2 * SPDX-License-Identifier: LGPL-2.1-or-later
4 * Copyright 2010-2011 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
5 * Copyright 2011 Lai Jiangshan <laijs@cn.fujitsu.com>
7 * Userspace RCU library - Lock-Free Resizable RCU Hash Table
11 * Based on the following articles:
12 * - Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free
13 * extensible hash tables. J. ACM 53, 3 (May 2006), 379-405.
14 * - Michael, M. M. High performance dynamic lock-free hash tables
15 * and list-based sets. In Proceedings of the fourteenth annual ACM
16 * symposium on Parallel algorithms and architectures, ACM Press,
19 * Some specificities of this Lock-Free Resizable RCU Hash Table
22 * - RCU read-side critical section allows readers to perform hash
23 * table lookups, as well as traversals, and use the returned objects
24 * safely by allowing memory reclaim to take place only after a grace
26 * - Add and remove operations are lock-free, and do not need to
27 * allocate memory. They need to be executed within RCU read-side
28 * critical section to ensure the objects they read are valid and to
29 * deal with the cmpxchg ABA problem.
30 * - add and add_unique operations are supported. add_unique checks if
31 * the node key already exists in the hash table. It ensures not to
32 * populate a duplicate key if the node key already exists in the hash
34 * - The resize operation executes concurrently with
35 * add/add_unique/add_replace/remove/lookup/traversal.
36 * - Hash table nodes are contained within a split-ordered list. This
37 * list is ordered by incrementing reversed-bits-hash value.
38 * - An index of bucket nodes is kept. These bucket nodes are the hash
39 * table "buckets". These buckets are internal nodes that allow to
40 * perform a fast hash lookup, similarly to a skip list. These
41 * buckets are chained together in the split-ordered list, which
42 * allows recursive expansion by inserting new buckets between the
43 * existing buckets. The split-ordered list allows adding new buckets
44 * between existing buckets as the table needs to grow.
45 * - The resize operation for small tables only allows expanding the
46 * hash table. It is triggered automatically by detecting long chains
47 * in the add operation.
48 * - The resize operation for larger tables (and available through an
49 * API) allows both expanding and shrinking the hash table.
50 * - Split-counters are used to keep track of the number of
51 * nodes within the hash table for automatic resize triggering.
52 * - Resize operation initiated by long chain detection is executed by a
53 * worker thread, which keeps lock-freedom of add and remove.
54 * - Resize operations are protected by a mutex.
55 * - The removal operation is split in two parts: first, a "removed"
56 * flag is set in the next pointer within the node to remove. Then,
57 * a "garbage collection" is performed in the bucket containing the
58 * removed node (from the start of the bucket up to the removed node).
59 * All encountered nodes with "removed" flag set in their next
60 * pointers are removed from the linked-list. If the cmpxchg used for
61 * removal fails (due to concurrent garbage-collection or concurrent
62 * add), we retry from the beginning of the bucket. This ensures that
63 * the node with "removed" flag set is removed from the hash table
64 * (not visible to lookups anymore) before the RCU read-side critical
65 * section held across removal ends. Furthermore, this ensures that
66 * the node with "removed" flag set is removed from the linked-list
67 * before its memory is reclaimed. After setting the "removal" flag,
68 * only the thread which removal is the first to set the "removal
69 * owner" flag (with an xchg) into a node's next pointer is considered
70 * to have succeeded its removal (and thus owns the node to reclaim).
71 * Because we garbage-collect starting from an invariant node (the
72 * start-of-bucket bucket node) up to the "removed" node (or find a
73 * reverse-hash that is higher), we are sure that a successful
74 * traversal of the chain leads to a chain that is present in the
75 * linked-list (the start node is never removed) and that it does not
76 * contain the "removed" node anymore, even if concurrent delete/add
77 * operations are changing the structure of the list concurrently.
78 * - The add operations perform garbage collection of buckets if they
79 * encounter nodes with removed flag set in the bucket where they want
80 * to add their new node. This ensures lock-freedom of add operation by
81 * helping the remover unlink nodes from the list rather than to wait
83 * - There are three memory backends for the hash table buckets: the
84 * "order table", the "chunks", and the "mmap".
85 * - These bucket containers contain a compact version of the hash table
87 * - The RCU "order table":
88 * - has a first level table indexed by log2(hash index) which is
89 * copied and expanded by the resize operation. This order table
90 * allows finding the "bucket node" tables.
91 * - There is one bucket node table per hash index order. The size of
92 * each bucket node table is half the number of hashes contained in
93 * this order (except for order 0).
94 * - The RCU "chunks" is best suited for close interaction with a page
95 * allocator. It uses a linear array as index to "chunks" containing
96 * each the same number of buckets.
97 * - The RCU "mmap" memory backend uses a single memory map to hold
99 * - synchronize_rcu is used to garbage-collect the old bucket node table.
101 * Ordering Guarantees:
103 * To discuss these guarantees, we first define "read" operation as any
104 * of the the basic lttng_ust_lfht_lookup, lttng_ust_lfht_next_duplicate,
105 * lttng_ust_lfht_first, lttng_ust_lfht_next operation, as well as
106 * lttng_ust_lfht_add_unique (failure).
