1 // SPDX-License-Identifier: GPL-2.0+
3 * Maple Tree implementation
4 * Copyright (c) 2018-2022 Oracle Corporation
5 * Authors: Liam R. Howlett <Liam.Howlett@oracle.com>
6 * Matthew Wilcox <willy@infradead.org>
10 * DOC: Interesting implementation details of the Maple Tree
12 * Each node type has a number of slots for entries and a number of slots for
13 * pivots. In the case of dense nodes, the pivots are implied by the position
14 * and are simply the slot index + the minimum of the node.
16 * In regular B-Tree terms, pivots are called keys. The term pivot is used to
17 * indicate that the tree is specifying ranges, Pivots may appear in the
18 * subtree with an entry attached to the value where as keys are unique to a
19 * specific position of a B-tree. Pivot values are inclusive of the slot with
23 * The following illustrates the layout of a range64 nodes slots and pivots.
26 * Slots -> | 0 | 1 | 2 | ... | 12 | 13 | 14 | 15 |
28 * │ │ │ │ │ │ │ │ └─ Implied maximum
29 * │ │ │ │ │ │ │ └─ Pivot 14
30 * │ │ │ │ │ │ └─ Pivot 13
31 * │ │ │ │ │ └─ Pivot 12
39 * Internal (non-leaf) nodes contain pointers to other nodes.
40 * Leaf nodes contain entries.
42 * The location of interest is often referred to as an offset. All offsets have
43 * a slot, but the last offset has an implied pivot from the node above (or
44 * UINT_MAX for the root node.
46 * Ranges complicate certain write activities. When modifying any of
47 * the B-tree variants, it is known that one entry will either be added or
48 * deleted. When modifying the Maple Tree, one store operation may overwrite
49 * the entire data set, or one half of the tree, or the middle half of the tree.
54 #include <linux/maple_tree.h>
55 #include <linux/xarray.h>
56 #include <linux/types.h>
57 #include <linux/export.h>
58 #include <linux/slab.h>
59 #include <linux/limits.h>
60 #include <asm/barrier.h>
62 #define CREATE_TRACE_POINTS
63 #include <trace/events/maple_tree.h>
65 #define MA_ROOT_PARENT 1
69 * * MA_STATE_BULK - Bulk insert mode
70 * * MA_STATE_REBALANCE - Indicate a rebalance during bulk insert
71 * * MA_STATE_PREALLOC - Preallocated nodes, WARN_ON allocation
73 #define MA_STATE_BULK 1
74 #define MA_STATE_REBALANCE 2
75 #define MA_STATE_PREALLOC 4
77 #define ma_parent_ptr(x) ((struct maple_pnode *)(x))
78 #define ma_mnode_ptr(x) ((struct maple_node *)(x))
79 #define ma_enode_ptr(x) ((struct maple_enode *)(x))
80 static struct kmem_cache *maple_node_cache;
82 #ifdef CONFIG_DEBUG_MAPLE_TREE
83 static const unsigned long mt_max[] = {
84 [maple_dense] = MAPLE_NODE_SLOTS,
85 [maple_leaf_64] = ULONG_MAX,
86 [maple_range_64] = ULONG_MAX,
87 [maple_arange_64] = ULONG_MAX,
89 #define mt_node_max(x) mt_max[mte_node_type(x)]
92 static const unsigned char mt_slots[] = {
93 [maple_dense] = MAPLE_NODE_SLOTS,
94 [maple_leaf_64] = MAPLE_RANGE64_SLOTS,
95 [maple_range_64] = MAPLE_RANGE64_SLOTS,
96 [maple_arange_64] = MAPLE_ARANGE64_SLOTS,
98 #define mt_slot_count(x) mt_slots[mte_node_type(x)]
100 static const unsigned char mt_pivots[] = {
102 [maple_leaf_64] = MAPLE_RANGE64_SLOTS - 1,
103 [maple_range_64] = MAPLE_RANGE64_SLOTS - 1,
104 [maple_arange_64] = MAPLE_ARANGE64_SLOTS - 1,
106 #define mt_pivot_count(x) mt_pivots[mte_node_type(x)]
108 static const unsigned char mt_min_slots[] = {
109 [maple_dense] = MAPLE_NODE_SLOTS / 2,
110 [maple_leaf_64] = (MAPLE_RANGE64_SLOTS / 2) - 2,
111 [maple_range_64] = (MAPLE_RANGE64_SLOTS / 2) - 2,
112 [maple_arange_64] = (MAPLE_ARANGE64_SLOTS / 2) - 1,
114 #define mt_min_slot_count(x) mt_min_slots[mte_node_type(x)]
116 #define MAPLE_BIG_NODE_SLOTS (MAPLE_RANGE64_SLOTS * 2 + 2)
117 #define MAPLE_BIG_NODE_GAPS (MAPLE_ARANGE64_SLOTS * 2 + 1)
119 struct maple_big_node {
120 struct maple_pnode *parent;
121 unsigned long pivot[MAPLE_BIG_NODE_SLOTS - 1];
123 struct maple_enode *slot[MAPLE_BIG_NODE_SLOTS];
125 unsigned long padding[MAPLE_BIG_NODE_GAPS];
126 unsigned long gap[MAPLE_BIG_NODE_GAPS];
130 enum maple_type type;
134 * The maple_subtree_state is used to build a tree to replace a segment of an
135 * existing tree in a more atomic way. Any walkers of the older tree will hit a
136 * dead node and restart on updates.
138 struct maple_subtree_state {
139 struct ma_state *orig_l; /* Original left side of subtree */
140 struct ma_state *orig_r; /* Original right side of subtree */
141 struct ma_state *l; /* New left side of subtree */
142 struct ma_state *m; /* New middle of subtree (rare) */
143 struct ma_state *r; /* New right side of subtree */
144 struct ma_topiary *free; /* nodes to be freed */
145 struct ma_topiary *destroy; /* Nodes to be destroyed (walked and freed) */
146 struct maple_big_node *bn;
150 static inline struct maple_node *mt_alloc_one(gfp_t gfp)
152 return kmem_cache_alloc(maple_node_cache, gfp | __GFP_ZERO);
155 static inline int mt_alloc_bulk(gfp_t gfp, size_t size, void **nodes)
157 return kmem_cache_alloc_bulk(maple_node_cache, gfp | __GFP_ZERO, size,
161 static inline void mt_free_bulk(size_t size, void __rcu **nodes)
163 kmem_cache_free_bulk(maple_node_cache, size, (void **)nodes);
166 static void mt_free_rcu(struct rcu_head *head)
168 struct maple_node *node = container_of(head, struct maple_node, rcu);
170 kmem_cache_free(maple_node_cache, node);
174 * ma_free_rcu() - Use rcu callback to free a maple node
175 * @node: The node to free
177 * The maple tree uses the parent pointer to indicate this node is no longer in
178 * use and will be freed.
180 static void ma_free_rcu(struct maple_node *node)
182 node->parent = ma_parent_ptr(node);
183 call_rcu(&node->rcu, mt_free_rcu);
187 static void mas_set_height(struct ma_state *mas)
189 unsigned int new_flags = mas->tree->ma_flags;
191 new_flags &= ~MT_FLAGS_HEIGHT_MASK;
192 BUG_ON(mas->depth > MAPLE_HEIGHT_MAX);
193 new_flags |= mas->depth << MT_FLAGS_HEIGHT_OFFSET;
194 mas->tree->ma_flags = new_flags;
197 static unsigned int mas_mt_height(struct ma_state *mas)
199 return mt_height(mas->tree);
202 static inline enum maple_type mte_node_type(const struct maple_enode *entry)
204 return ((unsigned long)entry >> MAPLE_NODE_TYPE_SHIFT) &
205 MAPLE_NODE_TYPE_MASK;
208 static inline bool ma_is_dense(const enum maple_type type)
210 return type < maple_leaf_64;
213 static inline bool ma_is_leaf(const enum maple_type type)
215 return type < maple_range_64;
218 static inline bool mte_is_leaf(const struct maple_enode *entry)
220 return ma_is_leaf(mte_node_type(entry));
224 * We also reserve values with the bottom two bits set to '10' which are
227 static inline bool mt_is_reserved(const void *entry)
229 return ((unsigned long)entry < MAPLE_RESERVED_RANGE) &&
230 xa_is_internal(entry);
233 static inline void mas_set_err(struct ma_state *mas, long err)
235 mas->node = MA_ERROR(err);
238 static inline bool mas_is_ptr(struct ma_state *mas)
240 return mas->node == MAS_ROOT;
243 static inline bool mas_is_start(struct ma_state *mas)
245 return mas->node == MAS_START;
248 bool mas_is_err(struct ma_state *mas)
250 return xa_is_err(mas->node);
253 static inline bool mas_searchable(struct ma_state *mas)
255 if (mas_is_none(mas))
264 static inline struct maple_node *mte_to_node(const struct maple_enode *entry)
266 return (struct maple_node *)((unsigned long)entry & ~MAPLE_NODE_MASK);
270 * mte_to_mat() - Convert a maple encoded node to a maple topiary node.
271 * @entry: The maple encoded node
273 * Return: a maple topiary pointer
275 static inline struct maple_topiary *mte_to_mat(const struct maple_enode *entry)
277 return (struct maple_topiary *)
278 ((unsigned long)entry & ~MAPLE_NODE_MASK);
282 * mas_mn() - Get the maple state node.
283 * @mas: The maple state
285 * Return: the maple node (not encoded - bare pointer).
287 static inline struct maple_node *mas_mn(const struct ma_state *mas)
289 return mte_to_node(mas->node);
293 * mte_set_node_dead() - Set a maple encoded node as dead.
294 * @mn: The maple encoded node.
296 static inline void mte_set_node_dead(struct maple_enode *mn)
298 mte_to_node(mn)->parent = ma_parent_ptr(mte_to_node(mn));
299 smp_wmb(); /* Needed for RCU */
302 /* Bit 1 indicates the root is a node */
303 #define MAPLE_ROOT_NODE 0x02
304 /* maple_type stored bit 3-6 */
305 #define MAPLE_ENODE_TYPE_SHIFT 0x03
306 /* Bit 2 means a NULL somewhere below */
307 #define MAPLE_ENODE_NULL 0x04
309 static inline struct maple_enode *mt_mk_node(const struct maple_node *node,
310 enum maple_type type)
312 return (void *)((unsigned long)node |
313 (type << MAPLE_ENODE_TYPE_SHIFT) | MAPLE_ENODE_NULL);
316 static inline void *mte_mk_root(const struct maple_enode *node)
318 return (void *)((unsigned long)node | MAPLE_ROOT_NODE);
321 static inline void *mte_safe_root(const struct maple_enode *node)
323 return (void *)((unsigned long)node & ~MAPLE_ROOT_NODE);
326 static inline void mte_set_full(const struct maple_enode *node)
328 node = (void *)((unsigned long)node & ~MAPLE_ENODE_NULL);
331 static inline void mte_clear_full(const struct maple_enode *node)
333 node = (void *)((unsigned long)node | MAPLE_ENODE_NULL);
336 static inline bool ma_is_root(struct maple_node *node)
338 return ((unsigned long)node->parent & MA_ROOT_PARENT);
341 static inline bool mte_is_root(const struct maple_enode *node)
343 return ma_is_root(mte_to_node(node));
346 static inline bool mas_is_root_limits(const struct ma_state *mas)
348 return !mas->min && mas->max == ULONG_MAX;
351 static inline bool mt_is_alloc(struct maple_tree *mt)
353 return (mt->ma_flags & MT_FLAGS_ALLOC_RANGE);
358 * Excluding root, the parent pointer is 256B aligned like all other tree nodes.
359 * When storing a 32 or 64 bit values, the offset can fit into 5 bits. The 16
360 * bit values need an extra bit to store the offset. This extra bit comes from
361 * a reuse of the last bit in the node type. This is possible by using bit 1 to
362 * indicate if bit 2 is part of the type or the slot.
366 * 0x?00 = 16 bit nodes
367 * 0x010 = 32 bit nodes
368 * 0x110 = 64 bit nodes
370 * Slot size and alignment
372 * 0b?00 : 16 bit values, type in 0-1, slot in 2-7
373 * 0b010 : 32 bit values, type in 0-2, slot in 3-7
374 * 0b110 : 64 bit values, type in 0-2, slot in 3-7
377 #define MAPLE_PARENT_ROOT 0x01
379 #define MAPLE_PARENT_SLOT_SHIFT 0x03
380 #define MAPLE_PARENT_SLOT_MASK 0xF8
382 #define MAPLE_PARENT_16B_SLOT_SHIFT 0x02
383 #define MAPLE_PARENT_16B_SLOT_MASK 0xFC
385 #define MAPLE_PARENT_RANGE64 0x06
386 #define MAPLE_PARENT_RANGE32 0x04
387 #define MAPLE_PARENT_NOT_RANGE16 0x02
390 * mte_parent_shift() - Get the parent shift for the slot storage.
391 * @parent: The parent pointer cast as an unsigned long
392 * Return: The shift into that pointer to the star to of the slot
394 static inline unsigned long mte_parent_shift(unsigned long parent)
396 /* Note bit 1 == 0 means 16B */
397 if (likely(parent & MAPLE_PARENT_NOT_RANGE16))
398 return MAPLE_PARENT_SLOT_SHIFT;
400 return MAPLE_PARENT_16B_SLOT_SHIFT;
404 * mte_parent_slot_mask() - Get the slot mask for the parent.
405 * @parent: The parent pointer cast as an unsigned long.
406 * Return: The slot mask for that parent.
408 static inline unsigned long mte_parent_slot_mask(unsigned long parent)
410 /* Note bit 1 == 0 means 16B */
411 if (likely(parent & MAPLE_PARENT_NOT_RANGE16))
412 return MAPLE_PARENT_SLOT_MASK;
414 return MAPLE_PARENT_16B_SLOT_MASK;
418 * mas_parent_enum() - Return the maple_type of the parent from the stored
420 * @mas: The maple state
421 * @node: The maple_enode to extract the parent's enum
422 * Return: The node->parent maple_type
425 enum maple_type mte_parent_enum(struct maple_enode *p_enode,
426 struct maple_tree *mt)
428 unsigned long p_type;
430 p_type = (unsigned long)p_enode;
431 if (p_type & MAPLE_PARENT_ROOT)
432 return 0; /* Validated in the caller. */
434 p_type &= MAPLE_NODE_MASK;
435 p_type = p_type & ~(MAPLE_PARENT_ROOT | mte_parent_slot_mask(p_type));
438 case MAPLE_PARENT_RANGE64: /* or MAPLE_PARENT_ARANGE64 */
440 return maple_arange_64;
441 return maple_range_64;
448 enum maple_type mas_parent_enum(struct ma_state *mas, struct maple_enode *enode)
450 return mte_parent_enum(ma_enode_ptr(mte_to_node(enode)->parent), mas->tree);
454 * mte_set_parent() - Set the parent node and encode the slot
455 * @enode: The encoded maple node.
456 * @parent: The encoded maple node that is the parent of @enode.
457 * @slot: The slot that @enode resides in @parent.
459 * Slot number is encoded in the enode->parent bit 3-6 or 2-6, depending on the
463 void mte_set_parent(struct maple_enode *enode, const struct maple_enode *parent,
466 unsigned long val = (unsigned long) parent;
469 enum maple_type p_type = mte_node_type(parent);
471 BUG_ON(p_type == maple_dense);
472 BUG_ON(p_type == maple_leaf_64);
476 case maple_arange_64:
477 shift = MAPLE_PARENT_SLOT_SHIFT;
478 type = MAPLE_PARENT_RANGE64;
487 val &= ~MAPLE_NODE_MASK; /* Clear all node metadata in parent */
488 val |= (slot << shift) | type;
489 mte_to_node(enode)->parent = ma_parent_ptr(val);
493 * mte_parent_slot() - get the parent slot of @enode.
494 * @enode: The encoded maple node.
496 * Return: The slot in the parent node where @enode resides.
498 static inline unsigned int mte_parent_slot(const struct maple_enode *enode)
500 unsigned long val = (unsigned long) mte_to_node(enode)->parent;
507 * Okay to use MAPLE_PARENT_16B_SLOT_MASK as the last bit will be lost
508 * by shift if the parent shift is MAPLE_PARENT_SLOT_SHIFT
510 return (val & MAPLE_PARENT_16B_SLOT_MASK) >> mte_parent_shift(val);
514 * mte_parent() - Get the parent of @node.
515 * @node: The encoded maple node.
517 * Return: The parent maple node.
519 static inline struct maple_node *mte_parent(const struct maple_enode *enode)
521 return (void *)((unsigned long)
522 (mte_to_node(enode)->parent) & ~MAPLE_NODE_MASK);
526 * ma_dead_node() - check if the @enode is dead.
527 * @enode: The encoded maple node
529 * Return: true if dead, false otherwise.
531 static inline bool ma_dead_node(const struct maple_node *node)
533 struct maple_node *parent = (void *)((unsigned long)
534 node->parent & ~MAPLE_NODE_MASK);
536 return (parent == node);
539 * mte_dead_node() - check if the @enode is dead.
540 * @enode: The encoded maple node
542 * Return: true if dead, false otherwise.
544 static inline bool mte_dead_node(const struct maple_enode *enode)
546 struct maple_node *parent, *node;
548 node = mte_to_node(enode);
549 parent = mte_parent(enode);
550 return (parent == node);
554 * mas_allocated() - Get the number of nodes allocated in a maple state.
555 * @mas: The maple state
557 * The ma_state alloc member is overloaded to hold a pointer to the first
558 * allocated node or to the number of requested nodes to allocate. If bit 0 is
559 * set, then the alloc contains the number of requested nodes. If there is an
560 * allocated node, then the total allocated nodes is in that node.
562 * Return: The total number of nodes allocated
564 static inline unsigned long mas_allocated(const struct ma_state *mas)
566 if (!mas->alloc || ((unsigned long)mas->alloc & 0x1))
569 return mas->alloc->total;
573 * mas_set_alloc_req() - Set the requested number of allocations.
574 * @mas: the maple state
575 * @count: the number of allocations.
577 * The requested number of allocations is either in the first allocated node,
578 * located in @mas->alloc->request_count, or directly in @mas->alloc if there is
579 * no allocated node. Set the request either in the node or do the necessary
580 * encoding to store in @mas->alloc directly.
582 static inline void mas_set_alloc_req(struct ma_state *mas, unsigned long count)
584 if (!mas->alloc || ((unsigned long)mas->alloc & 0x1)) {
588 mas->alloc = (struct maple_alloc *)(((count) << 1U) | 1U);
592 mas->alloc->request_count = count;
596 * mas_alloc_req() - get the requested number of allocations.
597 * @mas: The maple state
599 * The alloc count is either stored directly in @mas, or in
600 * @mas->alloc->request_count if there is at least one node allocated. Decode
601 * the request count if it's stored directly in @mas->alloc.
603 * Return: The allocation request count.
605 static inline unsigned int mas_alloc_req(const struct ma_state *mas)
607 if ((unsigned long)mas->alloc & 0x1)
608 return (unsigned long)(mas->alloc) >> 1;
610 return mas->alloc->request_count;
615 * ma_pivots() - Get a pointer to the maple node pivots.
616 * @node - the maple node
617 * @type - the node type
619 * Return: A pointer to the maple node pivots
621 static inline unsigned long *ma_pivots(struct maple_node *node,
622 enum maple_type type)
625 case maple_arange_64:
626 return node->ma64.pivot;
629 return node->mr64.pivot;
637 * ma_gaps() - Get a pointer to the maple node gaps.
638 * @node - the maple node
639 * @type - the node type
641 * Return: A pointer to the maple node gaps
643 static inline unsigned long *ma_gaps(struct maple_node *node,
644 enum maple_type type)
647 case maple_arange_64:
648 return node->ma64.gap;
658 * mte_pivot() - Get the pivot at @piv of the maple encoded node.
659 * @mn: The maple encoded node.
662 * Return: the pivot at @piv of @mn.
664 static inline unsigned long mte_pivot(const struct maple_enode *mn,
667 struct maple_node *node = mte_to_node(mn);
668 enum maple_type type = mte_node_type(mn);
670 if (piv >= mt_pivots[type]) {
675 case maple_arange_64:
676 return node->ma64.pivot[piv];
679 return node->mr64.pivot[piv];
687 * mas_safe_pivot() - get the pivot at @piv or mas->max.
688 * @mas: The maple state
689 * @pivots: The pointer to the maple node pivots
690 * @piv: The pivot to fetch
691 * @type: The maple node type
693 * Return: The pivot at @piv within the limit of the @pivots array, @mas->max
696 static inline unsigned long
697 mas_safe_pivot(const struct ma_state *mas, unsigned long *pivots,
698 unsigned char piv, enum maple_type type)
700 if (piv >= mt_pivots[type])
707 * mas_safe_min() - Return the minimum for a given offset.
708 * @mas: The maple state
709 * @pivots: The pointer to the maple node pivots
710 * @offset: The offset into the pivot array
712 * Return: The minimum range value that is contained in @offset.
714 static inline unsigned long
715 mas_safe_min(struct ma_state *mas, unsigned long *pivots, unsigned char offset)
718 return pivots[offset - 1] + 1;
724 * mas_logical_pivot() - Get the logical pivot of a given offset.
725 * @mas: The maple state
726 * @pivots: The pointer to the maple node pivots
727 * @offset: The offset into the pivot array
728 * @type: The maple node type
730 * When there is no value at a pivot (beyond the end of the data), then the
731 * pivot is actually @mas->max.
733 * Return: the logical pivot of a given @offset.
735 static inline unsigned long
736 mas_logical_pivot(struct ma_state *mas, unsigned long *pivots,
737 unsigned char offset, enum maple_type type)
739 unsigned long lpiv = mas_safe_pivot(mas, pivots, offset, type);
751 * mte_set_pivot() - Set a pivot to a value in an encoded maple node.
752 * @mn: The encoded maple node
753 * @piv: The pivot offset
754 * @val: The value of the pivot
756 static inline void mte_set_pivot(struct maple_enode *mn, unsigned char piv,
759 struct maple_node *node = mte_to_node(mn);
760 enum maple_type type = mte_node_type(mn);
762 BUG_ON(piv >= mt_pivots[type]);
767 node->mr64.pivot[piv] = val;
769 case maple_arange_64:
770 node->ma64.pivot[piv] = val;
779 * ma_slots() - Get a pointer to the maple node slots.
780 * @mn: The maple node
781 * @mt: The maple node type
783 * Return: A pointer to the maple node slots
785 static inline void __rcu **ma_slots(struct maple_node *mn, enum maple_type mt)
789 case maple_arange_64:
790 return mn->ma64.slot;
793 return mn->mr64.slot;
799 static inline bool mt_locked(const struct maple_tree *mt)
801 return mt_external_lock(mt) ? mt_lock_is_held(mt) :
802 lockdep_is_held(&mt->ma_lock);
805 static inline void *mt_slot(const struct maple_tree *mt,
806 void __rcu **slots, unsigned char offset)
808 return rcu_dereference_check(slots[offset], mt_locked(mt));
812 * mas_slot_locked() - Get the slot value when holding the maple tree lock.
813 * @mas: The maple state
814 * @slots: The pointer to the slots
815 * @offset: The offset into the slots array to fetch
817 * Return: The entry stored in @slots at the @offset.
819 static inline void *mas_slot_locked(struct ma_state *mas, void __rcu **slots,
820 unsigned char offset)
822 return rcu_dereference_protected(slots[offset], mt_locked(mas->tree));
826 * mas_slot() - Get the slot value when not holding the maple tree lock.
827 * @mas: The maple state
828 * @slots: The pointer to the slots
829 * @offset: The offset into the slots array to fetch
831 * Return: The entry stored in @slots at the @offset
833 static inline void *mas_slot(struct ma_state *mas, void __rcu **slots,
834 unsigned char offset)
836 return mt_slot(mas->tree, slots, offset);
840 * mas_root() - Get the maple tree root.
841 * @mas: The maple state.
843 * Return: The pointer to the root of the tree
845 static inline void *mas_root(struct ma_state *mas)
847 return rcu_dereference_check(mas->tree->ma_root, mt_locked(mas->tree));
850 static inline void *mt_root_locked(struct maple_tree *mt)
852 return rcu_dereference_protected(mt->ma_root, mt_locked(mt));
856 * mas_root_locked() - Get the maple tree root when holding the maple tree lock.
857 * @mas: The maple state.
859 * Return: The pointer to the root of the tree
861 static inline void *mas_root_locked(struct ma_state *mas)
863 return mt_root_locked(mas->tree);
866 static inline struct maple_metadata *ma_meta(struct maple_node *mn,
870 case maple_arange_64:
871 return &mn->ma64.meta;
873 return &mn->mr64.meta;
878 * ma_set_meta() - Set the metadata information of a node.
879 * @mn: The maple node
880 * @mt: The maple node type
881 * @offset: The offset of the highest sub-gap in this node.
