together in this node
732 * in the same slot, so we have to replace this node with a shortcut to
733 * skip over the identical parts of the key and then place a pair of
734 * nodes, one inside the other, at the end of the shortcut and
735 * distribute the keys between them.
736 *
737 * Firstly we need to work out where the leaves start diverging as a
738 * bit position into their keys so that we know how big the shortcut
739 * needs to be.
740 *
741 * We only need to make a single pass of N of the N+1 leaves because if
742 * any keys differ between themselves at bit X then at least one of
743 * them must also differ with the base key at bit X or before.
744 */
745 pr_devel("all leaves cluster together\n");
746 diff = INT_MAX;
747 for (i = 0; i < ASSOC_ARRAY_FAN_OUT; i++) {
David Howells23fd78d2013-12-02 11:24:18 +0000748 int x = ops->diff_objects(assoc_array_ptr_to_leaf(node->slots[i]),
749 index_key);
David Howells3cb98952013-09-24 10:35:17 +0100750 if (x < diff) {
751 BUG_ON(x < 0);
752 diff = x;
753 }
754 }
755 BUG_ON(diff == INT_MAX);
756 BUG_ON(diff < level + ASSOC_ARRAY_LEVEL_STEP);
757
758 keylen = round_up(diff, ASSOC_ARRAY_KEY_CHUNK_SIZE);
759 keylen >>= ASSOC_ARRAY_KEY_CHUNK_SHIFT;
760
761 new_s0 = kzalloc(sizeof(struct assoc_array_shortcut) +
762 keylen * sizeof(unsigned long), GFP_KERNEL);
763 if (!new_s0)
764 return false;
765 edit->new_meta[2] = assoc_array_shortcut_to_ptr(new_s0);
766
767 edit->set[0].to = assoc_array_shortcut_to_ptr(new_s0);
768 new_s0->back_pointer = node->back_pointer;
769 new_s0->parent_slot = node->parent_slot;
770 new_s0->next_node = assoc_array_node_to_ptr(new_n0);
771 new_n0->back_pointer = assoc_array_shortcut_to_ptr(new_s0);
772 new_n0->parent_slot = 0;
773 new_n1->back_pointer = assoc_array_node_to_ptr(new_n0);
774 new_n1->parent_slot = -1; /* Need to calculate this */
775
776 new_s0->skip_to_level = level = diff & ~ASSOC_ARRAY_LEVEL_STEP_MASK;
777 pr_devel("skip_to_level = %d [diff %d]\n", level, diff);
778 BUG_ON(level <= 0);
779
780 for (i = 0; i < keylen; i++)
781 new_s0->index_key[i] =
782 ops->get_key_chunk(index_key, i * ASSOC_ARRAY_KEY_CHUNK_SIZE);
783
David Howells740f4ae2019-02-14 16:20:15 +0000784 if (level & ASSOC_ARRAY_KEY_CHUNK_MASK) {
785 blank = ULONG_MAX << (level & ASSOC_ARRAY_KEY_CHUNK_MASK);
786 pr_devel("blank off [%zu] %d: %lx\n", keylen - 1, level, blank);
787 new_s0->index_key[keylen - 1] &= ~blank;
788 }
David Howells3cb98952013-09-24 10:35:17 +0100789
790 /* This now reduces to a node splitting exercise for which we'll need
791 * to regenerate the disparity table.
792 */
793 for (i = 0; i < ASSOC_ARRAY_FAN_OUT; i++) {
794 ptr = node->slots[i];
795 base_seg = ops->get_object_key_chunk(assoc_array_ptr_to_leaf(ptr),
796 level);
797 base_seg >>= level & ASSOC_ARRAY_KEY_CHUNK_MASK;
798 edit->segment_cache[i] = base_seg & ASSOC_ARRAY_FAN_MASK;
799 }
800
801 base_seg = ops->get_key_chunk(index_key, level);
802 base_seg >>= level & ASSOC_ARRAY_KEY_CHUNK_MASK;
803 edit->segment_cache[ASSOC_ARRAY_FAN_OUT] = base_seg & ASSOC_ARRAY_FAN_MASK;
804 goto do_split_node;
805}
806
807/*
808 * Handle insertion into the middle of a shortcut.
809 */
810static bool assoc_array_insert_mid_shortcut(struct assoc_array_edit *edit,
811 const struct assoc_array_ops *ops,
812 struct assoc_array_walk_result *result)
813{
814 struct assoc_array_shortcut *shortcut, *new_s0, *new_s1;
815 struct assoc_array_node *node, *new_n0, *side;
816 unsigned long sc_segments, dissimilarity, blank;
817 size_t keylen;
818 int level, sc_level, diff;
819 int sc_slot;
820
821 shortcut = result->wrong_shortcut.shortcut;
822 level = result->wrong_shortcut.level;
823 sc_level = result->wrong_shortcut.sc_level;
824 sc_segments = result->wrong_shortcut.sc_segments;
825 dissimilarity = result->wrong_shortcut.dissimilarity;
826
827 pr_devel("-->%s(ix=%d dis=%lx scix=%d)\n",
828 __func__, level, dissimilarity, sc_level);
829
830 /* We need to split a shortcut and insert a node between the two
831 * pieces. Zero-length pieces will be dispensed with entirely.
832 *
833 * First of all, we need to find out in which level the first
834 * difference was.