108 * We define "read traversal" operation as any of the following
109 * group of operations
110 * - lttng_ust_lfht_lookup followed by iteration with lttng_ust_lfht_next_duplicate
111 * (and/or lttng_ust_lfht_next, although less common).
112 * - lttng_ust_lfht_add_unique (failure) followed by iteration with
113 * lttng_ust_lfht_next_duplicate (and/or lttng_ust_lfht_next, although less
115 * - lttng_ust_lfht_first followed iteration with lttng_ust_lfht_next (and/or
116 * lttng_ust_lfht_next_duplicate, although less common).
118 * We define "write" operations as any of lttng_ust_lfht_add, lttng_ust_lfht_replace,
119 * lttng_ust_lfht_add_unique (success), lttng_ust_lfht_add_replace, lttng_ust_lfht_del.
121 * When lttng_ust_lfht_add_unique succeeds (returns the node passed as
122 * parameter), it acts as a "write" operation. When lttng_ust_lfht_add_unique
123 * fails (returns a node different from the one passed as parameter), it
124 * acts as a "read" operation. A lttng_ust_lfht_add_unique failure is a
125 * lttng_ust_lfht_lookup "read" operation, therefore, any ordering guarantee
126 * referring to "lookup" imply any of "lookup" or lttng_ust_lfht_add_unique
129 * We define "prior" and "later" node as nodes observable by reads and
130 * read traversals respectively before and after a write or sequence of
133 * Hash-table operations are often cascaded, for example, the pointer
134 * returned by a lttng_ust_lfht_lookup() might be passed to a lttng_ust_lfht_next(),
135 * whose return value might in turn be passed to another hash-table
136 * operation. This entire cascaded series of operations must be enclosed
137 * by a pair of matching rcu_read_lock() and rcu_read_unlock()
140 * The following ordering guarantees are offered by this hash table:
142 * A.1) "read" after "write": if there is ordering between a write and a
143 * later read, then the read is guaranteed to see the write or some
145 * A.2) "read traversal" after "write": given that there is dependency
146 * ordering between reads in a "read traversal", if there is
147 * ordering between a write and the first read of the traversal,
148 * then the "read traversal" is guaranteed to see the write or
150 * B.1) "write" after "read": if there is ordering between a read and a
151 * later write, then the read will never see the write.
152 * B.2) "write" after "read traversal": given that there is dependency
153 * ordering between reads in a "read traversal", if there is
154 * ordering between the last read of the traversal and a later
155 * write, then the "read traversal" will never see the write.
156 * C) "write" while "read traversal": if a write occurs during a "read
157 * traversal", the traversal may, or may not, see the write.
158 * D.1) "write" after "write": if there is ordering between a write and
159 * a later write, then the later write is guaranteed to see the
160 * effects of the first write.
161 * D.2) Concurrent "write" pairs: The system will assign an arbitrary
162 * order to any pair of concurrent conflicting writes.
163 * Non-conflicting writes (for example, to different keys) are
165 * E) If a grace period separates a "del" or "replace" operation
166 * and a subsequent operation, then that subsequent operation is
167 * guaranteed not to see the removed item.
168 * F) Uniqueness guarantee: given a hash table that does not contain
169 * duplicate items for a given key, there will only be one item in
170 * the hash table after an arbitrary sequence of add_unique and/or
171 * add_replace operations. Note, however, that a pair of
172 * concurrent read operations might well access two different items
174 * G.1) If a pair of lookups for a given key are ordered (e.g. by a
175 * memory barrier), then the second lookup will return the same
176 * node as the previous lookup, or some later node.
177 * G.2) A "read traversal" that starts after the end of a prior "read
178 * traversal" (ordered by memory barriers) is guaranteed to see the
179 * same nodes as the previous traversal, or some later nodes.
180 * G.3) Concurrent "read" pairs: concurrent reads are unordered. For
181 * example, if a pair of reads to the same key run concurrently
182 * with an insertion of that same key, the reads remain unordered
183 * regardless of their return values. In other words, you cannot
184 * rely on the values returned by the reads to deduce ordering.
186 * Progress guarantees:
188 * * Reads are wait-free. These operations always move forward in the
189 * hash table linked list, and this list has no loop.
190 * * Writes are lock-free. Any retry loop performed by a write operation
191 * is triggered by progress made within another update operation.
193 * Bucket node tables:
195 * hash table hash table the last all bucket node tables
196 * order size bucket node 0 1 2 3 4 5 6(index)
203 * 5 32 16 1 1 2 4 8 16
204 * 6 64 32 1 1 2 4 8 16 32
206 * When growing/shrinking, we only focus on the last bucket node table
207 * which size is (!order ? 1 : (1 << (order -1))).
209 * Example for growing/shrinking:
210 * grow hash table from order 5 to 6: init the index=6 bucket node table
211 * shrink hash table from order 6 to 5: fini the index=6 bucket node table
213 * A bit of ascii art explanation:
215 * The order index is the off-by-one compared to the actual power of 2
216 * because we use index 0 to deal with the 0 special-case.
