]>
git.proxmox.com Git - mirror_ovs.git/blob - lib/pvector.h
2 * Copyright (c) 2014, 2016 Nicira, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at:
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
26 /* Concurrent Priority Vector
27 * ==========================
29 * Concurrent priority vector holds non-NULL pointers to objects in a
30 * nondecreasing priority order and allows readers to traverse the vector
31 * without being concerned about writers modifying the vector as they are
34 * Multiple elements of a given priority are allowed.
36 * The priority order is maintained as a linear vector of elements to allow
37 * for efficient memory prefetching.
39 * Concurrency is implemented with OVS RCU so that the readers can assume
40 * that once they have taken a pointer to the vector with
41 * pvector_cursor_init(), the 'size' member will not decrease, so that
42 * they can safely read 'size' entries from 'vector', and find that each
43 * entry has a valid, non-NULL 'ptr', and the vector is in order from highest
44 * to lowest 'priority'. The 'priority' values can change any time, but only
45 * so that the order of the entries does not change, so readers can use
46 * 'priority' values read at any time after acquisition of the vector pointer.
48 * Writers can concurrently add entries to the end of the vector, incrementing
49 * 'size', or update the 'priority' value of an entry, but only if that does
50 * not change the ordering of the entries. Writers will never change the 'ptr'
51 * values, or decrement the 'size' on a copy that readers have access to.
53 * Most modifications are internally staged at the 'temp' vector, from which
54 * they can be published at 'impl' by calling pvector_publish(). This saves
55 * unnecessary memory allocations when many changes are done back-to-back.
56 * 'temp' may contain NULL pointers and it may be in unsorted order. It is
57 * sorted before it is published at 'impl', which also removes the NULLs from
58 * the published vector.
60 * Since the vector is RCU protected, the entry destruction after removal must
61 * be RCU postponed. Also, if it happens before changes published with
62 * pvector_publish(), destruction must be double postponed, i.e., the second
63 * ovsrcu_postpone() call to destruct the entry should be called from the first
64 * RCU callback. This is required because readers could still obtain the
65 * unmodified vector until updated version is published.
68 struct pvector_entry
{
74 atomic_size_t size
; /* Number of entries in the vector. */
75 size_t allocated
; /* Number of allocated entries. */
76 struct pvector_entry vector
[];
79 /* Concurrent priority vector. */
81 OVSRCU_TYPE(struct pvector_impl
*) impl
;
82 struct pvector_impl
*temp
;
86 void pvector_init(struct pvector
*);
87 void pvector_destroy(struct pvector
*);
89 /* Insertion and deletion. These work on 'temp'. */
90 void pvector_insert(struct pvector
*, void *, int priority
);
91 void pvector_change_priority(struct pvector
*, void *, int priority
);
92 void pvector_remove(struct pvector
*, void *);
94 /* Make the modified pvector available for iteration. */
95 static inline void pvector_publish(struct pvector
*);
97 /* Count. These operate on the published pvector. */
98 static inline size_t pvector_count(const struct pvector
*);
99 static inline bool pvector_is_empty(const struct pvector
*);
107 * Iteration is safe even in a pvector that is changing concurrently.
108 * Multiple writers must exclude each other via e.g., a mutex.
117 * struct my_node elem1, elem2, *iter;
118 * struct pvector my_pvector;
120 * pvector_init(&my_pvector);
122 * pvector_insert(&my_pvector, &elem1, 1);
123 * pvector_insert(&my_pvector, &elem2, 2);
125 * PVECTOR_FOR_EACH (iter, &my_pvector) {
126 * ...operate on '*iter'...
127 * ...elem2 to be seen before elem1...
130 * pvector_destroy(&my_pvector);
132 * There is no PVECTOR_FOR_EACH_SAFE variant as iteration is performed on RCU
133 * protected instance of the priority vector. Any concurrent modifications
134 * that would be disruptive for readers (such as deletions), will be performed
135 * on a new instance. To see any of the modifications, a new iteration loop
138 * The PVECTOR_FOR_EACH_PRIORITY limits the iteration to entries with higher
139 * than or equal to the given priority and allows for object lookahead.
141 * The iteration loop must be completed without entering the OVS RCU quiescent
142 * period. That is, an old iteration loop must not be continued after any
143 * blocking IO (VLOG is non-blocking, so that is OK).