835 */
836 diff = __ffs(dissimilarity);
837 diff &= ~ASSOC_ARRAY_LEVEL_STEP_MASK;
838 diff += sc_level & ~ASSOC_ARRAY_KEY_CHUNK_MASK;
839 pr_devel("diff=%d\n", diff);
840
841 if (!shortcut->back_pointer) {
842 edit->set[0].ptr = &edit->array->root;
843 } else if (assoc_array_ptr_is_node(shortcut->back_pointer)) {
844 node = assoc_array_ptr_to_node(shortcut->back_pointer);
845 edit->set[0].ptr = &node->slots[shortcut->parent_slot];
846 } else {
847 BUG();
848 }
849
850 edit->excised_meta[0] = assoc_array_shortcut_to_ptr(shortcut);
851
852 /* Create a new node now since we're going to need it anyway */
853 new_n0 = kzalloc(sizeof(struct assoc_array_node), GFP_KERNEL);
854 if (!new_n0)
855 return false;
856 edit->new_meta[0] = assoc_array_node_to_ptr(new_n0);
857 edit->adjust_count_on = new_n0;
858
859 /* Insert a new shortcut before the new node if this segment isn't of
860 * zero length - otherwise we just connect the new node directly to the
861 * parent.
862 */
863 level += ASSOC_ARRAY_LEVEL_STEP;
864 if (diff > level) {
865 pr_devel("pre-shortcut %d...%d\n", level, diff);
866 keylen = round_up(diff, ASSOC_ARRAY_KEY_CHUNK_SIZE);
867 keylen >>= ASSOC_ARRAY_KEY_CHUNK_SHIFT;
868
869 new_s0 = kzalloc(sizeof(struct assoc_array_shortcut) +
870 keylen * sizeof(unsigned long), GFP_KERNEL);
871 if (!new_s0)
872 return false;
873 edit->new_meta[1] = assoc_array_shortcut_to_ptr(new_s0);
874 edit->set[0].to = assoc_array_shortcut_to_ptr(new_s0);
875 new_s0->back_pointer = shortcut->back_pointer;
876 new_s0->parent_slot = shortcut->parent_slot;
877 new_s0->next_node = assoc_array_node_to_ptr(new_n0);
878 new_s0->skip_to_level = diff;
879
880 new_n0->back_pointer = assoc_array_shortcut_to_ptr(new_s0);
881 new_n0->parent_slot = 0;
882
883 memcpy(new_s0->index_key, shortcut->index_key,
884 keylen * sizeof(unsigned long));
885
886 blank = ULONG_MAX << (diff & ASSOC_ARRAY_KEY_CHUNK_MASK);
887 pr_devel("blank off [%zu] %d: %lx\n", keylen - 1, diff, blank);
888 new_s0->index_key[keylen - 1] &= ~blank;
889 } else {
890 pr_devel("no pre-shortcut\n");
891 edit->set[0].to = assoc_array_node_to_ptr(new_n0);
892 new_n0->back_pointer = shortcut->back_pointer;
893 new_n0->parent_slot = shortcut->parent_slot;
894 }
895
896 side = assoc_array_ptr_to_node(shortcut->next_node);
897 new_n0->nr_leaves_on_branch = side->nr_leaves_on_branch;
898
899 /* We need to know which slot in the new node is going to take a
900 * metadata pointer.
901 */
902 sc_slot = sc_segments >> (diff & ASSOC_ARRAY_KEY_CHUNK_MASK);
903 sc_slot &= ASSOC_ARRAY_FAN_MASK;
904
905 pr_devel("new slot %lx >> %d -> %d\n",
906 sc_segments, diff & ASSOC_ARRAY_KEY_CHUNK_MASK, sc_slot);
907
908 /* Determine whether we need to follow the new node with a replacement
909 * for the current shortcut. We could in theory reuse the current
910 * shortcut if its parent slot number doesn't change - but that's a
911 * 1-in-16 chance so not worth expending the code upon.
912 */
913 level = diff + ASSOC_ARRAY_LEVEL_STEP;
914 if (level < shortcut->skip_to_level) {
915 pr_devel("post-shortcut %d...%d\n", level, shortcut->skip_to_level);
916 keylen = round_up(shortcut->skip_to_level, ASSOC_ARRAY_KEY_CHUNK_SIZE);
917 keylen >>= ASSOC_ARRAY_KEY_CHUNK_SHIFT;
918
919 new_s1 = kzalloc(sizeof(struct assoc_array_shortcut) +
920 keylen * sizeof(unsigned long), GFP_KERNEL);
921 if (!new_s1)
922 return false;
923 edit->new_meta[2] = assoc_array_shortcut_to_ptr(new_s1);
924
925 new_s1->back_pointer = assoc_array_node_to_ptr(new_n0);
926 new_s1->parent_slot = sc_slot;
927 new_s1->next_node = shortcut->next_node;
928 new_s1->skip_to_level = shortcut->skip_to_level;
929
930 new_n0->slots[sc_slot] = assoc_array_shortcut_to_ptr(new_s1);
931
932 memcpy(new_s1->index_key, shortcut->index_key,
933 keylen * sizeof(unsigned long));
934
935 edit->set[1].ptr = &side->back_pointer;
936 edit->set[1].to = assoc_array_shortcut_to_ptr(new_s1);
937 } else {
938 pr_devel("no post-shortcut\n");
939
940 /* We don't have to replace the pointed-to node as long as we
941 * use memory barriers to make sure the parent slot number is
942 * changed before the back pointer (the parent slot number is
943 * irrelevant to the old parent shortcut).