218 * This shows the nodes for a small table ordered by reversed bits:
230 * This shows the nodes in order of non-reversed bits, linked by
231 * reversed-bit order.
236 * 2 | | 2 010 010 <- |
237 * | | | 3 011 110 | <- |
238 * 3 -> | | | 4 100 001 | |
245 * Note on port to lttng-ust: auto-resize and accounting features are
259 #include <lttng/ust-arch.h>
260 #include <lttng/urcu/pointer.h>
261 #include <urcu/arch.h>
262 #include <urcu/uatomic.h>
263 #include <urcu/compiler.h>
264 #include "rculfhash.h"
265 #include "rculfhash-internal.h"
271 * Split-counters lazily update the global counter each 1024
272 * addition/removal. It automatically keeps track of resize required.
273 * We use the bucket length as indicator for need to expand for small
274 * tables and machines lacking per-cpu data support.
276 #define COUNT_COMMIT_ORDER 10
279 * Define the minimum table size.
281 #define MIN_TABLE_ORDER 0
282 #define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
285 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
287 #define MIN_PARTITION_PER_THREAD_ORDER 12
288 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
291 * The removed flag needs to be updated atomically with the pointer.
292 * It indicates that no node must attach to the node scheduled for
293 * removal, and that node garbage collection must be performed.
294 * The bucket flag does not require to be updated atomically with the
295 * pointer, but it is added as a pointer low bit flag to save space.
296 * The "removal owner" flag is used to detect which of the "del"
297 * operation that has set the "removed flag" gets to return the removed
298 * node to its caller. Note that the replace operation does not need to
299 * iteract with the "removal owner" flag, because it validates that
300 * the "removed" flag is not set before performing its cmpxchg.
302 #define REMOVED_FLAG (1UL << 0)
303 #define BUCKET_FLAG (1UL << 1)
304 #define REMOVAL_OWNER_FLAG (1UL << 2)
305 #define FLAGS_MASK ((1UL << 3) - 1)
307 /* Value of the end pointer. Should not interact with flags. */
308 #define END_VALUE NULL
311 * ht_items_count: Split-counters counting the number of node addition
312 * and removal in the table. Only used if the LTTNG_UST_LFHT_ACCOUNTING flag
313 * is set at hash table creation.
315 * These are free-running counters, never reset to zero. They count the
316 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
317 * operations to update the global counter. We choose a power-of-2 value
318 * for the trigger to deal with 32 or 64-bit overflow of the counter.
320 struct ht_items_count
{
321 unsigned long add
, del
;
322 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
324 #ifdef CONFIG_LTTNG_UST_LFHT_ITER_DEBUG
327 void lttng_ust_lfht_iter_debug_set_ht(struct lttng_ust_lfht
*ht
, struct lttng_ust_lfht_iter
*iter
)
332 #define lttng_ust_lfht_iter_debug_assert(...) assert(__VA_ARGS__)
337 void lttng_ust_lfht_iter_debug_set_ht(struct lttng_ust_lfht
*ht
, struct lttng_ust_lfht_iter
*iter
)
341 #define lttng_ust_lfht_iter_debug_assert(...)
346 * Algorithm to reverse bits in a word by lookup table, extended to
349 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
350 * Originally from Public Domain.
353 static const uint8_t BitReverseTable256
[256] =
355 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
356 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
357 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
358 R6(0), R6(2), R6(1), R6(3)
365 uint8_t bit_reverse_u8(uint8_t v
)
367 return BitReverseTable256
[v
];
370 #if (CAA_BITS_PER_LONG == 32)
372 uint32_t bit_reverse_u32(uint32_t v
)
374 return ((uint32_t) bit_reverse_u8(v
) << 24) |
375 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
376 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
377 ((uint32_t) bit_reverse_u8(v
>> 24));
381 uint64_t bit_reverse_u64(uint64_t v
)
383 return ((uint64_t) bit_reverse_u8(v
) << 56) |
384 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
385 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
386 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
387 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
388 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
389 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
390 ((uint64_t) bit_reverse_u8(v
>> 56));
395 unsigned long bit_reverse_ulong(unsigned long v
)
397 #if (CAA_BITS_PER_LONG == 32)
398 return bit_reverse_u32(v
);
400 return bit_reverse_u64(v
);
405 * fls: returns the position of the most significant bit.
406 * Returns 0 if no bit is set, else returns the position of the most
407 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
409 #if defined(LTTNG_UST_ARCH_X86)
411 unsigned int fls_u32(uint32_t x
)
415 __asm__ ("bsrl %1,%0\n\t"
419 : "=r" (r
) : "rm" (x
));
425 #if defined(LTTNG_UST_ARCH_AMD64)
427 unsigned int fls_u64(uint64_t x
)
431 __asm__ ("bsrq %1,%0\n\t"
435 : "=r" (r
) : "rm" (x
));
443 unsigned int fls_u64(uint64_t x
)
444 __attribute__((unused
));
446 unsigned int fls_u64(uint64_t x
)
453 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
457 if (!(x
& 0xFFFF000000000000ULL
)) {
461 if (!(x
& 0xFF00000000000000ULL
)) {
465 if (!(x
& 0xF000000000000000ULL
)) {
469 if (!(x
& 0xC000000000000000ULL
)) {
473 if (!(x
& 0x8000000000000000ULL
)) {
483 unsigned int fls_u32(uint32_t x
)
484 __attribute__((unused
));
486 unsigned int fls_u32(uint32_t x
)
492 if (!(x
& 0xFFFF0000U
)) {
496 if (!(x
& 0xFF000000U
)) {
500 if (!(x
& 0xF0000000U
)) {
504 if (!(x
& 0xC0000000U
)) {
508 if (!(x
& 0x80000000U
)) {
516 unsigned int lttng_ust_lfht_fls_ulong(unsigned long x
)
518 #if (CAA_BITS_PER_LONG == 32)
526 * Return the minimum order for which x <= (1UL << order).