882 * @end: The end of the data in this node.
884 static inline void ma_set_meta(struct maple_node *mn, enum maple_type mt,
885 unsigned char offset, unsigned char end)
887 struct maple_metadata *meta = ma_meta(mn, mt);
894 * ma_meta_end() - Get the data end of a node from the metadata
895 * @mn: The maple node
896 * @mt: The maple node type
898 static inline unsigned char ma_meta_end(struct maple_node *mn,
901 struct maple_metadata *meta = ma_meta(mn, mt);
907 * ma_meta_gap() - Get the largest gap location of a node from the metadata
908 * @mn: The maple node
909 * @mt: The maple node type
911 static inline unsigned char ma_meta_gap(struct maple_node *mn,
914 BUG_ON(mt != maple_arange_64);
916 return mn->ma64.meta.gap;
920 * ma_set_meta_gap() - Set the largest gap location in a nodes metadata
921 * @mn: The maple node
922 * @mn: The maple node type
923 * @offset: The location of the largest gap.
925 static inline void ma_set_meta_gap(struct maple_node *mn, enum maple_type mt,
926 unsigned char offset)
929 struct maple_metadata *meta = ma_meta(mn, mt);
935 * mat_add() - Add a @dead_enode to the ma_topiary of a list of dead nodes.
936 * @mat - the ma_topiary, a linked list of dead nodes.
937 * @dead_enode - the node to be marked as dead and added to the tail of the list
939 * Add the @dead_enode to the linked list in @mat.
941 static inline void mat_add(struct ma_topiary *mat,
942 struct maple_enode *dead_enode)
944 mte_set_node_dead(dead_enode);
945 mte_to_mat(dead_enode)->next = NULL;
947 mat->tail = mat->head = dead_enode;
951 mte_to_mat(mat->tail)->next = dead_enode;
952 mat->tail = dead_enode;
955 static void mte_destroy_walk(struct maple_enode *, struct maple_tree *);
956 static inline void mas_free(struct ma_state *mas, struct maple_enode *used);
959 * mas_mat_free() - Free all nodes in a dead list.
960 * @mas - the maple state
961 * @mat - the ma_topiary linked list of dead nodes to free.
963 * Free walk a dead list.
965 static void mas_mat_free(struct ma_state *mas, struct ma_topiary *mat)
967 struct maple_enode *next;
970 next = mte_to_mat(mat->head)->next;
971 mas_free(mas, mat->head);
977 * mas_mat_destroy() - Free all nodes and subtrees in a dead list.
978 * @mas - the maple state
979 * @mat - the ma_topiary linked list of dead nodes to free.
981 * Destroy walk a dead list.
983 static void mas_mat_destroy(struct ma_state *mas, struct ma_topiary *mat)
985 struct maple_enode *next;
988 next = mte_to_mat(mat->head)->next;
989 mte_destroy_walk(mat->head, mat->mtree);
994 * mas_descend() - Descend into the slot stored in the ma_state.
995 * @mas - the maple state.
997 * Note: Not RCU safe, only use in write side or debug code.
999 static inline void mas_descend(struct ma_state *mas)
1001 enum maple_type type;
1002 unsigned long *pivots;
1003 struct maple_node *node;
1007 type = mte_node_type(mas->node);
1008 pivots = ma_pivots(node, type);
1009 slots = ma_slots(node, type);
1012 mas->min = pivots[mas->offset - 1] + 1;
1013 mas->max = mas_safe_pivot(mas, pivots, mas->offset, type);
1014 mas->node = mas_slot(mas, slots, mas->offset);
1018 * mte_set_gap() - Set a maple node gap.
1019 * @mn: The encoded maple node
1020 * @gap: The offset of the gap to set
1021 * @val: The gap value
1023 static inline void mte_set_gap(const struct maple_enode *mn,
1024 unsigned char gap, unsigned long val)
1026 switch (mte_node_type(mn)) {
1029 case maple_arange_64:
1030 mte_to_node(mn)->ma64.gap[gap] = val;
1036 * mas_ascend() - Walk up a level of the tree.
1037 * @mas: The maple state
1039 * Sets the @mas->max and @mas->min to the correct values when walking up. This
1040 * may cause several levels of walking up to find the correct min and max.
1041 * May find a dead node which will cause a premature return.
1042 * Return: 1 on dead node, 0 otherwise
1044 static int mas_ascend(struct ma_state *mas)
1046 struct maple_enode *p_enode; /* parent enode. */
1047 struct maple_enode *a_enode; /* ancestor enode. */
1048 struct maple_node *a_node; /* ancestor node. */
1049 struct maple_node *p_node; /* parent node. */
1050 unsigned char a_slot;
1051 enum maple_type a_type;
1052 unsigned long min, max;
1053 unsigned long *pivots;
1054 unsigned char offset;
1055 bool set_max = false, set_min = false;
1057 a_node = mas_mn(mas);
1058 if (ma_is_root(a_node)) {
1063 p_node = mte_parent(mas->node);
1064 if (unlikely(a_node == p_node))
1066 a_type = mas_parent_enum(mas, mas->node);
1067 offset = mte_parent_slot(mas->node);
1068 a_enode = mt_mk_node(p_node, a_type);
1070 /* Check to make sure all parent information is still accurate */
1071 if (p_node != mte_parent(mas->node))
1074 mas->node = a_enode;
1075 mas->offset = offset;
1077 if (mte_is_root(a_enode)) {
1078 mas->max = ULONG_MAX;
1087 a_type = mas_parent_enum(mas, p_enode);
1088 a_node = mte_parent(p_enode);
1089 a_slot = mte_parent_slot(p_enode);
1090 pivots = ma_pivots(a_node, a_type);
1091 a_enode = mt_mk_node(a_node, a_type);
1093 if (!set_min && a_slot) {
1095 min = pivots[a_slot - 1] + 1;
1098 if (!set_max && a_slot < mt_pivots[a_type]) {
1100 max = pivots[a_slot];
1103 if (unlikely(ma_dead_node(a_node)))
1106 if (unlikely(ma_is_root(a_node)))
1109 } while (!set_min || !set_max);
1117 * mas_pop_node() - Get a previously allocated maple node from the maple state.
1118 * @mas: The maple state
1120 * Return: A pointer to a maple node.
1122 static inline struct maple_node *mas_pop_node(struct ma_state *mas)
1124 struct maple_alloc *ret, *node = mas->alloc;
1125 unsigned long total = mas_allocated(mas);
1127 /* nothing or a request pending. */
1128 if (unlikely(!total))
1132 /* single allocation in this ma_state */
1138 if (!node->node_count) {
1139 /* Single allocation in this node. */
1140 mas->alloc = node->slot[0];
1141 node->slot[0] = NULL;
1142 mas->alloc->total = node->total - 1;
1148 ret = node->slot[node->node_count];
1149 node->slot[node->node_count--] = NULL;
1154 ret->node_count = 0;
1155 if (ret->request_count) {
1156 mas_set_alloc_req(mas, ret->request_count + 1);
1157 ret->request_count = 0;
1159 return (struct maple_node *)ret;
1163 * mas_push_node() - Push a node back on the maple state allocation.
1164 * @mas: The maple state
1165 * @used: The used maple node
1167 * Stores the maple node back into @mas->alloc for reuse. Updates allocated and
1168 * requested node count as necessary.
1170 static inline void mas_push_node(struct ma_state *mas, struct maple_node *used)
1172 struct maple_alloc *reuse = (struct maple_alloc *)used;
1173 struct maple_alloc *head = mas->alloc;
1174 unsigned long count;
1175 unsigned int requested = mas_alloc_req(mas);
1177 memset(reuse, 0, sizeof(*reuse));
1178 count = mas_allocated(mas);
1180 if (count && (head->node_count < MAPLE_ALLOC_SLOTS - 1)) {
1183 head->slot[head->node_count] = reuse;
1189 if ((head) && !((unsigned long)head & 0x1)) {
1190 head->request_count = 0;
1191 reuse->slot[0] = head;
1192 reuse->total += head->total;
1198 mas_set_alloc_req(mas, requested - 1);
1202 * mas_alloc_nodes() - Allocate nodes into a maple state
1203 * @mas: The maple state
1204 * @gfp: The GFP Flags
1206 static inline void mas_alloc_nodes(struct ma_state *mas, gfp_t gfp)
1208 struct maple_alloc *node;
1209 unsigned long allocated = mas_allocated(mas);
1210 unsigned long success = allocated;
1211 unsigned int requested = mas_alloc_req(mas);
1213 void **slots = NULL;
1214 unsigned int max_req = 0;
1219 mas_set_alloc_req(mas, 0);
1220 if (mas->mas_flags & MA_STATE_PREALLOC) {
1223 WARN_ON(!allocated);
1226 if (!allocated || mas->alloc->node_count == MAPLE_ALLOC_SLOTS - 1) {
1227 node = (struct maple_alloc *)mt_alloc_one(gfp);
1232 node->slot[0] = mas->alloc;
1241 max_req = MAPLE_ALLOC_SLOTS;
1242 if (node->slot[0]) {
1243 unsigned int offset = node->node_count + 1;
1245 slots = (void **)&node->slot[offset];
1248 slots = (void **)&node->slot;
1251 max_req = min(requested, max_req);
1252 count = mt_alloc_bulk(gfp, max_req, slots);
1256 node->node_count += count;
1258 if (slots == (void **)&node->slot)
1262 node = node->slot[0];
1265 mas->alloc->total = success;
1269 /* Clean up potential freed allocations on bulk failure */
1270 memset(slots, 0, max_req * sizeof(unsigned long));
1272 mas_set_alloc_req(mas, requested);
1273 if (mas->alloc && !(((unsigned long)mas->alloc & 0x1)))
1274 mas->alloc->total = success;
1275 mas_set_err(mas, -ENOMEM);
1281 * mas_free() - Free an encoded maple node
1282 * @mas: The maple state
1283 * @used: The encoded maple node to free.
1285 * Uses rcu free if necessary, pushes @used back on the maple state allocations
1288 static inline void mas_free(struct ma_state *mas, struct maple_enode *used)
1290 struct maple_node *tmp = mte_to_node(used);
1292 if (mt_in_rcu(mas->tree))
1295 mas_push_node(mas, tmp);
1299 * mas_node_count() - Check if enough nodes are allocated and request more if
1300 * there is not enough nodes.
1301 * @mas: The maple state
1302 * @count: The number of nodes needed
1303 * @gfp: the gfp flags
1305 static void mas_node_count_gfp(struct ma_state *mas, int count, gfp_t gfp)
1307 unsigned long allocated = mas_allocated(mas);
1309 if (allocated < count) {
1310 mas_set_alloc_req(mas, count - allocated);
1311 mas_alloc_nodes(mas, gfp);
1316 * mas_node_count() - Check if enough nodes are allocated and request more if
1317 * there is not enough nodes.
1318 * @mas: The maple state
1319 * @count: The number of nodes needed
1321 * Note: Uses GFP_NOWAIT | __GFP_NOWARN for gfp flags.
1323 static void mas_node_count(struct ma_state *mas, int count)
1325 return mas_node_count_gfp(mas, count, GFP_NOWAIT | __GFP_NOWARN);
1329 * mas_start() - Sets up maple state for operations.
1330 * @mas: The maple state.
1332 * If mas->node == MAS_START, then set the min, max, depth, and offset to
1336 * - If mas->node is an error or not MAS_START, return NULL.
1337 * - If it's an empty tree: NULL & mas->node == MAS_NONE
1338 * - If it's a single entry: The entry & mas->node == MAS_ROOT
1339 * - If it's a tree: NULL & mas->node == safe root node.
1341 static inline struct maple_enode *mas_start(struct ma_state *mas)
1343 if (likely(mas_is_start(mas))) {
1344 struct maple_enode *root;
1346 mas->node = MAS_NONE;
1348 mas->max = ULONG_MAX;
1352 root = mas_root(mas);
1353 /* Tree with nodes */
1354 if (likely(xa_is_node(root))) {
1356 mas->node = mte_safe_root(root);
1361 if (unlikely(!root)) {
1362 mas->offset = MAPLE_NODE_SLOTS;
1366 /* Single entry tree */
1367 mas->node = MAS_ROOT;
1368 mas->offset = MAPLE_NODE_SLOTS;
1370 /* Single entry tree. */
1381 * ma_data_end() - Find the end of the data in a node.
1382 * @node: The maple node
1383 * @type: The maple node type
1384 * @pivots: The array of pivots in the node
1385 * @max: The maximum value in the node
1387 * Uses metadata to find the end of the data when possible.
1388 * Return: The zero indexed last slot with data (may be null).
1390 static inline unsigned char ma_data_end(struct maple_node *node,
1391 enum maple_type type,
1392 unsigned long *pivots,
1395 unsigned char offset;
1397 if (type == maple_arange_64)
1398 return ma_meta_end(node, type);
1400 offset = mt_pivots[type] - 1;
1401 if (likely(!pivots[offset]))
1402 return ma_meta_end(node, type);
1404 if (likely(pivots[offset] == max))
1407 return mt_pivots[type];
1411 * mas_data_end() - Find the end of the data (slot).
1412 * @mas: the maple state
1414 * This method is optimized to check the metadata of a node if the node type
1415 * supports data end metadata.
1417 * Return: The zero indexed last slot with data (may be null).
1419 static inline unsigned char mas_data_end(struct ma_state *mas)
1421 enum maple_type type;
1422 struct maple_node *node;
1423 unsigned char offset;
1424 unsigned long *pivots;
1426 type = mte_node_type(mas->node);
1428 if (type == maple_arange_64)
1429 return ma_meta_end(node, type);
1431 pivots = ma_pivots(node, type);
1432 offset = mt_pivots[type] - 1;
1433 if (likely(!pivots[offset]))
1434 return ma_meta_end(node, type);
1436 if (likely(pivots[offset] == mas->max))
1439 return mt_pivots[type];
1443 * mas_leaf_max_gap() - Returns the largest gap in a leaf node
1444 * @mas - the maple state
1446 * Return: The maximum gap in the leaf.
1448 static unsigned long mas_leaf_max_gap(struct ma_state *mas)
1451 unsigned long pstart, gap, max_gap;
1452 struct maple_node *mn;
1453 unsigned long *pivots;
1456 unsigned char max_piv;
1458 mt = mte_node_type(mas->node);
1460 slots = ma_slots(mn, mt);
1462 if (unlikely(ma_is_dense(mt))) {
1464 for (i = 0; i < mt_slots[mt]; i++) {
1479 * Check the first implied pivot optimizes the loop below and slot 1 may
1480 * be skipped if there is a gap in slot 0.
1482 pivots = ma_pivots(mn, mt);
1483 if (likely(!slots[0])) {
1484 max_gap = pivots[0] - mas->min + 1;
1490 /* reduce max_piv as the special case is checked before the loop */
1491 max_piv = ma_data_end(mn, mt, pivots, mas->max) - 1;
1493 * Check end implied pivot which can only be a gap on the right most
1496 if (unlikely(mas->max == ULONG_MAX) && !slots[max_piv + 1]) {
1497 gap = ULONG_MAX - pivots[max_piv];
1502 for (; i <= max_piv; i++) {
1503 /* data == no gap. */
1504 if (likely(slots[i]))
1507 pstart = pivots[i - 1];
1508 gap = pivots[i] - pstart;
1512 /* There cannot be two gaps in a row. */
1519 * ma_max_gap() - Get the maximum gap in a maple node (non-leaf)
1520 * @node: The maple node
1521 * @gaps: The pointer to the gaps
1522 * @mt: The maple node type
1523 * @*off: Pointer to store the offset location of the gap.
1525 * Uses the metadata data end to scan backwards across set gaps.
1527 * Return: The maximum gap value
1529 static inline unsigned long
1530 ma_max_gap(struct maple_node *node, unsigned long *gaps, enum maple_type mt,
1533 unsigned char offset, i;
1534 unsigned long max_gap = 0;
1536 i = offset = ma_meta_end(node, mt);
1538 if (gaps[i] > max_gap) {
1549 * mas_max_gap() - find the largest gap in a non-leaf node and set the slot.
1550 * @mas: The maple state.
1552 * If the metadata gap is set to MAPLE_ARANGE64_META_MAX, there is no gap.
1554 * Return: The gap value.
1556 static inline unsigned long mas_max_gap(struct ma_state *mas)
1558 unsigned long *gaps;
1559 unsigned char offset;
1561 struct maple_node *node;
1563 mt = mte_node_type(mas->node);
1565 return mas_leaf_max_gap(mas);
1568 offset = ma_meta_gap(node, mt);
1569 if (offset == MAPLE_ARANGE64_META_MAX)
1572 gaps = ma_gaps(node, mt);
1573 return gaps[offset];
1577 * mas_parent_gap() - Set the parent gap and any gaps above, as needed
1578 * @mas: The maple state
1579 * @offset: The gap offset in the parent to set
1580 * @new: The new gap value.
1582 * Set the parent gap then continue to set the gap upwards, using the metadata
1583 * of the parent to see if it is necessary to check the node above.
1585 static inline void mas_parent_gap(struct ma_state *mas, unsigned char offset,
1588 unsigned long meta_gap = 0;
1589 struct maple_node *pnode;
1590 struct maple_enode *penode;
1591 unsigned long *pgaps;
1592 unsigned char meta_offset;
1593 enum maple_type pmt;
1595 pnode = mte_parent(mas->node);
1596 pmt = mas_parent_enum(mas, mas->node);
1597 penode = mt_mk_node(pnode, pmt);
1598 pgaps = ma_gaps(pnode, pmt);
1601 meta_offset = ma_meta_gap(pnode, pmt);
1602 if (meta_offset == MAPLE_ARANGE64_META_MAX)
1605 meta_gap = pgaps[meta_offset];
1607 pgaps[offset] = new;
1609 if (meta_gap == new)
1612 if (offset != meta_offset) {
1616 ma_set_meta_gap(pnode, pmt, offset);
1617 } else if (new < meta_gap) {
1619 new = ma_max_gap(pnode, pgaps, pmt, &meta_offset);
1620 ma_set_meta_gap(pnode, pmt, meta_offset);
1623 if (ma_is_root(pnode))
1626 /* Go to the parent node. */
1627 pnode = mte_parent(penode);
1628 pmt = mas_parent_enum(mas, penode);
1629 pgaps = ma_gaps(pnode, pmt);
1630 offset = mte_parent_slot(penode);
1631 penode = mt_mk_node(pnode, pmt);
1636 * mas_update_gap() - Update a nodes gaps and propagate up if necessary.
1637 * @mas - the maple state.
1639 static inline void mas_update_gap(struct ma_state *mas)
1641 unsigned char pslot;
1642 unsigned long p_gap;
1643 unsigned long max_gap;
1645 if (!mt_is_alloc(mas->tree))
1648 if (mte_is_root(mas->node))
1651 max_gap = mas_max_gap(mas);
1653 pslot = mte_parent_slot(mas->node);
1654 p_gap = ma_gaps(mte_parent(mas->node),
1655 mas_parent_enum(mas, mas->node))[pslot];
1657 if (p_gap != max_gap)
1658 mas_parent_gap(mas, pslot, max_gap);
1662 * mas_adopt_children() - Set the parent pointer of all nodes in @parent to
1663 * @parent with the slot encoded.
1664 * @mas - the maple state (for the tree)
1665 * @parent - the maple encoded node containing the children.
1667 static inline void mas_adopt_children(struct ma_state *mas,
1668 struct maple_enode *parent)
1670 enum maple_type type = mte_node_type(parent);
1671 struct maple_node *node = mas_mn(mas);
1672 void __rcu **slots = ma_slots(node, type);
1673 unsigned long *pivots = ma_pivots(node, type);
1674 struct maple_enode *child;
1675 unsigned char offset;
1677 offset = ma_data_end(node, type, pivots, mas->max);
1679 child = mas_slot_locked(mas, slots, offset);
1680 mte_set_parent(child, parent, offset);
1685 * mas_replace() - Replace a maple node in the tree with mas->node. Uses the
1686 * parent encoding to locate the maple node in the tree.
1687 * @mas - the ma_state to use for operations.
1688 * @advanced - boolean to adopt the child nodes and free the old node (false) or
1689 * leave the node (true) and handle the adoption and free elsewhere.
1691 static inline void mas_replace(struct ma_state *mas, bool advanced)
1692 __must_hold(mas->tree->lock)
1694 struct maple_node *mn = mas_mn(mas);
1695 struct maple_enode *old_enode;
1696 unsigned char offset = 0;
1697 void __rcu **slots = NULL;
1699 if (ma_is_root(mn)) {
1700 old_enode = mas_root_locked(mas);
1702 offset = mte_parent_slot(mas->node);
1703 slots = ma_slots(mte_parent(mas->node),
1704 mas_parent_enum(mas, mas->node));
1705 old_enode = mas_slot_locked(mas, slots, offset);
1708 if (!advanced && !mte_is_leaf(mas->node))
1709 mas_adopt_children(mas, mas->node);
1711 if (mte_is_root(mas->node)) {
1712 mn->parent = ma_parent_ptr(
1713 ((unsigned long)mas->tree | MA_ROOT_PARENT));
1714 rcu_assign_pointer(mas->tree->ma_root, mte_mk_root(mas->node));
1715 mas_set_height(mas);
1717 rcu_assign_pointer(slots[offset], mas->node);
1721 mas_free(mas, old_enode);
1725 * mas_new_child() - Find the new child of a node.
1726 * @mas: the maple state
1727 * @child: the maple state to store the child.
1729 static inline bool mas_new_child(struct ma_state *mas, struct ma_state *child)
1730 __must_hold(mas->tree->lock)
1733 unsigned char offset;
1735 unsigned long *pivots;
1736 struct maple_enode *entry;
1737 struct maple_node *node;
1740 mt = mte_node_type(mas->node);
1742 slots = ma_slots(node, mt);
1743 pivots = ma_pivots(node, mt);
1744 end = ma_data_end(node, mt, pivots, mas->max);
1745 for (offset = mas->offset; offset <= end; offset++) {
1746 entry = mas_slot_locked(mas, slots, offset);
1747 if (mte_parent(entry) == node) {
1749 mas->offset = offset + 1;
1750 child->offset = offset;
1760 * mab_shift_right() - Shift the data in mab right. Note, does not clean out the
1761 * old data or set b_node->b_end.
1762 * @b_node: the maple_big_node
1763 * @shift: the shift count
1765 static inline void mab_shift_right(struct maple_big_node *b_node,
1766 unsigned char shift)
1768 unsigned long size = b_node->b_end * sizeof(unsigned long);
1770 memmove(b_node->pivot + shift, b_node->pivot, size);
1771 memmove(b_node->slot + shift, b_node->slot, size);
1772 if (b_node->type == maple_arange_64)
1773 memmove(b_node->gap + shift, b_node->gap, size);
1777 * mab_middle_node() - Check if a middle node is needed (unlikely)
1778 * @b_node: the maple_big_node that contains the data.
1779 * @size: the amount of data in the b_node
1780 * @split: the potential split location
1781 * @slot_count: the size that can be stored in a single node being considered.
1783 * Return: true if a middle node is required.
1785 static inline bool mab_middle_node(struct maple_big_node *b_node, int split,
1786 unsigned char slot_count)
1788 unsigned char size = b_node->b_end;
1790 if (size >= 2 * slot_count)
1793 if (!b_node->slot[split] && (size >= 2 * slot_count - 1))
1800 * mab_no_null_split() - ensure the split doesn't fall on a NULL
1801 * @b_node: the maple_big_node with the data
1802 * @split: the suggested split location
1803 * @slot_count: the number of slots in the node being considered.
1805 * Return: the split location.
1807 static inline int mab_no_null_split(struct maple_big_node *b_node,
1808 unsigned char split, unsigned char slot_count)
1810 if (!b_node->slot[split]) {
1812 * If the split is less than the max slot && the right side will
1813 * still be sufficient, then increment the split on NULL.
1815 if ((split < slot_count - 1) &&
1816 (b_node->b_end - split) > (mt_min_slots[b_node->type]))
1825 * mab_calc_split() - Calculate the split location and if there needs to be two
1827 * @bn: The maple_big_node with the data
1828 * @mid_split: The second split, if required. 0 otherwise.
1830 * Return: The first split location. The middle split is set in @mid_split.
1832 static inline int mab_calc_split(struct ma_state *mas,
1833 struct maple_big_node *bn, unsigned char *mid_split, unsigned long min)
1835 unsigned char b_end = bn->b_end;
1836 int split = b_end / 2; /* Assume equal split. */
1837 unsigned char slot_min, slot_count = mt_slots[bn->type];
1840 * To support gap tracking, all NULL entries are kept together and a node cannot
1841 * end on a NULL entry, with the exception of the left-most leaf. The
1842 * limitation means that the split of a node must be checked for this condition
1843 * and be able to put more data in one direction or the other.