944 */
945 new_n0->slots[sc_slot] = shortcut->next_node;
946 edit->set_parent_slot[0].p = &side->parent_slot;
947 edit->set_parent_slot[0].to = sc_slot;
948 edit->set[1].ptr = &side->back_pointer;
949 edit->set[1].to = assoc_array_node_to_ptr(new_n0);
950 }
951
952 /* Install the new leaf in a spare slot in the new node. */
953 if (sc_slot == 0)
954 edit->leaf_p = &new_n0->slots[1];
955 else
956 edit->leaf_p = &new_n0->slots[0];
957
958 pr_devel("<--%s() = ok [split shortcut]\n", __func__);
959 return edit;
960}
961
962/**
963 * assoc_array_insert - Script insertion of an object into an associative array
964 * @array: The array to insert into.
965 * @ops: The operations to use.
966 * @index_key: The key to insert at.
967 * @object: The object to insert.
968 *
969 * Precalculate and preallocate a script for the insertion or replacement of an
970 * object in an associative array. This results in an edit script that can
971 * either be applied or cancelled.
972 *
973 * The function returns a pointer to an edit script or -ENOMEM.
974 *
975 * The caller should lock against other modifications and must continue to hold
976 * the lock until assoc_array_apply_edit() has been called.
977 *
978 * Accesses to the tree may take place concurrently with this function,
979 * provided they hold the RCU read lock.
980 */
981struct assoc_array_edit *assoc_array_insert(struct assoc_array *array,
982 const struct assoc_array_ops *ops,
983 const void *index_key,
984 void *object)
985{
986 struct assoc_array_walk_result result;
987 struct assoc_array_edit *edit;
988
989 pr_devel("-->%s()\n", __func__);
990
991 /* The leaf pointer we're given must not have the bottom bit set as we
992 * use those for type-marking the pointer. NULL pointers are also not
993 * allowed as they indicate an empty slot but we have to allow them
994 * here as they can be updated later.
995 */
996 BUG_ON(assoc_array_ptr_is_meta(object));
997
998 edit = kzalloc(sizeof(struct assoc_array_edit), GFP_KERNEL);
999 if (!edit)
1000 return ERR_PTR(-ENOMEM);
1001 edit->array = array;
1002 edit->ops = ops;
1003 edit->leaf = assoc_array_leaf_to_ptr(object);
1004 edit->adjust_count_by = 1;
1005
1006 switch (assoc_array_walk(array, ops, index_key, &result)) {
1007 case assoc_array_walk_tree_empty:
1008 /* Allocate a root node if there isn't one yet */
1009 if (!assoc_array_insert_in_empty_tree(edit))
1010 goto enomem;
1011 return edit;
1012
1013 case assoc_array_walk_found_terminal_node:
1014 /* We found a node that doesn't have a node/shortcut pointer in
1015 * the slot corresponding to the index key that we have to
1016 * follow.
1017 */
1018 if (!assoc_array_insert_into_terminal_node(edit, ops, index_key,
1019 &result))
1020 goto enomem;
1021 return edit;
1022
1023 case assoc_array_walk_found_wrong_shortcut:
1024 /* We found a shortcut that didn't match our key in a slot we
1025 * needed to follow.
1026 */
1027 if (!assoc_array_insert_mid_shortcut(edit, ops, &result))
1028 goto enomem;
1029 return edit;
1030 }
1031
1032enomem:
1033 /* Clean up after an out of memory error */
1034 pr_devel("enomem\n");
1035 assoc_array_cancel_edit(edit);
1036 return ERR_PTR(-ENOMEM);
1037}
1038
1039/**
1040 * assoc_array_insert_set_object - Set the new object pointer in an edit script
1041 * @edit: The edit script to modify.
1042 * @object: The object pointer to set.
1043 *
1044 * Change the object to be inserted in an edit script. The object pointed to
1045 * by the old object is not freed. This must be done prior to applying the
1046 * script.
1047 */
1048void assoc_array_insert_set_object(struct assoc_array_edit *edit, void *object)
1049{
1050 BUG_ON(!object);
1051 edit->leaf = assoc_array_leaf_to_ptr(object);
1052}
1053
1054struct assoc_array_delete_collapse_context {
1055 struct assoc_array_node *node;
1056 const void *skip_leaf;
1057 int slot;
1058};
1059
1060/*
1061 * Subtree collapse to node iterator.
1062 */
1063static int assoc_array_delete_collapse_iterator(const void *leaf,
1064 void *iterator_data)
1065{
1066 struct assoc_array_delete_collapse_context *collapse = iterator_data;
1067
1068 if (leaf == collapse->skip_leaf)
1069 return 0;
1070
1071 BUG_ON(collapse->slot >= ASSOC_ARRAY_FAN_OUT);
1072
1073 collapse->node->slots[collapse->slot++] = assoc_array_leaf_to_ptr(leaf);
1074 return 0;
1075}
1076
1077/**
1078 * assoc_array_delete - Script deletion of an object from an associative array
1079 * @array: The array to search.
1080 * @ops: The operations to use.
1081 * @index_key: The key to the object.
1082 *
1083 * Precalculate and preallocate a script for the deletion of an object from an
1084 * associative array. This results in an edit script that can either be
1085 * applied or cancelled.