527 * Return -1 if x is 0.
529 int lttng_ust_lfht_get_count_order_u32(uint32_t x
)
534 return fls_u32(x
- 1);
538 * Return the minimum order for which x <= (1UL << order).
539 * Return -1 if x is 0.
541 int lttng_ust_lfht_get_count_order_ulong(unsigned long x
)
546 return lttng_ust_lfht_fls_ulong(x
- 1);
550 struct lttng_ust_lfht_node
*clear_flag(struct lttng_ust_lfht_node
*node
)
552 return (struct lttng_ust_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
556 int is_removed(const struct lttng_ust_lfht_node
*node
)
558 return ((unsigned long) node
) & REMOVED_FLAG
;
562 int is_bucket(struct lttng_ust_lfht_node
*node
)
564 return ((unsigned long) node
) & BUCKET_FLAG
;
568 struct lttng_ust_lfht_node
*flag_bucket(struct lttng_ust_lfht_node
*node
)
570 return (struct lttng_ust_lfht_node
*) (((unsigned long) node
) | BUCKET_FLAG
);
574 int is_removal_owner(struct lttng_ust_lfht_node
*node
)
576 return ((unsigned long) node
) & REMOVAL_OWNER_FLAG
;
580 struct lttng_ust_lfht_node
*flag_removal_owner(struct lttng_ust_lfht_node
*node
)
582 return (struct lttng_ust_lfht_node
*) (((unsigned long) node
) | REMOVAL_OWNER_FLAG
);
586 struct lttng_ust_lfht_node
*flag_removed_or_removal_owner(struct lttng_ust_lfht_node
*node
)
588 return (struct lttng_ust_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
| REMOVAL_OWNER_FLAG
);
592 struct lttng_ust_lfht_node
*get_end(void)
594 return (struct lttng_ust_lfht_node
*) END_VALUE
;
598 int is_end(struct lttng_ust_lfht_node
*node
)
600 return clear_flag(node
) == (struct lttng_ust_lfht_node
*) END_VALUE
;
604 void lttng_ust_lfht_alloc_bucket_table(struct lttng_ust_lfht
*ht
, unsigned long order
)
606 return ht
->mm
->alloc_bucket_table(ht
, order
);
610 * lttng_ust_lfht_free_bucket_table() should be called with decreasing order.
611 * When lttng_ust_lfht_free_bucket_table(0) is called, it means the whole
615 void lttng_ust_lfht_free_bucket_table(struct lttng_ust_lfht
*ht
, unsigned long order
)
617 return ht
->mm
->free_bucket_table(ht
, order
);
621 struct lttng_ust_lfht_node
*bucket_at(struct lttng_ust_lfht
*ht
, unsigned long index
)
623 return ht
->bucket_at(ht
, index
);
627 struct lttng_ust_lfht_node
*lookup_bucket(struct lttng_ust_lfht
*ht
, unsigned long size
,
631 return bucket_at(ht
, hash
& (size
- 1));
635 * Remove all logically deleted nodes from a bucket up to a certain node key.
638 void _lttng_ust_lfht_gc_bucket(struct lttng_ust_lfht_node
*bucket
, struct lttng_ust_lfht_node
*node
)
640 struct lttng_ust_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
642 assert(!is_bucket(bucket
));
643 assert(!is_removed(bucket
));
644 assert(!is_removal_owner(bucket
));
645 assert(!is_bucket(node
));
646 assert(!is_removed(node
));
647 assert(!is_removal_owner(node
));
650 /* We can always skip the bucket node initially */
651 iter
= lttng_ust_rcu_dereference(iter_prev
->next
);
652 assert(!is_removed(iter
));
653 assert(!is_removal_owner(iter
));
654 assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
656 * We should never be called with bucket (start of chain)
657 * and logically removed node (end of path compression
658 * marker) being the actual same node. This would be a
659 * bug in the algorithm implementation.
