Commit | Line | Data |
---|---|---|
ec16227e AG |
1 | #ifndef _RDS_IB_H |
2 | #define _RDS_IB_H | |
3 | ||
4 | #include <rdma/ib_verbs.h> | |
5 | #include <rdma/rdma_cm.h> | |
6 | #include "rds.h" | |
7 | #include "rdma_transport.h" | |
8 | ||
9 | #define RDS_FMR_SIZE 256 | |
10 | #define RDS_FMR_POOL_SIZE 4096 | |
11 | ||
12 | #define RDS_IB_MAX_SGE 8 | |
13 | #define RDS_IB_RECV_SGE 2 | |
14 | ||
15 | #define RDS_IB_DEFAULT_RECV_WR 1024 | |
16 | #define RDS_IB_DEFAULT_SEND_WR 256 | |
17 | ||
3ba23ade AG |
18 | #define RDS_IB_DEFAULT_RETRY_COUNT 2 |
19 | ||
ec16227e AG |
20 | #define RDS_IB_SUPPORTED_PROTOCOLS 0x00000003 /* minor versions supported */ |
21 | ||
22 | extern struct list_head rds_ib_devices; | |
23 | ||
24 | /* | |
25 | * IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to | |
26 | * try and minimize the amount of memory tied up both the device and | |
27 | * socket receive queues. | |
28 | */ | |
29 | /* page offset of the final full frag that fits in the page */ | |
30 | #define RDS_PAGE_LAST_OFF (((PAGE_SIZE / RDS_FRAG_SIZE) - 1) * RDS_FRAG_SIZE) | |
31 | struct rds_page_frag { | |
32 | struct list_head f_item; | |
33 | struct page *f_page; | |
34 | unsigned long f_offset; | |
35 | dma_addr_t f_mapped; | |
36 | }; | |
37 | ||
38 | struct rds_ib_incoming { | |
39 | struct list_head ii_frags; | |
40 | struct rds_incoming ii_inc; | |
41 | }; | |
42 | ||
43 | struct rds_ib_connect_private { | |
44 | /* Add new fields at the end, and don't permute existing fields. */ | |
45 | __be32 dp_saddr; | |
46 | __be32 dp_daddr; | |
47 | u8 dp_protocol_major; | |
48 | u8 dp_protocol_minor; | |
49 | __be16 dp_protocol_minor_mask; /* bitmask */ | |
50 | __be32 dp_reserved1; | |
51 | __be64 dp_ack_seq; | |
52 | __be32 dp_credit; /* non-zero enables flow ctl */ | |
53 | }; | |
54 | ||
55 | struct rds_ib_send_work { | |
56 | struct rds_message *s_rm; | |
f8b3aaf2 | 57 | struct rm_rdma_op *s_op; |
ec16227e AG |
58 | struct ib_send_wr s_wr; |
59 | struct ib_sge s_sge[RDS_IB_MAX_SGE]; | |
60 | unsigned long s_queued; | |
61 | }; | |
62 | ||
63 | struct rds_ib_recv_work { | |
64 | struct rds_ib_incoming *r_ibinc; | |
65 | struct rds_page_frag *r_frag; | |
66 | struct ib_recv_wr r_wr; | |
67 | struct ib_sge r_sge[2]; | |
68 | }; | |
69 | ||
70 | struct rds_ib_work_ring { | |
71 | u32 w_nr; | |
72 | u32 w_alloc_ptr; | |
73 | u32 w_alloc_ctr; | |
74 | u32 w_free_ptr; | |
75 | atomic_t w_free_ctr; | |
76 | }; | |
77 | ||
78 | struct rds_ib_device; | |
79 | ||
80 | struct rds_ib_connection { | |
81 | ||
82 | struct list_head ib_node; | |
83 | struct rds_ib_device *rds_ibdev; | |
84 | struct rds_connection *conn; | |
85 | ||
86 | /* alphabet soup, IBTA style */ | |
87 | struct rdma_cm_id *i_cm_id; | |
88 | struct ib_pd *i_pd; | |
89 | struct ib_mr *i_mr; | |
90 | struct ib_cq *i_send_cq; | |
91 | struct ib_cq *i_recv_cq; | |
92 | ||
93 | /* tx */ | |
94 | struct rds_ib_work_ring i_send_ring; | |
95 | struct rds_message *i_rm; | |
96 | struct rds_header *i_send_hdrs; | |
97 | u64 i_send_hdrs_dma; | |
98 | struct rds_ib_send_work *i_sends; | |
99 | ||
100 | /* rx */ | |
d521b63b | 101 | struct tasklet_struct i_recv_tasklet; |
ec16227e AG |
102 | struct mutex i_recv_mutex; |
103 | struct rds_ib_work_ring i_recv_ring; | |
104 | struct rds_ib_incoming *i_ibinc; | |
105 | u32 i_recv_data_rem; | |
106 | struct rds_header *i_recv_hdrs; | |
107 | u64 i_recv_hdrs_dma; | |
108 | struct rds_ib_recv_work *i_recvs; | |
109 | struct rds_page_frag i_frag; | |
110 | u64 i_ack_recv; /* last ACK received */ | |
111 | ||
112 | /* sending acks */ | |
113 | unsigned long i_ack_flags; | |
8cbd9606 AG |
