xref: /linux/include/linux/objpool.h (revision 70a663205d5085f1d82f7058e9419ff7612e9396)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 
3 #ifndef _LINUX_OBJPOOL_H
4 #define _LINUX_OBJPOOL_H
5 
6 #include <linux/types.h>
7 #include <linux/refcount.h>
8 #include <linux/atomic.h>
9 #include <linux/cpumask.h>
10 #include <linux/irqflags.h>
11 #include <linux/smp.h>
12 
13 /*
14  * objpool: ring-array based lockless MPMC queue
15  *
16  * Copyright: wuqiang.matt@bytedance.com,mhiramat@kernel.org
17  *
18  * objpool is a scalable implementation of high performance queue for
19  * object allocation and reclamation, such as kretprobe instances.
20  *
21  * With leveraging percpu ring-array to mitigate hot spots of memory
22  * contention, it delivers near-linear scalability for high parallel
23  * scenarios. The objpool is best suited for the following cases:
24  * 1) Memory allocation or reclamation are prohibited or too expensive
25  * 2) Consumers are of different priorities, such as irqs and threads
26  *
27  * Limitations:
28  * 1) Maximum objects (capacity) is fixed after objpool creation
29  * 2) All pre-allocated objects are managed in percpu ring array,
30  *    which consumes more memory than linked lists
31  */
32 
33 /**
34  * struct objpool_slot - percpu ring array of objpool
35  * @head: head sequence of the local ring array (to retrieve at)
36  * @tail: tail sequence of the local ring array (to append at)
37  * @last: the last sequence number marked as ready for retrieve
38  * @mask: bits mask for modulo capacity to compute array indexes
39  * @entries: object entries on this slot
40  *
41  * Represents a cpu-local array-based ring buffer, its size is specialized
42  * during initialization of object pool. The percpu objpool node is to be
43  * allocated from local memory for NUMA system, and to be kept compact in
44  * continuous memory: CPU assigned number of objects are stored just after
45  * the body of objpool_node.
46  *
47  * Real size of the ring array is far too smaller than the value range of
48  * head and tail, typed as uint32_t: [0, 2^32), so only lower bits (mask)
49  * of head and tail are used as the actual position in the ring array. In
50  * general the ring array is acting like a small sliding window, which is
51  * always moving forward in the loop of [0, 2^32).
52  */
53 struct objpool_slot {
54 	uint32_t            head;
55 	uint32_t            tail;
56 	uint32_t            last;
57 	uint32_t            mask;
58 	void               *entries[];
59 } __packed;
60 
61 struct objpool_head;
62 
63 /*
64  * caller-specified callback for object initial setup, it's only called
65  * once for each object (just after the memory allocation of the object)
66  */
67 typedef int (*objpool_init_obj_cb)(void *obj, void *context);
68 
69 /* caller-specified cleanup callback for objpool destruction */
70 typedef int (*objpool_fini_cb)(struct objpool_head *head, void *context);
71 
72 /**
73  * struct objpool_head - object pooling metadata
74  * @obj_size:   object size, aligned to sizeof(void *)
75  * @nr_objs:    total objs (to be pre-allocated with objpool)
76  * @nr_possible_cpus: cached value of num_possible_cpus()
77  * @capacity:   max objs can be managed by one objpool_slot
78  * @gfp:        gfp flags for kmalloc & vmalloc
79  * @ref:        refcount of objpool
80  * @flags:      flags for objpool management
81  * @cpu_slots:  pointer to the array of objpool_slot
82  * @release:    resource cleanup callback
83  * @context:    caller-provided context
84  */
85 struct objpool_head {
86 	int                     obj_size;
87 	int                     nr_objs;
88 	int                     nr_possible_cpus;
89 	int                     capacity;
90 	gfp_t                   gfp;
91 	refcount_t              ref;
92 	unsigned long           flags;
93 	struct objpool_slot   **cpu_slots;
94 	objpool_fini_cb         release;
95 	void                   *context;
96 };
97 
98 #define OBJPOOL_NR_OBJECT_MAX	(1UL << 24) /* maximum numbers of total objects */
99 #define OBJPOOL_OBJECT_SIZE_MAX	(1UL << 16) /* maximum size of an object */
100 
101 /**
102  * objpool_init() - initialize objpool and pre-allocated objects
103  * @pool:    the object pool to be initialized, declared by caller
104  * @nr_objs: total objects to be pre-allocated by this object pool
105  * @object_size: size of an object (should be > 0)
106  * @gfp:     flags for memory allocation (via kmalloc or vmalloc)
107  * @context: user context for object initialization callback
108  * @objinit: object initialization callback for extra setup
109  * @release: cleanup callback for extra cleanup task
110  *
111  * return value: 0 for success, otherwise error code
112  *
113  * All pre-allocated objects are to be zeroed after memory allocation.
114  * Caller could do extra initialization in objinit callback. objinit()
115  * will be called just after slot allocation and called only once for
116  * each object. After that the objpool won't touch any content of the
117  * objects. It's caller's duty to perform reinitialization after each
118  * pop (object allocation) or do clearance before each push (object
119  * reclamation).
