Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 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 | |
| 18 | #define RDS_IB_SUPPORTED_PROTOCOLS 0x00000003 /* minor versions supported */ |
| 19 | |
| 20 | extern struct list_head rds_ib_devices; |
| 21 | |
| 22 | /* |
| 23 | * IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to |
| 24 | * try and minimize the amount of memory tied up both the device and |
| 25 | * socket receive queues. |
| 26 | */ |
| 27 | /* page offset of the final full frag that fits in the page */ |
| 28 | #define RDS_PAGE_LAST_OFF (((PAGE_SIZE / RDS_FRAG_SIZE) - 1) * RDS_FRAG_SIZE) |
| 29 | struct rds_page_frag { |
| 30 | struct list_head f_item; |
| 31 | struct page *f_page; |
| 32 | unsigned long f_offset; |
| 33 | dma_addr_t f_mapped; |
| 34 | }; |
| 35 | |
| 36 | struct rds_ib_incoming { |
| 37 | struct list_head ii_frags; |
| 38 | struct rds_incoming ii_inc; |
| 39 | }; |
| 40 | |
| 41 | struct rds_ib_connect_private { |
| 42 | /* Add new fields at the end, and don't permute existing fields. */ |
| 43 | __be32 dp_saddr; |
| 44 | __be32 dp_daddr; |
| 45 | u8 dp_protocol_major; |
| 46 | u8 dp_protocol_minor; |
| 47 | __be16 dp_protocol_minor_mask; /* bitmask */ |
| 48 | __be32 dp_reserved1; |
| 49 | __be64 dp_ack_seq; |
| 50 | __be32 dp_credit; /* non-zero enables flow ctl */ |
| 51 | }; |
| 52 | |
| 53 | struct rds_ib_send_work { |
| 54 | struct rds_message *s_rm; |
| 55 | struct rds_rdma_op *s_op; |
| 56 | struct ib_send_wr s_wr; |
| 57 | struct ib_sge s_sge[RDS_IB_MAX_SGE]; |
| 58 | unsigned long s_queued; |
| 59 | }; |
| 60 | |
| 61 | struct rds_ib_recv_work { |
| 62 | struct rds_ib_incoming *r_ibinc; |
| 63 | struct rds_page_frag *r_frag; |
| 64 | struct ib_recv_wr r_wr; |
| 65 | struct ib_sge r_sge[2]; |
| 66 | }; |
| 67 | |
| 68 | struct rds_ib_work_ring { |
| 69 | u32 w_nr; |
| 70 | u32 w_alloc_ptr; |
| 71 | u32 w_alloc_ctr; |
| 72 | u32 w_free_ptr; |
| 73 | atomic_t w_free_ctr; |
| 74 | }; |
| 75 | |
| 76 | struct rds_ib_device; |
| 77 | |
| 78 | struct rds_ib_connection { |
| 79 | |
| 80 | struct list_head ib_node; |
| 81 | struct rds_ib_device *rds_ibdev; |
| 82 | struct rds_connection *conn; |
| 83 | |
| 84 | /* alphabet soup, IBTA style */ |
| 85 | struct rdma_cm_id *i_cm_id; |
| 86 | struct ib_pd *i_pd; |
| 87 | struct ib_mr *i_mr; |
| 88 | struct ib_cq *i_send_cq; |
| 89 | struct ib_cq *i_recv_cq; |
| 90 | |
| 91 | /* tx */ |
| 92 | struct rds_ib_work_ring i_send_ring; |
| 93 | struct rds_message *i_rm; |
| 94 | struct rds_header *i_send_hdrs; |
| 95 | u64 i_send_hdrs_dma; |
| 96 | struct rds_ib_send_work *i_sends; |
| 97 | |
| 98 | /* rx */ |
| 99 | struct mutex i_recv_mutex; |
| 100 | struct rds_ib_work_ring i_recv_ring; |
| 101 | struct rds_ib_incoming *i_ibinc; |
| 102 | u32 i_recv_data_rem; |
| 103 | struct rds_header *i_recv_hdrs; |
| 104 | u64 i_recv_hdrs_dma; |
| 105 | struct rds_ib_recv_work *i_recvs; |
| 106 | struct rds_page_frag i_frag; |
| 107 | u64 i_ack_recv; /* last ACK received */ |
| 108 | |
| 109 | /* sending acks */ |
| 110 | unsigned long i_ack_flags; |
Andy Grover | 8cbd960 | 2009-04-01 08:20:20 +0000 | [diff] [blame] | 111 | #ifdef KERNEL_HAS_ATOMIC64 |
| 112 | atomic64_t i_ack_next; /* next ACK to send */ |
| 113 | #else |
| 114 | spinlock_t i_ack_lock; /* protect i_ack_next */ |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 115 | u64 i_ack_next; /* next ACK to send */ |
Andy Grover | 8cbd960 | 2009-04-01 08:20:20 +0000 | [diff] [blame] | 116 | #endif |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 117 | struct rds_header *i_ack; |
| 118 | struct ib_send_wr i_ack_wr; |
| 119 | struct ib_sge i_ack_sge; |
| 120 | u64 i_ack_dma; |
| 121 | unsigned long i_ack_queued; |
| 122 | |
| 123 | /* Flow control related information |
| 124 | * |
| 125 | * Our algorithm uses a pair variables that we need to access |