114 | #ifdef KERNEL_HAS_ATOMIC64 |
115 | atomic64_t i_ack_next; /* next ACK to send */ | |
116 | #else | |
117 | spinlock_t i_ack_lock; /* protect i_ack_next */ | |
ec16227e | 118 | u64 i_ack_next; /* next ACK to send */ |
8cbd9606 | 119 | #endif |
ec16227e AG |
120 | struct rds_header *i_ack; |
121 | struct ib_send_wr i_ack_wr; | |
122 | struct ib_sge i_ack_sge; | |
123 | u64 i_ack_dma; | |
124 | unsigned long i_ack_queued; | |
125 | ||
126 | /* Flow control related information | |
127 | * | |
128 | * Our algorithm uses a pair variables that we need to access | |
129 | * atomically - one for the send credits, and one posted | |
130 | * recv credits we need to transfer to remote. | |
131 | * Rather than protect them using a slow spinlock, we put both into | |
132 | * a single atomic_t and update it using cmpxchg | |
133 | */ | |
134 | atomic_t i_credits; | |
135 | ||
136 | /* Protocol version specific information */ | |
137 | unsigned int i_flowctl:1; /* enable/disable flow ctl */ | |
138 | ||
139 | /* Batched completions */ | |
140 | unsigned int i_unsignaled_wrs; | |
ec16227e AG |
141 | }; |
142 | ||
143 | /* This assumes that atomic_t is at least 32 bits */ | |
144 | #define IB_GET_SEND_CREDITS(v) ((v) & 0xffff) | |
145 | #define IB_GET_POST_CREDITS(v) ((v) >> 16) | |
146 | #define IB_SET_SEND_CREDITS(v) ((v) & 0xffff) | |
147 | #define IB_SET_POST_CREDITS(v) ((v) << 16) | |
148 | ||
149 | struct rds_ib_ipaddr { | |
150 | struct list_head list; | |
151 | __be32 ipaddr; | |
152 | }; | |
153 | ||
154 | struct rds_ib_device { | |
155 | struct list_head list; | |
156 | struct list_head ipaddr_list; | |
157 | struct list_head conn_list; | |
158 | struct ib_device *dev; | |
159 | struct ib_pd *pd; | |
160 | struct ib_mr *mr; | |
161 | struct rds_ib_mr_pool *mr_pool; | |
ec16227e AG |
162 | unsigned int fmr_max_remaps; |
163 | unsigned int max_fmrs; | |
164 | int max_sge; | |
165 | unsigned int max_wrs; | |
40589e74 AG |
166 | unsigned int max_initiator_depth; |
167 | unsigned int max_responder_resources; | |
ec16227e AG |
168 | spinlock_t spinlock; /* protect the above */ |
169 | }; | |
170 | ||
171 | /* bits for i_ack_flags */ | |
172 | #define IB_ACK_IN_FLIGHT 0 | |
173 | #define IB_ACK_REQUESTED 1 | |
174 | ||
175 | /* Magic WR_ID for ACKs */ | |
176 | #define RDS_IB_ACK_WR_ID (~(u64) 0) | |
177 | ||
178 | struct rds_ib_statistics { | |
179 | uint64_t s_ib_connect_raced; | |
180 | uint64_t s_ib_listen_closed_stale; | |
181 | uint64_t s_ib_tx_cq_call; | |
182 | uint64_t s_ib_tx_cq_event; | |
183 | uint64_t s_ib_tx_ring_full; | |
184 | uint64_t s_ib_tx_throttle; | |
185 | uint64_t s_ib_tx_sg_mapping_failure; | |
186 | uint64_t s_ib_tx_stalled; | |
187 | uint64_t s_ib_tx_credit_updates; | |
188 | uint64_t s_ib_rx_cq_call; | |
189 | uint64_t s_ib_rx_cq_event; | |
190 | uint64_t s_ib_rx_ring_empty; | |
191 | uint64_t s_ib_rx_refill_from_cq; | |
192 | uint64_t s_ib_rx_refill_from_thread; | |
193 | uint64_t s_ib_rx_alloc_limit; | |
194 | uint64_t s_ib_rx_credit_updates; | |
195 | uint64_t s_ib_ack_sent; | |
196 | uint64_t s_ib_ack_send_failure; | |
197 | uint64_t s_ib_ack_send_delayed; | |
198 | uint64_t s_ib_ack_send_piggybacked; | |
199 | uint64_t s_ib_ack_received; | |
200 | uint64_t s_ib_rdma_mr_alloc; | |
201 | uint64_t s_ib_rdma_mr_free; | |
202 | uint64_t s_ib_rdma_mr_used; | |
203 | uint64_t s_ib_rdma_mr_pool_flush; | |
204 | uint64_t s_ib_rdma_mr_pool_wait; | |
205 | uint64_t s_ib_rdma_mr_pool_depleted; | |
206 | }; | |
207 | ||
208 | extern struct workqueue_struct *rds_ib_wq; | |
209 | ||
210 | /* | |
211 | * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h | |
