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; | |
57 | struct rds_rdma_op *s_op; | |
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 */ | |
101 | struct mutex i_recv_mutex; | |
102 | struct rds_ib_work_ring i_recv_ring; | |
103 | struct rds_ib_incoming *i_ibinc; | |
104 | u32 i_recv_data_rem; | |
105 | struct rds_header *i_recv_hdrs; | |
106 | u64 i_recv_hdrs_dma; | |
107 | struct rds_ib_recv_work *i_recvs; | |
108 | struct rds_page_frag i_frag; | |
109 | u64 i_ack_recv; /* last ACK received */ | |
110 | ||
111 | /* sending acks */ | |
112 | unsigned long i_ack_flags; | |
8cbd9606 AG |
113 | #ifdef KERNEL_HAS_ATOMIC64 |
114 | atomic64_t i_ack_next; /* next ACK to send */ | |
115 | #else | |
116 | spinlock_t i_ack_lock; /* protect i_ack_next */ | |
ec16227e | 117 | u64 i_ack_next; /* next ACK to send */ |
8cbd9606 | 118 | #endif |
ec16227e AG |
119 | struct rds_header *i_ack; |
120 | struct ib_send_wr i_ack_wr; | |
121 | struct ib_sge i_ack_sge; | |
122 | u64 i_ack_dma; | |
123 | unsigned long i_ack_queued; | |
124 | ||
125 | /* Flow control related information | |
126 | * | |
127 | * Our algorithm uses a pair variables that we need to access | |
128 | * atomically - one for the send credits, and one posted | |
129 | * recv credits we need to transfer to remote. | |
130 | * Rather than protect them using a slow spinlock, we put both into | |
131 | * a single atomic_t and update it using cmpxchg | |
132 | */ | |
133 | atomic_t i_credits; | |
134 | ||
135 | /* Protocol version specific information */ | |
136 | unsigned int i_flowctl:1; /* enable/disable flow ctl */ | |
137 | ||
138 | /* Batched completions */ | |
139 | unsigned int i_unsignaled_wrs; | |
140 | long i_unsignaled_bytes; | |
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; | |
166 | spinlock_t spinlock; /* protect the above */ | |
167 | }; | |
168 | ||
169 | /* bits for i_ack_flags */ | |
170 | #define IB_ACK_IN_FLIGHT 0 | |
171 | #define IB_ACK_REQUESTED 1 | |
172 | ||
173 | /* Magic WR_ID for ACKs */ | |
174 | #define RDS_IB_ACK_WR_ID (~(u64) 0) | |
175 | ||
176 | struct rds_ib_statistics { | |
177 | uint64_t s_ib_connect_raced; | |
178 | uint64_t s_ib_listen_closed_stale; | |
179 | uint64_t s_ib_tx_cq_call; | |
180 | uint64_t s_ib_tx_cq_event; | |
181 | uint64_t s_ib_tx_ring_full; | |
182 | uint64_t s_ib_tx_throttle; | |
183 | uint64_t s_ib_tx_sg_mapping_failure; | |
184 | uint64_t s_ib_tx_stalled; | |
185 | uint64_t s_ib_tx_credit_updates; | |
186 | uint64_t s_ib_rx_cq_call; | |
187 | uint64_t s_ib_rx_cq_event; | |
188 | uint64_t s_ib_rx_ring_empty; | |
189 | uint64_t s_ib_rx_refill_from_cq; | |
190 | uint64_t s_ib_rx_refill_from_thread; | |
191 | uint64_t s_ib_rx_alloc_limit; | |
192 | uint64_t s_ib_rx_credit_updates; | |
193 | uint64_t s_ib_ack_sent; | |
194 | uint64_t s_ib_ack_send_failure; | |
195 | uint64_t s_ib_ack_send_delayed; | |
196 | uint64_t s_ib_ack_send_piggybacked; | |
197 | uint64_t s_ib_ack_received; | |
198 | uint64_t s_ib_rdma_mr_alloc; | |
199 | uint64_t s_ib_rdma_mr_free; | |
200 | uint64_t s_ib_rdma_mr_used; | |
201 | uint64_t s_ib_rdma_mr_pool_flush; | |
202 | uint64_t s_ib_rdma_mr_pool_wait; | |
203 | uint64_t s_ib_rdma_mr_pool_depleted; | |
204 | }; | |
205 | ||
206 | extern struct workqueue_struct *rds_ib_wq; | |
207 | ||
208 | /* | |
209 | * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h | |
210 | * doesn't define it. | |
211 | */ | |
212 | static inline void rds_ib_dma_sync_sg_for_cpu(struct ib_device *dev, | |
213 | struct scatterlist *sg, unsigned int sg_dma_len, int direction) | |
214 | { | |
215 | unsigned int i; | |
216 | ||
217 | for (i = 0; i < sg_dma_len; ++i) { | |
218 | ib_dma_sync_single_for_cpu(dev, | |
219 | ib_sg_dma_address(dev, &sg[i]), | |
220 | ib_sg_dma_len(dev, &sg[i]), | |