1845 if (unlikely((mas->mas_flags & MA_STATE_BULK))) {
1847 split = b_end - mt_min_slots[bn->type];
1849 if (!ma_is_leaf(bn->type))
1852 mas->mas_flags |= MA_STATE_REBALANCE;
1853 if (!bn->slot[split])
1859 * Although extremely rare, it is possible to enter what is known as the 3-way
1860 * split scenario. The 3-way split comes about by means of a store of a range
1861 * that overwrites the end and beginning of two full nodes. The result is a set
1862 * of entries that cannot be stored in 2 nodes. Sometimes, these two nodes can
1863 * also be located in different parent nodes which are also full. This can
1864 * carry upwards all the way to the root in the worst case.
1866 if (unlikely(mab_middle_node(bn, split, slot_count))) {
1868 *mid_split = split * 2;
1870 slot_min = mt_min_slots[bn->type];
1874 * Avoid having a range less than the slot count unless it
1875 * causes one node to be deficient.
1876 * NOTE: mt_min_slots is 1 based, b_end and split are zero.
1878 while (((bn->pivot[split] - min) < slot_count - 1) &&
1879 (split < slot_count - 1) && (b_end - split > slot_min))
1883 /* Avoid ending a node on a NULL entry */
1884 split = mab_no_null_split(bn, split, slot_count);
1888 *mid_split = mab_no_null_split(bn, *mid_split, slot_count);
1894 * mas_mab_cp() - Copy data from a maple state inclusively to a maple_big_node
1895 * and set @b_node->b_end to the next free slot.
1896 * @mas: The maple state
1897 * @mas_start: The starting slot to copy
1898 * @mas_end: The end slot to copy (inclusively)
1899 * @b_node: The maple_big_node to place the data
1900 * @mab_start: The starting location in maple_big_node to store the data.
1902 static inline void mas_mab_cp(struct ma_state *mas, unsigned char mas_start,
1903 unsigned char mas_end, struct maple_big_node *b_node,
1904 unsigned char mab_start)
1907 struct maple_node *node;
1909 unsigned long *pivots, *gaps;
1910 int i = mas_start, j = mab_start;
1911 unsigned char piv_end;
1914 mt = mte_node_type(mas->node);
1915 pivots = ma_pivots(node, mt);
1917 b_node->pivot[j] = pivots[i++];
1918 if (unlikely(i > mas_end))
1923 piv_end = min(mas_end, mt_pivots[mt]);
1924 for (; i < piv_end; i++, j++) {
1925 b_node->pivot[j] = pivots[i];
1926 if (unlikely(!b_node->pivot[j]))
1929 if (unlikely(mas->max == b_node->pivot[j]))
1933 if (likely(i <= mas_end))
1934 b_node->pivot[j] = mas_safe_pivot(mas, pivots, i, mt);
1937 b_node->b_end = ++j;
1939 slots = ma_slots(node, mt);
1940 memcpy(b_node->slot + mab_start, slots + mas_start, sizeof(void *) * j);
1941 if (!ma_is_leaf(mt) && mt_is_alloc(mas->tree)) {
1942 gaps = ma_gaps(node, mt);
1943 memcpy(b_node->gap + mab_start, gaps + mas_start,
1944 sizeof(unsigned long) * j);
1949 * mas_leaf_set_meta() - Set the metadata of a leaf if possible.
1950 * @mas: The maple state
1951 * @node: The maple node
1952 * @pivots: pointer to the maple node pivots
1953 * @mt: The maple type
1954 * @end: The assumed end
1956 * Note, end may be incremented within this function but not modified at the
1957 * source. This is fine since the metadata is the last thing to be stored in a
1958 * node during a write.
1960 static inline void mas_leaf_set_meta(struct ma_state *mas,
1961 struct maple_node *node, unsigned long *pivots,
1962 enum maple_type mt, unsigned char end)
1964 /* There is no room for metadata already */
1965 if (mt_pivots[mt] <= end)
1968 if (pivots[end] && pivots[end] < mas->max)
1971 if (end < mt_slots[mt] - 1)
1972 ma_set_meta(node, mt, 0, end);
1976 * mab_mas_cp() - Copy data from maple_big_node to a maple encoded node.
1977 * @b_node: the maple_big_node that has the data
1978 * @mab_start: the start location in @b_node.
1979 * @mab_end: The end location in @b_node (inclusively)
1980 * @mas: The maple state with the maple encoded node.
1982 static inline void mab_mas_cp(struct maple_big_node *b_node,
1983 unsigned char mab_start, unsigned char mab_end,
1984 struct ma_state *mas, bool new_max)
1987 enum maple_type mt = mte_node_type(mas->node);
1988 struct maple_node *node = mte_to_node(mas->node);
1989 void __rcu **slots = ma_slots(node, mt);
1990 unsigned long *pivots = ma_pivots(node, mt);
1991 unsigned long *gaps = NULL;
1994 if (mab_end - mab_start > mt_pivots[mt])
1997 if (!pivots[mt_pivots[mt] - 1])
1998 slots[mt_pivots[mt]] = NULL;
2002 pivots[j++] = b_node->pivot[i++];
2003 } while (i <= mab_end && likely(b_node->pivot[i]));
2005 memcpy(slots, b_node->slot + mab_start,
2006 sizeof(void *) * (i - mab_start));
2009 mas->max = b_node->pivot[i - 1];
2012 if (likely(!ma_is_leaf(mt) && mt_is_alloc(mas->tree))) {
2013 unsigned long max_gap = 0;
2014 unsigned char offset = 15;
2016 gaps = ma_gaps(node, mt);
2018 gaps[--j] = b_node->gap[--i];
2019 if (gaps[j] > max_gap) {
2025 ma_set_meta(node, mt, offset, end);
2027 mas_leaf_set_meta(mas, node, pivots, mt, end);
2032 * mas_descend_adopt() - Descend through a sub-tree and adopt children.
2033 * @mas: the maple state with the maple encoded node of the sub-tree.
2035 * Descend through a sub-tree and adopt children who do not have the correct
2036 * parents set. Follow the parents which have the correct parents as they are
2037 * the new entries which need to be followed to find other incorrectly set
2040 static inline void mas_descend_adopt(struct ma_state *mas)
2042 struct ma_state list[3], next[3];
2046 * At each level there may be up to 3 correct parent pointers which indicates
2047 * the new nodes which need to be walked to find any new nodes at a lower level.
2050 for (i = 0; i < 3; i++) {
2057 while (!mte_is_leaf(list[0].node)) {
2059 for (i = 0; i < 3; i++) {
2060 if (mas_is_none(&list[i]))
2063 if (i && list[i-1].node == list[i].node)
2066 while ((n < 3) && (mas_new_child(&list[i], &next[n])))
2069 mas_adopt_children(&list[i], list[i].node);
2073 next[n++].node = MAS_NONE;
2075 /* descend by setting the list to the children */
2076 for (i = 0; i < 3; i++)
2082 * mas_bulk_rebalance() - Rebalance the end of a tree after a bulk insert.
2083 * @mas: The maple state
2084 * @end: The maple node end
2085 * @mt: The maple node type
2087 static inline void mas_bulk_rebalance(struct ma_state *mas, unsigned char end,
2090 if (!(mas->mas_flags & MA_STATE_BULK))
2093 if (mte_is_root(mas->node))
2096 if (end > mt_min_slots[mt]) {
2097 mas->mas_flags &= ~MA_STATE_REBALANCE;
2103 * mas_store_b_node() - Store an @entry into the b_node while also copying the
2104 * data from a maple encoded node.
2105 * @wr_mas: the maple write state
2106 * @b_node: the maple_big_node to fill with data
2107 * @offset_end: the offset to end copying
2109 * Return: The actual end of the data stored in @b_node
2111 static inline void mas_store_b_node(struct ma_wr_state *wr_mas,
2112 struct maple_big_node *b_node, unsigned char offset_end)
2115 unsigned char b_end;
2116 /* Possible underflow of piv will wrap back to 0 before use. */
2118 struct ma_state *mas = wr_mas->mas;
2120 b_node->type = wr_mas->type;
2124 /* Copy start data up to insert. */
2125 mas_mab_cp(mas, 0, slot - 1, b_node, 0);
2126 b_end = b_node->b_end;
2127 piv = b_node->pivot[b_end - 1];
2131 if (piv + 1 < mas->index) {
2132 /* Handle range starting after old range */
2133 b_node->slot[b_end] = wr_mas->content;
2134 if (!wr_mas->content)
2135 b_node->gap[b_end] = mas->index - 1 - piv;
2136 b_node->pivot[b_end++] = mas->index - 1;
2139 /* Store the new entry. */
2140 mas->offset = b_end;
2141 b_node->slot[b_end] = wr_mas->entry;
2142 b_node->pivot[b_end] = mas->last;
2145 if (mas->last >= mas->max)
2148 /* Handle new range ending before old range ends */
2149 piv = mas_logical_pivot(mas, wr_mas->pivots, offset_end, wr_mas->type);
2150 if (piv > mas->last) {
2151 if (piv == ULONG_MAX)
2152 mas_bulk_rebalance(mas, b_node->b_end, wr_mas->type);
2154 if (offset_end != slot)
2155 wr_mas->content = mas_slot_locked(mas, wr_mas->slots,
2158 b_node->slot[++b_end] = wr_mas->content;
2159 if (!wr_mas->content)
2160 b_node->gap[b_end] = piv - mas->last + 1;
2161 b_node->pivot[b_end] = piv;
2164 slot = offset_end + 1;
2165 if (slot > wr_mas->node_end)
2168 /* Copy end data to the end of the node. */
2169 mas_mab_cp(mas, slot, wr_mas->node_end + 1, b_node, ++b_end);
2174 b_node->b_end = b_end;
2178 * mas_prev_sibling() - Find the previous node with the same parent.
2179 * @mas: the maple state
2181 * Return: True if there is a previous sibling, false otherwise.
2183 static inline bool mas_prev_sibling(struct ma_state *mas)
2185 unsigned int p_slot = mte_parent_slot(mas->node);
2187 if (mte_is_root(mas->node))
2194 mas->offset = p_slot - 1;
2200 * mas_next_sibling() - Find the next node with the same parent.
2201 * @mas: the maple state
2203 * Return: true if there is a next sibling, false otherwise.
2205 static inline bool mas_next_sibling(struct ma_state *mas)
2207 MA_STATE(parent, mas->tree, mas->index, mas->last);
2209 if (mte_is_root(mas->node))
2213 mas_ascend(&parent);
2214 parent.offset = mte_parent_slot(mas->node) + 1;
2215 if (parent.offset > mas_data_end(&parent))
2224 * mte_node_or_node() - Return the encoded node or MAS_NONE.
2225 * @enode: The encoded maple node.
2227 * Shorthand to avoid setting %NULLs in the tree or maple_subtree_state.
2229 * Return: @enode or MAS_NONE
2231 static inline struct maple_enode *mte_node_or_none(struct maple_enode *enode)
2236 return ma_enode_ptr(MAS_NONE);
2240 * mas_wr_node_walk() - Find the correct offset for the index in the @mas.
2241 * @wr_mas: The maple write state
2243 * Uses mas_slot_locked() and does not need to worry about dead nodes.
2245 static inline void mas_wr_node_walk(struct ma_wr_state *wr_mas)
2247 struct ma_state *mas = wr_mas->mas;
2248 unsigned char count;
2249 unsigned char offset;
2250 unsigned long index, min, max;
2252 if (unlikely(ma_is_dense(wr_mas->type))) {
2253 wr_mas->r_max = wr_mas->r_min = mas->index;
2254 mas->offset = mas->index = mas->min;
2258 wr_mas->node = mas_mn(wr_mas->mas);
2259 wr_mas->pivots = ma_pivots(wr_mas->node, wr_mas->type);
2260 count = wr_mas->node_end = ma_data_end(wr_mas->node, wr_mas->type,
2261 wr_mas->pivots, mas->max);
2262 offset = mas->offset;
2263 min = mas_safe_min(mas, wr_mas->pivots, offset);
2264 if (unlikely(offset == count))
2267 max = wr_mas->pivots[offset];
2269 if (unlikely(index <= max))
2272 if (unlikely(!max && offset))
2276 while (++offset < count) {
2277 max = wr_mas->pivots[offset];
2280 else if (unlikely(!max))
2289 wr_mas->r_max = max;
2290 wr_mas->r_min = min;
2291 wr_mas->offset_end = mas->offset = offset;
2295 * mas_topiary_range() - Add a range of slots to the topiary.
2296 * @mas: The maple state
2297 * @destroy: The topiary to add the slots (usually destroy)
2298 * @start: The starting slot inclusively
2299 * @end: The end slot inclusively
2301 static inline void mas_topiary_range(struct ma_state *mas,
2302 struct ma_topiary *destroy, unsigned char start, unsigned char end)
2305 unsigned char offset;
2307 MT_BUG_ON(mas->tree, mte_is_leaf(mas->node));
2308 slots = ma_slots(mas_mn(mas), mte_node_type(mas->node));
2309 for (offset = start; offset <= end; offset++) {
2310 struct maple_enode *enode = mas_slot_locked(mas, slots, offset);
2312 if (mte_dead_node(enode))
2315 mat_add(destroy, enode);
2320 * mast_topiary() - Add the portions of the tree to the removal list; either to
2321 * be freed or discarded (destroy walk).
2322 * @mast: The maple_subtree_state.
2324 static inline void mast_topiary(struct maple_subtree_state *mast)
2326 MA_WR_STATE(wr_mas, mast->orig_l, NULL);
2327 unsigned char r_start, r_end;
2328 unsigned char l_start, l_end;
2329 void __rcu **l_slots, **r_slots;
2331 wr_mas.type = mte_node_type(mast->orig_l->node);
2332 mast->orig_l->index = mast->orig_l->last;
2333 mas_wr_node_walk(&wr_mas);
2334 l_start = mast->orig_l->offset + 1;
2335 l_end = mas_data_end(mast->orig_l);
2337 r_end = mast->orig_r->offset;
2342 l_slots = ma_slots(mas_mn(mast->orig_l),
2343 mte_node_type(mast->orig_l->node));
2345 r_slots = ma_slots(mas_mn(mast->orig_r),
2346 mte_node_type(mast->orig_r->node));
2348 if ((l_start < l_end) &&
2349 mte_dead_node(mas_slot_locked(mast->orig_l, l_slots, l_start))) {
2353 if (mte_dead_node(mas_slot_locked(mast->orig_r, r_slots, r_end))) {
2358 if ((l_start > r_end) && (mast->orig_l->node == mast->orig_r->node))
2361 /* At the node where left and right sides meet, add the parts between */
2362 if (mast->orig_l->node == mast->orig_r->node) {
2363 return mas_topiary_range(mast->orig_l, mast->destroy,
2367 /* mast->orig_r is different and consumed. */
2368 if (mte_is_leaf(mast->orig_r->node))
2371 if (mte_dead_node(mas_slot_locked(mast->orig_l, l_slots, l_end)))
2375 if (l_start <= l_end)
2376 mas_topiary_range(mast->orig_l, mast->destroy, l_start, l_end);
2378 if (mte_dead_node(mas_slot_locked(mast->orig_r, r_slots, r_start)))
2381 if (r_start <= r_end)
2382 mas_topiary_range(mast->orig_r, mast->destroy, 0, r_end);
2386 * mast_rebalance_next() - Rebalance against the next node
2387 * @mast: The maple subtree state
2388 * @old_r: The encoded maple node to the right (next node).
2390 static inline void mast_rebalance_next(struct maple_subtree_state *mast)
2392 unsigned char b_end = mast->bn->b_end;
2394 mas_mab_cp(mast->orig_r, 0, mt_slot_count(mast->orig_r->node),
2396 mast->orig_r->last = mast->orig_r->max;
2400 * mast_rebalance_prev() - Rebalance against the previous node
2401 * @mast: The maple subtree state
2402 * @old_l: The encoded maple node to the left (previous node)
2404 static inline void mast_rebalance_prev(struct maple_subtree_state *mast)
2406 unsigned char end = mas_data_end(mast->orig_l) + 1;
2407 unsigned char b_end = mast->bn->b_end;
2409 mab_shift_right(mast->bn, end);
2410 mas_mab_cp(mast->orig_l, 0, end - 1, mast->bn, 0);
2411 mast->l->min = mast->orig_l->min;
2412 mast->orig_l->index = mast->orig_l->min;
2413 mast->bn->b_end = end + b_end;
2414 mast->l->offset += end;
2418 * mast_spanning_rebalance() - Rebalance nodes with nearest neighbour favouring
2419 * the node to the right. Checking the nodes to the right then the left at each
2420 * level upwards until root is reached. Free and destroy as needed.
2421 * Data is copied into the @mast->bn.
2422 * @mast: The maple_subtree_state.
2425 bool mast_spanning_rebalance(struct maple_subtree_state *mast)
2427 struct ma_state r_tmp = *mast->orig_r;
2428 struct ma_state l_tmp = *mast->orig_l;
2429 struct maple_enode *ancestor = NULL;
2430 unsigned char start, end;
2431 unsigned char depth = 0;
2433 r_tmp = *mast->orig_r;
2434 l_tmp = *mast->orig_l;
2436 mas_ascend(mast->orig_r);
2437 mas_ascend(mast->orig_l);
2440 (mast->orig_r->node == mast->orig_l->node)) {
2441 ancestor = mast->orig_r->node;
2442 end = mast->orig_r->offset - 1;
2443 start = mast->orig_l->offset + 1;
2446 if (mast->orig_r->offset < mas_data_end(mast->orig_r)) {
2448 ancestor = mast->orig_r->node;
2452 mast->orig_r->offset++;
2454 mas_descend(mast->orig_r);
2455 mast->orig_r->offset = 0;
2459 mast_rebalance_next(mast);
2461 unsigned char l_off = 0;
2462 struct maple_enode *child = r_tmp.node;
2465 if (ancestor == r_tmp.node)
2471 if (l_off < r_tmp.offset)
2472 mas_topiary_range(&r_tmp, mast->destroy,
2473 l_off, r_tmp.offset);
2475 if (l_tmp.node != child)
2476 mat_add(mast->free, child);
2478 } while (r_tmp.node != ancestor);
2480 *mast->orig_l = l_tmp;
2483 } else if (mast->orig_l->offset != 0) {
2485 ancestor = mast->orig_l->node;
2486 end = mas_data_end(mast->orig_l);
2489 mast->orig_l->offset--;
2491 mas_descend(mast->orig_l);
2492 mast->orig_l->offset =
2493 mas_data_end(mast->orig_l);
2497 mast_rebalance_prev(mast);
2499 unsigned char r_off;
2500 struct maple_enode *child = l_tmp.node;
2503 if (ancestor == l_tmp.node)
2506 r_off = mas_data_end(&l_tmp);
2508 if (l_tmp.offset < r_off)
2511 if (l_tmp.offset < r_off)
2512 mas_topiary_range(&l_tmp, mast->destroy,
2513 l_tmp.offset, r_off);
2515 if (r_tmp.node != child)
2516 mat_add(mast->free, child);
2518 } while (l_tmp.node != ancestor);
2520 *mast->orig_r = r_tmp;
2523 } while (!mte_is_root(mast->orig_r->node));
2525 *mast->orig_r = r_tmp;
2526 *mast->orig_l = l_tmp;
2531 * mast_ascend_free() - Add current original maple state nodes to the free list
2533 * @mast: the maple subtree state.
2535 * Ascend the original left and right sides and add the previous nodes to the
2536 * free list. Set the slots to point to the correct location in the new nodes.
2539 mast_ascend_free(struct maple_subtree_state *mast)
2541 MA_WR_STATE(wr_mas, mast->orig_r, NULL);
2542 struct maple_enode *left = mast->orig_l->node;
2543 struct maple_enode *right = mast->orig_r->node;
2545 mas_ascend(mast->orig_l);
2546 mas_ascend(mast->orig_r);
2547 mat_add(mast->free, left);
2550 mat_add(mast->free, right);
2552 mast->orig_r->offset = 0;
2553 mast->orig_r->index = mast->r->max;
2554 /* last should be larger than or equal to index */
2555 if (mast->orig_r->last < mast->orig_r->index)
2556 mast->orig_r->last = mast->orig_r->index;
2558 * The node may not contain the value so set slot to ensure all
2559 * of the nodes contents are freed or destroyed.
2561 wr_mas.type = mte_node_type(mast->orig_r->node);
2562 mas_wr_node_walk(&wr_mas);
2563 /* Set up the left side of things */
2564 mast->orig_l->offset = 0;
2565 mast->orig_l->index = mast->l->min;
2566 wr_mas.mas = mast->orig_l;
2567 wr_mas.type = mte_node_type(mast->orig_l->node);
2568 mas_wr_node_walk(&wr_mas);
2570 mast->bn->type = wr_mas.type;
2574 * mas_new_ma_node() - Create and return a new maple node. Helper function.
2575 * @mas: the maple state with the allocations.
2576 * @b_node: the maple_big_node with the type encoding.
2578 * Use the node type from the maple_big_node to allocate a new node from the
2579 * ma_state. This function exists mainly for code readability.
2581 * Return: A new maple encoded node
2583 static inline struct maple_enode
2584 *mas_new_ma_node(struct ma_state *mas, struct maple_big_node *b_node)
2586 return mt_mk_node(ma_mnode_ptr(mas_pop_node(mas)), b_node->type);
2590 * mas_mab_to_node() - Set up right and middle nodes
2592 * @mas: the maple state that contains the allocations.
2593 * @b_node: the node which contains the data.
2594 * @left: The pointer which will have the left node
2595 * @right: The pointer which may have the right node
2596 * @middle: the pointer which may have the middle node (rare)
2597 * @mid_split: the split location for the middle node
2599 * Return: the split of left.
2601 static inline unsigned char mas_mab_to_node(struct ma_state *mas,
2602 struct maple_big_node *b_node, struct maple_enode **left,
2603 struct maple_enode **right, struct maple_enode **middle,
2604 unsigned char *mid_split, unsigned long min)
2606 unsigned char split = 0;
2607 unsigned char slot_count = mt_slots[b_node->type];
2609 *left = mas_new_ma_node(mas, b_node);
2614 if (b_node->b_end < slot_count) {
2615 split = b_node->b_end;
2617 split = mab_calc_split(mas, b_node, mid_split, min);
2618 *right = mas_new_ma_node(mas, b_node);
2622 *middle = mas_new_ma_node(mas, b_node);
2629 * mab_set_b_end() - Add entry to b_node at b_node->b_end and increment the end
2631 * @b_node - the big node to add the entry
2632 * @mas - the maple state to get the pivot (mas->max)
2633 * @entry - the entry to add, if NULL nothing happens.
2635 static inline void mab_set_b_end(struct maple_big_node *b_node,
2636 struct ma_state *mas,
2642 b_node->slot[b_node->b_end] = entry;
2643 if (mt_is_alloc(mas->tree))
2644 b_node->gap[b_node->b_end] = mas_max_gap(mas);
2645 b_node->pivot[b_node->b_end++] = mas->max;
2649 * mas_set_split_parent() - combine_then_separate helper function. Sets the parent
2650 * of @mas->node to either @left or @right, depending on @slot and @split
2652 * @mas - the maple state with the node that needs a parent
2653 * @left - possible parent 1
2654 * @right - possible parent 2
2655 * @slot - the slot the mas->node was placed
2656 * @split - the split location between @left and @right
2658 static inline void mas_set_split_parent(struct ma_state *mas,
2659 struct maple_enode *left,
2660 struct maple_enode *right,
2661 unsigned char *slot, unsigned char split)
2663 if (mas_is_none(mas))
2666 if ((*slot) <= split)
2667 mte_set_parent(mas->node, left, *slot);
2669 mte_set_parent(mas->node, right, (*slot) - split - 1);
2675 * mte_mid_split_check() - Check if the next node passes the mid-split
2676 * @**l: Pointer to left encoded maple node.
2677 * @**m: Pointer to middle encoded maple node.
2678 * @**r: Pointer to right encoded maple node.
2680 * @*split: The split location.
2681 * @mid_split: The middle split.
2683 static inline void mte_mid_split_check(struct maple_enode **l,
2684 struct maple_enode **r,
2685 struct maple_enode *right,
2687 unsigned char *split,
2688 unsigned char mid_split)
2693 if (slot < mid_split)
2702 * mast_set_split_parents() - Helper function to set three nodes parents. Slot
2703 * is taken from @mast->l.
2704 * @mast - the maple subtree state
2705 * @left - the left node
2706 * @right - the right node
2707 * @split - the split location.