212 | * doesn't define it. | |
213 | */ | |
214 | static inline void rds_ib_dma_sync_sg_for_cpu(struct ib_device *dev, | |
215 | struct scatterlist *sg, unsigned int sg_dma_len, int direction) | |
216 | { | |
217 | unsigned int i; | |
218 | ||
219 | for (i = 0; i < sg_dma_len; ++i) { | |
220 | ib_dma_sync_single_for_cpu(dev, | |
221 | ib_sg_dma_address(dev, &sg[i]), | |
222 | ib_sg_dma_len(dev, &sg[i]), | |
223 | direction); | |
224 | } | |
225 | } | |
226 | #define ib_dma_sync_sg_for_cpu rds_ib_dma_sync_sg_for_cpu | |
227 | ||
228 | static inline void rds_ib_dma_sync_sg_for_device(struct ib_device *dev, | |
229 | struct scatterlist *sg, unsigned int sg_dma_len, int direction) | |
230 | { | |
231 | unsigned int i; | |
232 | ||
233 | for (i = 0; i < sg_dma_len; ++i) { | |
234 | ib_dma_sync_single_for_device(dev, | |
235 | ib_sg_dma_address(dev, &sg[i]), | |
236 | ib_sg_dma_len(dev, &sg[i]), | |
237 | direction); | |
238 | } | |
239 | } | |
240 | #define ib_dma_sync_sg_for_device rds_ib_dma_sync_sg_for_device | |
241 | ||
242 | ||
243 | /* ib.c */ | |
244 | extern struct rds_transport rds_ib_transport; | |
245 | extern void rds_ib_add_one(struct ib_device *device); | |
246 | extern void rds_ib_remove_one(struct ib_device *device); | |
247 | extern struct ib_client rds_ib_client; | |
248 | ||
249 | extern unsigned int fmr_pool_size; | |
250 | extern unsigned int fmr_message_size; | |
3ba23ade | 251 | extern unsigned int rds_ib_retry_count; |
ec16227e AG |
252 | |
253 | extern spinlock_t ib_nodev_conns_lock; | |
254 | extern struct list_head ib_nodev_conns; | |
255 | ||
256 | /* ib_cm.c */ | |
257 | int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp); | |
258 | void rds_ib_conn_free(void *arg); | |
259 | int rds_ib_conn_connect(struct rds_connection *conn); | |
260 | void rds_ib_conn_shutdown(struct rds_connection *conn); | |
261 | void rds_ib_state_change(struct sock *sk); | |
262 | int __init rds_ib_listen_init(void); | |
263 | void rds_ib_listen_stop(void); | |
264 | void __rds_ib_conn_error(struct rds_connection *conn, const char *, ...); | |
265 | int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id, | |
266 | struct rdma_cm_event *event); | |
267 | int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id); | |
268 | void rds_ib_cm_connect_complete(struct rds_connection *conn, | |
269 | struct rdma_cm_event *event); | |
270 | ||
271 | ||
272 | #define rds_ib_conn_error(conn, fmt...) \ | |
273 | __rds_ib_conn_error(conn, KERN_WARNING "RDS/IB: " fmt) | |
274 | ||
275 | /* ib_rdma.c */ | |
276 | int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr); | |
745cbcca AG |
277 | void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn); |
278 | void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn); | |
279 | void __rds_ib_destroy_conns(struct list_head *list, spinlock_t *list_lock); | |
280 | static inline void rds_ib_destroy_nodev_conns(void) | |
281 | { | |
282 | __rds_ib_destroy_conns(&ib_nodev_conns, &ib_nodev_conns_lock); | |
283 | } | |
284 | static inline void rds_ib_destroy_conns(struct rds_ib_device *rds_ibdev) | |
285 | { | |
286 | __rds_ib_destroy_conns(&rds_ibdev->conn_list, &rds_ibdev->spinlock); | |
287 | } | |
ec16227e AG |
288 | struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *); |
289 | void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo); | |
290 | void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *); | |
291 | void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents, | |
292 | struct rds_sock *rs, u32 *key_ret); | |
293 | void rds_ib_sync_mr(void *trans_private, int dir); | |