221 | direction); | |
222 | } | |
223 | } | |
224 | #define ib_dma_sync_sg_for_cpu rds_ib_dma_sync_sg_for_cpu | |
225 | ||
226 | static inline void rds_ib_dma_sync_sg_for_device(struct ib_device *dev, | |
227 | struct scatterlist *sg, unsigned int sg_dma_len, int direction) | |
228 | { | |
229 | unsigned int i; | |
230 | ||
231 | for (i = 0; i < sg_dma_len; ++i) { | |
232 | ib_dma_sync_single_for_device(dev, | |
233 | ib_sg_dma_address(dev, &sg[i]), | |
234 | ib_sg_dma_len(dev, &sg[i]), | |
235 | direction); | |
236 | } | |
237 | } | |
238 | #define ib_dma_sync_sg_for_device rds_ib_dma_sync_sg_for_device | |
239 | ||
240 | ||
241 | /* ib.c */ | |
242 | extern struct rds_transport rds_ib_transport; | |
243 | extern void rds_ib_add_one(struct ib_device *device); | |
244 | extern void rds_ib_remove_one(struct ib_device *device); | |
245 | extern struct ib_client rds_ib_client; | |
246 | ||
247 | extern unsigned int fmr_pool_size; | |
248 | extern unsigned int fmr_message_size; | |
3ba23ade | 249 | extern unsigned int rds_ib_retry_count; |
ec16227e AG |
250 | |
251 | extern spinlock_t ib_nodev_conns_lock; | |
252 | extern struct list_head ib_nodev_conns; | |
253 | ||
254 | /* ib_cm.c */ | |
255 | int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp); | |
256 | void rds_ib_conn_free(void *arg); | |
257 | int rds_ib_conn_connect(struct rds_connection *conn); | |
258 | void rds_ib_conn_shutdown(struct rds_connection *conn); | |
259 | void rds_ib_state_change(struct sock *sk); | |
260 | int __init rds_ib_listen_init(void); | |
261 | void rds_ib_listen_stop(void); | |
262 | void __rds_ib_conn_error(struct rds_connection *conn, const char *, ...); | |
263 | int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id, | |
264 | struct rdma_cm_event *event); | |
265 | int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id); | |
266 | void rds_ib_cm_connect_complete(struct rds_connection *conn, | |
267 | struct rdma_cm_event *event); | |
268 | ||
269 | ||
270 | #define rds_ib_conn_error(conn, fmt...) \ | |
271 | __rds_ib_conn_error(conn, KERN_WARNING "RDS/IB: " fmt) | |
272 | ||
273 | /* ib_rdma.c */ | |
274 | int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr); | |
745cbcca AG |
275 | void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn); |
276 | void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn); | |
277 | void __rds_ib_destroy_conns(struct list_head *list, spinlock_t *list_lock); | |
278 | static inline void rds_ib_destroy_nodev_conns(void) | |
279 | { | |
280 | __rds_ib_destroy_conns(&ib_nodev_conns, &ib_nodev_conns_lock); | |
281 | } | |
282 | static inline void rds_ib_destroy_conns(struct rds_ib_device *rds_ibdev) | |
283 | { | |
284 | __rds_ib_destroy_conns(&rds_ibdev->conn_list, &rds_ibdev->spinlock); | |
285 | } | |
ec16227e AG |
286 | struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *); |
287 | void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo); | |
288 | void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *); | |
289 | void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents, | |
290 | struct rds_sock *rs, u32 *key_ret); | |
291 | void rds_ib_sync_mr(void *trans_private, int dir); | |
292 | void rds_ib_free_mr(void *trans_private, int invalidate); | |
293 | void rds_ib_flush_mrs(void); | |
294 | ||
295 | /* ib_recv.c */ | |
296 | int __init rds_ib_recv_init(void); | |
297 | void rds_ib_recv_exit(void); | |
298 | int rds_ib_recv(struct rds_connection *conn); | |
299 | int rds_ib_recv_refill(struct rds_connection *conn, gfp_t kptr_gfp, | |
300 | gfp_t page_gfp, int prefill); | |
301 | void rds_ib_inc_purge(struct rds_incoming *inc); | |
302 | void rds_ib_inc_free(struct rds_incoming *inc); | |
303 | int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iovec *iov, | |
