1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3 * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved.
4 * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved.
5 */
6
7 #include "rxe.h"
8
9 #define RXE_POOL_TIMEOUT (200)
10 #define RXE_POOL_ALIGN (16)
11
12 static const struct rxe_type_info {
13 const char *name;
14 size_t size;
15 size_t elem_offset;
16 void (*cleanup)(struct rxe_pool_elem *elem);
17 u32 min_index;
18 u32 max_index;
19 u32 max_elem;
20 } rxe_type_info[RXE_NUM_TYPES] = {
21 [RXE_TYPE_UC] = {
22 .name = "uc",
23 .size = sizeof(struct rxe_ucontext),
24 .elem_offset = offsetof(struct rxe_ucontext, elem),
25 .min_index = 1,
26 .max_index = RXE_MAX_UCONTEXT,
27 .max_elem = RXE_MAX_UCONTEXT,
28 },
29 [RXE_TYPE_PD] = {
30 .name = "pd",
31 .size = sizeof(struct rxe_pd),
32 .elem_offset = offsetof(struct rxe_pd, elem),
33 .min_index = 1,
34 .max_index = RXE_MAX_PD,
35 .max_elem = RXE_MAX_PD,
36 },
37 [RXE_TYPE_AH] = {
38 .name = "ah",
39 .size = sizeof(struct rxe_ah),
40 .elem_offset = offsetof(struct rxe_ah, elem),
41 .min_index = RXE_MIN_AH_INDEX,
42 .max_index = RXE_MAX_AH_INDEX,
43 .max_elem = RXE_MAX_AH,
44 },
45 [RXE_TYPE_SRQ] = {
46 .name = "srq",
47 .size = sizeof(struct rxe_srq),
48 .elem_offset = offsetof(struct rxe_srq, elem),
49 .cleanup = rxe_srq_cleanup,
50 .min_index = RXE_MIN_SRQ_INDEX,
51 .max_index = RXE_MAX_SRQ_INDEX,
52 .max_elem = RXE_MAX_SRQ,
53 },
54 [RXE_TYPE_QP] = {
55 .name = "qp",
56 .size = sizeof(struct rxe_qp),
57 .elem_offset = offsetof(struct rxe_qp, elem),
58 .cleanup = rxe_qp_cleanup,
59 .min_index = RXE_MIN_QP_INDEX,
60 .max_index = RXE_MAX_QP_INDEX,
61 .max_elem = RXE_MAX_QP,
62 },
63 [RXE_TYPE_CQ] = {
64 .name = "cq",
65 .size = sizeof(struct rxe_cq),
66 .elem_offset = offsetof(struct rxe_cq, elem),
67 .cleanup = rxe_cq_cleanup,
68 .min_index = 1,
69 .max_index = RXE_MAX_CQ,
70 .max_elem = RXE_MAX_CQ,
71 },
72 [RXE_TYPE_MR] = {
73 .name = "mr",
74 .size = sizeof(struct rxe_mr),
75 .elem_offset = offsetof(struct rxe_mr, elem),
76 .cleanup = rxe_mr_cleanup,
77 .min_index = RXE_MIN_MR_INDEX,
78 .max_index = RXE_MAX_MR_INDEX,
79 .max_elem = RXE_MAX_MR,
80 },
81 [RXE_TYPE_MW] = {
82 .name = "mw",
83 .size = sizeof(struct rxe_mw),
84 .elem_offset = offsetof(struct rxe_mw, elem),
85 .cleanup = rxe_mw_cleanup,
86 .min_index = RXE_MIN_MW_INDEX,
87 .max_index = RXE_MAX_MW_INDEX,
88 .max_elem = RXE_MAX_MW,
89 },
90 };
91
rxe_pool_init(struct rxe_dev * rxe,struct rxe_pool * pool,enum rxe_elem_type type)92 void rxe_pool_init(struct rxe_dev *rxe, struct rxe_pool *pool,
93 enum rxe_elem_type type)
94 {
95 const struct rxe_type_info *info = &rxe_type_info[type];
96
97 memset(pool, 0, sizeof(*pool));
98
99 pool->rxe = rxe;
100 pool->name = info->name;
101 pool->type = type;
102 pool->max_elem = info->max_elem;
103 pool->elem_size = ALIGN(info->size, RXE_POOL_ALIGN);
104 pool->elem_offset = info->elem_offset;
105 pool->cleanup = info->cleanup;
106
107 atomic_set(&pool->num_elem, 0);
108
109 xa_init_flags(&pool->xa, XA_FLAGS_ALLOC);
110 pool->limit.min = info->min_index;
111 pool->limit.max = info->max_index;
112 }
113
rxe_pool_cleanup(struct rxe_pool * pool)114 void rxe_pool_cleanup(struct rxe_pool *pool)
115 {
116 WARN_ON(!xa_empty(&pool->xa));
117 }
118
__rxe_add_to_pool(struct rxe_pool * pool,struct rxe_pool_elem * elem,bool sleepable)119 int __rxe_add_to_pool(struct rxe_pool *pool, struct rxe_pool_elem *elem,
120 bool sleepable)
121 {
122 int err;
123 gfp_t gfp_flags;
124
125 if (atomic_inc_return(&pool->num_elem) > pool->max_elem)
126 goto err_cnt;
127
128 elem->pool = pool;
129 elem->obj = (u8 *)elem - pool->elem_offset;
130 kref_init(&elem->ref_cnt);
131 init_completion(&elem->complete);
132
133 /* AH objects are unique in that the create_ah verb
134 * can be called in atomic context. If the create_ah
135 * call is not sleepable use GFP_ATOMIC.