2709 static inline void mast_set_split_parents(struct maple_subtree_state *mast,
2710 struct maple_enode *left,
2711 struct maple_enode *middle,
2712 struct maple_enode *right,
2713 unsigned char split,
2714 unsigned char mid_split)
2717 struct maple_enode *l = left;
2718 struct maple_enode *r = right;
2720 if (mas_is_none(mast->l))
2726 slot = mast->l->offset;
2728 mte_mid_split_check(&l, &r, right, slot, &split, mid_split);
2729 mas_set_split_parent(mast->l, l, r, &slot, split);
2731 mte_mid_split_check(&l, &r, right, slot, &split, mid_split);
2732 mas_set_split_parent(mast->m, l, r, &slot, split);
2734 mte_mid_split_check(&l, &r, right, slot, &split, mid_split);
2735 mas_set_split_parent(mast->r, l, r, &slot, split);
2739 * mas_wmb_replace() - Write memory barrier and replace
2740 * @mas: The maple state
2741 * @free: the maple topiary list of nodes to free
2742 * @destroy: The maple topiary list of nodes to destroy (walk and free)
2744 * Updates gap as necessary.
2746 static inline void mas_wmb_replace(struct ma_state *mas,
2747 struct ma_topiary *free,
2748 struct ma_topiary *destroy)
2750 /* All nodes must see old data as dead prior to replacing that data */
2751 smp_wmb(); /* Needed for RCU */
2753 /* Insert the new data in the tree */
2754 mas_replace(mas, true);
2756 if (!mte_is_leaf(mas->node))
2757 mas_descend_adopt(mas);
2759 mas_mat_free(mas, free);
2762 mas_mat_destroy(mas, destroy);
2764 if (mte_is_leaf(mas->node))
2767 mas_update_gap(mas);
2771 * mast_new_root() - Set a new tree root during subtree creation
2772 * @mast: The maple subtree state
2773 * @mas: The maple state
2775 static inline void mast_new_root(struct maple_subtree_state *mast,
2776 struct ma_state *mas)
2778 mas_mn(mast->l)->parent =
2779 ma_parent_ptr(((unsigned long)mas->tree | MA_ROOT_PARENT));
2780 if (!mte_dead_node(mast->orig_l->node) &&
2781 !mte_is_root(mast->orig_l->node)) {
2783 mast_ascend_free(mast);
2785 } while (!mte_is_root(mast->orig_l->node));
2787 if ((mast->orig_l->node != mas->node) &&
2788 (mast->l->depth > mas_mt_height(mas))) {
2789 mat_add(mast->free, mas->node);
2794 * mast_cp_to_nodes() - Copy data out to nodes.
2795 * @mast: The maple subtree state
2796 * @left: The left encoded maple node
2797 * @middle: The middle encoded maple node
2798 * @right: The right encoded maple node
2799 * @split: The location to split between left and (middle ? middle : right)
2800 * @mid_split: The location to split between middle and right.
2802 static inline void mast_cp_to_nodes(struct maple_subtree_state *mast,
2803 struct maple_enode *left, struct maple_enode *middle,
2804 struct maple_enode *right, unsigned char split, unsigned char mid_split)
2806 bool new_lmax = true;
2808 mast->l->node = mte_node_or_none(left);
2809 mast->m->node = mte_node_or_none(middle);
2810 mast->r->node = mte_node_or_none(right);
2812 mast->l->min = mast->orig_l->min;
2813 if (split == mast->bn->b_end) {
2814 mast->l->max = mast->orig_r->max;
2818 mab_mas_cp(mast->bn, 0, split, mast->l, new_lmax);
2821 mab_mas_cp(mast->bn, 1 + split, mid_split, mast->m, true);
2822 mast->m->min = mast->bn->pivot[split] + 1;
2826 mast->r->max = mast->orig_r->max;
2828 mab_mas_cp(mast->bn, 1 + split, mast->bn->b_end, mast->r, false);
2829 mast->r->min = mast->bn->pivot[split] + 1;
2834 * mast_combine_cp_left - Copy in the original left side of the tree into the
2835 * combined data set in the maple subtree state big node.
2836 * @mast: The maple subtree state
2838 static inline void mast_combine_cp_left(struct maple_subtree_state *mast)
2840 unsigned char l_slot = mast->orig_l->offset;
2845 mas_mab_cp(mast->orig_l, 0, l_slot - 1, mast->bn, 0);
2849 * mast_combine_cp_right: Copy in the original right side of the tree into the
2850 * combined data set in the maple subtree state big node.
2851 * @mast: The maple subtree state
2853 static inline void mast_combine_cp_right(struct maple_subtree_state *mast)
2855 if (mast->bn->pivot[mast->bn->b_end - 1] >= mast->orig_r->max)
2858 mas_mab_cp(mast->orig_r, mast->orig_r->offset + 1,
2859 mt_slot_count(mast->orig_r->node), mast->bn,
2861 mast->orig_r->last = mast->orig_r->max;
2865 * mast_sufficient: Check if the maple subtree state has enough data in the big
2866 * node to create at least one sufficient node
2867 * @mast: the maple subtree state
2869 static inline bool mast_sufficient(struct maple_subtree_state *mast)
2871 if (mast->bn->b_end > mt_min_slot_count(mast->orig_l->node))
2878 * mast_overflow: Check if there is too much data in the subtree state for a
2880 * @mast: The maple subtree state
2882 static inline bool mast_overflow(struct maple_subtree_state *mast)
2884 if (mast->bn->b_end >= mt_slot_count(mast->orig_l->node))
2890 static inline void *mtree_range_walk(struct ma_state *mas)
2892 unsigned long *pivots;
2893 unsigned char offset;
2894 struct maple_node *node;
2895 struct maple_enode *next, *last;
2896 enum maple_type type;
2899 unsigned long max, min;
2900 unsigned long prev_max, prev_min;
2908 node = mte_to_node(next);
2909 type = mte_node_type(next);
2910 pivots = ma_pivots(node, type);
2911 end = ma_data_end(node, type, pivots, max);
2912 if (unlikely(ma_dead_node(node)))
2915 if (pivots[offset] >= mas->index) {
2918 max = pivots[offset];
2924 } while ((offset < end) && (pivots[offset] < mas->index));
2927 min = pivots[offset - 1] + 1;
2929 if (likely(offset < end && pivots[offset]))
2930 max = pivots[offset];
2933 slots = ma_slots(node, type);
2934 next = mt_slot(mas->tree, slots, offset);
2935 if (unlikely(ma_dead_node(node)))
2937 } while (!ma_is_leaf(type));
2939 mas->offset = offset;
2942 mas->min = prev_min;
2943 mas->max = prev_max;
2945 return (void *) next;
2953 * mas_spanning_rebalance() - Rebalance across two nodes which may not be peers.
2954 * @mas: The starting maple state
2955 * @mast: The maple_subtree_state, keeps track of 4 maple states.
2956 * @count: The estimated count of iterations needed.
2958 * Follow the tree upwards from @l_mas and @r_mas for @count, or until the root
2959 * is hit. First @b_node is split into two entries which are inserted into the
2960 * next iteration of the loop. @b_node is returned populated with the final
2961 * iteration. @mas is used to obtain allocations. orig_l_mas keeps track of the
2962 * nodes that will remain active by using orig_l_mas->index and orig_l_mas->last
2963 * to account of what has been copied into the new sub-tree. The update of
2964 * orig_l_mas->last is used in mas_consume to find the slots that will need to
2965 * be either freed or destroyed. orig_l_mas->depth keeps track of the height of
2966 * the new sub-tree in case the sub-tree becomes the full tree.
2968 * Return: the number of elements in b_node during the last loop.
2970 static int mas_spanning_rebalance(struct ma_state *mas,
2971 struct maple_subtree_state *mast, unsigned char count)
2973 unsigned char split, mid_split;
2974 unsigned char slot = 0;
2975 struct maple_enode *left = NULL, *middle = NULL, *right = NULL;
2977 MA_STATE(l_mas, mas->tree, mas->index, mas->index);
2978 MA_STATE(r_mas, mas->tree, mas->index, mas->last);
2979 MA_STATE(m_mas, mas->tree, mas->index, mas->index);
2980 MA_TOPIARY(free, mas->tree);
2981 MA_TOPIARY(destroy, mas->tree);
2984 * The tree needs to be rebalanced and leaves need to be kept at the same level.
2985 * Rebalancing is done by use of the ``struct maple_topiary``.
2991 mast->destroy = &destroy;
2992 l_mas.node = r_mas.node = m_mas.node = MAS_NONE;
2994 /* Check if this is not root and has sufficient data. */
2995 if (((mast->orig_l->min != 0) || (mast->orig_r->max != ULONG_MAX)) &&
2996 unlikely(mast->bn->b_end <= mt_min_slots[mast->bn->type]))
2997 mast_spanning_rebalance(mast);
2999 mast->orig_l->depth = 0;
3002 * Each level of the tree is examined and balanced, pushing data to the left or
3003 * right, or rebalancing against left or right nodes is employed to avoid
3004 * rippling up the tree to limit the amount of churn. Once a new sub-section of
3005 * the tree is created, there may be a mix of new and old nodes. The old nodes
3006 * will have the incorrect parent pointers and currently be in two trees: the
3007 * original tree and the partially new tree. To remedy the parent pointers in
3008 * the old tree, the new data is swapped into the active tree and a walk down
3009 * the tree is performed and the parent pointers are updated.
3010 * See mas_descend_adopt() for more information..
3014 mast->bn->type = mte_node_type(mast->orig_l->node);
3015 split = mas_mab_to_node(mas, mast->bn, &left, &right, &middle,
3016 &mid_split, mast->orig_l->min);
3017 mast_set_split_parents(mast, left, middle, right, split,
3019 mast_cp_to_nodes(mast, left, middle, right, split, mid_split);
3022 * Copy data from next level in the tree to mast->bn from next
3025 memset(mast->bn, 0, sizeof(struct maple_big_node));
3026 mast->bn->type = mte_node_type(left);
3027 mast->orig_l->depth++;
3029 /* Root already stored in l->node. */
3030 if (mas_is_root_limits(mast->l))
3033 mast_ascend_free(mast);
3034 mast_combine_cp_left(mast);
3035 l_mas.offset = mast->bn->b_end;
3036 mab_set_b_end(mast->bn, &l_mas, left);
3037 mab_set_b_end(mast->bn, &m_mas, middle);
3038 mab_set_b_end(mast->bn, &r_mas, right);
3040 /* Copy anything necessary out of the right node. */
3041 mast_combine_cp_right(mast);
3043 mast->orig_l->last = mast->orig_l->max;
3045 if (mast_sufficient(mast))
3048 if (mast_overflow(mast))
3051 /* May be a new root stored in mast->bn */
3052 if (mas_is_root_limits(mast->orig_l))
3055 mast_spanning_rebalance(mast);
3057 /* rebalancing from other nodes may require another loop. */
3062 l_mas.node = mt_mk_node(ma_mnode_ptr(mas_pop_node(mas)),
3063 mte_node_type(mast->orig_l->node));
3064 mast->orig_l->depth++;
3065 mab_mas_cp(mast->bn, 0, mt_slots[mast->bn->type] - 1, &l_mas, true);
3066 mte_set_parent(left, l_mas.node, slot);
3068 mte_set_parent(middle, l_mas.node, ++slot);
3071 mte_set_parent(right, l_mas.node, ++slot);
3073 if (mas_is_root_limits(mast->l)) {
3075 mast_new_root(mast, mas);
3077 mas_mn(&l_mas)->parent = mas_mn(mast->orig_l)->parent;
3080 if (!mte_dead_node(mast->orig_l->node))
3081 mat_add(&free, mast->orig_l->node);
3083 mas->depth = mast->orig_l->depth;
3084 *mast->orig_l = l_mas;
3085 mte_set_node_dead(mas->node);
3087 /* Set up mas for insertion. */
3088 mast->orig_l->depth = mas->depth;
3089 mast->orig_l->alloc = mas->alloc;
3090 *mas = *mast->orig_l;
3091 mas_wmb_replace(mas, &free, &destroy);
3092 mtree_range_walk(mas);
3093 return mast->bn->b_end;
3097 * mas_rebalance() - Rebalance a given node.
3098 * @mas: The maple state
3099 * @b_node: The big maple node.
3101 * Rebalance two nodes into a single node or two new nodes that are sufficient.
3102 * Continue upwards until tree is sufficient.
3104 * Return: the number of elements in b_node during the last loop.
3106 static inline int mas_rebalance(struct ma_state *mas,
3107 struct maple_big_node *b_node)
3109 char empty_count = mas_mt_height(mas);
3110 struct maple_subtree_state mast;
3111 unsigned char shift, b_end = ++b_node->b_end;
3113 MA_STATE(l_mas, mas->tree, mas->index, mas->last);
3114 MA_STATE(r_mas, mas->tree, mas->index, mas->last);
3116 trace_ma_op(__func__, mas);
3119 * Rebalancing occurs if a node is insufficient. Data is rebalanced
3120 * against the node to the right if it exists, otherwise the node to the
3121 * left of this node is rebalanced against this node. If rebalancing
3122 * causes just one node to be produced instead of two, then the parent
3123 * is also examined and rebalanced if it is insufficient. Every level
3124 * tries to combine the data in the same way. If one node contains the
3125 * entire range of the tree, then that node is used as a new root node.
3127 mas_node_count(mas, 1 + empty_count * 3);
3128 if (mas_is_err(mas))
3131 mast.orig_l = &l_mas;
3132 mast.orig_r = &r_mas;
3134 mast.bn->type = mte_node_type(mas->node);
3136 l_mas = r_mas = *mas;
3138 if (mas_next_sibling(&r_mas)) {
3139 mas_mab_cp(&r_mas, 0, mt_slot_count(r_mas.node), b_node, b_end);
3140 r_mas.last = r_mas.index = r_mas.max;
3142 mas_prev_sibling(&l_mas);
3143 shift = mas_data_end(&l_mas) + 1;
3144 mab_shift_right(b_node, shift);
3145 mas->offset += shift;
3146 mas_mab_cp(&l_mas, 0, shift - 1, b_node, 0);
3147 b_node->b_end = shift + b_end;
3148 l_mas.index = l_mas.last = l_mas.min;
3151 return mas_spanning_rebalance(mas, &mast, empty_count);
3155 * mas_destroy_rebalance() - Rebalance left-most node while destroying the maple
3157 * @mas: The maple state
3158 * @end: The end of the left-most node.
3160 * During a mass-insert event (such as forking), it may be necessary to
3161 * rebalance the left-most node when it is not sufficient.
3163 static inline void mas_destroy_rebalance(struct ma_state *mas, unsigned char end)
3165 enum maple_type mt = mte_node_type(mas->node);
3166 struct maple_node reuse, *newnode, *parent, *new_left, *left, *node;
3167 struct maple_enode *eparent;
3168 unsigned char offset, tmp, split = mt_slots[mt] / 2;
3169 void __rcu **l_slots, **slots;
3170 unsigned long *l_pivs, *pivs, gap;
3171 bool in_rcu = mt_in_rcu(mas->tree);
3173 MA_STATE(l_mas, mas->tree, mas->index, mas->last);
3176 mas_prev_sibling(&l_mas);
3180 /* Allocate for both left and right as well as parent. */
3181 mas_node_count(mas, 3);
3182 if (mas_is_err(mas))
3185 newnode = mas_pop_node(mas);
3191 newnode->parent = node->parent;
3192 slots = ma_slots(newnode, mt);
3193 pivs = ma_pivots(newnode, mt);
3194 left = mas_mn(&l_mas);
3195 l_slots = ma_slots(left, mt);
3196 l_pivs = ma_pivots(left, mt);
3197 if (!l_slots[split])
3199 tmp = mas_data_end(&l_mas) - split;
3201 memcpy(slots, l_slots + split + 1, sizeof(void *) * tmp);
3202 memcpy(pivs, l_pivs + split + 1, sizeof(unsigned long) * tmp);
3203 pivs[tmp] = l_mas.max;
3204 memcpy(slots + tmp, ma_slots(node, mt), sizeof(void *) * end);
3205 memcpy(pivs + tmp, ma_pivots(node, mt), sizeof(unsigned long) * end);
3207 l_mas.max = l_pivs[split];
3208 mas->min = l_mas.max + 1;
3209 eparent = mt_mk_node(mte_parent(l_mas.node),
3210 mas_parent_enum(&l_mas, l_mas.node));
3213 unsigned char max_p = mt_pivots[mt];
3214 unsigned char max_s = mt_slots[mt];
3217 memset(pivs + tmp, 0,
3218 sizeof(unsigned long *) * (max_p - tmp));
3220 if (tmp < mt_slots[mt])
3221 memset(slots + tmp, 0, sizeof(void *) * (max_s - tmp));
3223 memcpy(node, newnode, sizeof(struct maple_node));
3224 ma_set_meta(node, mt, 0, tmp - 1);
3225 mte_set_pivot(eparent, mte_parent_slot(l_mas.node),
3228 /* Remove data from l_pivs. */
3230 memset(l_pivs + tmp, 0, sizeof(unsigned long) * (max_p - tmp));
3231 memset(l_slots + tmp, 0, sizeof(void *) * (max_s - tmp));
3232 ma_set_meta(left, mt, 0, split);
3237 /* RCU requires replacing both l_mas, mas, and parent. */
3238 mas->node = mt_mk_node(newnode, mt);
3239 ma_set_meta(newnode, mt, 0, tmp);
3241 new_left = mas_pop_node(mas);
3242 new_left->parent = left->parent;
3243 mt = mte_node_type(l_mas.node);
3244 slots = ma_slots(new_left, mt);
3245 pivs = ma_pivots(new_left, mt);
3246 memcpy(slots, l_slots, sizeof(void *) * split);
3247 memcpy(pivs, l_pivs, sizeof(unsigned long) * split);
3248 ma_set_meta(new_left, mt, 0, split);
3249 l_mas.node = mt_mk_node(new_left, mt);
3251 /* replace parent. */
3252 offset = mte_parent_slot(mas->node);
3253 mt = mas_parent_enum(&l_mas, l_mas.node);
3254 parent = mas_pop_node(mas);
3255 slots = ma_slots(parent, mt);
3256 pivs = ma_pivots(parent, mt);
3257 memcpy(parent, mte_to_node(eparent), sizeof(struct maple_node));
3258 rcu_assign_pointer(slots[offset], mas->node);
3259 rcu_assign_pointer(slots[offset - 1], l_mas.node);
3260 pivs[offset - 1] = l_mas.max;
3261 eparent = mt_mk_node(parent, mt);
3263 gap = mas_leaf_max_gap(mas);
3264 mte_set_gap(eparent, mte_parent_slot(mas->node), gap);
3265 gap = mas_leaf_max_gap(&l_mas);
3266 mte_set_gap(eparent, mte_parent_slot(l_mas.node), gap);
3270 mas_replace(mas, false);
3272 mas_update_gap(mas);
3276 * mas_split_final_node() - Split the final node in a subtree operation.
3277 * @mast: the maple subtree state
3278 * @mas: The maple state
3279 * @height: The height of the tree in case it's a new root.
3281 static inline bool mas_split_final_node(struct maple_subtree_state *mast,
3282 struct ma_state *mas, int height)
3284 struct maple_enode *ancestor;
3286 if (mte_is_root(mas->node)) {
3287 if (mt_is_alloc(mas->tree))
3288 mast->bn->type = maple_arange_64;
3290 mast->bn->type = maple_range_64;
3291 mas->depth = height;
3294 * Only a single node is used here, could be root.
3295 * The Big_node data should just fit in a single node.
3297 ancestor = mas_new_ma_node(mas, mast->bn);
3298 mte_set_parent(mast->l->node, ancestor, mast->l->offset);
3299 mte_set_parent(mast->r->node, ancestor, mast->r->offset);
3300 mte_to_node(ancestor)->parent = mas_mn(mas)->parent;
3302 mast->l->node = ancestor;
3303 mab_mas_cp(mast->bn, 0, mt_slots[mast->bn->type] - 1, mast->l, true);
3304 mas->offset = mast->bn->b_end - 1;
3309 * mast_fill_bnode() - Copy data into the big node in the subtree state
3310 * @mast: The maple subtree state
3311 * @mas: the maple state
3312 * @skip: The number of entries to skip for new nodes insertion.
3314 static inline void mast_fill_bnode(struct maple_subtree_state *mast,
3315 struct ma_state *mas,
3319 struct maple_enode *old = mas->node;
3320 unsigned char split;
3322 memset(mast->bn->gap, 0, sizeof(unsigned long) * ARRAY_SIZE(mast->bn->gap));
3323 memset(mast->bn->slot, 0, sizeof(unsigned long) * ARRAY_SIZE(mast->bn->slot));
3324 memset(mast->bn->pivot, 0, sizeof(unsigned long) * ARRAY_SIZE(mast->bn->pivot));
3325 mast->bn->b_end = 0;
3327 if (mte_is_root(mas->node)) {
3331 mat_add(mast->free, old);
3332 mas->offset = mte_parent_slot(mas->node);
3335 if (cp && mast->l->offset)
3336 mas_mab_cp(mas, 0, mast->l->offset - 1, mast->bn, 0);
3338 split = mast->bn->b_end;
3339 mab_set_b_end(mast->bn, mast->l, mast->l->node);
3340 mast->r->offset = mast->bn->b_end;
3341 mab_set_b_end(mast->bn, mast->r, mast->r->node);
3342 if (mast->bn->pivot[mast->bn->b_end - 1] == mas->max)
3346 mas_mab_cp(mas, split + skip, mt_slot_count(mas->node) - 1,
3347 mast->bn, mast->bn->b_end);
3350 mast->bn->type = mte_node_type(mas->node);
3354 * mast_split_data() - Split the data in the subtree state big node into regular
3356 * @mast: The maple subtree state
3357 * @mas: The maple state
3358 * @split: The location to split the big node
3360 static inline void mast_split_data(struct maple_subtree_state *mast,
3361 struct ma_state *mas, unsigned char split)
3363 unsigned char p_slot;
3365 mab_mas_cp(mast->bn, 0, split, mast->l, true);
3366 mte_set_pivot(mast->r->node, 0, mast->r->max);
3367 mab_mas_cp(mast->bn, split + 1, mast->bn->b_end, mast->r, false);
3368 mast->l->offset = mte_parent_slot(mas->node);
3369 mast->l->max = mast->bn->pivot[split];
3370 mast->r->min = mast->l->max + 1;
3371 if (mte_is_leaf(mas->node))
3374 p_slot = mast->orig_l->offset;
3375 mas_set_split_parent(mast->orig_l, mast->l->node, mast->r->node,
3377 mas_set_split_parent(mast->orig_r, mast->l->node, mast->r->node,
3382 * mas_push_data() - Instead of splitting a node, it is beneficial to push the
3383 * data to the right or left node if there is room.
3384 * @mas: The maple state
3385 * @height: The current height of the maple state
3386 * @mast: The maple subtree state
3387 * @left: Push left or not.
3389 * Keeping the height of the tree low means faster lookups.
3391 * Return: True if pushed, false otherwise.
3393 static inline bool mas_push_data(struct ma_state *mas, int height,
3394 struct maple_subtree_state *mast, bool left)
3396 unsigned char slot_total = mast->bn->b_end;
3397 unsigned char end, space, split;
3399 MA_STATE(tmp_mas, mas->tree, mas->index, mas->last);
3401 tmp_mas.depth = mast->l->depth;
3403 if (left && !mas_prev_sibling(&tmp_mas))
3405 else if (!left && !mas_next_sibling(&tmp_mas))
3408 end = mas_data_end(&tmp_mas);
3410 space = 2 * mt_slot_count(mas->node) - 2;
3411 /* -2 instead of -1 to ensure there isn't a triple split */
3412 if (ma_is_leaf(mast->bn->type))
3415 if (mas->max == ULONG_MAX)
3418 if (slot_total >= space)
3421 /* Get the data; Fill mast->bn */
3424 mab_shift_right(mast->bn, end + 1);
3425 mas_mab_cp(&tmp_mas, 0, end, mast->bn, 0);
3426 mast->bn->b_end = slot_total + 1;
3428 mas_mab_cp(&tmp_mas, 0, end, mast->bn, mast->bn->b_end);
3431 /* Configure mast for splitting of mast->bn */
3432 split = mt_slots[mast->bn->type] - 2;
3434 /* Switch mas to prev node */
3435 mat_add(mast->free, mas->node);
3437 /* Start using mast->l for the left side. */
3438 tmp_mas.node = mast->l->node;
3441 mat_add(mast->free, tmp_mas.node);
3442 tmp_mas.node = mast->r->node;
3444 split = slot_total - split;
3446 split = mab_no_null_split(mast->bn, split, mt_slots[mast->bn->type]);
3447 /* Update parent slot for split calculation. */
3449 mast->orig_l->offset += end + 1;
3451 mast_split_data(mast, mas, split);
3452 mast_fill_bnode(mast, mas, 2);
3453 mas_split_final_node(mast, mas, height + 1);
3458 * mas_split() - Split data that is too big for one node into two.
3459 * @mas: The maple state
3460 * @b_node: The maple big node
3461 * Return: 1 on success, 0 on failure.