661 assert(bucket
!= node
);
663 if (caa_unlikely(is_end(iter
)))
665 if (caa_likely(clear_flag(iter
)->reverse_hash
> node
->reverse_hash
))
667 next
= lttng_ust_rcu_dereference(clear_flag(iter
)->next
);
668 if (caa_likely(is_removed(next
)))
670 iter_prev
= clear_flag(iter
);
673 assert(!is_removed(iter
));
674 assert(!is_removal_owner(iter
));
676 new_next
= flag_bucket(clear_flag(next
));
678 new_next
= clear_flag(next
);
679 (void) uatomic_cmpxchg(&iter_prev
->next
, iter
, new_next
);
684 int _lttng_ust_lfht_replace(struct lttng_ust_lfht
*ht
, unsigned long size
,
685 struct lttng_ust_lfht_node
*old_node
,
686 struct lttng_ust_lfht_node
*old_next
,
687 struct lttng_ust_lfht_node
*new_node
)
689 struct lttng_ust_lfht_node
*bucket
, *ret_next
;
691 if (!old_node
) /* Return -ENOENT if asked to replace NULL node */
694 assert(!is_removed(old_node
));
695 assert(!is_removal_owner(old_node
));
696 assert(!is_bucket(old_node
));
697 assert(!is_removed(new_node
));
698 assert(!is_removal_owner(new_node
));
699 assert(!is_bucket(new_node
));
700 assert(new_node
!= old_node
);
702 /* Insert after node to be replaced */
703 if (is_removed(old_next
)) {
705 * Too late, the old node has been removed under us
706 * between lookup and replace. Fail.
710 assert(old_next
== clear_flag(old_next
));
711 assert(new_node
!= old_next
);
713 * REMOVAL_OWNER flag is _NEVER_ set before the REMOVED
714 * flag. It is either set atomically at the same time
715 * (replace) or after (del).
717 assert(!is_removal_owner(old_next
));
718 new_node
->next
= old_next
;
720 * Here is the whole trick for lock-free replace: we add
721 * the replacement node _after_ the node we want to
722 * replace by atomically setting its next pointer at the
723 * same time we set its removal flag. Given that
724 * the lookups/get next use an iterator aware of the
725 * next pointer, they will either skip the old node due
726 * to the removal flag and see the new node, or use
727 * the old node, but will not see the new one.
728 * This is a replacement of a node with another node
729 * that has the same value: we are therefore not
730 * removing a value from the hash table. We set both the
731 * REMOVED and REMOVAL_OWNER flags atomically so we own
732 * the node after successful cmpxchg.
734 ret_next
= uatomic_cmpxchg(&old_node
->next
,
735 old_next
, flag_removed_or_removal_owner(new_node
));
736 if (ret_next
== old_next
)
737 break; /* We performed the replacement. */
742 * Ensure that the old node is not visible to readers anymore:
743 * lookup for the node, and remove it (along with any other
744 * logically removed node) if found.
746 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(old_node
->reverse_hash
));
747 _lttng_ust_lfht_gc_bucket(bucket
, new_node
);
749 assert(is_removed(CMM_LOAD_SHARED(old_node
->next
)));
754 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
755 * mode. A NULL unique_ret allows creation of duplicate keys.
758 void _lttng_ust_lfht_add(struct lttng_ust_lfht
*ht
,
760 lttng_ust_lfht_match_fct match
,
763 struct lttng_ust_lfht_node
*node
,
764 struct lttng_ust_lfht_iter
*unique_ret
,
767 struct lttng_ust_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
769 struct lttng_ust_lfht_node
*bucket
;
771 assert(!is_bucket(node
));
772 assert(!is_removed(node
));
773 assert(!is_removal_owner(node
));
774 bucket
= lookup_bucket(ht
, size
, hash
);
777 * iter_prev points to the non-removed node prior to the
781 /* We can always skip the bucket node initially */
782 iter
= lttng_ust_rcu_dereference(iter_prev
->next
);
783 assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
785 if (caa_unlikely(is_end(iter
)))
787 if (caa_likely(clear_flag(iter
)->reverse_hash
> node
->reverse_hash
))
790 /* bucket node is the first node of the identical-hash-value chain */
791 if (bucket_flag
&& clear_flag(iter
)->reverse_hash
== node
->reverse_hash
)
794 next
= lttng_ust_rcu_dereference(clear_flag(iter
)->next
);
795 if (caa_unlikely(is_removed(next
)))
801 && clear_flag(iter
)->reverse_hash
== node
->reverse_hash
) {
802 struct lttng_ust_lfht_iter d_iter
= {
805 #ifdef CONFIG_LTTNG_UST_LFHT_ITER_DEBUG
811 * uniquely adding inserts the node as the first
812 * node of the identical-hash-value node chain.