1086 *
1087 * The function returns a pointer to an edit script if the object was found,
1088 * NULL if the object was not found or -ENOMEM.
1089 *
1090 * The caller should lock against other modifications and must continue to hold
1091 * the lock until assoc_array_apply_edit() has been called.
1092 *
1093 * Accesses to the tree may take place concurrently with this function,
1094 * provided they hold the RCU read lock.
1095 */
1096struct assoc_array_edit *assoc_array_delete(struct assoc_array *array,
1097 const struct assoc_array_ops *ops,
1098 const void *index_key)
1099{
1100 struct assoc_array_delete_collapse_context collapse;
1101 struct assoc_array_walk_result result;
1102 struct assoc_array_node *node, *new_n0;
1103 struct assoc_array_edit *edit;
1104 struct assoc_array_ptr *ptr;
1105 bool has_meta;
1106 int slot, i;
1107
1108 pr_devel("-->%s()\n", __func__);
1109
1110 edit = kzalloc(sizeof(struct assoc_array_edit), GFP_KERNEL);
1111 if (!edit)
1112 return ERR_PTR(-ENOMEM);
1113 edit->array = array;
1114 edit->ops = ops;
1115 edit->adjust_count_by = -1;
1116
1117 switch (assoc_array_walk(array, ops, index_key, &result)) {
1118 case assoc_array_walk_found_terminal_node:
1119 /* We found a node that should contain the leaf we've been
1120 * asked to remove - *if* it's in the tree.
1121 */
1122 pr_devel("terminal_node\n");
1123 node = result.terminal_node.node;
1124
1125 for (slot = 0; slot < ASSOC_ARRAY_FAN_OUT; slot++) {
1126 ptr = node->slots[slot];
1127 if (ptr &&
1128 assoc_array_ptr_is_leaf(ptr) &&
1129 ops->compare_object(assoc_array_ptr_to_leaf(ptr),
1130 index_key))
1131 goto found_leaf;
1132 }
1133 case assoc_array_walk_tree_empty:
1134 case assoc_array_walk_found_wrong_shortcut:
1135 default:
1136 assoc_array_cancel_edit(edit);
1137 pr_devel("not found\n");
1138 return NULL;
1139 }
1140
1141found_leaf:
1142 BUG_ON(array->nr_leaves_on_tree <= 0);
1143
1144 /* In the simplest form of deletion we just clear the slot and release
1145 * the leaf after a suitable interval.
1146 */
1147 edit->dead_leaf = node->slots[slot];
1148 edit->set[0].ptr = &node->slots[slot];
1149 edit->set[0].to = NULL;
1150 edit->adjust_count_on = node;
1151
1152 /* If that concludes erasure of the last leaf, then delete the entire
1153 * internal array.
1154 */
1155 if (array->nr_leaves_on_tree == 1) {
1156 edit->set[1].ptr = &array->root;
1157 edit->set[1].to = NULL;
1158 edit->adjust_count_on = NULL;
1159 edit->excised_subtree = array->root;
1160 pr_devel("all gone\n");
1161 return edit;
1162 }
1163
1164 /* However, we'd also like to clear up some metadata blocks if we
1165 * possibly can.
1166 *
1167 * We go for a simple algorithm of: if this node has FAN_OUT or fewer
1168 * leaves in it, then attempt to collapse it - and attempt to
1169 * recursively collapse up the tree.
1170 *
1171 * We could also try and collapse in partially filled subtrees to take
1172 * up space in this node.
1173 */
1174 if (node->nr_leaves_on_branch <= ASSOC_ARRAY_FAN_OUT + 1) {
1175 struct assoc_array_node *parent, *grandparent;
1176 struct assoc_array_ptr *ptr;
1177
1178 /* First of all, we need to know if this node has metadata so
1179 * that we don't try collapsing if all the leaves are already
1180 * here.
1181 */
1182 has_meta = false;
1183 for (i = 0; i < ASSOC_ARRAY_FAN_OUT; i++) {
1184 ptr = node->slots[i];
1185 if (assoc_array_ptr_is_meta(ptr)) {
1186 has_meta = true;
1187 break;
1188 }
1189 }
1190
1191 pr_devel("leaves: %ld [m=%d]\n",
1192 node->nr_leaves_on_branch - 1, has_meta);
1193
1194 /* Look further up the tree to see if we can collapse this node
1195 * into a more proximal node too.
1196 */
1197 parent = node;
1198 collapse_up:
1199 pr_devel("collapse subtree: %ld\n", parent->nr_leaves_on_branch);
1200
1201 ptr = parent->back_pointer;
1202 if (!ptr)
1203 goto do_collapse;
1204 if (assoc_array_ptr_is_shortcut(ptr)) {
1205 struct assoc_array_shortcut *s = assoc_array_ptr_to_shortcut(ptr);
1206 ptr = s->back_pointer;
1207 if (!ptr)
1208 goto do_collapse;
1209 }
1210
1211 grandparent = assoc_array_ptr_to_node(ptr);
1212 if (grandparent->nr_leaves_on_branch <= ASSOC_ARRAY_FAN_OUT + 1) {
1213 parent = grandparent;
1214 goto collapse_up;
1215 }
1216
1217 do_collapse:
1218 /* There's no point collapsing if the original node has no meta
1219 * pointers to discard and if we didn't merge into one of that
1220 * node's ancestry.