3463 static int mas_split(struct ma_state *mas, struct maple_big_node *b_node)
3466 struct maple_subtree_state mast;
3468 unsigned char mid_split, split = 0;
3471 * Splitting is handled differently from any other B-tree; the Maple
3472 * Tree splits upwards. Splitting up means that the split operation
3473 * occurs when the walk of the tree hits the leaves and not on the way
3474 * down. The reason for splitting up is that it is impossible to know
3475 * how much space will be needed until the leaf is (or leaves are)
3476 * reached. Since overwriting data is allowed and a range could
3477 * overwrite more than one range or result in changing one entry into 3
3478 * entries, it is impossible to know if a split is required until the
3481 * Splitting is a balancing act between keeping allocations to a minimum
3482 * and avoiding a 'jitter' event where a tree is expanded to make room
3483 * for an entry followed by a contraction when the entry is removed. To
3484 * accomplish the balance, there are empty slots remaining in both left
3485 * and right nodes after a split.
3487 MA_STATE(l_mas, mas->tree, mas->index, mas->last);
3488 MA_STATE(r_mas, mas->tree, mas->index, mas->last);
3489 MA_STATE(prev_l_mas, mas->tree, mas->index, mas->last);
3490 MA_STATE(prev_r_mas, mas->tree, mas->index, mas->last);
3491 MA_TOPIARY(mat, mas->tree);
3493 trace_ma_op(__func__, mas);
3494 mas->depth = mas_mt_height(mas);
3495 /* Allocation failures will happen early. */
3496 mas_node_count(mas, 1 + mas->depth * 2);
3497 if (mas_is_err(mas))
3502 mast.orig_l = &prev_l_mas;
3503 mast.orig_r = &prev_r_mas;
3507 while (height++ <= mas->depth) {
3508 if (mt_slots[b_node->type] > b_node->b_end) {
3509 mas_split_final_node(&mast, mas, height);
3513 l_mas = r_mas = *mas;
3514 l_mas.node = mas_new_ma_node(mas, b_node);
3515 r_mas.node = mas_new_ma_node(mas, b_node);
3517 * Another way that 'jitter' is avoided is to terminate a split up early if the
3518 * left or right node has space to spare. This is referred to as "pushing left"
3519 * or "pushing right" and is similar to the B* tree, except the nodes left or
3520 * right can rarely be reused due to RCU, but the ripple upwards is halted which
3521 * is a significant savings.
3523 /* Try to push left. */
3524 if (mas_push_data(mas, height, &mast, true))
3527 /* Try to push right. */
3528 if (mas_push_data(mas, height, &mast, false))
3531 split = mab_calc_split(mas, b_node, &mid_split, prev_l_mas.min);
3532 mast_split_data(&mast, mas, split);
3534 * Usually correct, mab_mas_cp in the above call overwrites
3537 mast.r->max = mas->max;
3538 mast_fill_bnode(&mast, mas, 1);
3539 prev_l_mas = *mast.l;
3540 prev_r_mas = *mast.r;
3543 /* Set the original node as dead */
3544 mat_add(mast.free, mas->node);
3545 mas->node = l_mas.node;
3546 mas_wmb_replace(mas, mast.free, NULL);
3547 mtree_range_walk(mas);
3552 * mas_reuse_node() - Reuse the node to store the data.
3553 * @wr_mas: The maple write state
3554 * @bn: The maple big node
3555 * @end: The end of the data.
3557 * Will always return false in RCU mode.
3559 * Return: True if node was reused, false otherwise.
3561 static inline bool mas_reuse_node(struct ma_wr_state *wr_mas,
3562 struct maple_big_node *bn, unsigned char end)
3564 /* Need to be rcu safe. */
3565 if (mt_in_rcu(wr_mas->mas->tree))
3568 if (end > bn->b_end) {
3569 int clear = mt_slots[wr_mas->type] - bn->b_end;
3571 memset(wr_mas->slots + bn->b_end, 0, sizeof(void *) * clear--);
3572 memset(wr_mas->pivots + bn->b_end, 0, sizeof(void *) * clear);
3574 mab_mas_cp(bn, 0, bn->b_end, wr_mas->mas, false);
3579 * mas_commit_b_node() - Commit the big node into the tree.
3580 * @wr_mas: The maple write state
3581 * @b_node: The maple big node
3582 * @end: The end of the data.
3584 static inline int mas_commit_b_node(struct ma_wr_state *wr_mas,
3585 struct maple_big_node *b_node, unsigned char end)
3587 struct maple_node *node;
3588 unsigned char b_end = b_node->b_end;
3589 enum maple_type b_type = b_node->type;
3591 if ((b_end < mt_min_slots[b_type]) &&
3592 (!mte_is_root(wr_mas->mas->node)) &&
3593 (mas_mt_height(wr_mas->mas) > 1))
3594 return mas_rebalance(wr_mas->mas, b_node);
3596 if (b_end >= mt_slots[b_type])
3597 return mas_split(wr_mas->mas, b_node);
3599 if (mas_reuse_node(wr_mas, b_node, end))
3602 mas_node_count(wr_mas->mas, 1);
3603 if (mas_is_err(wr_mas->mas))
3606 node = mas_pop_node(wr_mas->mas);
3607 node->parent = mas_mn(wr_mas->mas)->parent;
3608 wr_mas->mas->node = mt_mk_node(node, b_type);
3609 mab_mas_cp(b_node, 0, b_end, wr_mas->mas, false);
3610 mas_replace(wr_mas->mas, false);
3612 mas_update_gap(wr_mas->mas);
3617 * mas_root_expand() - Expand a root to a node
3618 * @mas: The maple state
3619 * @entry: The entry to store into the tree
3621 static inline int mas_root_expand(struct ma_state *mas, void *entry)
3623 void *contents = mas_root_locked(mas);
3624 enum maple_type type = maple_leaf_64;
3625 struct maple_node *node;
3627 unsigned long *pivots;
3630 mas_node_count(mas, 1);
3631 if (unlikely(mas_is_err(mas)))
3634 node = mas_pop_node(mas);
3635 pivots = ma_pivots(node, type);
3636 slots = ma_slots(node, type);
3637 node->parent = ma_parent_ptr(
3638 ((unsigned long)mas->tree | MA_ROOT_PARENT));
3639 mas->node = mt_mk_node(node, type);
3643 rcu_assign_pointer(slots[slot], contents);
3644 if (likely(mas->index > 1))
3647 pivots[slot++] = mas->index - 1;
3650 rcu_assign_pointer(slots[slot], entry);
3652 pivots[slot] = mas->last;
3653 if (mas->last != ULONG_MAX)
3656 mas_set_height(mas);
3657 ma_set_meta(node, maple_leaf_64, 0, slot);
3658 /* swap the new root into the tree */
3659 rcu_assign_pointer(mas->tree->ma_root, mte_mk_root(mas->node));
3663 static inline void mas_store_root(struct ma_state *mas, void *entry)
3665 if (likely((mas->last != 0) || (mas->index != 0)))
3666 mas_root_expand(mas, entry);
3667 else if (((unsigned long) (entry) & 3) == 2)
3668 mas_root_expand(mas, entry);
3670 rcu_assign_pointer(mas->tree->ma_root, entry);
3671 mas->node = MAS_START;
3676 * mas_is_span_wr() - Check if the write needs to be treated as a write that
3678 * @mas: The maple state
3679 * @piv: The pivot value being written
3680 * @type: The maple node type
3681 * @entry: The data to write
3683 * Spanning writes are writes that start in one node and end in another OR if
3684 * the write of a %NULL will cause the node to end with a %NULL.
3686 * Return: True if this is a spanning write, false otherwise.
3688 static bool mas_is_span_wr(struct ma_wr_state *wr_mas)
3691 unsigned long last = wr_mas->mas->last;
3692 unsigned long piv = wr_mas->r_max;
3693 enum maple_type type = wr_mas->type;
3694 void *entry = wr_mas->entry;
3696 /* Contained in this pivot */
3700 max = wr_mas->mas->max;
3701 if (unlikely(ma_is_leaf(type))) {
3702 /* Fits in the node, but may span slots. */
3706 /* Writes to the end of the node but not null. */
3707 if ((last == max) && entry)
3711 * Writing ULONG_MAX is not a spanning write regardless of the
3712 * value being written as long as the range fits in the node.
3714 if ((last == ULONG_MAX) && (last == max))
3716 } else if (piv == last) {
3720 /* Detect spanning store wr walk */
3721 if (last == ULONG_MAX)
3725 trace_ma_write(__func__, wr_mas->mas, piv, entry);
3730 static inline void mas_wr_walk_descend(struct ma_wr_state *wr_mas)
3732 wr_mas->type = mte_node_type(wr_mas->mas->node);
3733 mas_wr_node_walk(wr_mas);
3734 wr_mas->slots = ma_slots(wr_mas->node, wr_mas->type);
3737 static inline void mas_wr_walk_traverse(struct ma_wr_state *wr_mas)
3739 wr_mas->mas->max = wr_mas->r_max;
3740 wr_mas->mas->min = wr_mas->r_min;
3741 wr_mas->mas->node = wr_mas->content;
3742 wr_mas->mas->offset = 0;
3743 wr_mas->mas->depth++;
3746 * mas_wr_walk() - Walk the tree for a write.
3747 * @wr_mas: The maple write state
3749 * Uses mas_slot_locked() and does not need to worry about dead nodes.
3751 * Return: True if it's contained in a node, false on spanning write.
3753 static bool mas_wr_walk(struct ma_wr_state *wr_mas)
3755 struct ma_state *mas = wr_mas->mas;
3758 mas_wr_walk_descend(wr_mas);
3759 if (unlikely(mas_is_span_wr(wr_mas)))
3762 wr_mas->content = mas_slot_locked(mas, wr_mas->slots,
3764 if (ma_is_leaf(wr_mas->type))
3767 mas_wr_walk_traverse(wr_mas);
3773 static bool mas_wr_walk_index(struct ma_wr_state *wr_mas)
3775 struct ma_state *mas = wr_mas->mas;
3778 mas_wr_walk_descend(wr_mas);
3779 wr_mas->content = mas_slot_locked(mas, wr_mas->slots,
3781 if (ma_is_leaf(wr_mas->type))
3783 mas_wr_walk_traverse(wr_mas);
3789 * mas_extend_spanning_null() - Extend a store of a %NULL to include surrounding %NULLs.
3790 * @l_wr_mas: The left maple write state
3791 * @r_wr_mas: The right maple write state
3793 static inline void mas_extend_spanning_null(struct ma_wr_state *l_wr_mas,
3794 struct ma_wr_state *r_wr_mas)
3796 struct ma_state *r_mas = r_wr_mas->mas;
3797 struct ma_state *l_mas = l_wr_mas->mas;
3798 unsigned char l_slot;
3800 l_slot = l_mas->offset;
3801 if (!l_wr_mas->content)
3802 l_mas->index = l_wr_mas->r_min;
3804 if ((l_mas->index == l_wr_mas->r_min) &&
3806 !mas_slot_locked(l_mas, l_wr_mas->slots, l_slot - 1))) {
3808 l_mas->index = l_wr_mas->pivots[l_slot - 2] + 1;
3810 l_mas->index = l_mas->min;
3812 l_mas->offset = l_slot - 1;
3815 if (!r_wr_mas->content) {
3816 if (r_mas->last < r_wr_mas->r_max)
3817 r_mas->last = r_wr_mas->r_max;
3819 } else if ((r_mas->last == r_wr_mas->r_max) &&
3820 (r_mas->last < r_mas->max) &&
3821 !mas_slot_locked(r_mas, r_wr_mas->slots, r_mas->offset + 1)) {
3822 r_mas->last = mas_safe_pivot(r_mas, r_wr_mas->pivots,
3823 r_wr_mas->type, r_mas->offset + 1);
3828 static inline void *mas_state_walk(struct ma_state *mas)
3832 entry = mas_start(mas);
3833 if (mas_is_none(mas))
3836 if (mas_is_ptr(mas))
3839 return mtree_range_walk(mas);
3843 * mtree_lookup_walk() - Internal quick lookup that does not keep maple state up
3846 * @mas: The maple state.
3848 * Note: Leaves mas in undesirable state.
3849 * Return: The entry for @mas->index or %NULL on dead node.
3851 static inline void *mtree_lookup_walk(struct ma_state *mas)
3853 unsigned long *pivots;
3854 unsigned char offset;
3855 struct maple_node *node;
3856 struct maple_enode *next;
3857 enum maple_type type;
3866 node = mte_to_node(next);
3867 type = mte_node_type(next);
3868 pivots = ma_pivots(node, type);
3869 end = ma_data_end(node, type, pivots, max);
3870 if (unlikely(ma_dead_node(node)))
3873 if (pivots[offset] >= mas->index) {
3874 max = pivots[offset];
3877 } while (++offset < end);
3879 slots = ma_slots(node, type);
3880 next = mt_slot(mas->tree, slots, offset);
3881 if (unlikely(ma_dead_node(node)))
3883 } while (!ma_is_leaf(type));
3885 return (void *) next;
3893 * mas_new_root() - Create a new root node that only contains the entry passed
3895 * @mas: The maple state
3896 * @entry: The entry to store.
3898 * Only valid when the index == 0 and the last == ULONG_MAX
3900 * Return 0 on error, 1 on success.
3902 static inline int mas_new_root(struct ma_state *mas, void *entry)
3904 struct maple_enode *root = mas_root_locked(mas);
3905 enum maple_type type = maple_leaf_64;
3906 struct maple_node *node;
3908 unsigned long *pivots;
3910 if (!entry && !mas->index && mas->last == ULONG_MAX) {
3912 mas_set_height(mas);
3913 rcu_assign_pointer(mas->tree->ma_root, entry);
3914 mas->node = MAS_START;
3918 mas_node_count(mas, 1);
3919 if (mas_is_err(mas))
3922 node = mas_pop_node(mas);
3923 pivots = ma_pivots(node, type);
3924 slots = ma_slots(node, type);
3925 node->parent = ma_parent_ptr(
3926 ((unsigned long)mas->tree | MA_ROOT_PARENT));
3927 mas->node = mt_mk_node(node, type);
3928 rcu_assign_pointer(slots[0], entry);
3929 pivots[0] = mas->last;
3931 mas_set_height(mas);
3932 rcu_assign_pointer(mas->tree->ma_root, mte_mk_root(mas->node));
3935 if (xa_is_node(root))
3936 mte_destroy_walk(root, mas->tree);
3941 * mas_wr_spanning_store() - Create a subtree with the store operation completed
3942 * and new nodes where necessary, then place the sub-tree in the actual tree.
3943 * Note that mas is expected to point to the node which caused the store to
3945 * @wr_mas: The maple write state
3947 * Return: 0 on error, positive on success.
3949 static inline int mas_wr_spanning_store(struct ma_wr_state *wr_mas)
3951 struct maple_subtree_state mast;
3952 struct maple_big_node b_node;
3953 struct ma_state *mas;
3954 unsigned char height;
3956 /* Left and Right side of spanning store */
3957 MA_STATE(l_mas, NULL, 0, 0);
3958 MA_STATE(r_mas, NULL, 0, 0);
3960 MA_WR_STATE(r_wr_mas, &r_mas, wr_mas->entry);
3961 MA_WR_STATE(l_wr_mas, &l_mas, wr_mas->entry);
3964 * A store operation that spans multiple nodes is called a spanning
3965 * store and is handled early in the store call stack by the function
3966 * mas_is_span_wr(). When a spanning store is identified, the maple
3967 * state is duplicated. The first maple state walks the left tree path
3968 * to ``index``, the duplicate walks the right tree path to ``last``.
3969 * The data in the two nodes are combined into a single node, two nodes,
3970 * or possibly three nodes (see the 3-way split above). A ``NULL``
3971 * written to the last entry of a node is considered a spanning store as
3972 * a rebalance is required for the operation to complete and an overflow
3973 * of data may happen.
3976 trace_ma_op(__func__, mas);
3978 if (unlikely(!mas->index && mas->last == ULONG_MAX))
3979 return mas_new_root(mas, wr_mas->entry);
3981 * Node rebalancing may occur due to this store, so there may be three new
3982 * entries per level plus a new root.
3984 height = mas_mt_height(mas);
3985 mas_node_count(mas, 1 + height * 3);
3986 if (mas_is_err(mas))
3990 * Set up right side. Need to get to the next offset after the spanning
3991 * store to ensure it's not NULL and to combine both the next node and
3992 * the node with the start together.
3995 /* Avoid overflow, walk to next slot in the tree. */
3999 r_mas.index = r_mas.last;
4000 mas_wr_walk_index(&r_wr_mas);
4001 r_mas.last = r_mas.index = mas->last;
4003 /* Set up left side. */
4005 mas_wr_walk_index(&l_wr_mas);
4007 if (!wr_mas->entry) {
4008 mas_extend_spanning_null(&l_wr_mas, &r_wr_mas);
4009 mas->offset = l_mas.offset;
4010 mas->index = l_mas.index;
4011 mas->last = l_mas.last = r_mas.last;
4014 /* expanding NULLs may make this cover the entire range */
4015 if (!l_mas.index && r_mas.last == ULONG_MAX) {
4016 mas_set_range(mas, 0, ULONG_MAX);
4017 return mas_new_root(mas, wr_mas->entry);
4020 memset(&b_node, 0, sizeof(struct maple_big_node));
4021 /* Copy l_mas and store the value in b_node. */
4022 mas_store_b_node(&l_wr_mas, &b_node, l_wr_mas.node_end);
4023 /* Copy r_mas into b_node. */
4024 if (r_mas.offset <= r_wr_mas.node_end)
4025 mas_mab_cp(&r_mas, r_mas.offset, r_wr_mas.node_end,
4026 &b_node, b_node.b_end + 1);
4030 /* Stop spanning searches by searching for just index. */
4031 l_mas.index = l_mas.last = mas->index;
4034 mast.orig_l = &l_mas;
4035 mast.orig_r = &r_mas;
4036 /* Combine l_mas and r_mas and split them up evenly again. */
4037 return mas_spanning_rebalance(mas, &mast, height + 1);
4041 * mas_wr_node_store() - Attempt to store the value in a node
4042 * @wr_mas: The maple write state
4044 * Attempts to reuse the node, but may allocate.
4046 * Return: True if stored, false otherwise
4048 static inline bool mas_wr_node_store(struct ma_wr_state *wr_mas)
4050 struct ma_state *mas = wr_mas->mas;
4051 void __rcu **dst_slots;
4052 unsigned long *dst_pivots;
4053 unsigned char dst_offset;
4054 unsigned char new_end = wr_mas->node_end;
4055 unsigned char offset;
4056 unsigned char node_slots = mt_slots[wr_mas->type];
4057 struct maple_node reuse, *newnode;
4058 unsigned char copy_size, max_piv = mt_pivots[wr_mas->type];
4059 bool in_rcu = mt_in_rcu(mas->tree);
4061 offset = mas->offset;
4062 if (mas->last == wr_mas->r_max) {
4063 /* runs right to the end of the node */
4064 if (mas->last == mas->max)
4066 /* don't copy this offset */
4067 wr_mas->offset_end++;
4068 } else if (mas->last < wr_mas->r_max) {
4069 /* new range ends in this range */
4070 if (unlikely(wr_mas->r_max == ULONG_MAX))
4071 mas_bulk_rebalance(mas, wr_mas->node_end, wr_mas->type);
4075 if (wr_mas->end_piv == mas->last)
4076 wr_mas->offset_end++;
4078 new_end -= wr_mas->offset_end - offset - 1;
4081 /* new range starts within a range */
4082 if (wr_mas->r_min < mas->index)
4085 /* Not enough room */
4086 if (new_end >= node_slots)
4089 /* Not enough data. */
4090 if (!mte_is_root(mas->node) && (new_end <= mt_min_slots[wr_mas->type]) &&
4091 !(mas->mas_flags & MA_STATE_BULK))
4096 mas_node_count(mas, 1);
4097 if (mas_is_err(mas))
4100 newnode = mas_pop_node(mas);
4102 memset(&reuse, 0, sizeof(struct maple_node));
4106 newnode->parent = mas_mn(mas)->parent;
4107 dst_pivots = ma_pivots(newnode, wr_mas->type);
4108 dst_slots = ma_slots(newnode, wr_mas->type);
4109 /* Copy from start to insert point */
4110 memcpy(dst_pivots, wr_mas->pivots, sizeof(unsigned long) * (offset + 1));
4111 memcpy(dst_slots, wr_mas->slots, sizeof(void *) * (offset + 1));
4112 dst_offset = offset;
4114 /* Handle insert of new range starting after old range */
4115 if (wr_mas->r_min < mas->index) {
4117 rcu_assign_pointer(dst_slots[dst_offset], wr_mas->content);
4118 dst_pivots[dst_offset++] = mas->index - 1;
4121 /* Store the new entry and range end. */
4122 if (dst_offset < max_piv)
4123 dst_pivots[dst_offset] = mas->last;
4124 mas->offset = dst_offset;
4125 rcu_assign_pointer(dst_slots[dst_offset], wr_mas->entry);
4128 * this range wrote to the end of the node or it overwrote the rest of
4131 if (wr_mas->offset_end > wr_mas->node_end || mas->last >= mas->max) {
4132 new_end = dst_offset;
4137 /* Copy to the end of node if necessary. */
4138 copy_size = wr_mas->node_end - wr_mas->offset_end + 1;
4139 memcpy(dst_slots + dst_offset, wr_mas->slots + wr_mas->offset_end,
4140 sizeof(void *) * copy_size);
4141 if (dst_offset < max_piv) {
4142 if (copy_size > max_piv - dst_offset)
4143 copy_size = max_piv - dst_offset;
4145 memcpy(dst_pivots + dst_offset,
4146 wr_mas->pivots + wr_mas->offset_end,
4147 sizeof(unsigned long) * copy_size);
4150 if ((wr_mas->node_end == node_slots - 1) && (new_end < node_slots - 1))
4151 dst_pivots[new_end] = mas->max;
4154 mas_leaf_set_meta(mas, newnode, dst_pivots, maple_leaf_64, new_end);
4156 mas->node = mt_mk_node(newnode, wr_mas->type);
4157 mas_replace(mas, false);
4159 memcpy(wr_mas->node, newnode, sizeof(struct maple_node));
4161 trace_ma_write(__func__, mas, 0, wr_mas->entry);
4162 mas_update_gap(mas);
4167 * mas_wr_slot_store: Attempt to store a value in a slot.
4168 * @wr_mas: the maple write state
4170 * Return: True if stored, false otherwise
4172 static inline bool mas_wr_slot_store(struct ma_wr_state *wr_mas)
4174 struct ma_state *mas = wr_mas->mas;
4175 unsigned long lmax; /* Logical max. */
4176 unsigned char offset = mas->offset;
4178 if ((wr_mas->r_max > mas->last) && ((wr_mas->r_min != mas->index) ||
4179 (offset != wr_mas->node_end)))
4182 if (offset == wr_mas->node_end - 1)
4185 lmax = wr_mas->pivots[offset + 1];
4187 /* going to overwrite too many slots. */
4188 if (lmax < mas->last)
4191 if (wr_mas->r_min == mas->index) {
4192 /* overwriting two or more ranges with one. */
4193 if (lmax == mas->last)
4196 /* Overwriting all of offset and a portion of offset + 1. */
4197 rcu_assign_pointer(wr_mas->slots[offset], wr_mas->entry);
4198 wr_mas->pivots[offset] = mas->last;
4202 /* Doesn't end on the next range end. */
4203 if (lmax != mas->last)
4206 /* Overwriting a portion of offset and all of offset + 1 */
4207 if ((offset + 1 < mt_pivots[wr_mas->type]) &&
4208 (wr_mas->entry || wr_mas->pivots[offset + 1]))
4209 wr_mas->pivots[offset + 1] = mas->last;
4211 rcu_assign_pointer(wr_mas->slots[offset + 1], wr_mas->entry);
4212 wr_mas->pivots[offset] = mas->index - 1;
4213 mas->offset++; /* Keep mas accurate. */
4216 trace_ma_write(__func__, mas, 0, wr_mas->entry);
4217 mas_update_gap(mas);
4221 static inline void mas_wr_end_piv(struct ma_wr_state *wr_mas)
4223 while ((wr_mas->mas->last > wr_mas->end_piv) &&
4224 (wr_mas->offset_end < wr_mas->node_end))
4225 wr_mas->end_piv = wr_mas->pivots[++wr_mas->offset_end];
4227 if (wr_mas->mas->last > wr_mas->end_piv)
4228 wr_mas->end_piv = wr_mas->mas->max;
4231 static inline void mas_wr_extend_null(struct ma_wr_state *wr_mas)
4233 struct ma_state *mas = wr_mas->mas;
4235 if (mas->last < wr_mas->end_piv && !wr_mas->slots[wr_mas->offset_end])
4236 mas->last = wr_mas->end_piv;
4238 /* Check next slot(s) if we are overwriting the end */
4239 if ((mas->last == wr_mas->end_piv) &&
4240 (wr_mas->node_end != wr_mas->offset_end) &&
4241 !wr_mas->slots[wr_mas->offset_end + 1]) {
4242 wr_mas->offset_end++;
4243 if (wr_mas->offset_end == wr_mas->node_end)
4244 mas->last = mas->max;
4246 mas->last = wr_mas->pivots[wr_mas->offset_end];
4247 wr_mas->end_piv = mas->last;
4250 if (!wr_mas->content) {
4251 /* If this one is null, the next and prev are not */
4252 mas->index = wr_mas->r_min;
4254 /* Check prev slot if we are overwriting the start */
4255 if (mas->index == wr_mas->r_min && mas->offset &&
4256 !wr_mas->slots[mas->offset - 1]) {
4258 wr_mas->r_min = mas->index =
4259 mas_safe_min(mas, wr_mas->pivots, mas->offset);
4260 wr_mas->r_max = wr_mas->pivots[mas->offset];
4265 static inline bool mas_wr_append(struct ma_wr_state *wr_mas)
4267 unsigned char end = wr_mas->node_end;
4268 unsigned char new_end = end + 1;
4269 struct ma_state *mas = wr_mas->mas;
4270 unsigned char node_pivots = mt_pivots[wr_mas->type];
4272 if ((mas->index != wr_mas->r_min) && (mas->last == wr_mas->r_max)) {
4273 if (new_end < node_pivots)
4274 wr_mas->pivots[new_end] = wr_mas->pivots[end];
4276 if (new_end < node_pivots)
4277 ma_set_meta(wr_mas->node, maple_leaf_64, 0, new_end);
4279 rcu_assign_pointer(wr_mas->slots[new_end], wr_mas->entry);
4280 mas->offset = new_end;
4281 wr_mas->pivots[end] = mas->index - 1;
4286 if ((mas->index == wr_mas->r_min) && (mas->last < wr_mas->r_max)) {
4287 if (new_end < node_pivots)
4288 wr_mas->pivots[new_end] = wr_mas->pivots[end];
4290 rcu_assign_pointer(wr_mas->slots[new_end], wr_mas->content);
4291 if (new_end < node_pivots)
4292 ma_set_meta(wr_mas->node, maple_leaf_64, 0, new_end);
4294 wr_mas->pivots[end] = mas->last;
4295 rcu_assign_pointer(wr_mas->slots[end], wr_mas->entry);
4303 * mas_wr_bnode() - Slow path for a modification.