136 */
137 gfp_flags = sleepable ? GFP_KERNEL : GFP_ATOMIC;
138
139 if (sleepable)
140 might_sleep();
141 err = xa_alloc_cyclic(&pool->xa, &elem->index, NULL, pool->limit,
142 &pool->next, gfp_flags);
143 if (err < 0)
144 goto err_cnt;
145
146 return 0;
147
148 err_cnt:
149 atomic_dec(&pool->num_elem);
150 return -EINVAL;
151 }
152
rxe_pool_get_index(struct rxe_pool * pool,u32 index)153 void *rxe_pool_get_index(struct rxe_pool *pool, u32 index)
154 {
155 struct rxe_pool_elem *elem;
156 struct xarray *xa = &pool->xa;
157 void *obj;
158
159 rcu_read_lock();
160 elem = xa_load(xa, index);
161 if (elem && kref_get_unless_zero(&elem->ref_cnt))
162 obj = elem->obj;
163 else
164 obj = NULL;
165 rcu_read_unlock();
166
167 return obj;
168 }
169
rxe_elem_release(struct kref * kref)170 static void rxe_elem_release(struct kref *kref)
171 {
172 struct rxe_pool_elem *elem = container_of(kref, typeof(*elem), ref_cnt);
173
174 complete(&elem->complete);
175 }
176
__rxe_cleanup(struct rxe_pool_elem * elem,bool sleepable)177 int __rxe_cleanup(struct rxe_pool_elem *elem, bool sleepable)
178 {
179 struct rxe_pool *pool = elem->pool;
180 struct xarray *xa = &pool->xa;
181 static int timeout = RXE_POOL_TIMEOUT;
182 int ret, err = 0;
183 void *xa_ret;
184
185 if (sleepable)
186 might_sleep();
187
188 /* erase xarray entry to prevent looking up
189 * the pool elem from its index
190 */
191 xa_ret = xa_erase(xa, elem->index);
192 WARN_ON(xa_err(xa_ret));
193
194 /* if this is the last call to rxe_put complete the
195 * object. It is safe to touch obj->elem after this since
196 * it is freed below
197 */
198 __rxe_put(elem);
199
200 /* wait until all references to the object have been
201 * dropped before final object specific cleanup and
202 * return to rdma-core
203 */
204 if (sleepable) {
205 if (!completion_done(&elem->complete) && timeout) {
206 ret = wait_for_completion_timeout(&elem->complete,
207 timeout);
208
209 /* Shouldn't happen. There are still references to
210 * the object but, rather than deadlock, free the
211 * object or pass back to rdma-core.
212 */
213 if (WARN_ON(!ret))
214 err = -EINVAL;
215 }
216 } else {
217 unsigned long until = jiffies + timeout;
218
219 /* AH objects are unique in that the destroy_ah verb
220 * can be called in atomic context. This delay
221 * replaces the wait_for_completion call above
222 * when the destroy_ah call is not sleepable
223 */
224 while (!completion_done(&elem->complete) &&
225 time_before(jiffies, until))
226 mdelay(1);
227
228 if (WARN_ON(!completion_done(&elem->complete)))
229 err = -EINVAL;
230 }
231
232 if (pool->cleanup)
233 pool->cleanup(elem);
234
235 atomic_dec(&pool->num_elem);
236
237 return err;
238 }
239
__rxe_get(struct rxe_pool_elem * elem)240 int __rxe_get(struct rxe_pool_elem *elem)
241 {
242 return kref_get_unless_zero(&elem->ref_cnt);
243 }
244
__rxe_put(struct rxe_pool_elem * elem)245 int __rxe_put(struct rxe_pool_elem *elem)
246 {
247 return kref_put(&elem->ref_cnt, rxe_elem_release);
248 }
249
__rxe_finalize(struct rxe_pool_elem * elem)250 void __rxe_finalize(struct rxe_pool_elem *elem)
251 {
252 void *xa_ret;
253
254 xa_ret = xa_store(&elem->pool->xa, elem->index, elem, GFP_KERNEL);
255 WARN_ON(xa_err(xa_ret));
256 }
257