1221 */
1222 if (has_meta || parent != node) {
1223 node = parent;
1224
1225 /* Create a new node to collapse into */
1226 new_n0 = kzalloc(sizeof(struct assoc_array_node), GFP_KERNEL);
1227 if (!new_n0)
1228 goto enomem;
1229 edit->new_meta[0] = assoc_array_node_to_ptr(new_n0);
1230
1231 new_n0->back_pointer = node->back_pointer;
1232 new_n0->parent_slot = node->parent_slot;
1233 new_n0->nr_leaves_on_branch = node->nr_leaves_on_branch;
1234 edit->adjust_count_on = new_n0;
1235
1236 collapse.node = new_n0;
1237 collapse.skip_leaf = assoc_array_ptr_to_leaf(edit->dead_leaf);
1238 collapse.slot = 0;
1239 assoc_array_subtree_iterate(assoc_array_node_to_ptr(node),
1240 node->back_pointer,
1241 assoc_array_delete_collapse_iterator,
1242 &collapse);
1243 pr_devel("collapsed %d,%lu\n", collapse.slot, new_n0->nr_leaves_on_branch);
1244 BUG_ON(collapse.slot != new_n0->nr_leaves_on_branch - 1);
1245
1246 if (!node->back_pointer) {
1247 edit->set[1].ptr = &array->root;
1248 } else if (assoc_array_ptr_is_leaf(node->back_pointer)) {
1249 BUG();
1250 } else if (assoc_array_ptr_is_node(node->back_pointer)) {
1251 struct assoc_array_node *p =
1252 assoc_array_ptr_to_node(node->back_pointer);
1253 edit->set[1].ptr = &p->slots[node->parent_slot];
1254 } else if (assoc_array_ptr_is_shortcut(node->back_pointer)) {
1255 struct assoc_array_shortcut *s =
1256 assoc_array_ptr_to_shortcut(node->back_pointer);
1257 edit->set[1].ptr = &s->next_node;
1258 }
1259 edit->set[1].to = assoc_array_node_to_ptr(new_n0);
1260 edit->excised_subtree = assoc_array_node_to_ptr(node);
1261 }
1262 }
1263
1264 return edit;
1265
1266enomem:
1267 /* Clean up after an out of memory error */
1268 pr_devel("enomem\n");
1269 assoc_array_cancel_edit(edit);
1270 return ERR_PTR(-ENOMEM);
1271}
1272
1273/**
1274 * assoc_array_clear - Script deletion of all objects from an associative array
1275 * @array: The array to clear.
1276 * @ops: The operations to use.
1277 *
1278 * Precalculate and preallocate a script for the deletion of all the objects
1279 * from an associative array. This results in an edit script that can either
1280 * be applied or cancelled.
1281 *
1282 * The function returns a pointer to an edit script if there are objects to be
1283 * deleted, NULL if there are no objects in the array or -ENOMEM.
1284 *
1285 * The caller should lock against other modifications and must continue to hold
1286 * the lock until assoc_array_apply_edit() has been called.
1287 *
1288 * Accesses to the tree may take place concurrently with this function,
1289 * provided they hold the RCU read lock.
1290 */
1291struct assoc_array_edit *assoc_array_clear(struct assoc_array *array,
1292 const struct assoc_array_ops *ops)
1293{
1294 struct assoc_array_edit *edit;
1295
1296 pr_devel("-->%s()\n", __func__);
1297
1298 if (!array->root)
1299 return NULL;
1300
1301 edit = kzalloc(sizeof(struct assoc_array_edit), GFP_KERNEL);
1302 if (!edit)
1303 return ERR_PTR(-ENOMEM);
1304 edit->array = array;
1305 edit->ops = ops;
1306 edit->set[1].ptr = &array->root;
1307 edit->set[1].to = NULL;
1308 edit->excised_subtree = array->root;
1309 edit->ops_for_excised_subtree = ops;
1310 pr_devel("all gone\n");
1311 return edit;
1312}
1313
1314/*
1315 * Handle the deferred destruction after an applied edit.
1316 */
1317static void assoc_array_rcu_cleanup(struct rcu_head *head)
1318{
1319 struct assoc_array_edit *edit =
1320 container_of(head, struct assoc_array_edit, rcu);
1321 int i;
1322
1323 pr_devel("-->%s()\n", __func__);
1324
1325 if (edit->dead_leaf)
1326 edit->ops->free_object(assoc_array_ptr_to_leaf(edit->dead_leaf));
1327 for (i = 0; i < ARRAY_SIZE(edit->excised_meta); i++)
1328 if (edit->excised_meta[i])
1329 kfree(assoc_array_ptr_to_node(edit->excised_meta[i]));
1330
1331 if (edit->excised_subtree) {
1332 BUG_ON(assoc_array_ptr_is_leaf(edit->excised_subtree));
1333 if (assoc_array_ptr_is_node(edit->excised_subtree)) {
1334 struct assoc_array_node *n =
1335 assoc_array_ptr_to_node(edit->excised_subtree);
1336 n->back_pointer = NULL;
1337 } else {
1338 struct assoc_array_shortcut *s =
1339 assoc_array_ptr_to_shortcut(edit->excised_subtree);
1340 s->back_pointer = NULL;
1341 }
1342 assoc_array_destroy_subtree(edit->excised_subtree,
1343 edit->ops_for_excised_subtree);
1344 }
1345
1346 kfree(edit);
1347}
1348
1349/**
1350 * assoc_array_apply_edit - Apply an edit script to an associative array
1351 * @edit: The script to apply.