4304 * @wr_mas: The write maple state
4306 * This is where split, rebalance end up.
4308 static void mas_wr_bnode(struct ma_wr_state *wr_mas)
4310 struct maple_big_node b_node;
4312 trace_ma_write(__func__, wr_mas->mas, 0, wr_mas->entry);
4313 memset(&b_node, 0, sizeof(struct maple_big_node));
4314 mas_store_b_node(wr_mas, &b_node, wr_mas->offset_end);
4315 mas_commit_b_node(wr_mas, &b_node, wr_mas->node_end);
4318 static inline void mas_wr_modify(struct ma_wr_state *wr_mas)
4320 unsigned char node_slots;
4321 unsigned char node_size;
4322 struct ma_state *mas = wr_mas->mas;
4324 /* Direct replacement */
4325 if (wr_mas->r_min == mas->index && wr_mas->r_max == mas->last) {
4326 rcu_assign_pointer(wr_mas->slots[mas->offset], wr_mas->entry);
4327 if (!!wr_mas->entry ^ !!wr_mas->content)
4328 mas_update_gap(mas);
4332 /* Attempt to append */
4333 node_slots = mt_slots[wr_mas->type];
4334 node_size = wr_mas->node_end - wr_mas->offset_end + mas->offset + 2;
4335 if (mas->max == ULONG_MAX)
4338 /* slot and node store will not fit, go to the slow path */
4339 if (unlikely(node_size >= node_slots))
4342 if (wr_mas->entry && (wr_mas->node_end < node_slots - 1) &&
4343 (mas->offset == wr_mas->node_end) && mas_wr_append(wr_mas)) {
4344 if (!wr_mas->content || !wr_mas->entry)
4345 mas_update_gap(mas);
4349 if ((wr_mas->offset_end - mas->offset <= 1) && mas_wr_slot_store(wr_mas))
4351 else if (mas_wr_node_store(wr_mas))
4354 if (mas_is_err(mas))
4358 mas_wr_bnode(wr_mas);
4362 * mas_wr_store_entry() - Internal call to store a value
4363 * @mas: The maple state
4364 * @entry: The entry to store.
4366 * Return: The contents that was stored at the index.
4368 static inline void *mas_wr_store_entry(struct ma_wr_state *wr_mas)
4370 struct ma_state *mas = wr_mas->mas;
4372 wr_mas->content = mas_start(mas);
4373 if (mas_is_none(mas) || mas_is_ptr(mas)) {
4374 mas_store_root(mas, wr_mas->entry);
4375 return wr_mas->content;
4378 if (unlikely(!mas_wr_walk(wr_mas))) {
4379 mas_wr_spanning_store(wr_mas);
4380 return wr_mas->content;
4383 /* At this point, we are at the leaf node that needs to be altered. */
4384 wr_mas->end_piv = wr_mas->r_max;
4385 mas_wr_end_piv(wr_mas);
4388 mas_wr_extend_null(wr_mas);
4390 /* New root for a single pointer */
4391 if (unlikely(!mas->index && mas->last == ULONG_MAX)) {
4392 mas_new_root(mas, wr_mas->entry);
4393 return wr_mas->content;
4396 mas_wr_modify(wr_mas);
4397 return wr_mas->content;
4401 * mas_insert() - Internal call to insert a value
4402 * @mas: The maple state
4403 * @entry: The entry to store
4405 * Return: %NULL or the contents that already exists at the requested index
4406 * otherwise. The maple state needs to be checked for error conditions.
4408 static inline void *mas_insert(struct ma_state *mas, void *entry)
4410 MA_WR_STATE(wr_mas, mas, entry);
4413 * Inserting a new range inserts either 0, 1, or 2 pivots within the
4414 * tree. If the insert fits exactly into an existing gap with a value
4415 * of NULL, then the slot only needs to be written with the new value.
4416 * If the range being inserted is adjacent to another range, then only a
4417 * single pivot needs to be inserted (as well as writing the entry). If
4418 * the new range is within a gap but does not touch any other ranges,
4419 * then two pivots need to be inserted: the start - 1, and the end. As
4420 * usual, the entry must be written. Most operations require a new node
4421 * to be allocated and replace an existing node to ensure RCU safety,
4422 * when in RCU mode. The exception to requiring a newly allocated node
4423 * is when inserting at the end of a node (appending). When done
4424 * carefully, appending can reuse the node in place.
4426 wr_mas.content = mas_start(mas);
4430 if (mas_is_none(mas) || mas_is_ptr(mas)) {
4431 mas_store_root(mas, entry);
4435 /* spanning writes always overwrite something */
4436 if (!mas_wr_walk(&wr_mas))
4439 /* At this point, we are at the leaf node that needs to be altered. */
4440 wr_mas.offset_end = mas->offset;
4441 wr_mas.end_piv = wr_mas.r_max;
4443 if (wr_mas.content || (mas->last > wr_mas.r_max))
4449 mas_wr_modify(&wr_mas);
4450 return wr_mas.content;
4453 mas_set_err(mas, -EEXIST);
4454 return wr_mas.content;
4459 * mas_prev_node() - Find the prev non-null entry at the same level in the
4460 * tree. The prev value will be mas->node[mas->offset] or MAS_NONE.
4461 * @mas: The maple state
4462 * @min: The lower limit to search
4464 * The prev node value will be mas->node[mas->offset] or MAS_NONE.
4465 * Return: 1 if the node is dead, 0 otherwise.
4467 static inline int mas_prev_node(struct ma_state *mas, unsigned long min)
4472 struct maple_node *node;
4473 struct maple_enode *enode;
4474 unsigned long *pivots;
4476 if (mas_is_none(mas))
4482 if (ma_is_root(node))
4486 if (unlikely(mas_ascend(mas)))
4488 offset = mas->offset;
4493 mt = mte_node_type(mas->node);
4495 slots = ma_slots(node, mt);
4496 pivots = ma_pivots(node, mt);
4497 mas->max = pivots[offset];
4499 mas->min = pivots[offset - 1] + 1;
4500 if (unlikely(ma_dead_node(node)))
4508 enode = mas_slot(mas, slots, offset);
4509 if (unlikely(ma_dead_node(node)))
4513 mt = mte_node_type(mas->node);
4515 slots = ma_slots(node, mt);
4516 pivots = ma_pivots(node, mt);
4517 offset = ma_data_end(node, mt, pivots, mas->max);
4519 mas->min = pivots[offset - 1] + 1;
4521 if (offset < mt_pivots[mt])
4522 mas->max = pivots[offset];
4528 mas->node = mas_slot(mas, slots, offset);
4529 if (unlikely(ma_dead_node(node)))
4532 mas->offset = mas_data_end(mas);
4533 if (unlikely(mte_dead_node(mas->node)))
4539 mas->offset = offset;
4541 mas->min = pivots[offset - 1] + 1;
4543 if (unlikely(ma_dead_node(node)))
4546 mas->node = MAS_NONE;
4551 * mas_next_node() - Get the next node at the same level in the tree.
4552 * @mas: The maple state
4553 * @max: The maximum pivot value to check.
4555 * The next value will be mas->node[mas->offset] or MAS_NONE.
4556 * Return: 1 on dead node, 0 otherwise.
4558 static inline int mas_next_node(struct ma_state *mas, struct maple_node *node,
4561 unsigned long min, pivot;
4562 unsigned long *pivots;
4563 struct maple_enode *enode;
4565 unsigned char offset;
4569 if (mas->max >= max)
4574 if (ma_is_root(node))
4581 if (unlikely(mas_ascend(mas)))
4584 offset = mas->offset;
4587 mt = mte_node_type(mas->node);
4588 pivots = ma_pivots(node, mt);
4589 } while (unlikely(offset == ma_data_end(node, mt, pivots, mas->max)));
4591 slots = ma_slots(node, mt);
4592 pivot = mas_safe_pivot(mas, pivots, ++offset, mt);
4593 while (unlikely(level > 1)) {
4594 /* Descend, if necessary */
4595 enode = mas_slot(mas, slots, offset);
4596 if (unlikely(ma_dead_node(node)))
4602 mt = mte_node_type(mas->node);
4603 slots = ma_slots(node, mt);
4604 pivots = ma_pivots(node, mt);
4609 enode = mas_slot(mas, slots, offset);
4610 if (unlikely(ma_dead_node(node)))
4619 if (unlikely(ma_dead_node(node)))
4622 mas->node = MAS_NONE;
4627 * mas_next_nentry() - Get the next node entry
4628 * @mas: The maple state
4629 * @max: The maximum value to check
4630 * @*range_start: Pointer to store the start of the range.
4632 * Sets @mas->offset to the offset of the next node entry, @mas->last to the
4633 * pivot of the entry.
4635 * Return: The next entry, %NULL otherwise
4637 static inline void *mas_next_nentry(struct ma_state *mas,
4638 struct maple_node *node, unsigned long max, enum maple_type type)
4640 unsigned char count;
4641 unsigned long pivot;
4642 unsigned long *pivots;
4646 if (mas->last == mas->max) {
4647 mas->index = mas->max;
4651 pivots = ma_pivots(node, type);
4652 slots = ma_slots(node, type);
4653 mas->index = mas_safe_min(mas, pivots, mas->offset);
4654 if (ma_dead_node(node))
4657 if (mas->index > max)
4660 count = ma_data_end(node, type, pivots, mas->max);
4661 if (mas->offset > count)
4664 while (mas->offset < count) {
4665 pivot = pivots[mas->offset];
4666 entry = mas_slot(mas, slots, mas->offset);
4667 if (ma_dead_node(node))
4676 mas->index = pivot + 1;
4680 if (mas->index > mas->max) {
4681 mas->index = mas->last;
4685 pivot = mas_safe_pivot(mas, pivots, mas->offset, type);
4686 entry = mas_slot(mas, slots, mas->offset);
4687 if (ma_dead_node(node))
4701 static inline void mas_rewalk(struct ma_state *mas, unsigned long index)
4705 mas_set(mas, index);
4706 mas_state_walk(mas);
4707 if (mas_is_start(mas))
4715 * mas_next_entry() - Internal function to get the next entry.
4716 * @mas: The maple state
4717 * @limit: The maximum range start.
4719 * Set the @mas->node to the next entry and the range_start to
4720 * the beginning value for the entry. Does not check beyond @limit.
4721 * Sets @mas->index and @mas->last to the limit if it is hit.
4722 * Restarts on dead nodes.
4724 * Return: the next entry or %NULL.
4726 static inline void *mas_next_entry(struct ma_state *mas, unsigned long limit)
4729 struct maple_enode *prev_node;
4730 struct maple_node *node;
4731 unsigned char offset;
4737 offset = mas->offset;
4738 prev_node = mas->node;
4740 mt = mte_node_type(mas->node);
4742 if (unlikely(mas->offset >= mt_slots[mt])) {
4743 mas->offset = mt_slots[mt] - 1;
4747 while (!mas_is_none(mas)) {
4748 entry = mas_next_nentry(mas, node, limit, mt);
4749 if (unlikely(ma_dead_node(node))) {
4750 mas_rewalk(mas, last);
4757 if (unlikely((mas->index > limit)))
4761 prev_node = mas->node;
4762 offset = mas->offset;
4763 if (unlikely(mas_next_node(mas, node, limit))) {
4764 mas_rewalk(mas, last);
4769 mt = mte_node_type(mas->node);
4772 mas->index = mas->last = limit;
4773 mas->offset = offset;
4774 mas->node = prev_node;
4779 * mas_prev_nentry() - Get the previous node entry.
4780 * @mas: The maple state.
4781 * @limit: The lower limit to check for a value.
4783 * Return: the entry, %NULL otherwise.
4785 static inline void *mas_prev_nentry(struct ma_state *mas, unsigned long limit,
4786 unsigned long index)
4788 unsigned long pivot, min;
4789 unsigned char offset;
4790 struct maple_node *mn;
4792 unsigned long *pivots;
4801 mt = mte_node_type(mas->node);
4802 offset = mas->offset - 1;
4803 if (offset >= mt_slots[mt])
4804 offset = mt_slots[mt] - 1;
4806 slots = ma_slots(mn, mt);
4807 pivots = ma_pivots(mn, mt);
4808 if (offset == mt_pivots[mt])
4811 pivot = pivots[offset];
4813 if (unlikely(ma_dead_node(mn))) {
4814 mas_rewalk(mas, index);
4818 while (offset && ((!mas_slot(mas, slots, offset) && pivot >= limit) ||
4820 pivot = pivots[--offset];
4822 min = mas_safe_min(mas, pivots, offset);
4823 entry = mas_slot(mas, slots, offset);
4824 if (unlikely(ma_dead_node(mn))) {
4825 mas_rewalk(mas, index);
4829 if (likely(entry)) {
4830 mas->offset = offset;
4837 static inline void *mas_prev_entry(struct ma_state *mas, unsigned long min)
4842 while (likely(!mas_is_none(mas))) {
4843 entry = mas_prev_nentry(mas, min, mas->index);
4844 if (unlikely(mas->last < min))
4850 if (unlikely(mas_prev_node(mas, min))) {
4851 mas_rewalk(mas, mas->index);
4860 mas->index = mas->last = min;
4865 * mas_rev_awalk() - Internal function. Reverse allocation walk. Find the
4866 * highest gap address of a given size in a given node and descend.
4867 * @mas: The maple state
4868 * @size: The needed size.
4870 * Return: True if found in a leaf, false otherwise.
4873 static bool mas_rev_awalk(struct ma_state *mas, unsigned long size)
4875 enum maple_type type = mte_node_type(mas->node);
4876 struct maple_node *node = mas_mn(mas);
4877 unsigned long *pivots, *gaps;
4879 unsigned long gap = 0;
4880 unsigned long max, min;
4881 unsigned char offset;
4883 if (unlikely(mas_is_err(mas)))
4886 if (ma_is_dense(type)) {
4888 mas->offset = (unsigned char)(mas->index - mas->min);
4892 pivots = ma_pivots(node, type);
4893 slots = ma_slots(node, type);
4894 gaps = ma_gaps(node, type);
4895 offset = mas->offset;
4896 min = mas_safe_min(mas, pivots, offset);
4897 /* Skip out of bounds. */
4898 while (mas->last < min)
4899 min = mas_safe_min(mas, pivots, --offset);
4901 max = mas_safe_pivot(mas, pivots, offset, type);
4902 while (mas->index <= max) {
4906 else if (!mas_slot(mas, slots, offset))
4907 gap = max - min + 1;
4910 if ((size <= gap) && (size <= mas->last - min + 1))
4914 /* Skip the next slot, it cannot be a gap. */
4919 max = pivots[offset];
4920 min = mas_safe_min(mas, pivots, offset);
4930 min = mas_safe_min(mas, pivots, offset);
4933 if (unlikely((mas->index > max) || (size - 1 > max - mas->index)))
4936 if (unlikely(ma_is_leaf(type))) {
4937 mas->offset = offset;
4939 mas->max = min + gap - 1;
4943 /* descend, only happens under lock. */
4944 mas->node = mas_slot(mas, slots, offset);
4947 mas->offset = mas_data_end(mas);
4951 if (!mte_is_root(mas->node))
4955 mas_set_err(mas, -EBUSY);
4959 static inline bool mas_anode_descend(struct ma_state *mas, unsigned long size)
4961 enum maple_type type = mte_node_type(mas->node);
4962 unsigned long pivot, min, gap = 0;
4963 unsigned char offset;
4964 unsigned long *gaps;
4965 unsigned long *pivots = ma_pivots(mas_mn(mas), type);
4966 void __rcu **slots = ma_slots(mas_mn(mas), type);
4969 if (ma_is_dense(type)) {
4970 mas->offset = (unsigned char)(mas->index - mas->min);
4974 gaps = ma_gaps(mte_to_node(mas->node), type);
4975 offset = mas->offset;
4976 min = mas_safe_min(mas, pivots, offset);
4977 for (; offset < mt_slots[type]; offset++) {
4978 pivot = mas_safe_pivot(mas, pivots, offset, type);
4979 if (offset && !pivot)
4982 /* Not within lower bounds */
4983 if (mas->index > pivot)
4988 else if (!mas_slot(mas, slots, offset))
4989 gap = min(pivot, mas->last) - max(mas->index, min) + 1;
4994 if (ma_is_leaf(type)) {
4998 if (mas->index <= pivot) {
4999 mas->node = mas_slot(mas, slots, offset);
5008 if (mas->last <= pivot) {
5009 mas_set_err(mas, -EBUSY);
5014 if (mte_is_root(mas->node))
5017 mas->offset = offset;
5022 * mas_walk() - Search for @mas->index in the tree.
5023 * @mas: The maple state.
5025 * mas->index and mas->last will be set to the range if there is a value. If
5026 * mas->node is MAS_NONE, reset to MAS_START.
5028 * Return: the entry at the location or %NULL.
5030 void *mas_walk(struct ma_state *mas)
5035 entry = mas_state_walk(mas);
5036 if (mas_is_start(mas))
5039 if (mas_is_ptr(mas)) {
5044 mas->last = ULONG_MAX;
5049 if (mas_is_none(mas)) {
5051 mas->last = ULONG_MAX;
5056 EXPORT_SYMBOL_GPL(mas_walk);
5058 static inline bool mas_rewind_node(struct ma_state *mas)
5063 if (mte_is_root(mas->node)) {
5073 mas->offset = --slot;
5078 * mas_skip_node() - Internal function. Skip over a node.
5079 * @mas: The maple state.
5081 * Return: true if there is another node, false otherwise.
5083 static inline bool mas_skip_node(struct ma_state *mas)
5085 if (mas_is_err(mas))
5089 if (mte_is_root(mas->node)) {
5090 if (mas->offset >= mas_data_end(mas)) {
5091 mas_set_err(mas, -EBUSY);
5097 } while (mas->offset >= mas_data_end(mas));
5104 * mas_awalk() - Allocation walk. Search from low address to high, for a gap of
5106 * @mas: The maple state
5107 * @size: The size of the gap required
5109 * Search between @mas->index and @mas->last for a gap of @size.
5111 static inline void mas_awalk(struct ma_state *mas, unsigned long size)
5113 struct maple_enode *last = NULL;
5116 * There are 4 options:
5117 * go to child (descend)
5118 * go back to parent (ascend)
5119 * no gap found. (return, slot == MAPLE_NODE_SLOTS)
5120 * found the gap. (return, slot != MAPLE_NODE_SLOTS)
5122 while (!mas_is_err(mas) && !mas_anode_descend(mas, size)) {
5123 if (last == mas->node)
5131 * mas_fill_gap() - Fill a located gap with @entry.
5132 * @mas: The maple state
5133 * @entry: The value to store
5134 * @slot: The offset into the node to store the @entry
5135 * @size: The size of the entry
5136 * @index: The start location
5138 static inline void mas_fill_gap(struct ma_state *mas, void *entry,
5139 unsigned char slot, unsigned long size, unsigned long *index)
5141 MA_WR_STATE(wr_mas, mas, entry);
5142 unsigned char pslot = mte_parent_slot(mas->node);
5143 struct maple_enode *mn = mas->node;
5144 unsigned long *pivots;
5145 enum maple_type ptype;
5147 * mas->index is the start address for the search
5148 * which may no longer be needed.
5149 * mas->last is the end address for the search
5152 *index = mas->index;
5153 mas->last = mas->index + size - 1;
5156 * It is possible that using mas->max and mas->min to correctly
5157 * calculate the index and last will cause an issue in the gap
5158 * calculation, so fix the ma_state here
5161 ptype = mte_node_type(mas->node);
5162 pivots = ma_pivots(mas_mn(mas), ptype);
5163 mas->max = mas_safe_pivot(mas, pivots, pslot, ptype);
5164 mas->min = mas_safe_min(mas, pivots, pslot);
5167 mas_wr_store_entry(&wr_mas);
5171 * mas_sparse_area() - Internal function. Return upper or lower limit when
5172 * searching for a gap in an empty tree.
5173 * @mas: The maple state
5174 * @min: the minimum range
5175 * @max: The maximum range
5176 * @size: The size of the gap
5177 * @fwd: Searching forward or back
5179 static inline void mas_sparse_area(struct ma_state *mas, unsigned long min,
5180 unsigned long max, unsigned long size, bool fwd)
5182 unsigned long start = 0;
5184 if (!unlikely(mas_is_none(mas)))
5193 mas->last = start + size - 1;
5201 * mas_empty_area() - Get the lowest address within the range that is
5202 * sufficient for the size requested.
5203 * @mas: The maple state
5204 * @min: The lowest value of the range
5205 * @max: The highest value of the range
5206 * @size: The size needed
5208 int mas_empty_area(struct ma_state *mas, unsigned long min,
5209 unsigned long max, unsigned long size)
5211 unsigned char offset;
5212 unsigned long *pivots;
5215 if (mas_is_start(mas))
5217 else if (mas->offset >= 2)
5219 else if (!mas_skip_node(mas))
5223 if (mas_is_none(mas) || mas_is_ptr(mas)) {
5224 mas_sparse_area(mas, min, max, size, true);
5228 /* The start of the window can only be within these values */
5231 mas_awalk(mas, size);
5233 if (unlikely(mas_is_err(mas)))
5234 return xa_err(mas->node);
5236 offset = mas->offset;
5237 if (unlikely(offset == MAPLE_NODE_SLOTS))
5240 mt = mte_node_type(mas->node);
5241 pivots = ma_pivots(mas_mn(mas), mt);
5243 mas->min = pivots[offset - 1] + 1;
5245 if (offset < mt_pivots[mt])
5246 mas->max = pivots[offset];
5248 if (mas->index < mas->min)
5249 mas->index = mas->min;
5251 mas->last = mas->index + size - 1;
5254 EXPORT_SYMBOL_GPL(mas_empty_area);
5257 * mas_empty_area_rev() - Get the highest address within the range that is
5258 * sufficient for the size requested.
5259 * @mas: The maple state
5260 * @min: The lowest value of the range
5261 * @max: The highest value of the range
5262 * @size: The size needed
5264 int mas_empty_area_rev(struct ma_state *mas, unsigned long min,
5265 unsigned long max, unsigned long size)
5267 struct maple_enode *last = mas->node;
5269 if (mas_is_start(mas)) {
5271 mas->offset = mas_data_end(mas);
5272 } else if (mas->offset >= 2) {
5274 } else if (!mas_rewind_node(mas)) {
5279 if (mas_is_none(mas) || mas_is_ptr(mas)) {
5280 mas_sparse_area(mas, min, max, size, false);
5284 /* The start of the window can only be within these values. */
5288 while (!mas_rev_awalk(mas, size)) {
5289 if (last == mas->node) {
5290 if (!mas_rewind_node(mas))
5297 if (mas_is_err(mas))
5298 return xa_err(mas->node);
5300 if (unlikely(mas->offset == MAPLE_NODE_SLOTS))
5304 * mas_rev_awalk() has set mas->min and mas->max to the gap values. If
5305 * the maximum is outside the window we are searching, then use the last
5306 * location in the search.
