1 // SPDX-License-Identifier: GPL-2.0
2
3 #include <linux/objpool.h>
4 #include <linux/slab.h>
5 #include <linux/vmalloc.h>
6 #include <linux/atomic.h>
7 #include <linux/irqflags.h>
8 #include <linux/cpumask.h>
9 #include <linux/log2.h>
10
11 /*
12 * objpool: ring-array based lockless MPMC/FIFO queues
13 *
14 * Copyright: wuqiang.matt@bytedance.com,mhiramat@kernel.org
15 */
16
17 /* initialize percpu objpool_slot */
18 static int
19 objpool_init_percpu_slot(struct objpool_head *pool,
20 struct objpool_slot *slot,
21 int nodes, void *context,
22 objpool_init_obj_cb objinit)
23 {
24 void *obj = (void *)&slot->entries[pool->capacity];
25 int i;
26
27 /* initialize elements of percpu objpool_slot */
28 slot->mask = pool->capacity - 1;
29
30 for (i = 0; i < nodes; i++) {
31 if (objinit) {
32 int rc = objinit(obj, context);
33 if (rc)
34 return rc;
35 }
36 slot->entries[slot->tail & slot->mask] = obj;
37 obj = obj + pool->obj_size;
38 slot->tail++;
39 slot->last = slot->tail;
40 pool->nr_objs++;
41 }
42
43 return 0;
44 }
45
46 /* allocate and initialize percpu slots */
47 static int
48 objpool_init_percpu_slots(struct objpool_head *pool, int nr_objs,
49 void *context, objpool_init_obj_cb objinit)
50 {
51 int i, cpu_count = 0;
52
53 for (i = 0; i < nr_cpu_ids; i++) {
54
55 struct objpool_slot *slot;
56 int nodes, size, rc;
57
58 /* skip the cpu node which could never be present */
59 if (!cpu_possible(i))
60 continue;
61
62 /* compute how many objects to be allocated with this slot */
63 nodes = nr_objs / pool->nr_possible_cpus;
64 if (cpu_count < (nr_objs % pool->nr_possible_cpus))
65 nodes++;
66 cpu_count++;
67
68 size = struct_size(slot, entries, pool->capacity) +
69 pool->obj_size * nodes;
70
71 /*
72 * here we allocate percpu-slot & objs together in a single
73 * allocation to make it more compact, taking advantage of
74 * warm caches and TLB hits. in default vmalloc is used to
75 * reduce the pressure of kernel slab system. as we know,
76 * mimimal size of vmalloc is one page since vmalloc would
77 * always align the requested size to page size
78 */
79 if (pool->gfp & GFP_ATOMIC)
80 slot = kmalloc_node(size, pool->gfp, cpu_to_node(i));
81 else
82 slot = __vmalloc_node(size, sizeof(void *), pool->gfp,
83 cpu_to_node(i), __builtin_return_address(0));
84 if (!slot)
85 return -ENOMEM;
86 memset(slot, 0, size);
87 pool->cpu_slots[i] = slot;
88
89 /* initialize the objpool_slot of cpu node i */
90 rc = objpool_init_percpu_slot(pool, slot, nodes, context, objinit);
91 if (rc)
92 return rc;
93 }
94
95 return 0;
96 }
97
98 /* cleanup all percpu slots of the object pool */
99 static void objpool_fini_percpu_slots(struct objpool_head *pool)
100 {
101 int i;
102
103 if (!pool->cpu_slots)
104 return;
105
106 for (i = 0; i < nr_cpu_ids; i++)
107 kvfree(pool->cpu_slots[i]);
108 kfree(pool->cpu_slots);
109 }
110
111 /* initialize object pool and pre-allocate objects */
112 int objpool_init(struct objpool_head *pool, int nr_objs, int object_size,
113 gfp_t gfp, void *context, objpool_init_obj_cb objinit,
114 objpool_fini_cb release)
115 {
116 int rc, capacity, slot_size;
117
118 /* check input parameters */
119 if (nr_objs <= 0 || nr_objs > OBJPOOL_NR_OBJECT_MAX ||
120 object_size <= 0 || object_size > OBJPOOL_OBJECT_SIZE_MAX)
121 return -EINVAL;
122
123 /* align up to unsigned long size */
124 object_size = ALIGN(object_size, sizeof(long));
125
126 /* calculate capacity of percpu objpool_slot */
127 capacity = roundup_pow_of_two(nr_objs);
128 if (!capacity)
129 return -EINVAL;
130
131 /* initialize objpool pool */
132 memset(pool, 0, sizeof(struct objpool_head));
133 pool->nr_possible_cpus = num_possible_cpus();
134 pool->obj_size = object_size;
135 pool->capacity = capacity;
136 pool->gfp = gfp & ~__GFP_ZERO;
137 pool->context = context;
138 pool->release = release;
139 slot_size = nr_cpu_ids * sizeof(struct objpool_slot);
140 pool->cpu_slots = kzalloc(slot_size, pool->gfp);
141 if (!pool->cpu_slots)
142 return -ENOMEM;
143
144 /* initialize per-cpu slots */
145 rc = objpool_init_percpu_slots(pool, nr_objs, context, objinit);
146 if (rc)
147 objpool_fini_percpu_slots(pool);
148 else
149 refcount_set(&pool->ref, pool->nr_objs + 1);
150
151 return rc;
152 }
153 EXPORT_SYMBOL_GPL(objpool_init);
154
155 /* release whole objpool forcely */
156 void objpool_free(struct objpool_head *pool)
157 {
158 if (!pool->cpu_slots)
159 return;
160
161 /* release percpu slots */
162 objpool_fini_percpu_slots(pool);
163
164 /* call user's cleanup callback if provided */
165 if (pool->release)
166 pool->release(pool, pool->context);
167 }
168 EXPORT_SYMBOL_GPL(objpool_free);
169
170 /* drop the allocated object, rather reclaim it to objpool */
171 int objpool_drop(void *obj, struct objpool_head *pool)
172 {
173 if (!obj || !pool)
174 return -EINVAL;
175
176 if (refcount_dec_and_test(&pool->ref)) {
177 objpool_free(pool);
178 return 0;
179 }
180
181 return -EAGAIN;
182 }
183 EXPORT_SYMBOL_GPL(objpool_drop);
184
185 /* drop unused objects and defref objpool for releasing */
186 void objpool_fini(struct objpool_head *pool)
187 {
188 int count = 1; /* extra ref for objpool itself */
189
190 /* drop all remained objects from objpool */
191 while (objpool_pop(pool))
192 count++;
193
194 if (refcount_sub_and_test(count, &pool->ref))
195 objpool_free(pool);
196 }
197 EXPORT_SYMBOL_GPL(objpool_fini);
198
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.