1352 *
1353 * Apply an edit script to an associative array to effect an insertion,
1354 * deletion or clearance. As the edit script includes preallocated memory,
1355 * this is guaranteed not to fail.
1356 *
1357 * The edit script, dead objects and dead metadata will be scheduled for
1358 * destruction after an RCU grace period to permit those doing read-only
1359 * accesses on the array to continue to do so under the RCU read lock whilst
1360 * the edit is taking place.
1361 */
1362void assoc_array_apply_edit(struct assoc_array_edit *edit)
1363{
1364 struct assoc_array_shortcut *shortcut;
1365 struct assoc_array_node *node;
1366 struct assoc_array_ptr *ptr;
1367 int i;
1368
1369 pr_devel("-->%s()\n", __func__);
1370
1371 smp_wmb();
1372 if (edit->leaf_p)
1373 *edit->leaf_p = edit->leaf;
1374
1375 smp_wmb();
1376 for (i = 0; i < ARRAY_SIZE(edit->set_parent_slot); i++)
1377 if (edit->set_parent_slot[i].p)
1378 *edit->set_parent_slot[i].p = edit->set_parent_slot[i].to;
1379
1380 smp_wmb();
1381 for (i = 0; i < ARRAY_SIZE(edit->set_backpointers); i++)
1382 if (edit->set_backpointers[i])
1383 *edit->set_backpointers[i] = edit->set_backpointers_to;
1384
1385 smp_wmb();
1386 for (i = 0; i < ARRAY_SIZE(edit->set); i++)
1387 if (edit->set[i].ptr)
1388 *edit->set[i].ptr = edit->set[i].to;
1389
1390 if (edit->array->root == NULL) {
1391 edit->array->nr_leaves_on_tree = 0;
1392 } else if (edit->adjust_count_on) {
1393 node = edit->adjust_count_on;
1394 for (;;) {
1395 node->nr_leaves_on_branch += edit->adjust_count_by;
1396
1397 ptr = node->back_pointer;
1398 if (!ptr)
1399 break;
1400 if (assoc_array_ptr_is_shortcut(ptr)) {
1401 shortcut = assoc_array_ptr_to_shortcut(ptr);
1402 ptr = shortcut->back_pointer;
1403 if (!ptr)
1404 break;
1405 }
1406 BUG_ON(!assoc_array_ptr_is_node(ptr));
1407 node = assoc_array_ptr_to_node(ptr);
1408 }
1409
1410 edit->array->nr_leaves_on_tree += edit->adjust_count_by;
1411 }
1412
1413 call_rcu(&edit->rcu, assoc_array_rcu_cleanup);
1414}
1415
1416/**
1417 * assoc_array_cancel_edit - Discard an edit script.
1418 * @edit: The script to discard.
1419 *
1420 * Free an edit script and all the preallocated data it holds without making
1421 * any changes to the associative array it was intended for.
1422 *
1423 * NOTE! In the case of an insertion script, this does _not_ release the leaf
1424 * that was to be inserted. That is left to the caller.
1425 */
1426void assoc_array_cancel_edit(struct assoc_array_edit *edit)
1427{
1428 struct assoc_array_ptr *ptr;
1429 int i;
1430
1431 pr_devel("-->%s()\n", __func__);
1432
1433 /* Clean up after an out of memory error */
1434 for (i = 0; i < ARRAY_SIZE(edit->new_meta); i++) {
1435 ptr = edit->new_meta[i];
1436 if (ptr) {
1437 if (assoc_array_ptr_is_node(ptr))
1438 kfree(assoc_array_ptr_to_node(ptr));
1439 else
1440 kfree(assoc_array_ptr_to_shortcut(ptr));
1441 }
1442 }
1443 kfree(edit);
1444}
1445
1446/**
1447 * assoc_array_gc - Garbage collect an associative array.
1448 * @array: The array to clean.
1449 * @ops: The operations to use.
1450 * @iterator: A callback function to pass judgement on each object.
1451 * @iterator_data: Private data for the callback function.
1452 *
1453 * Collect garbage from an associative array and pack down the internal tree to
1454 * save memory.
1455 *
1456 * The iterator function is asked to pass judgement upon each object in the
1457 * array. If it returns false, the object is discard and if it returns true,
1458 * the object is kept. If it returns true, it must increment the object's
1459 * usage count (or whatever it needs to do to retain it) before returning.
1460 *
1461 * This function returns 0 if successful or -ENOMEM if out of memory. In the
1462 * latter case, the array is not changed.
1463 *
1464 * The caller should lock against other modifications and must continue to hold
1465 * the lock until assoc_array_apply_edit() has been called.
1466 *
1467 * Accesses to the tree may take place concurrently with this function,
1468 * provided they hold the RCU read lock.