814 * This semantic ensures no duplicated keys
815 * should ever be observable in the table
816 * (including traversing the table node by
817 * node by forward iterations)
819 lttng_ust_lfht_next_duplicate(ht
, match
, key
, &d_iter
);
823 *unique_ret
= d_iter
;
827 iter_prev
= clear_flag(iter
);
832 assert(node
!= clear_flag(iter
));
833 assert(!is_removed(iter_prev
));
834 assert(!is_removal_owner(iter_prev
));
835 assert(!is_removed(iter
));
836 assert(!is_removal_owner(iter
));
837 assert(iter_prev
!= node
);
839 node
->next
= clear_flag(iter
);
841 node
->next
= flag_bucket(clear_flag(iter
));
843 new_node
= flag_bucket(node
);
846 if (uatomic_cmpxchg(&iter_prev
->next
, iter
,
848 continue; /* retry */
855 assert(!is_removed(iter
));
856 assert(!is_removal_owner(iter
));
858 new_next
= flag_bucket(clear_flag(next
));
860 new_next
= clear_flag(next
);
861 (void) uatomic_cmpxchg(&iter_prev
->next
, iter
, new_next
);
866 unique_ret
->node
= return_node
;
867 /* unique_ret->next left unset, never used. */
872 int _lttng_ust_lfht_del(struct lttng_ust_lfht
*ht
, unsigned long size
,
873 struct lttng_ust_lfht_node
*node
)
875 struct lttng_ust_lfht_node
*bucket
, *next
;
877 if (!node
) /* Return -ENOENT if asked to delete NULL node */
880 /* logically delete the node */
881 assert(!is_bucket(node
));
882 assert(!is_removed(node
));
883 assert(!is_removal_owner(node
));
886 * We are first checking if the node had previously been
887 * logically removed (this check is not atomic with setting the
888 * logical removal flag). Return -ENOENT if the node had
889 * previously been removed.
891 next
= CMM_LOAD_SHARED(node
->next
); /* next is not dereferenced */
892 if (caa_unlikely(is_removed(next
)))
894 assert(!is_bucket(next
));
896 * The del operation semantic guarantees a full memory barrier
897 * before the uatomic_or atomic commit of the deletion flag.
899 cmm_smp_mb__before_uatomic_or();
901 * We set the REMOVED_FLAG unconditionally. Note that there may
902 * be more than one concurrent thread setting this flag.
903 * Knowing which wins the race will be known after the garbage
904 * collection phase, stay tuned!
906 uatomic_or(&node
->next
, REMOVED_FLAG
);
907 /* We performed the (logical) deletion. */
910 * Ensure that the node is not visible to readers anymore: lookup for
911 * the node, and remove it (along with any other logically removed node)
914 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->reverse_hash
));
915 _lttng_ust_lfht_gc_bucket(bucket
, node
);
917 assert(is_removed(CMM_LOAD_SHARED(node
->next
)));
919 * Last phase: atomically exchange node->next with a version
920 * having "REMOVAL_OWNER_FLAG" set. If the returned node->next
921 * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own
922 * the node and win the removal race.
923 * It is interesting to note that all "add" paths are forbidden
924 * to change the next pointer starting from the point where the
925 * REMOVED_FLAG is set, so here using a read, followed by a
926 * xchg() suffice to guarantee that the xchg() will ever only
927 * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag
930 if (!is_removal_owner(uatomic_xchg(&node
->next
,
931 flag_removal_owner(node
->next
))))
938 * Never called with size < 1.
941 void lttng_ust_lfht_create_bucket(struct lttng_ust_lfht
*ht
, unsigned long size
)
943 struct lttng_ust_lfht_node
*prev
, *node
;
944 unsigned long order
, len
, i
;
947 lttng_ust_lfht_alloc_bucket_table(ht
, 0);
949 dbg_printf("create bucket: order 0 index 0 hash 0\n");
950 node
= bucket_at(ht
, 0);
951 node
->next
= flag_bucket(get_end());
952 node
->reverse_hash
= 0;
954 bucket_order
= lttng_ust_lfht_get_count_order_ulong(size
);
955 assert(bucket_order
>= 0);
957 for (order
= 1; order
< (unsigned long) bucket_order
+ 1; order
++) {
958 len
= 1UL << (order
- 1);
959 lttng_ust_lfht_alloc_bucket_table(ht
, order
);
961 for (i
= 0; i
< len
; i
++) {
963 * Now, we are trying to init the node with the
964 * hash=(len+i) (which is also a bucket with the
965 * index=(len+i)) and insert it into the hash table,
966 * so this node has to be inserted after the bucket
967 * with the index=(len+i)&(len-1)=i. And because there
968 * is no other non-bucket node nor bucket node with
969 * larger index/hash inserted, so the bucket node
970 * being inserted should be inserted directly linked
971 * after the bucket node with index=i.
973 prev
= bucket_at(ht
, i
);
974 node
= bucket_at(ht
, len
+ i
);
976 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
977 order
, len
+ i
, len
+ i
);
978 node
->reverse_hash
= bit_reverse_ulong(len
+ i
);
980 /* insert after prev */
981 assert(is_bucket(prev
->next
));
982 node
->next
= prev
->next
;
983 prev
->next
= flag_bucket(node
);
988 #if (CAA_BITS_PER_LONG > 32)
990 * For 64-bit architectures, with max number of buckets small enough not to
991 * use the entire 64-bit memory mapping space (and allowing a fair number of
992 * hash table instances), use the mmap allocator, which is faster. Otherwise,
993 * fallback to the order allocator.
996 const struct lttng_ust_lfht_mm_type
*get_mm_type(unsigned long max_nr_buckets
)
998 if (max_nr_buckets
&& max_nr_buckets
<= (1ULL << 32))
999 return <tng_ust_lfht_mm_mmap
;
1001 return <tng_ust_lfht_mm_order
;
1005 * For 32-bit architectures, use the order allocator.
