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 */
18static int
19objpool_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 */
47static int
48objpool_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 < pool->nr_cpus; 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(cpu: i))
60 continue;
61
62 /* compute how many objects to be allocated with this slot */
63 nodes = nr_objs / num_possible_cpus();
64 if (cpu_count < (nr_objs % num_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, flags: pool->gfp, cpu_to_node(cpu: i));
81 else
82 slot = __vmalloc_node(size, align: sizeof(void *), gfp_mask: pool->gfp,
83 cpu_to_node(cpu: i), caller: __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 */
99static 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 < pool->nr_cpus; i++)
107 kvfree(addr: pool->cpu_slots[i]);
108 kfree(objp: pool->cpu_slots);
109}
110
111/* initialize object pool and pre-allocate objects */
112int 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_cpus = nr_cpu_ids;
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 = pool->nr_cpus * sizeof(struct objpool_slot);
140 pool->cpu_slots = kzalloc(size: slot_size, flags: 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(r: &pool->ref, n: pool->nr_objs + 1);
150
151 return rc;
152}
153EXPORT_SYMBOL_GPL(objpool_init);
154
155/* adding object to slot, abort if the slot was already full */
156static inline int
157objpool_try_add_slot(void *obj, struct objpool_head *pool, int cpu)
158{
159 struct objpool_slot *slot = pool->cpu_slots[cpu];
160 uint32_t head, tail;
161
162 /* loading tail and head as a local snapshot, tail first */
163 tail = READ_ONCE(slot->tail);
164
165 do {
166 head = READ_ONCE(slot->head);
167 /* fault caught: something must be wrong */
168 WARN_ON_ONCE(tail - head > pool->nr_objs);
169 } while (!try_cmpxchg_acquire(&slot->tail, &tail, tail + 1));
170
171 /* now the tail position is reserved for the given obj */
172 WRITE_ONCE(slot->entries[tail & slot->mask], obj);
173 /* update sequence to make this obj available for pop() */
174 smp_store_release(&slot->last, tail + 1);
175
176 return 0;
177}
178
179/* reclaim an object to object pool */
180int objpool_push(void *obj, struct objpool_head *pool)
181{
182 unsigned long flags;
183 int rc;
184
185 /* disable local irq to avoid preemption & interruption */
186 raw_local_irq_save(flags);
187 rc = objpool_try_add_slot(obj, pool, raw_smp_processor_id());
188 raw_local_irq_restore(flags);
189
190 return rc;
191}
192EXPORT_SYMBOL_GPL(objpool_push);
193
194/* try to retrieve object from slot */
195static inline void *objpool_try_get_slot(struct objpool_head *pool, int cpu)
196{
197 struct objpool_slot *slot = pool->cpu_slots[cpu];
198 /* load head snapshot, other cpus may change it */
199 uint32_t head = smp_load_acquire(&slot->head);
200
201 while (head != READ_ONCE(slot->last)) {
202 void *obj;
203
204 /* obj must be retrieved before moving forward head */
205 obj = READ_ONCE(slot->entries[head & slot->mask]);
206
207 /* move head forward to mark it's consumption */
208 if (try_cmpxchg_release(&slot->head, &head, head + 1))
209 return obj;
210 }
211
212 return NULL;
213}
214
215/* allocate an object from object pool */
216void *objpool_pop(struct objpool_head *pool)
217{
218 void *obj = NULL;
219 unsigned long flags;
220 int i, cpu;
221
222 /* disable local irq to avoid preemption & interruption */
223 raw_local_irq_save(flags);
224
225 cpu = raw_smp_processor_id();
226 for (i = 0; i < num_possible_cpus(); i++) {
227 obj = objpool_try_get_slot(pool, cpu);
228 if (obj)
229 break;
230 cpu = cpumask_next_wrap(n: cpu, cpu_possible_mask, start: -1, wrap: 1);
231 }
232 raw_local_irq_restore(flags);
233
234 return obj;
235}
236EXPORT_SYMBOL_GPL(objpool_pop);
237
238/* release whole objpool forcely */
239void objpool_free(struct objpool_head *pool)
240{
241 if (!pool->cpu_slots)
242 return;
243
244 /* release percpu slots */
245 objpool_fini_percpu_slots(pool);
246
247 /* call user's cleanup callback if provided */
248 if (pool->release)
249 pool->release(pool, pool->context);
250}
251EXPORT_SYMBOL_GPL(objpool_free);
252
253/* drop the allocated object, rather reclaim it to objpool */
254int objpool_drop(void *obj, struct objpool_head *pool)
255{
256 if (!obj || !pool)
257 return -EINVAL;
258
259 if (refcount_dec_and_test(r: &pool->ref)) {
260 objpool_free(pool);
261 return 0;
262 }
263
264 return -EAGAIN;
265}
266EXPORT_SYMBOL_GPL(objpool_drop);
267
268/* drop unused objects and defref objpool for releasing */
269void objpool_fini(struct objpool_head *pool)
270{
271 int count = 1; /* extra ref for objpool itself */
272
273 /* drop all remained objects from objpool */
274 while (objpool_pop(pool))
275 count++;
276
277 if (refcount_sub_and_test(i: count, r: &pool->ref))
278 objpool_free(pool);
279}
280EXPORT_SYMBOL_GPL(objpool_fini);
281

source code of linux/lib/objpool.c