1469 */
1470int assoc_array_gc(struct assoc_array *array,
1471 const struct assoc_array_ops *ops,
1472 bool (*iterator)(void *object, void *iterator_data),
1473 void *iterator_data)
1474{
1475 struct assoc_array_shortcut *shortcut, *new_s;
1476 struct assoc_array_node *node, *new_n;
1477 struct assoc_array_edit *edit;
1478 struct assoc_array_ptr *cursor, *ptr;
1479 struct assoc_array_ptr *new_root, *new_parent, **new_ptr_pp;
1480 unsigned long nr_leaves_on_tree;
1481 int keylen, slot, nr_free, next_slot, i;
1482
1483 pr_devel("-->%s()\n", __func__);
1484
1485 if (!array->root)
1486 return 0;
1487
1488 edit = kzalloc(sizeof(struct assoc_array_edit), GFP_KERNEL);
1489 if (!edit)
1490 return -ENOMEM;
1491 edit->array = array;
1492 edit->ops = ops;
1493 edit->ops_for_excised_subtree = ops;
1494 edit->set[0].ptr = &array->root;
1495 edit->excised_subtree = array->root;
1496
1497 new_root = new_parent = NULL;
1498 new_ptr_pp = &new_root;
1499 cursor = array->root;
1500
1501descend:
1502 /* If this point is a shortcut, then we need to duplicate it and
1503 * advance the target cursor.
1504 */
1505 if (assoc_array_ptr_is_shortcut(cursor)) {
1506 shortcut = assoc_array_ptr_to_shortcut(cursor);
1507 keylen = round_up(shortcut->skip_to_level, ASSOC_ARRAY_KEY_CHUNK_SIZE);
1508 keylen >>= ASSOC_ARRAY_KEY_CHUNK_SHIFT;
1509 new_s = kmalloc(sizeof(struct assoc_array_shortcut) +
1510 keylen * sizeof(unsigned long), GFP_KERNEL);
1511 if (!new_s)
1512 goto enomem;
1513 pr_devel("dup shortcut %p -> %p\n", shortcut, new_s);
1514 memcpy(new_s, shortcut, (sizeof(struct assoc_array_shortcut) +
1515 keylen * sizeof(unsigned long)));
1516 new_s->back_pointer = new_parent;
1517 new_s->parent_slot = shortcut->parent_slot;
1518 *new_ptr_pp = new_parent = assoc_array_shortcut_to_ptr(new_s);
1519 new_ptr_pp = &new_s->next_node;
1520 cursor = shortcut->next_node;
1521 }
1522
1523 /* Duplicate the node at this position */
1524 node = assoc_array_ptr_to_node(cursor);
1525 new_n = kzalloc(sizeof(struct assoc_array_node), GFP_KERNEL);
1526 if (!new_n)
1527 goto enomem;
1528 pr_devel("dup node %p -> %p\n", node, new_n);
1529 new_n->back_pointer = new_parent;
1530 new_n->parent_slot = node->parent_slot;
1531 *new_ptr_pp = new_parent = assoc_array_node_to_ptr(new_n);
1532 new_ptr_pp = NULL;
1533 slot = 0;
1534
1535continue_node:
1536 /* Filter across any leaves and gc any subtrees */
1537 for (; slot < ASSOC_ARRAY_FAN_OUT; slot++) {
1538 ptr = node->slots[slot];
1539 if (!ptr)
1540 continue;
1541
1542 if (assoc_array_ptr_is_leaf(ptr)) {
1543 if (iterator(assoc_array_ptr_to_leaf(ptr),
1544 iterator_data))
1545 /* The iterator will have done any reference
1546 * counting on the object for us.
1547 */
1548 new_n->slots[slot] = ptr;
1549 continue;
1550 }
1551
1552 new_ptr_pp = &new_n->slots[slot];
1553 cursor = ptr;
1554 goto descend;
1555 }
1556
1557 pr_devel("-- compress node %p --\n", new_n);
1558
1559 /* Count up the number of empty slots in this node and work out the
1560 * subtree leaf count.
1561 */
1562 new_n->nr_leaves_on_branch = 0;
1563 nr_free = 0;
1564 for (slot = 0; slot < ASSOC_ARRAY_FAN_OUT; slot++) {
1565 ptr = new_n->slots[slot];
1566 if (!ptr)
1567 nr_free++;
1568 else if (assoc_array_ptr_is_leaf(ptr))
1569 new_n->nr_leaves_on_branch++;
1570 }
1571 pr_devel("free=%d, leaves=%lu\n", nr_free, new_n->nr_leaves_on_branch);
1572
1573 /* See what we can fold in */
1574 next_slot = 0;
1575 for (slot = 0; slot < ASSOC_ARRAY_FAN_OUT; slot++) {
1576 struct assoc_array_shortcut *s;
1577 struct assoc_array_node *child;
1578
1579 ptr = new_n->slots[slot];
1580 if (!ptr || assoc_array_ptr_is_leaf(ptr))
1581 continue;
1582
1583 s = NULL;
1584 if (assoc_array_ptr_is_shortcut(ptr)) {
1585 s = assoc_array_ptr_to_shortcut(ptr);
1586 ptr = s->next_node;
1587 }
1588
1589 child = assoc_array_ptr_to_node(ptr);
1590 new_n->nr_leaves_on_branch += child->nr_leaves_on_branch;
1591
1592 if (child->nr_leaves_on_branch <= nr_free + 1) {
1593 /* Fold the child node into this one */
1594 pr_devel("[%d] fold node %lu/%d [nx %d]\n",
1595 slot, child->nr_leaves_on_branch, nr_free + 1,
1596 next_slot);
1597
1598 /* We would already have reaped an intervening shortcut
1599 * on the way back up the tree.