5307 * mas->max and mas->min is the range of the gap.
5308 * mas->index and mas->last are currently set to the search range.
5311 /* Trim the upper limit to the max. */
5312 if (mas->max <= mas->last)
5313 mas->last = mas->max;
5315 mas->index = mas->last - size + 1;
5318 EXPORT_SYMBOL_GPL(mas_empty_area_rev);
5320 static inline int mas_alloc(struct ma_state *mas, void *entry,
5321 unsigned long size, unsigned long *index)
5326 if (mas_is_none(mas) || mas_is_ptr(mas)) {
5327 mas_root_expand(mas, entry);
5328 if (mas_is_err(mas))
5329 return xa_err(mas->node);
5332 return mte_pivot(mas->node, 0);
5333 return mte_pivot(mas->node, 1);
5336 /* Must be walking a tree. */
5337 mas_awalk(mas, size);
5338 if (mas_is_err(mas))
5339 return xa_err(mas->node);
5341 if (mas->offset == MAPLE_NODE_SLOTS)
5345 * At this point, mas->node points to the right node and we have an
5346 * offset that has a sufficient gap.
5350 min = mte_pivot(mas->node, mas->offset - 1) + 1;
5352 if (mas->index < min)
5355 mas_fill_gap(mas, entry, mas->offset, size, index);
5362 static inline int mas_rev_alloc(struct ma_state *mas, unsigned long min,
5363 unsigned long max, void *entry,
5364 unsigned long size, unsigned long *index)
5368 ret = mas_empty_area_rev(mas, min, max, size);
5372 if (mas_is_err(mas))
5373 return xa_err(mas->node);
5375 if (mas->offset == MAPLE_NODE_SLOTS)
5378 mas_fill_gap(mas, entry, mas->offset, size, index);
5386 * mas_dead_leaves() - Mark all leaves of a node as dead.
5387 * @mas: The maple state
5388 * @slots: Pointer to the slot array
5390 * Must hold the write lock.
5392 * Return: The number of leaves marked as dead.
5395 unsigned char mas_dead_leaves(struct ma_state *mas, void __rcu **slots)
5397 struct maple_node *node;
5398 enum maple_type type;
5402 for (offset = 0; offset < mt_slot_count(mas->node); offset++) {
5403 entry = mas_slot_locked(mas, slots, offset);
5404 type = mte_node_type(entry);
5405 node = mte_to_node(entry);
5406 /* Use both node and type to catch LE & BE metadata */
5410 mte_set_node_dead(entry);
5411 smp_wmb(); /* Needed for RCU */
5413 rcu_assign_pointer(slots[offset], node);
5419 static void __rcu **mas_dead_walk(struct ma_state *mas, unsigned char offset)
5421 struct maple_node *node, *next;
5422 void __rcu **slots = NULL;
5426 mas->node = ma_enode_ptr(next);
5428 slots = ma_slots(node, node->type);
5429 next = mas_slot_locked(mas, slots, offset);
5431 } while (!ma_is_leaf(next->type));
5436 static void mt_free_walk(struct rcu_head *head)
5439 struct maple_node *node, *start;
5440 struct maple_tree mt;
5441 unsigned char offset;
5442 enum maple_type type;
5443 MA_STATE(mas, &mt, 0, 0);
5445 node = container_of(head, struct maple_node, rcu);
5447 if (ma_is_leaf(node->type))
5450 mt_init_flags(&mt, node->ma_flags);
5453 mas.node = mt_mk_node(node, node->type);
5454 slots = mas_dead_walk(&mas, 0);
5455 node = mas_mn(&mas);
5457 mt_free_bulk(node->slot_len, slots);
5458 offset = node->parent_slot + 1;
5459 mas.node = node->piv_parent;
5460 if (mas_mn(&mas) == node)
5461 goto start_slots_free;
5463 type = mte_node_type(mas.node);
5464 slots = ma_slots(mte_to_node(mas.node), type);
5465 if ((offset < mt_slots[type]) && (slots[offset]))
5466 slots = mas_dead_walk(&mas, offset);
5468 node = mas_mn(&mas);
5469 } while ((node != start) || (node->slot_len < offset));
5471 slots = ma_slots(node, node->type);
5472 mt_free_bulk(node->slot_len, slots);
5477 mt_free_rcu(&node->rcu);
5480 static inline void __rcu **mas_destroy_descend(struct ma_state *mas,
5481 struct maple_enode *prev, unsigned char offset)
5483 struct maple_node *node;
5484 struct maple_enode *next = mas->node;
5485 void __rcu **slots = NULL;
5490 slots = ma_slots(node, mte_node_type(mas->node));
5491 next = mas_slot_locked(mas, slots, 0);
5492 if ((mte_dead_node(next)))
5493 next = mas_slot_locked(mas, slots, 1);
5495 mte_set_node_dead(mas->node);
5496 node->type = mte_node_type(mas->node);
5497 node->piv_parent = prev;
5498 node->parent_slot = offset;
5501 } while (!mte_is_leaf(next));
5506 static void mt_destroy_walk(struct maple_enode *enode, unsigned char ma_flags,
5510 struct maple_node *node = mte_to_node(enode);
5511 struct maple_enode *start;
5512 struct maple_tree mt;
5514 MA_STATE(mas, &mt, 0, 0);
5516 if (mte_is_leaf(enode))
5519 mt_init_flags(&mt, ma_flags);
5522 mas.node = start = enode;
5523 slots = mas_destroy_descend(&mas, start, 0);
5524 node = mas_mn(&mas);
5526 enum maple_type type;
5527 unsigned char offset;
5528 struct maple_enode *parent, *tmp;
5530 node->slot_len = mas_dead_leaves(&mas, slots);
5532 mt_free_bulk(node->slot_len, slots);
5533 offset = node->parent_slot + 1;
5534 mas.node = node->piv_parent;
5535 if (mas_mn(&mas) == node)
5536 goto start_slots_free;
5538 type = mte_node_type(mas.node);
5539 slots = ma_slots(mte_to_node(mas.node), type);
5540 if (offset >= mt_slots[type])
5543 tmp = mas_slot_locked(&mas, slots, offset);
5544 if (mte_node_type(tmp) && mte_to_node(tmp)) {
5547 slots = mas_destroy_descend(&mas, parent, offset);
5550 node = mas_mn(&mas);
5551 } while (start != mas.node);
5553 node = mas_mn(&mas);
5554 node->slot_len = mas_dead_leaves(&mas, slots);
5556 mt_free_bulk(node->slot_len, slots);
5563 mt_free_rcu(&node->rcu);
5567 * mte_destroy_walk() - Free a tree or sub-tree.
5568 * @enode - the encoded maple node (maple_enode) to start
5569 * @mn - the tree to free - needed for node types.
5571 * Must hold the write lock.
5573 static inline void mte_destroy_walk(struct maple_enode *enode,
5574 struct maple_tree *mt)
5576 struct maple_node *node = mte_to_node(enode);
5578 if (mt_in_rcu(mt)) {
5579 mt_destroy_walk(enode, mt->ma_flags, false);
5580 call_rcu(&node->rcu, mt_free_walk);
5582 mt_destroy_walk(enode, mt->ma_flags, true);
5586 static void mas_wr_store_setup(struct ma_wr_state *wr_mas)
5588 if (!mas_is_start(wr_mas->mas)) {
5589 if (mas_is_none(wr_mas->mas)) {
5590 mas_reset(wr_mas->mas);
5592 wr_mas->r_max = wr_mas->mas->max;
5593 wr_mas->type = mte_node_type(wr_mas->mas->node);
5594 if (mas_is_span_wr(wr_mas))
5595 mas_reset(wr_mas->mas);
5604 * mas_store() - Store an @entry.
5605 * @mas: The maple state.
5606 * @entry: The entry to store.
5608 * The @mas->index and @mas->last is used to set the range for the @entry.
5609 * Note: The @mas should have pre-allocated entries to ensure there is memory to
5610 * store the entry. Please see mas_expected_entries()/mas_destroy() for more details.
5612 * Return: the first entry between mas->index and mas->last or %NULL.
5614 void *mas_store(struct ma_state *mas, void *entry)
5616 MA_WR_STATE(wr_mas, mas, entry);
5618 trace_ma_write(__func__, mas, 0, entry);
5619 #ifdef CONFIG_DEBUG_MAPLE_TREE
5620 if (mas->index > mas->last)
5621 pr_err("Error %lu > %lu %p\n", mas->index, mas->last, entry);
5622 MT_BUG_ON(mas->tree, mas->index > mas->last);
5623 if (mas->index > mas->last) {
5624 mas_set_err(mas, -EINVAL);
5631 * Storing is the same operation as insert with the added caveat that it
5632 * can overwrite entries. Although this seems simple enough, one may
5633 * want to examine what happens if a single store operation was to
5634 * overwrite multiple entries within a self-balancing B-Tree.
5636 mas_wr_store_setup(&wr_mas);
5637 mas_wr_store_entry(&wr_mas);
5638 return wr_mas.content;
5640 EXPORT_SYMBOL_GPL(mas_store);
5643 * mas_store_gfp() - Store a value into the tree.
5644 * @mas: The maple state
5645 * @entry: The entry to store
5646 * @gfp: The GFP_FLAGS to use for allocations if necessary.
5648 * Return: 0 on success, -EINVAL on invalid request, -ENOMEM if memory could not
5651 int mas_store_gfp(struct ma_state *mas, void *entry, gfp_t gfp)
5653 MA_WR_STATE(wr_mas, mas, entry);
5655 mas_wr_store_setup(&wr_mas);
5656 trace_ma_write(__func__, mas, 0, entry);
5658 mas_wr_store_entry(&wr_mas);
5659 if (unlikely(mas_nomem(mas, gfp)))
5662 if (unlikely(mas_is_err(mas)))
5663 return xa_err(mas->node);
5667 EXPORT_SYMBOL_GPL(mas_store_gfp);
5670 * mas_store_prealloc() - Store a value into the tree using memory
5671 * preallocated in the maple state.
5672 * @mas: The maple state
5673 * @entry: The entry to store.
5675 void mas_store_prealloc(struct ma_state *mas, void *entry)
5677 MA_WR_STATE(wr_mas, mas, entry);
5679 mas_wr_store_setup(&wr_mas);
5680 trace_ma_write(__func__, mas, 0, entry);
5681 mas_wr_store_entry(&wr_mas);
5682 BUG_ON(mas_is_err(mas));
5685 EXPORT_SYMBOL_GPL(mas_store_prealloc);
5688 * mas_preallocate() - Preallocate enough nodes for a store operation
5689 * @mas: The maple state
5690 * @entry: The entry that will be stored
5691 * @gfp: The GFP_FLAGS to use for allocations.
5693 * Return: 0 on success, -ENOMEM if memory could not be allocated.
5695 int mas_preallocate(struct ma_state *mas, void *entry, gfp_t gfp)
5699 mas_node_count_gfp(mas, 1 + mas_mt_height(mas) * 3, gfp);
5700 mas->mas_flags |= MA_STATE_PREALLOC;
5701 if (likely(!mas_is_err(mas)))
5704 mas_set_alloc_req(mas, 0);
5705 ret = xa_err(mas->node);
5713 * mas_destroy() - destroy a maple state.
5714 * @mas: The maple state
5716 * Upon completion, check the left-most node and rebalance against the node to
5717 * the right if necessary. Frees any allocated nodes associated with this maple
5720 void mas_destroy(struct ma_state *mas)
5722 struct maple_alloc *node;
5725 * When using mas_for_each() to insert an expected number of elements,
5726 * it is possible that the number inserted is less than the expected
5727 * number. To fix an invalid final node, a check is performed here to
5728 * rebalance the previous node with the final node.
5730 if (mas->mas_flags & MA_STATE_REBALANCE) {
5733 if (mas_is_start(mas))
5736 mtree_range_walk(mas);
5737 end = mas_data_end(mas) + 1;
5738 if (end < mt_min_slot_count(mas->node) - 1)
5739 mas_destroy_rebalance(mas, end);
5741 mas->mas_flags &= ~MA_STATE_REBALANCE;
5743 mas->mas_flags &= ~(MA_STATE_BULK|MA_STATE_PREALLOC);
5745 while (mas->alloc && !((unsigned long)mas->alloc & 0x1)) {
5747 mas->alloc = node->slot[0];
5748 if (node->node_count > 0)
5749 mt_free_bulk(node->node_count,
5750 (void __rcu **)&node->slot[1]);
5751 kmem_cache_free(maple_node_cache, node);
5755 EXPORT_SYMBOL_GPL(mas_destroy);
5758 * mas_expected_entries() - Set the expected number of entries that will be inserted.
5759 * @mas: The maple state
5760 * @nr_entries: The number of expected entries.
5762 * This will attempt to pre-allocate enough nodes to store the expected number
5763 * of entries. The allocations will occur using the bulk allocator interface
5764 * for speed. Please call mas_destroy() on the @mas after inserting the entries
5765 * to ensure any unused nodes are freed.
5767 * Return: 0 on success, -ENOMEM if memory could not be allocated.
5769 int mas_expected_entries(struct ma_state *mas, unsigned long nr_entries)
5771 int nonleaf_cap = MAPLE_ARANGE64_SLOTS - 2;
5772 struct maple_enode *enode = mas->node;
5777 * Sometimes it is necessary to duplicate a tree to a new tree, such as
5778 * forking a process and duplicating the VMAs from one tree to a new
5779 * tree. When such a situation arises, it is known that the new tree is
5780 * not going to be used until the entire tree is populated. For
5781 * performance reasons, it is best to use a bulk load with RCU disabled.
5782 * This allows for optimistic splitting that favours the left and reuse
5783 * of nodes during the operation.
5786 /* Optimize splitting for bulk insert in-order */
5787 mas->mas_flags |= MA_STATE_BULK;
5790 * Avoid overflow, assume a gap between each entry and a trailing null.
5791 * If this is wrong, it just means allocation can happen during
5792 * insertion of entries.
5794 nr_nodes = max(nr_entries, nr_entries * 2 + 1);
5795 if (!mt_is_alloc(mas->tree))
5796 nonleaf_cap = MAPLE_RANGE64_SLOTS - 2;
5798 /* Leaves; reduce slots to keep space for expansion */
5799 nr_nodes = DIV_ROUND_UP(nr_nodes, MAPLE_RANGE64_SLOTS - 2);
5800 /* Internal nodes */
5801 nr_nodes += DIV_ROUND_UP(nr_nodes, nonleaf_cap);
5802 /* Add working room for split (2 nodes) + new parents */
5803 mas_node_count(mas, nr_nodes + 3);
5805 /* Detect if allocations run out */
5806 mas->mas_flags |= MA_STATE_PREALLOC;
5808 if (!mas_is_err(mas))
5811 ret = xa_err(mas->node);
5817 EXPORT_SYMBOL_GPL(mas_expected_entries);
5820 * mas_next() - Get the next entry.
5821 * @mas: The maple state
5822 * @max: The maximum index to check.
5824 * Returns the next entry after @mas->index.
5825 * Must hold rcu_read_lock or the write lock.
5826 * Can return the zero entry.
5828 * Return: The next entry or %NULL
5830 void *mas_next(struct ma_state *mas, unsigned long max)
5832 if (mas_is_none(mas) || mas_is_paused(mas))
5833 mas->node = MAS_START;
5835 if (mas_is_start(mas))
5836 mas_walk(mas); /* Retries on dead nodes handled by mas_walk */
5838 if (mas_is_ptr(mas)) {
5841 mas->last = ULONG_MAX;
5846 if (mas->last == ULONG_MAX)
5849 /* Retries on dead nodes handled by mas_next_entry */
5850 return mas_next_entry(mas, max);
5852 EXPORT_SYMBOL_GPL(mas_next);
5855 * mt_next() - get the next value in the maple tree
5856 * @mt: The maple tree
5857 * @index: The start index
5858 * @max: The maximum index to check
5860 * Return: The entry at @index or higher, or %NULL if nothing is found.
5862 void *mt_next(struct maple_tree *mt, unsigned long index, unsigned long max)
5865 MA_STATE(mas, mt, index, index);
5868 entry = mas_next(&mas, max);
5872 EXPORT_SYMBOL_GPL(mt_next);
5875 * mas_prev() - Get the previous entry
5876 * @mas: The maple state
5877 * @min: The minimum value to check.
5879 * Must hold rcu_read_lock or the write lock.
5880 * Will reset mas to MAS_START if the node is MAS_NONE. Will stop on not
5883 * Return: the previous value or %NULL.
5885 void *mas_prev(struct ma_state *mas, unsigned long min)
5888 /* Nothing comes before 0 */
5893 if (unlikely(mas_is_ptr(mas)))
5896 if (mas_is_none(mas) || mas_is_paused(mas))
5897 mas->node = MAS_START;
5899 if (mas_is_start(mas)) {
5905 if (mas_is_ptr(mas)) {
5911 mas->index = mas->last = 0;
5912 return mas_root_locked(mas);
5914 return mas_prev_entry(mas, min);
5916 EXPORT_SYMBOL_GPL(mas_prev);
5919 * mt_prev() - get the previous value in the maple tree
5920 * @mt: The maple tree
5921 * @index: The start index
5922 * @min: The minimum index to check
5924 * Return: The entry at @index or lower, or %NULL if nothing is found.
5926 void *mt_prev(struct maple_tree *mt, unsigned long index, unsigned long min)
5929 MA_STATE(mas, mt, index, index);
5932 entry = mas_prev(&mas, min);
5936 EXPORT_SYMBOL_GPL(mt_prev);
5939 * mas_pause() - Pause a mas_find/mas_for_each to drop the lock.
5940 * @mas: The maple state to pause
5942 * Some users need to pause a walk and drop the lock they're holding in
5943 * order to yield to a higher priority thread or carry out an operation
5944 * on an entry. Those users should call this function before they drop
5945 * the lock. It resets the @mas to be suitable for the next iteration
5946 * of the loop after the user has reacquired the lock. If most entries
5947 * found during a walk require you to call mas_pause(), the mt_for_each()
5948 * iterator may be more appropriate.
5951 void mas_pause(struct ma_state *mas)
5953 mas->node = MAS_PAUSE;
5955 EXPORT_SYMBOL_GPL(mas_pause);
5958 * mas_find() - On the first call, find the entry at or after mas->index up to
5959 * %max. Otherwise, find the entry after mas->index.
5960 * @mas: The maple state
5961 * @max: The maximum value to check.
5963 * Must hold rcu_read_lock or the write lock.
5964 * If an entry exists, last and index are updated accordingly.
5965 * May set @mas->node to MAS_NONE.
5967 * Return: The entry or %NULL.
5969 void *mas_find(struct ma_state *mas, unsigned long max)
5971 if (unlikely(mas_is_paused(mas))) {
5972 if (unlikely(mas->last == ULONG_MAX)) {
5973 mas->node = MAS_NONE;
5976 mas->node = MAS_START;
5977 mas->index = ++mas->last;
5980 if (unlikely(mas_is_start(mas))) {
5981 /* First run or continue */
5984 if (mas->index > max)
5987 entry = mas_walk(mas);
5992 if (unlikely(!mas_searchable(mas)))
5995 /* Retries on dead nodes handled by mas_next_entry */
5996 return mas_next_entry(mas, max);
5998 EXPORT_SYMBOL_GPL(mas_find);
6001 * mas_find_rev: On the first call, find the first non-null entry at or below
6002 * mas->index down to %min. Otherwise find the first non-null entry below
6003 * mas->index down to %min.
6004 * @mas: The maple state
6005 * @min: The minimum value to check.
6007 * Must hold rcu_read_lock or the write lock.
6008 * If an entry exists, last and index are updated accordingly.
6009 * May set @mas->node to MAS_NONE.
6011 * Return: The entry or %NULL.
6013 void *mas_find_rev(struct ma_state *mas, unsigned long min)
6015 if (unlikely(mas_is_paused(mas))) {
6016 if (unlikely(mas->last == ULONG_MAX)) {
6017 mas->node = MAS_NONE;
6020 mas->node = MAS_START;
6021 mas->last = --mas->index;
6024 if (unlikely(mas_is_start(mas))) {
6025 /* First run or continue */
6028 if (mas->index < min)
6031 entry = mas_walk(mas);
6036 if (unlikely(!mas_searchable(mas)))
6039 if (mas->index < min)
6042 /* Retries on dead nodes handled by mas_next_entry */
6043 return mas_prev_entry(mas, min);
6045 EXPORT_SYMBOL_GPL(mas_find_rev);
6048 * mas_erase() - Find the range in which index resides and erase the entire
6050 * @mas: The maple state
6052 * Must hold the write lock.
6053 * Searches for @mas->index, sets @mas->index and @mas->last to the range and
6054 * erases that range.
6056 * Return: the entry that was erased or %NULL, @mas->index and @mas->last are updated.
6058 void *mas_erase(struct ma_state *mas)
6061 MA_WR_STATE(wr_mas, mas, NULL);
6063 if (mas_is_none(mas) || mas_is_paused(mas))
6064 mas->node = MAS_START;
6066 /* Retry unnecessary when holding the write lock. */
6067 entry = mas_state_walk(mas);
6072 /* Must reset to ensure spanning writes of last slot are detected */
6074 mas_wr_store_setup(&wr_mas);
6075 mas_wr_store_entry(&wr_mas);
6076 if (mas_nomem(mas, GFP_KERNEL))
6081 EXPORT_SYMBOL_GPL(mas_erase);
6084 * mas_nomem() - Check if there was an error allocating and do the allocation
6085 * if necessary If there are allocations, then free them.
6086 * @mas: The maple state
6087 * @gfp: The GFP_FLAGS to use for allocations
6088 * Return: true on allocation, false otherwise.
6090 bool mas_nomem(struct ma_state *mas, gfp_t gfp)
6091 __must_hold(mas->tree->lock)
6093 if (likely(mas->node != MA_ERROR(-ENOMEM))) {
6098 if (gfpflags_allow_blocking(gfp) && !mt_external_lock(mas->tree)) {
6099 mtree_unlock(mas->tree);
6100 mas_alloc_nodes(mas, gfp);
6101 mtree_lock(mas->tree);
6103 mas_alloc_nodes(mas, gfp);
6106 if (!mas_allocated(mas))
6109 mas->node = MAS_START;
6113 void __init maple_tree_init(void)
6115 maple_node_cache = kmem_cache_create("maple_node",
6116 sizeof(struct maple_node), sizeof(struct maple_node),
6121 * mtree_load() - Load a value stored in a maple tree
6122 * @mt: The maple tree
6123 * @index: The index to load
6125 * Return: the entry or %NULL
6127 void *mtree_load(struct maple_tree *mt, unsigned long index)
6129 MA_STATE(mas, mt, index, index);
6132 trace_ma_read(__func__, &mas);
6135 entry = mas_start(&mas);
6136 if (unlikely(mas_is_none(&mas)))
6139 if (unlikely(mas_is_ptr(&mas))) {
6146 entry = mtree_lookup_walk(&mas);
6147 if (!entry && unlikely(mas_is_start(&mas)))
6151 if (xa_is_zero(entry))
6156 EXPORT_SYMBOL(mtree_load);
6159 * mtree_store_range() - Store an entry at a given range.
6160 * @mt: The maple tree
6161 * @index: The start of the range
6162 * @last: The end of the range
6163 * @entry: The entry to store
6164 * @gfp: The GFP_FLAGS to use for allocations
6166 * Return: 0 on success, -EINVAL on invalid request, -ENOMEM if memory could not
6169 int mtree_store_range(struct maple_tree *mt, unsigned long index,
6170 unsigned long last, void *entry, gfp_t gfp)
6172 MA_STATE(mas, mt, index, last);
6173 MA_WR_STATE(wr_mas, &mas, entry);
6175 trace_ma_write(__func__, &mas, 0, entry);
6176 if (WARN_ON_ONCE(xa_is_advanced(entry)))
6184 mas_wr_store_entry(&wr_mas);
6185 if (mas_nomem(&mas, gfp))
6189 if (mas_is_err(&mas))
6190 return xa_err(mas.node);
6194 EXPORT_SYMBOL(mtree_store_range);
6197 * mtree_store() - Store an entry at a given index.
6198 * @mt: The maple tree
6199 * @index: The index to store the value
6200 * @entry: The entry to store
6201 * @gfp: The GFP_FLAGS to use for allocations
6203 * Return: 0 on success, -EINVAL on invalid request, -ENOMEM if memory could not
6206 int mtree_store(struct maple_tree *mt, unsigned long index, void *entry,
6209 return mtree_store_range(mt, index, index, entry, gfp);
6211 EXPORT_SYMBOL(mtree_store);
6214 * mtree_insert_range() - Insert an entry at a give range if there is no value.
6215 * @mt: The maple tree
6216 * @first: The start of the range
6217 * @last: The end of the range
6218 * @entry: The entry to store
6219 * @gfp: The GFP_FLAGS to use for allocations.