304 | size_t size); | |
305 | void rds_ib_recv_cq_comp_handler(struct ib_cq *cq, void *context); | |
306 | void rds_ib_recv_init_ring(struct rds_ib_connection *ic); | |
307 | void rds_ib_recv_clear_ring(struct rds_ib_connection *ic); | |
308 | void rds_ib_recv_init_ack(struct rds_ib_connection *ic); | |
309 | void rds_ib_attempt_ack(struct rds_ib_connection *ic); | |
310 | void rds_ib_ack_send_complete(struct rds_ib_connection *ic); | |
311 | u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic); | |
312 | ||
313 | /* ib_ring.c */ | |
314 | void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr); | |
315 | void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr); | |
316 | u32 rds_ib_ring_alloc(struct rds_ib_work_ring *ring, u32 val, u32 *pos); | |
317 | void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val); | |
318 | void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val); | |
319 | int rds_ib_ring_empty(struct rds_ib_work_ring *ring); | |
320 | int rds_ib_ring_low(struct rds_ib_work_ring *ring); | |
321 | u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring); | |
322 | u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest); | |
323 | extern wait_queue_head_t rds_ib_ring_empty_wait; | |
324 | ||
325 | /* ib_send.c */ | |
326 | void rds_ib_xmit_complete(struct rds_connection *conn); | |
327 | int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm, | |
328 | unsigned int hdr_off, unsigned int sg, unsigned int off); | |
329 | void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context); | |
330 | void rds_ib_send_init_ring(struct rds_ib_connection *ic); | |
331 | void rds_ib_send_clear_ring(struct rds_ib_connection *ic); | |
332 | int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op); | |
333 | void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits); | |
334 | void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted); | |
335 | int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted, | |
7b70d033 | 336 | u32 *adv_credits, int need_posted, int max_posted); |
ec16227e AG |
337 | |
338 | /* ib_stats.c */ | |
339 | DECLARE_PER_CPU(struct rds_ib_statistics, rds_ib_stats); | |
340 | #define rds_ib_stats_inc(member) rds_stats_inc_which(rds_ib_stats, member) | |
341 | unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter, | |
342 | unsigned int avail); | |
343 | ||
344 | /* ib_sysctl.c */ | |
345 | int __init rds_ib_sysctl_init(void); | |
346 | void rds_ib_sysctl_exit(void); | |
347 | extern unsigned long rds_ib_sysctl_max_send_wr; | |
348 | extern unsigned long rds_ib_sysctl_max_recv_wr; | |
349 | extern unsigned long rds_ib_sysctl_max_unsig_wrs; | |
350 | extern unsigned long rds_ib_sysctl_max_unsig_bytes; | |
351 | extern unsigned long rds_ib_sysctl_max_recv_allocation; | |
352 | extern unsigned int rds_ib_sysctl_flow_control; | |
353 | extern ctl_table rds_ib_sysctl_table[]; | |
354 | ||
355 | /* | |
356 | * Helper functions for getting/setting the header and data SGEs in | |
357 | * RDS packets (not RDMA) | |
02a6a259 AG |
358 | * |
359 | * From version 3.1 onwards, header is in front of data in the sge. | |
ec16227e AG |
360 | */ |
361 | static inline struct ib_sge * | |
362 | rds_ib_header_sge(struct rds_ib_connection *ic, struct ib_sge *sge) | |
363 | { | |
02a6a259 AG |
364 | if (ic->conn->c_version > RDS_PROTOCOL_3_0) |
365 | return &sge[0]; | |
366 | else | |
367 | return &sge[1]; | |
ec16227e AG |
368 | } |
369 | ||
370 | static inline struct ib_sge * | |
371 | rds_ib_data_sge(struct rds_ib_connection *ic, struct ib_sge *sge) | |
372 | { | |
02a6a259 AG |
373 | if (ic->conn->c_version > RDS_PROTOCOL_3_0) |
374 | return &sge[1]; | |
375 | else | |
376 | return &sge[0]; | |
ec16227e AG |
377 | } |
378 | ||
ec16227e | 379 | #endif |