1600 */
1601 BUG_ON(s);
1602
1603 new_n->slots[slot] = NULL;
1604 nr_free++;
1605 if (slot < next_slot)
1606 next_slot = slot;
1607 for (i = 0; i < ASSOC_ARRAY_FAN_OUT; i++) {
1608 struct assoc_array_ptr *p = child->slots[i];
1609 if (!p)
1610 continue;
1611 BUG_ON(assoc_array_ptr_is_meta(p));
1612 while (new_n->slots[next_slot])
1613 next_slot++;
1614 BUG_ON(next_slot >= ASSOC_ARRAY_FAN_OUT);
1615 new_n->slots[next_slot++] = p;
1616 nr_free--;
1617 }
1618 kfree(child);
1619 } else {
1620 pr_devel("[%d] retain node %lu/%d [nx %d]\n",
1621 slot, child->nr_leaves_on_branch, nr_free + 1,
1622 next_slot);
1623 }
1624 }
1625
1626 pr_devel("after: %lu\n", new_n->nr_leaves_on_branch);
1627
1628 nr_leaves_on_tree = new_n->nr_leaves_on_branch;
1629
1630 /* Excise this node if it is singly occupied by a shortcut */
1631 if (nr_free == ASSOC_ARRAY_FAN_OUT - 1) {
1632 for (slot = 0; slot < ASSOC_ARRAY_FAN_OUT; slot++)
1633 if ((ptr = new_n->slots[slot]))
1634 break;
1635
1636 if (assoc_array_ptr_is_meta(ptr) &&
1637 assoc_array_ptr_is_shortcut(ptr)) {
1638 pr_devel("excise node %p with 1 shortcut\n", new_n);
1639 new_s = assoc_array_ptr_to_shortcut(ptr);
1640 new_parent = new_n->back_pointer;
1641 slot = new_n->parent_slot;
1642 kfree(new_n);
1643 if (!new_parent) {
1644 new_s->back_pointer = NULL;
1645 new_s->parent_slot = 0;
1646 new_root = ptr;
1647 goto gc_complete;
1648 }
1649
1650 if (assoc_array_ptr_is_shortcut(new_parent)) {
1651 /* We can discard any preceding shortcut also */
1652 struct assoc_array_shortcut *s =
1653 assoc_array_ptr_to_shortcut(new_parent);
1654
1655 pr_devel("excise preceding shortcut\n");
1656
1657 new_parent = new_s->back_pointer = s->back_pointer;
1658 slot = new_s->parent_slot = s->parent_slot;
1659 kfree(s);
1660 if (!new_parent) {
1661 new_s->back_pointer = NULL;
1662 new_s->parent_slot = 0;
1663 new_root = ptr;
1664 goto gc_complete;
1665 }
1666 }
1667
1668 new_s->back_pointer = new_parent;
1669 new_s->parent_slot = slot;
1670 new_n = assoc_array_ptr_to_node(new_parent);
1671 new_n->slots[slot] = ptr;
1672 goto ascend_old_tree;
1673 }
1674 }
1675
1676 /* Excise any shortcuts we might encounter that point to nodes that
1677 * only contain leaves.
1678 */
1679 ptr = new_n->back_pointer;
1680 if (!ptr)
1681 goto gc_complete;
1682
1683 if (assoc_array_ptr_is_shortcut(ptr)) {
1684 new_s = assoc_array_ptr_to_shortcut(ptr);
1685 new_parent = new_s->back_pointer;
1686 slot = new_s->parent_slot;
1687
1688 if (new_n->nr_leaves_on_branch <= ASSOC_ARRAY_FAN_OUT) {
1689 struct assoc_array_node *n;
1690
1691 pr_devel("excise shortcut\n");
1692 new_n->back_pointer = new_parent;
1693 new_n->parent_slot = slot;
1694 kfree(new_s);
1695 if (!new_parent) {
1696 new_root = assoc_array_node_to_ptr(new_n);
1697 goto gc_complete;
1698 }
1699
1700 n = assoc_array_ptr_to_node(new_parent);
1701 n->slots[slot] = assoc_array_node_to_ptr(new_n);
1702 }
1703 } else {
1704 new_parent = ptr;
1705 }
1706 new_n = assoc_array_ptr_to_node(new_parent);
1707
1708ascend_old_tree:
1709 ptr = node->back_pointer;
1710 if (assoc_array_ptr_is_shortcut(ptr)) {
1711 shortcut = assoc_array_ptr_to_shortcut(ptr);
1712 slot = shortcut->parent_slot;
1713 cursor = shortcut->back_pointer;
David Howells95389b02014-09-10 22:22:00 +01001714 if (!cursor)
1715 goto gc_complete;
David Howells3cb98952013-09-24 10:35:17 +01001716 } else {
1717 slot = node->parent_slot;
1718 cursor = ptr;
1719 }
David Howells95389b02014-09-10 22:22:00 +01001720 BUG_ON(!cursor);
David Howells3cb98952013-09-24 10:35:17 +01001721 node = assoc_array_ptr_to_node(cursor);
1722 slot++;
1723 goto continue_node;
1724
1725gc_complete:
1726 edit->set[0].to = new_root;
1727 assoc_array_apply_edit(edit);
David Howells27419602014-09-02 13:52:20 +01001728 array->nr_leaves_on_tree = nr_leaves_on_tree;
David Howells3cb98952013-09-24 10:35:17 +01001729 return 0;
1730
1731enomem:
1732 pr_devel("enomem\n");
1733 assoc_array_destroy_subtree(new_root, edit->ops);
1734 kfree(edit);
1735 return -ENOMEM;
1736}