6221 * Return: 0 on success, -EEXISTS if the range is occupied, -EINVAL on invalid
6222 * request, -ENOMEM if memory could not be allocated.
6224 int mtree_insert_range(struct maple_tree *mt, unsigned long first,
6225 unsigned long last, void *entry, gfp_t gfp)
6227 MA_STATE(ms, mt, first, last);
6229 if (WARN_ON_ONCE(xa_is_advanced(entry)))
6237 mas_insert(&ms, entry);
6238 if (mas_nomem(&ms, gfp))
6242 if (mas_is_err(&ms))
6243 return xa_err(ms.node);
6247 EXPORT_SYMBOL(mtree_insert_range);
6250 * mtree_insert() - Insert an entry at a give index if there is no value.
6251 * @mt: The maple tree
6252 * @index : The index to store the value
6253 * @entry: The entry to store
6254 * @gfp: The FGP_FLAGS to use for allocations.
6256 * Return: 0 on success, -EEXISTS if the range is occupied, -EINVAL on invalid
6257 * request, -ENOMEM if memory could not be allocated.
6259 int mtree_insert(struct maple_tree *mt, unsigned long index, void *entry,
6262 return mtree_insert_range(mt, index, index, entry, gfp);
6264 EXPORT_SYMBOL(mtree_insert);
6266 int mtree_alloc_range(struct maple_tree *mt, unsigned long *startp,
6267 void *entry, unsigned long size, unsigned long min,
6268 unsigned long max, gfp_t gfp)
6272 MA_STATE(mas, mt, min, max - size);
6273 if (!mt_is_alloc(mt))
6276 if (WARN_ON_ONCE(mt_is_reserved(entry)))
6292 mas.last = max - size;
6293 ret = mas_alloc(&mas, entry, size, startp);
6294 if (mas_nomem(&mas, gfp))
6300 EXPORT_SYMBOL(mtree_alloc_range);
6302 int mtree_alloc_rrange(struct maple_tree *mt, unsigned long *startp,
6303 void *entry, unsigned long size, unsigned long min,
6304 unsigned long max, gfp_t gfp)
6308 MA_STATE(mas, mt, min, max - size);
6309 if (!mt_is_alloc(mt))
6312 if (WARN_ON_ONCE(mt_is_reserved(entry)))
6326 ret = mas_rev_alloc(&mas, min, max, entry, size, startp);
6327 if (mas_nomem(&mas, gfp))
6333 EXPORT_SYMBOL(mtree_alloc_rrange);
6336 * mtree_erase() - Find an index and erase the entire range.
6337 * @mt: The maple tree
6338 * @index: The index to erase
6340 * Erasing is the same as a walk to an entry then a store of a NULL to that
6341 * ENTIRE range. In fact, it is implemented as such using the advanced API.
6343 * Return: The entry stored at the @index or %NULL
6345 void *mtree_erase(struct maple_tree *mt, unsigned long index)
6349 MA_STATE(mas, mt, index, index);
6350 trace_ma_op(__func__, &mas);
6353 entry = mas_erase(&mas);
6358 EXPORT_SYMBOL(mtree_erase);
6361 * __mt_destroy() - Walk and free all nodes of a locked maple tree.
6362 * @mt: The maple tree
6364 * Note: Does not handle locking.
6366 void __mt_destroy(struct maple_tree *mt)
6368 void *root = mt_root_locked(mt);
6370 rcu_assign_pointer(mt->ma_root, NULL);
6371 if (xa_is_node(root))
6372 mte_destroy_walk(root, mt);
6376 EXPORT_SYMBOL_GPL(__mt_destroy);
6379 * mtree_destroy() - Destroy a maple tree
6380 * @mt: The maple tree
6382 * Frees all resources used by the tree. Handles locking.
6384 void mtree_destroy(struct maple_tree *mt)
6390 EXPORT_SYMBOL(mtree_destroy);
6393 * mt_find() - Search from the start up until an entry is found.
6394 * @mt: The maple tree
6395 * @index: Pointer which contains the start location of the search
6396 * @max: The maximum value to check
6398 * Handles locking. @index will be incremented to one beyond the range.
6400 * Return: The entry at or after the @index or %NULL
6402 void *mt_find(struct maple_tree *mt, unsigned long *index, unsigned long max)
6404 MA_STATE(mas, mt, *index, *index);
6406 #ifdef CONFIG_DEBUG_MAPLE_TREE
6407 unsigned long copy = *index;
6410 trace_ma_read(__func__, &mas);
6417 entry = mas_state_walk(&mas);
6418 if (mas_is_start(&mas))
6421 if (unlikely(xa_is_zero(entry)))
6427 while (mas_searchable(&mas) && (mas.index < max)) {
6428 entry = mas_next_entry(&mas, max);
6429 if (likely(entry && !xa_is_zero(entry)))
6433 if (unlikely(xa_is_zero(entry)))
6437 if (likely(entry)) {
6438 *index = mas.last + 1;
6439 #ifdef CONFIG_DEBUG_MAPLE_TREE
6440 if ((*index) && (*index) <= copy)
6441 pr_err("index not increased! %lx <= %lx\n",
6443 MT_BUG_ON(mt, (*index) && ((*index) <= copy));
6449 EXPORT_SYMBOL(mt_find);
6452 * mt_find_after() - Search from the start up until an entry is found.
6453 * @mt: The maple tree
6454 * @index: Pointer which contains the start location of the search
6455 * @max: The maximum value to check
6457 * Handles locking, detects wrapping on index == 0
6459 * Return: The entry at or after the @index or %NULL
6461 void *mt_find_after(struct maple_tree *mt, unsigned long *index,
6467 return mt_find(mt, index, max);
6469 EXPORT_SYMBOL(mt_find_after);
6471 #ifdef CONFIG_DEBUG_MAPLE_TREE
6472 atomic_t maple_tree_tests_run;
6473 EXPORT_SYMBOL_GPL(maple_tree_tests_run);
6474 atomic_t maple_tree_tests_passed;
6475 EXPORT_SYMBOL_GPL(maple_tree_tests_passed);
6478 extern void kmem_cache_set_non_kernel(struct kmem_cache *, unsigned int);
6479 void mt_set_non_kernel(unsigned int val)
6481 kmem_cache_set_non_kernel(maple_node_cache, val);
6484 extern unsigned long kmem_cache_get_alloc(struct kmem_cache *);
6485 unsigned long mt_get_alloc_size(void)
6487 return kmem_cache_get_alloc(maple_node_cache);
6490 extern void kmem_cache_zero_nr_tallocated(struct kmem_cache *);
6491 void mt_zero_nr_tallocated(void)
6493 kmem_cache_zero_nr_tallocated(maple_node_cache);
6496 extern unsigned int kmem_cache_nr_tallocated(struct kmem_cache *);
6497 unsigned int mt_nr_tallocated(void)
6499 return kmem_cache_nr_tallocated(maple_node_cache);
6502 extern unsigned int kmem_cache_nr_allocated(struct kmem_cache *);
6503 unsigned int mt_nr_allocated(void)
6505 return kmem_cache_nr_allocated(maple_node_cache);
6509 * mas_dead_node() - Check if the maple state is pointing to a dead node.
6510 * @mas: The maple state
6511 * @index: The index to restore in @mas.
6513 * Used in test code.
6514 * Return: 1 if @mas has been reset to MAS_START, 0 otherwise.
6516 static inline int mas_dead_node(struct ma_state *mas, unsigned long index)
6518 if (unlikely(!mas_searchable(mas) || mas_is_start(mas)))
6521 if (likely(!mte_dead_node(mas->node)))
6524 mas_rewalk(mas, index);
6528 void mt_cache_shrink(void)
6533 * mt_cache_shrink() - For testing, don't use this.
6535 * Certain testcases can trigger an OOM when combined with other memory
6536 * debugging configuration options. This function is used to reduce the
6537 * possibility of an out of memory even due to kmem_cache objects remaining
6538 * around for longer than usual.
6540 void mt_cache_shrink(void)
6542 kmem_cache_shrink(maple_node_cache);
6545 EXPORT_SYMBOL_GPL(mt_cache_shrink);
6547 #endif /* not defined __KERNEL__ */
6549 * mas_get_slot() - Get the entry in the maple state node stored at @offset.
6550 * @mas: The maple state
6551 * @offset: The offset into the slot array to fetch.
6553 * Return: The entry stored at @offset.
6555 static inline struct maple_enode *mas_get_slot(struct ma_state *mas,
6556 unsigned char offset)
6558 return mas_slot(mas, ma_slots(mas_mn(mas), mte_node_type(mas->node)),
6564 * mas_first_entry() - Go the first leaf and find the first entry.
6565 * @mas: the maple state.
6566 * @limit: the maximum index to check.
6567 * @*r_start: Pointer to set to the range start.
6569 * Sets mas->offset to the offset of the entry, r_start to the range minimum.
6571 * Return: The first entry or MAS_NONE.
6573 static inline void *mas_first_entry(struct ma_state *mas, struct maple_node *mn,
6574 unsigned long limit, enum maple_type mt)
6578 unsigned long *pivots;
6582 mas->index = mas->min;
6583 if (mas->index > limit)
6588 while (likely(!ma_is_leaf(mt))) {
6589 MT_BUG_ON(mas->tree, mte_dead_node(mas->node));
6590 slots = ma_slots(mn, mt);
6591 pivots = ma_pivots(mn, mt);
6593 entry = mas_slot(mas, slots, 0);
6594 if (unlikely(ma_dead_node(mn)))
6598 mt = mte_node_type(mas->node);
6600 MT_BUG_ON(mas->tree, mte_dead_node(mas->node));
6603 slots = ma_slots(mn, mt);
6604 entry = mas_slot(mas, slots, 0);
6605 if (unlikely(ma_dead_node(mn)))
6608 /* Slot 0 or 1 must be set */
6609 if (mas->index > limit)
6615 pivots = ma_pivots(mn, mt);
6616 mas->index = pivots[0] + 1;
6618 entry = mas_slot(mas, slots, 1);
6619 if (unlikely(ma_dead_node(mn)))
6622 if (mas->index > limit)
6629 if (likely(!ma_dead_node(mn)))
6630 mas->node = MAS_NONE;
6634 /* Depth first search, post-order */
6635 static void mas_dfs_postorder(struct ma_state *mas, unsigned long max)
6638 struct maple_enode *p = MAS_NONE, *mn = mas->node;
6639 unsigned long p_min, p_max;
6641 mas_next_node(mas, mas_mn(mas), max);
6642 if (!mas_is_none(mas))
6645 if (mte_is_root(mn))
6650 while (mas->node != MAS_NONE) {
6654 mas_prev_node(mas, 0);
6665 /* Tree validations */
6666 static void mt_dump_node(const struct maple_tree *mt, void *entry,
6667 unsigned long min, unsigned long max, unsigned int depth);
6668 static void mt_dump_range(unsigned long min, unsigned long max,
6671 static const char spaces[] = " ";
6674 pr_info("%.*s%lu: ", depth * 2, spaces, min);
6676 pr_info("%.*s%lu-%lu: ", depth * 2, spaces, min, max);
6679 static void mt_dump_entry(void *entry, unsigned long min, unsigned long max,
6682 mt_dump_range(min, max, depth);
6684 if (xa_is_value(entry))
6685 pr_cont("value %ld (0x%lx) [%p]\n", xa_to_value(entry),
6686 xa_to_value(entry), entry);
6687 else if (xa_is_zero(entry))
6688 pr_cont("zero (%ld)\n", xa_to_internal(entry));
6689 else if (mt_is_reserved(entry))
6690 pr_cont("UNKNOWN ENTRY (%p)\n", entry);
6692 pr_cont("%p\n", entry);
6695 static void mt_dump_range64(const struct maple_tree *mt, void *entry,
6696 unsigned long min, unsigned long max, unsigned int depth)
6698 struct maple_range_64 *node = &mte_to_node(entry)->mr64;
6699 bool leaf = mte_is_leaf(entry);
6700 unsigned long first = min;
6703 pr_cont(" contents: ");
6704 for (i = 0; i < MAPLE_RANGE64_SLOTS - 1; i++)
6705 pr_cont("%p %lu ", node->slot[i], node->pivot[i]);
6706 pr_cont("%p\n", node->slot[i]);
6707 for (i = 0; i < MAPLE_RANGE64_SLOTS; i++) {
6708 unsigned long last = max;
6710 if (i < (MAPLE_RANGE64_SLOTS - 1))
6711 last = node->pivot[i];
6712 else if (!node->slot[i] && max != mt_max[mte_node_type(entry)])
6714 if (last == 0 && i > 0)
6717 mt_dump_entry(mt_slot(mt, node->slot, i),
6718 first, last, depth + 1);
6719 else if (node->slot[i])
6720 mt_dump_node(mt, mt_slot(mt, node->slot, i),
6721 first, last, depth + 1);
6726 pr_err("node %p last (%lu) > max (%lu) at pivot %d!\n",
6727 node, last, max, i);
6734 static void mt_dump_arange64(const struct maple_tree *mt, void *entry,
6735 unsigned long min, unsigned long max, unsigned int depth)
6737 struct maple_arange_64 *node = &mte_to_node(entry)->ma64;
6738 bool leaf = mte_is_leaf(entry);
6739 unsigned long first = min;
6742 pr_cont(" contents: ");
6743 for (i = 0; i < MAPLE_ARANGE64_SLOTS; i++)
6744 pr_cont("%lu ", node->gap[i]);
6745 pr_cont("| %02X %02X| ", node->meta.end, node->meta.gap);
6746 for (i = 0; i < MAPLE_ARANGE64_SLOTS - 1; i++)
6747 pr_cont("%p %lu ", node->slot[i], node->pivot[i]);
6748 pr_cont("%p\n", node->slot[i]);
6749 for (i = 0; i < MAPLE_ARANGE64_SLOTS; i++) {
6750 unsigned long last = max;
6752 if (i < (MAPLE_ARANGE64_SLOTS - 1))
6753 last = node->pivot[i];
6754 else if (!node->slot[i])
6756 if (last == 0 && i > 0)
6759 mt_dump_entry(mt_slot(mt, node->slot, i),
6760 first, last, depth + 1);
6761 else if (node->slot[i])
6762 mt_dump_node(mt, mt_slot(mt, node->slot, i),
6763 first, last, depth + 1);
6768 pr_err("node %p last (%lu) > max (%lu) at pivot %d!\n",
6769 node, last, max, i);
6776 static void mt_dump_node(const struct maple_tree *mt, void *entry,
6777 unsigned long min, unsigned long max, unsigned int depth)
6779 struct maple_node *node = mte_to_node(entry);
6780 unsigned int type = mte_node_type(entry);
6783 mt_dump_range(min, max, depth);
6785 pr_cont("node %p depth %d type %d parent %p", node, depth, type,
6786 node ? node->parent : NULL);
6790 for (i = 0; i < MAPLE_NODE_SLOTS; i++) {
6792 pr_cont("OUT OF RANGE: ");
6793 mt_dump_entry(mt_slot(mt, node->slot, i),
6794 min + i, min + i, depth);
6798 case maple_range_64:
6799 mt_dump_range64(mt, entry, min, max, depth);
6801 case maple_arange_64:
6802 mt_dump_arange64(mt, entry, min, max, depth);
6806 pr_cont(" UNKNOWN TYPE\n");
6810 void mt_dump(const struct maple_tree *mt)
6812 void *entry = rcu_dereference_check(mt->ma_root, mt_locked(mt));
6814 pr_info("maple_tree(%p) flags %X, height %u root %p\n",
6815 mt, mt->ma_flags, mt_height(mt), entry);
6816 if (!xa_is_node(entry))
6817 mt_dump_entry(entry, 0, 0, 0);
6819 mt_dump_node(mt, entry, 0, mt_max[mte_node_type(entry)], 0);
6821 EXPORT_SYMBOL_GPL(mt_dump);
6824 * Calculate the maximum gap in a node and check if that's what is reported in
6825 * the parent (unless root).
6827 static void mas_validate_gaps(struct ma_state *mas)
6829 struct maple_enode *mte = mas->node;
6830 struct maple_node *p_mn;
6831 unsigned long gap = 0, max_gap = 0;
6832 unsigned long p_end, p_start = mas->min;
6833 unsigned char p_slot;
6834 unsigned long *gaps = NULL;
6835 unsigned long *pivots = ma_pivots(mte_to_node(mte), mte_node_type(mte));
6838 if (ma_is_dense(mte_node_type(mte))) {
6839 for (i = 0; i < mt_slot_count(mte); i++) {
6840 if (mas_get_slot(mas, i)) {
6851 gaps = ma_gaps(mte_to_node(mte), mte_node_type(mte));
6852 for (i = 0; i < mt_slot_count(mte); i++) {
6853 p_end = mas_logical_pivot(mas, pivots, i, mte_node_type(mte));
6856 if (mas_get_slot(mas, i)) {
6861 gap += p_end - p_start + 1;
6863 void *entry = mas_get_slot(mas, i);
6867 if (gap != p_end - p_start + 1) {
6868 pr_err("%p[%u] -> %p %lu != %lu - %lu + 1\n",
6870 mas_get_slot(mas, i), gap,
6874 MT_BUG_ON(mas->tree,
6875 gap != p_end - p_start + 1);
6878 if (gap > p_end - p_start + 1) {
6879 pr_err("%p[%u] %lu >= %lu - %lu + 1 (%lu)\n",
6880 mas_mn(mas), i, gap, p_end, p_start,
6881 p_end - p_start + 1);
6882 MT_BUG_ON(mas->tree,
6883 gap > p_end - p_start + 1);
6891 p_start = p_end + 1;
6892 if (p_end >= mas->max)
6897 if (mte_is_root(mte))
6900 p_slot = mte_parent_slot(mas->node);
6901 p_mn = mte_parent(mte);
6902 MT_BUG_ON(mas->tree, max_gap > mas->max);
6903 if (ma_gaps(p_mn, mas_parent_enum(mas, mte))[p_slot] != max_gap) {
6904 pr_err("gap %p[%u] != %lu\n", p_mn, p_slot, max_gap);
6908 MT_BUG_ON(mas->tree,
6909 ma_gaps(p_mn, mas_parent_enum(mas, mte))[p_slot] != max_gap);
6912 static void mas_validate_parent_slot(struct ma_state *mas)
6914 struct maple_node *parent;
6915 struct maple_enode *node;
6916 enum maple_type p_type = mas_parent_enum(mas, mas->node);
6917 unsigned char p_slot = mte_parent_slot(mas->node);
6921 if (mte_is_root(mas->node))
6924 parent = mte_parent(mas->node);
6925 slots = ma_slots(parent, p_type);
6926 MT_BUG_ON(mas->tree, mas_mn(mas) == parent);
6928 /* Check prev/next parent slot for duplicate node entry */
6930 for (i = 0; i < mt_slots[p_type]; i++) {
6931 node = mas_slot(mas, slots, i);
6933 if (node != mas->node)
6934 pr_err("parent %p[%u] does not have %p\n",
6935 parent, i, mas_mn(mas));
6936 MT_BUG_ON(mas->tree, node != mas->node);
6937 } else if (node == mas->node) {
6938 pr_err("Invalid child %p at parent %p[%u] p_slot %u\n",
6939 mas_mn(mas), parent, i, p_slot);
6940 MT_BUG_ON(mas->tree, node == mas->node);
6945 static void mas_validate_child_slot(struct ma_state *mas)
6947 enum maple_type type = mte_node_type(mas->node);
6948 void __rcu **slots = ma_slots(mte_to_node(mas->node), type);
6949 unsigned long *pivots = ma_pivots(mte_to_node(mas->node), type);
6950 struct maple_enode *child;
6953 if (mte_is_leaf(mas->node))
6956 for (i = 0; i < mt_slots[type]; i++) {
6957 child = mas_slot(mas, slots, i);
6958 if (!pivots[i] || pivots[i] == mas->max)
6964 if (mte_parent_slot(child) != i) {
6965 pr_err("Slot error at %p[%u]: child %p has pslot %u\n",
6966 mas_mn(mas), i, mte_to_node(child),
6967 mte_parent_slot(child));
6968 MT_BUG_ON(mas->tree, 1);
6971 if (mte_parent(child) != mte_to_node(mas->node)) {
6972 pr_err("child %p has parent %p not %p\n",
6973 mte_to_node(child), mte_parent(child),
6974 mte_to_node(mas->node));
6975 MT_BUG_ON(mas->tree, 1);
6981 * Validate all pivots are within mas->min and mas->max.
6983 static void mas_validate_limits(struct ma_state *mas)
6986 unsigned long prev_piv = 0;
6987 enum maple_type type = mte_node_type(mas->node);
6988 void __rcu **slots = ma_slots(mte_to_node(mas->node), type);
6989 unsigned long *pivots = ma_pivots(mas_mn(mas), type);
6991 /* all limits are fine here. */
6992 if (mte_is_root(mas->node))
6995 for (i = 0; i < mt_slots[type]; i++) {
6998 piv = mas_safe_pivot(mas, pivots, i, type);
7000 if (!piv && (i != 0))
7003 if (!mte_is_leaf(mas->node)) {
7004 void *entry = mas_slot(mas, slots, i);
7007 pr_err("%p[%u] cannot be null\n",
7010 MT_BUG_ON(mas->tree, !entry);
7013 if (prev_piv > piv) {
7014 pr_err("%p[%u] piv %lu < prev_piv %lu\n",
7015 mas_mn(mas), i, piv, prev_piv);
7016 MT_BUG_ON(mas->tree, piv < prev_piv);
7019 if (piv < mas->min) {
7020 pr_err("%p[%u] %lu < %lu\n", mas_mn(mas), i,
7022 MT_BUG_ON(mas->tree, piv < mas->min);
7024 if (piv > mas->max) {
7025 pr_err("%p[%u] %lu > %lu\n", mas_mn(mas), i,
7027 MT_BUG_ON(mas->tree, piv > mas->max);
7030 if (piv == mas->max)
7033 for (i += 1; i < mt_slots[type]; i++) {
7034 void *entry = mas_slot(mas, slots, i);
7036 if (entry && (i != mt_slots[type] - 1)) {
7037 pr_err("%p[%u] should not have entry %p\n", mas_mn(mas),
7039 MT_BUG_ON(mas->tree, entry != NULL);
7042 if (i < mt_pivots[type]) {
7043 unsigned long piv = pivots[i];
7048 pr_err("%p[%u] should not have piv %lu\n",
7049 mas_mn(mas), i, piv);
7050 MT_BUG_ON(mas->tree, i < mt_pivots[type] - 1);
7055 static void mt_validate_nulls(struct maple_tree *mt)
7057 void *entry, *last = (void *)1;
7058 unsigned char offset = 0;
7060 MA_STATE(mas, mt, 0, 0);
7063 if (mas_is_none(&mas) || (mas.node == MAS_ROOT))
7066 while (!mte_is_leaf(mas.node))
7069 slots = ma_slots(mte_to_node(mas.node), mte_node_type(mas.node));
7071 entry = mas_slot(&mas, slots, offset);
7072 if (!last && !entry) {
7073 pr_err("Sequential nulls end at %p[%u]\n",
7074 mas_mn(&mas), offset);
7076 MT_BUG_ON(mt, !last && !entry);
7078 if (offset == mas_data_end(&mas)) {
7079 mas_next_node(&mas, mas_mn(&mas), ULONG_MAX);
7080 if (mas_is_none(&mas))
7083 slots = ma_slots(mte_to_node(mas.node),
7084 mte_node_type(mas.node));
7089 } while (!mas_is_none(&mas));
7093 * validate a maple tree by checking:
7094 * 1. The limits (pivots are within mas->min to mas->max)
7095 * 2. The gap is correctly set in the parents
7097 void mt_validate(struct maple_tree *mt)
7101 MA_STATE(mas, mt, 0, 0);
7104 if (!mas_searchable(&mas))
7107 mas_first_entry(&mas, mas_mn(&mas), ULONG_MAX, mte_node_type(mas.node));
7108 while (!mas_is_none(&mas)) {
7109 MT_BUG_ON(mas.tree, mte_dead_node(mas.node));
7110 if (!mte_is_root(mas.node)) {
7111 end = mas_data_end(&mas);
7112 if ((end < mt_min_slot_count(mas.node)) &&
7113 (mas.max != ULONG_MAX)) {
7114 pr_err("Invalid size %u of %p\n", end,
7116 MT_BUG_ON(mas.tree, 1);
7120 mas_validate_parent_slot(&mas);
7121 mas_validate_child_slot(&mas);
7122 mas_validate_limits(&mas);
7123 if (mt_is_alloc(mt))
7124 mas_validate_gaps(&mas);
7125 mas_dfs_postorder(&mas, ULONG_MAX);
7127 mt_validate_nulls(mt);
7132 EXPORT_SYMBOL_GPL(mt_validate);
7134 #endif /* CONFIG_DEBUG_MAPLE_TREE */