1008 const struct lttng_ust_lfht_mm_type
*get_mm_type(unsigned long max_nr_buckets
)
1010 return <tng_ust_lfht_mm_order
;
1014 struct lttng_ust_lfht
*lttng_ust_lfht_new(unsigned long init_size
,
1015 unsigned long min_nr_alloc_buckets
,
1016 unsigned long max_nr_buckets
,
1018 const struct lttng_ust_lfht_mm_type
*mm
)
1020 struct lttng_ust_lfht
*ht
;
1021 unsigned long order
;
1023 /* min_nr_alloc_buckets must be power of two */
1024 if (!min_nr_alloc_buckets
|| (min_nr_alloc_buckets
& (min_nr_alloc_buckets
- 1)))
1027 /* init_size must be power of two */
1028 if (!init_size
|| (init_size
& (init_size
- 1)))
1032 * Memory management plugin default.
1035 mm
= get_mm_type(max_nr_buckets
);
1037 /* max_nr_buckets == 0 for order based mm means infinite */
1038 if (mm
== <tng_ust_lfht_mm_order
&& !max_nr_buckets
)
1039 max_nr_buckets
= 1UL << (MAX_TABLE_ORDER
- 1);
1041 /* max_nr_buckets must be power of two */
1042 if (!max_nr_buckets
|| (max_nr_buckets
& (max_nr_buckets
- 1)))
1045 if (flags
& LTTNG_UST_LFHT_AUTO_RESIZE
)
1048 min_nr_alloc_buckets
= max(min_nr_alloc_buckets
, MIN_TABLE_SIZE
);
1049 init_size
= max(init_size
, MIN_TABLE_SIZE
);
1050 max_nr_buckets
= max(max_nr_buckets
, min_nr_alloc_buckets
);
1051 init_size
= min(init_size
, max_nr_buckets
);
1053 ht
= mm
->alloc_lttng_ust_lfht(min_nr_alloc_buckets
, max_nr_buckets
);
1055 assert(ht
->mm
== mm
);
1056 assert(ht
->bucket_at
== mm
->bucket_at
);
1059 /* this mutex should not nest in read-side C.S. */
1060 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1061 order
= lttng_ust_lfht_get_count_order_ulong(init_size
);
1062 ht
->resize_target
= 1UL << order
;
1063 lttng_ust_lfht_create_bucket(ht
, 1UL << order
);
1064 ht
->size
= 1UL << order
;
1068 void lttng_ust_lfht_lookup(struct lttng_ust_lfht
*ht
, unsigned long hash
,
1069 lttng_ust_lfht_match_fct match
, const void *key
,
1070 struct lttng_ust_lfht_iter
*iter
)
1072 struct lttng_ust_lfht_node
*node
, *next
, *bucket
;
1073 unsigned long reverse_hash
, size
;
1075 lttng_ust_lfht_iter_debug_set_ht(ht
, iter
);
1077 reverse_hash
= bit_reverse_ulong(hash
);
1079 size
= lttng_ust_rcu_dereference(ht
->size
);
1080 bucket
= lookup_bucket(ht
, size
, hash
);
1081 /* We can always skip the bucket node initially */
1082 node
= lttng_ust_rcu_dereference(bucket
->next
);
1083 node
= clear_flag(node
);
1085 if (caa_unlikely(is_end(node
))) {
1089 if (caa_unlikely(node
->reverse_hash
> reverse_hash
)) {
1093 next
= lttng_ust_rcu_dereference(node
->next
);
1094 assert(node
== clear_flag(node
));
1095 if (caa_likely(!is_removed(next
))
1097 && node
->reverse_hash
== reverse_hash
1098 && caa_likely(match(node
, key
))) {
1101 node
= clear_flag(next
);
1103 assert(!node
|| !is_bucket(CMM_LOAD_SHARED(node
->next
)));
1108 void lttng_ust_lfht_next_duplicate(struct lttng_ust_lfht
*ht
, lttng_ust_lfht_match_fct match
,
1109 const void *key
, struct lttng_ust_lfht_iter
*iter
)
1111 struct lttng_ust_lfht_node
*node
, *next
;
1112 unsigned long reverse_hash
;
1114 lttng_ust_lfht_iter_debug_assert(ht
== iter
->lfht
);
1116 reverse_hash
= node
->reverse_hash
;
1118 node
= clear_flag(next
);
1121 if (caa_unlikely(is_end(node
))) {
1125 if (caa_unlikely(node
->reverse_hash
> reverse_hash
)) {
1129 next
= lttng_ust_rcu_dereference(node
->next
);
1130 if (caa_likely(!is_removed(next
))
1132 && caa_likely(match(node
, key
))) {
1135 node
= clear_flag(next
);
1137 assert(!node
|| !is_bucket(CMM_LOAD_SHARED(node
->next
)));
1142 void lttng_ust_lfht_next(struct lttng_ust_lfht
*ht
, struct lttng_ust_lfht_iter
*iter
)
1144 struct lttng_ust_lfht_node
*node
, *next
;
1146 lttng_ust_lfht_iter_debug_assert(ht
== iter
->lfht
);
1147 node
= clear_flag(iter
->next
);
1149 if (caa_unlikely(is_end(node
))) {
1153 next
= lttng_ust_rcu_dereference(node
->next
);
1154 if (caa_likely(!is_removed(next
))
1155 && !is_bucket(next
)) {
1158 node
= clear_flag(next
);
1160 assert(!node
|| !is_bucket(CMM_LOAD_SHARED(node
->next
)));
1165 void lttng_ust_lfht_first(struct lttng_ust_lfht
*ht
, struct lttng_ust_lfht_iter
*iter
)
1167 lttng_ust_lfht_iter_debug_set_ht(ht
, iter
);
1169 * Get next after first bucket node. The first bucket node is the
1170 * first node of the linked list.