145 struct pvector_cursor
{
146 size_t size
; /* Number of entries in the vector. */
147 size_t entry_idx
; /* Current index. */
148 const struct pvector_entry
*vector
;
151 static inline struct pvector_cursor
pvector_cursor_init(const struct pvector
*,
154 static inline void *pvector_cursor_next(struct pvector_cursor
*,
156 size_t n_ahead
, size_t obj_size
);
157 static inline void pvector_cursor_lookahead(const struct pvector_cursor
*,
160 #define PVECTOR_FOR_EACH(PTR, PVECTOR) \
161 for (struct pvector_cursor cursor__ = pvector_cursor_init(PVECTOR, 0, 0); \
162 ((PTR) = pvector_cursor_next(&cursor__, INT_MIN, 0, 0)) != NULL; )
164 /* Loop while priority is higher than or equal to 'PRIORITY' and prefetch
165 * objects of size 'SZ' 'N' objects ahead from the current object. */
166 #define PVECTOR_FOR_EACH_PRIORITY(PTR, PRIORITY, N, SZ, PVECTOR) \
167 for (struct pvector_cursor cursor__ = pvector_cursor_init(PVECTOR, N, SZ); \
168 ((PTR) = pvector_cursor_next(&cursor__, PRIORITY, N, SZ)) != NULL; )
170 #define PVECTOR_CURSOR_FOR_EACH(PTR, CURSOR, PVECTOR) \
171 for (*(CURSOR) = pvector_cursor_init(PVECTOR, 0, 0); \
172 ((PTR) = pvector_cursor_next(CURSOR, INT_MIN, 0, 0)) != NULL; )
174 #define PVECTOR_CURSOR_FOR_EACH_CONTINUE(PTR, CURSOR) \
175 for (; ((PTR) = pvector_cursor_next(CURSOR, INT_MIN, 0, 0)) != NULL; )
178 /* Inline implementations. */
180 static inline struct pvector_cursor
181 pvector_cursor_init(const struct pvector
*pvec
,
182 size_t n_ahead
, size_t obj_size
)
184 const struct pvector_impl
*impl
;
185 struct pvector_cursor cursor
;
188 impl
= ovsrcu_get(struct pvector_impl
*, &pvec
->impl
);
190 /* Use memory_order_acquire to ensure entry access can not be
191 * reordered to happen before size read. */
192 atomic_read_explicit(&CONST_CAST(struct pvector_impl
*, impl
)->size
,
193 &size
, memory_order_acquire
);
194 ovs_prefetch_range(impl
->vector
, size
* sizeof impl
->vector
[0]);
197 cursor
.vector
= impl
->vector
;
198 cursor
.entry_idx
= -1;
200 for (size_t i
= 0; i
< n_ahead
; i
++) {
201 /* Prefetch the first objects. */
202 pvector_cursor_lookahead(&cursor
, i
, obj_size
);
207 static inline void *pvector_cursor_next(struct pvector_cursor
*cursor
,
209 size_t n_ahead
, size_t obj_size
)
211 if (++cursor
->entry_idx
< cursor
->size
&&
212 cursor
->vector
[cursor
->entry_idx
].priority
>= lowest_priority
) {
214 pvector_cursor_lookahead(cursor
, n_ahead
, obj_size
);
216 return cursor
->vector
[cursor
->entry_idx
].ptr
;
221 static inline void pvector_cursor_lookahead(const struct pvector_cursor
*cursor
,
224 if (cursor
->entry_idx
+ n
< cursor
->size
) {
225 ovs_prefetch_range(cursor
->vector
[cursor
->entry_idx
+ n
].ptr
, size
);
229 static inline size_t pvector_count(const struct pvector
*pvec
)
231 return ovsrcu_get(struct pvector_impl
*, &pvec
->impl
)->size
;
234 static inline bool pvector_is_empty(const struct pvector
*pvec
)
236 return pvector_count(pvec
) == 0;
239 void pvector_publish__(struct pvector
*);
241 /* Make the modified pvector available for iteration. */
242 static inline void pvector_publish(struct pvector
*pvec
)
245 pvector_publish__(pvec
);
249 #endif /* pvector.h */