120  */
121 int objpool_init(struct objpool_head *pool, int nr_objs, int object_size,
122 		 gfp_t gfp, void *context, objpool_init_obj_cb objinit,
123 		 objpool_fini_cb release);
124 
125 /* try to retrieve object from slot */
__objpool_try_get_slot(struct objpool_head * pool,int cpu)126 static inline void *__objpool_try_get_slot(struct objpool_head *pool, int cpu)
127 {
128 	struct objpool_slot *slot = pool->cpu_slots[cpu];
129 	/* load head snapshot, other cpus may change it */
130 	uint32_t head = smp_load_acquire(&slot->head);
131 
132 	while (head != READ_ONCE(slot->last)) {
133 		void *obj;
134 
135 		/*
136 		 * data visibility of 'last' and 'head' could be out of
137 		 * order since memory updating of 'last' and 'head' are
138 		 * performed in push() and pop() independently
139 		 *
140 		 * before any retrieving attempts, pop() must guarantee
141 		 * 'last' is behind 'head', that is to say, there must
142 		 * be available objects in slot, which could be ensured
143 		 * by condition 'last != head && last - head <= nr_objs'
144 		 * that is equivalent to 'last - head - 1 < nr_objs' as
145 		 * 'last' and 'head' are both unsigned int32
146 		 */
147 		if (READ_ONCE(slot->last) - head - 1 >= pool->nr_objs) {
148 			head = READ_ONCE(slot->head);
149 			continue;
150 		}
151 
152 		/* obj must be retrieved before moving forward head */
153 		obj = READ_ONCE(slot->entries[head & slot->mask]);
154 
155 		/* move head forward to mark it's consumption */
156 		if (try_cmpxchg_release(&slot->head, &head, head + 1))
157 			return obj;
158 	}
159 
160 	return NULL;
161 }
162 
163 /**
164  * objpool_pop() - allocate an object from objpool
165  * @pool: object pool
166  *
167  * return value: object ptr or NULL if failed
168  */
objpool_pop(struct objpool_head * pool)169 static inline void *objpool_pop(struct objpool_head *pool)
170 {
171 	void *obj = NULL;
172 	unsigned long flags;
173 	int i, cpu;
174 
175 	/* disable local irq to avoid preemption & interruption */
176 	raw_local_irq_save(flags);
177 
178 	cpu = raw_smp_processor_id();
179 	for (i = 0; i < pool->nr_possible_cpus; i++) {
180 		obj = __objpool_try_get_slot(pool, cpu);
181 		if (obj)
182 			break;
183 		cpu = cpumask_next_wrap(cpu, cpu_possible_mask, -1, 1);
184 	}
185 	raw_local_irq_restore(flags);
186 
187 	return obj;
188 }
189 
190 /* adding object to slot, abort if the slot was already full */
191 static inline int
__objpool_try_add_slot(void * obj,struct objpool_head * pool,int cpu)192 __objpool_try_add_slot(void *obj, struct objpool_head *pool, int cpu)
193 {
194 	struct objpool_slot *slot = pool->cpu_slots[cpu];
195 	uint32_t head, tail;
196 
197 	/* loading tail and head as a local snapshot, tail first */
198 	tail = READ_ONCE(slot->tail);
199 
200 	do {
201 		head = READ_ONCE(slot->head);
202 		/* fault caught: something must be wrong */
203 		WARN_ON_ONCE(tail - head > pool->nr_objs);
204 	} while (!try_cmpxchg_acquire(&slot->tail, &tail, tail + 1));
205 
206 	/* now the tail position is reserved for the given obj */
207 	WRITE_ONCE(slot->entries[tail & slot->mask], obj);
208 	/* update sequence to make this obj available for pop() */
209 	smp_store_release(&slot->last, tail + 1);
210 
211 	return 0;
212 }
213 
214 /**
215  * objpool_push() - reclaim the object and return back to objpool
216  * @obj:  object ptr to be pushed to objpool
217  * @pool: object pool
218  *
219  * return: 0 or error code (it fails only when user tries to push
220  * the same object multiple times or wrong "objects" into objpool)
221  */
objpool_push(void * obj,struct objpool_head * pool)222 static inline int objpool_push(void *obj, struct objpool_head *pool)
223 {
224 	unsigned long flags;
225 	int rc;
226 
227 	/* disable local irq to avoid preemption & interruption */
228 	raw_local_irq_save(flags);
229 	rc = __objpool_try_add_slot(obj, pool, raw_smp_processor_id());
230 	raw_local_irq_restore(flags);
231 
232 	return rc;
233 }
234 
235 
236 /**
237  * objpool_drop() - discard the object and deref objpool
238  * @obj:  object ptr to be discarded
239  * @pool: object pool
240  *
241  * return: 0 if objpool was released; -EAGAIN if there are still
242  *         outstanding objects
243  *
244  * objpool_drop is normally for the release of outstanding objects
245  * after objpool cleanup (objpool_fini). Thinking of this example:
246  * kretprobe is unregistered and objpool_fini() is called to release
247  * all remained objects, but there are still objects being used by
248  * unfinished kretprobes (like blockable function: sys_accept). So
249  * only when the last outstanding object is dropped could the whole
250  * objpool be released along with the call of objpool_drop()
251  */
252 int objpool_drop(void *obj, struct objpool_head *pool);
253 
254 /**
255  * objpool_free() - release objpool forcely (all objects to be freed)
256  * @pool: object pool to be released
257  */
258 void objpool_free(struct objpool_head *pool);
259 
260 /**
261  * objpool_fini() - deref object pool (also releasing unused objects)
262  * @pool: object pool to be dereferenced
263  *
264  * objpool_fini() will try to release all remained free objects and
265  * then drop an extra reference of the objpool. If all objects are
266  * already returned to objpool (so called synchronous use cases),
267  * the objpool itself will be freed together. But if there are still
268  * outstanding objects (so called asynchronous use cases, such like
269  * blockable kretprobe), the objpool won't be released until all
270  * the outstanding objects are dropped, but the caller must assure
271  * there are no concurrent objpool_push() on the fly. Normally RCU
272  * is being required to make sure all ongoing objpool_push() must
273  * be finished before calling objpool_fini(), so does test_objpool,
274  * kretprobe or rethook
275  */
276 void objpool_fini(struct objpool_head *pool);
277 
278 #endif /* _LINUX_OBJPOOL_H */
279