1172 iter
->next
= bucket_at(ht
, 0)->next
;
1173 lttng_ust_lfht_next(ht
, iter
);
1176 void lttng_ust_lfht_add(struct lttng_ust_lfht
*ht
, unsigned long hash
,
1177 struct lttng_ust_lfht_node
*node
)
1181 node
->reverse_hash
= bit_reverse_ulong(hash
);
1182 size
= lttng_ust_rcu_dereference(ht
->size
);
1183 _lttng_ust_lfht_add(ht
, hash
, NULL
, NULL
, size
, node
, NULL
, 0);
1186 struct lttng_ust_lfht_node
*lttng_ust_lfht_add_unique(struct lttng_ust_lfht
*ht
,
1188 lttng_ust_lfht_match_fct match
,
1190 struct lttng_ust_lfht_node
*node
)
1193 struct lttng_ust_lfht_iter iter
;
1195 node
->reverse_hash
= bit_reverse_ulong(hash
);
1196 size
= lttng_ust_rcu_dereference(ht
->size
);
1197 _lttng_ust_lfht_add(ht
, hash
, match
, key
, size
, node
, &iter
, 0);
1201 struct lttng_ust_lfht_node
*lttng_ust_lfht_add_replace(struct lttng_ust_lfht
*ht
,
1203 lttng_ust_lfht_match_fct match
,
1205 struct lttng_ust_lfht_node
*node
)
1208 struct lttng_ust_lfht_iter iter
;
1210 node
->reverse_hash
= bit_reverse_ulong(hash
);
1211 size
= lttng_ust_rcu_dereference(ht
->size
);
1213 _lttng_ust_lfht_add(ht
, hash
, match
, key
, size
, node
, &iter
, 0);
1214 if (iter
.node
== node
) {
1218 if (!_lttng_ust_lfht_replace(ht
, size
, iter
.node
, iter
.next
, node
))
1223 int lttng_ust_lfht_replace(struct lttng_ust_lfht
*ht
,
1224 struct lttng_ust_lfht_iter
*old_iter
,
1226 lttng_ust_lfht_match_fct match
,
1228 struct lttng_ust_lfht_node
*new_node
)
1232 new_node
->reverse_hash
= bit_reverse_ulong(hash
);
1233 if (!old_iter
->node
)
1235 if (caa_unlikely(old_iter
->node
->reverse_hash
!= new_node
->reverse_hash
))
1237 if (caa_unlikely(!match(old_iter
->node
, key
)))
1239 size
= lttng_ust_rcu_dereference(ht
->size
);
1240 return _lttng_ust_lfht_replace(ht
, size
, old_iter
->node
, old_iter
->next
,
1244 int lttng_ust_lfht_del(struct lttng_ust_lfht
*ht
, struct lttng_ust_lfht_node
*node
)
1248 size
= lttng_ust_rcu_dereference(ht
->size
);
1249 return _lttng_ust_lfht_del(ht
, size
, node
);
1252 int lttng_ust_lfht_is_node_deleted(const struct lttng_ust_lfht_node
*node
)
1254 return is_removed(CMM_LOAD_SHARED(node
->next
));
1258 int lttng_ust_lfht_delete_bucket(struct lttng_ust_lfht
*ht
)
1260 struct lttng_ust_lfht_node
*node
;
1261 unsigned long order
, i
, size
;
1263 /* Check that the table is empty */
1264 node
= bucket_at(ht
, 0);
1266 node
= clear_flag(node
)->next
;
1267 if (!is_bucket(node
))
1269 assert(!is_removed(node
));
1270 assert(!is_removal_owner(node
));
1271 } while (!is_end(node
));
1273 * size accessed without lttng_ust_rcu_dereference because hash table is
1277 /* Internal sanity check: all nodes left should be buckets */
1278 for (i
= 0; i
< size
; i
++) {
1279 node
= bucket_at(ht
, i
);
1280 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1281 i
, i
, bit_reverse_ulong(node
->reverse_hash
));
1282 assert(is_bucket(node
->next
));
1285 for (order
= lttng_ust_lfht_get_count_order_ulong(size
); (long)order
>= 0; order
--)
1286 lttng_ust_lfht_free_bucket_table(ht
, order
);
1292 * Should only be called when no more concurrent readers nor writers can
1293 * possibly access the table.
1295 int lttng_ust_lfht_destroy(struct lttng_ust_lfht
*ht
)
1299 ret
= lttng_ust_lfht_delete_bucket(ht
);
1302 ret
= pthread_mutex_destroy(&ht
->resize_mutex
);