294 | void rds_ib_free_mr(void *trans_private, int invalidate); | |
295 | void rds_ib_flush_mrs(void); | |
296 | ||
297 | /* ib_recv.c */ | |
298 | int __init rds_ib_recv_init(void); | |
299 | void rds_ib_recv_exit(void); | |
300 | int rds_ib_recv(struct rds_connection *conn); | |
301 | int rds_ib_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp, | |
302 | gfp_t page_gfp, int prefill); | |
ec16227e AG |
303 | void rds_ib_inc_free(struct rds_incoming *inc); |
304 | int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov, | |
305 | size_t size); | |
306 | void rds_ib_recv_cq_comp_handler(struct ib_cq *cq, void *context); | |
d521b63b | 307 | void rds_ib_recv_tasklet_fn(unsigned long data); |
ec16227e AG |
308 | void rds_ib_recv_init_ring(struct rds_ib_connection *ic); |
309 | void rds_ib_recv_clear_ring(struct rds_ib_connection *ic); | |
310 | void rds_ib_recv_init_ack(struct rds_ib_connection *ic); | |
311 | void rds_ib_attempt_ack(struct rds_ib_connection *ic); | |
312 | void rds_ib_ack_send_complete(struct rds_ib_connection *ic); | |
313 | u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic); | |
314 | ||
315 | /* ib_ring.c */ | |
316 | void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr); | |
317 | void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr); | |
318 | u32 rds_ib_ring_alloc(struct rds_ib_work_ring *ring, u32 val, u32 *pos); | |
319 | void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val); | |
320 | void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val); | |
321 | int rds_ib_ring_empty(struct rds_ib_work_ring *ring); | |
322 | int rds_ib_ring_low(struct rds_ib_work_ring *ring); | |
323 | u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring); | |
324 | u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest); | |
325 | extern wait_queue_head_t rds_ib_ring_empty_wait; | |
326 | ||
327 | /* ib_send.c */ | |
328 | void rds_ib_xmit_complete(struct rds_connection *conn); | |
329 | int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm, | |
330 | unsigned int hdr_off, unsigned int sg, unsigned int off); | |
331 | void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context); | |
332 | void rds_ib_send_init_ring(struct rds_ib_connection *ic); | |
333 | void rds_ib_send_clear_ring(struct rds_ib_connection *ic); | |
f8b3aaf2 | 334 | int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op); |
ec16227e AG |
335 | void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits); |
336 | void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted); | |
337 | int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted, | |
7b70d033 | 338 | u32 *adv_credits, int need_posted, int max_posted); |
241eef3e | 339 | int rds_ib_xmit_atomic(struct rds_connection *conn, struct rds_message *rm); |
ec16227e AG |
340 | |
341 | /* ib_stats.c */ | |
342 | DECLARE_PER_CPU(struct rds_ib_statistics, rds_ib_stats); | |
343 | #define rds_ib_stats_inc(member) rds_stats_inc_which(rds_ib_stats, member) | |
344 | unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter, | |
345 | unsigned int avail); | |
346 | ||
347 | /* ib_sysctl.c */ | |
348 | int __init rds_ib_sysctl_init(void); | |
349 | void rds_ib_sysctl_exit(void); | |
350 | extern unsigned long rds_ib_sysctl_max_send_wr; | |
351 | extern unsigned long rds_ib_sysctl_max_recv_wr; | |
352 | extern unsigned long rds_ib_sysctl_max_unsig_wrs; | |
353 | extern unsigned long rds_ib_sysctl_max_unsig_bytes; | |
354 | extern unsigned long rds_ib_sysctl_max_recv_allocation; | |
355 | extern unsigned int rds_ib_sysctl_flow_control; | |
356 | extern ctl_table rds_ib_sysctl_table[]; | |
357 | ||
ec16227e | 358 | #endif |