| 126 | * atomically - one for the send credits, and one posted |
| 127 | * recv credits we need to transfer to remote. |
| 128 | * Rather than protect them using a slow spinlock, we put both into |
| 129 | * a single atomic_t and update it using cmpxchg |
| 130 | */ |
| 131 | atomic_t i_credits; |
| 132 | |
| 133 | /* Protocol version specific information */ |
| 134 | unsigned int i_flowctl:1; /* enable/disable flow ctl */ |
| 135 | |
| 136 | /* Batched completions */ |
| 137 | unsigned int i_unsignaled_wrs; |
| 138 | long i_unsignaled_bytes; |
| 139 | }; |
| 140 | |
| 141 | /* This assumes that atomic_t is at least 32 bits */ |
| 142 | #define IB_GET_SEND_CREDITS(v) ((v) & 0xffff) |
| 143 | #define IB_GET_POST_CREDITS(v) ((v) >> 16) |
| 144 | #define IB_SET_SEND_CREDITS(v) ((v) & 0xffff) |
| 145 | #define IB_SET_POST_CREDITS(v) ((v) << 16) |
| 146 | |
| 147 | struct rds_ib_ipaddr { |
| 148 | struct list_head list; |
| 149 | __be32 ipaddr; |
| 150 | }; |
| 151 | |
| 152 | struct rds_ib_device { |
| 153 | struct list_head list; |
| 154 | struct list_head ipaddr_list; |
| 155 | struct list_head conn_list; |
| 156 | struct ib_device *dev; |
| 157 | struct ib_pd *pd; |
| 158 | struct ib_mr *mr; |
| 159 | struct rds_ib_mr_pool *mr_pool; |
| 160 | int fmr_page_shift; |
| 161 | int fmr_page_size; |
| 162 | u64 fmr_page_mask; |
| 163 | unsigned int fmr_max_remaps; |
| 164 | unsigned int max_fmrs; |
| 165 | int max_sge; |
| 166 | unsigned int max_wrs; |
| 167 | spinlock_t spinlock; /* protect the above */ |
| 168 | }; |
| 169 | |
| 170 | /* bits for i_ack_flags */ |
| 171 | #define IB_ACK_IN_FLIGHT 0 |
| 172 | #define IB_ACK_REQUESTED 1 |
| 173 | |
| 174 | /* Magic WR_ID for ACKs */ |
| 175 | #define RDS_IB_ACK_WR_ID (~(u64) 0) |
| 176 | |
| 177 | struct rds_ib_statistics { |
| 178 | uint64_t s_ib_connect_raced; |
| 179 | uint64_t s_ib_listen_closed_stale; |
| 180 | uint64_t s_ib_tx_cq_call; |
| 181 | uint64_t s_ib_tx_cq_event; |
| 182 | uint64_t s_ib_tx_ring_full; |
| 183 | uint64_t s_ib_tx_throttle; |
| 184 | uint64_t s_ib_tx_sg_mapping_failure; |
| 185 | uint64_t s_ib_tx_stalled; |
| 186 | uint64_t s_ib_tx_credit_updates; |
| 187 | uint64_t s_ib_rx_cq_call; |
| 188 | uint64_t s_ib_rx_cq_event; |
| 189 | uint64_t s_ib_rx_ring_empty; |
| 190 | uint64_t s_ib_rx_refill_from_cq; |
| 191 | uint64_t s_ib_rx_refill_from_thread; |
| 192 | uint64_t s_ib_rx_alloc_limit; |
| 193 | uint64_t s_ib_rx_credit_updates; |
| 194 | uint64_t s_ib_ack_sent; |
| 195 | uint64_t s_ib_ack_send_failure; |
| 196 | uint64_t s_ib_ack_send_delayed; |
| 197 | uint64_t s_ib_ack_send_piggybacked; |
| 198 | uint64_t s_ib_ack_received; |
| 199 | uint64_t s_ib_rdma_mr_alloc; |
| 200 | uint64_t s_ib_rdma_mr_free; |
| 201 | uint64_t s_ib_rdma_mr_used; |
| 202 | uint64_t s_ib_rdma_mr_pool_flush; |
| 203 | uint64_t s_ib_rdma_mr_pool_wait; |
| 204 | uint64_t s_ib_rdma_mr_pool_depleted; |
| 205 | }; |
| 206 | |
| 207 | extern struct workqueue_struct *rds_ib_wq; |
| 208 | |
| 209 | /* |
| 210 | * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h |
| 211 | * doesn't define it. |
| 212 | */ |
| 213 | static inline void rds_ib_dma_sync_sg_for_cpu(struct ib_device *dev, |
| 214 | struct scatterlist *sg, unsigned int sg_dma_len, int direction) |
| 215 | { |
| 216 | unsigned int i; |
| 217 | |
| 218 | for (i = 0; i < sg_dma_len; ++i) { |
| 219 | ib_dma_sync_single_for_cpu(dev, |
| 220 | ib_sg_dma_address(dev, &sg[i]), |
| 221 | ib_sg_dma_len(dev, &sg[i]), |
| 222 | direction); |
| 223 | } |
| 224 | } |
| 225 | #define ib_dma_sync_sg_for_cpu rds_ib_dma_sync_sg_for_cpu |
| 226 | |
| 227 | static inline void rds_ib_dma_sync_sg_for_device(struct ib_device *dev, |
| 228 | struct scatterlist *sg, unsigned int sg_dma_len, int direction) |
| 229 | { |
| 230 | unsigned int i; |
| 231 | |
| 232 | for (i = 0; i < sg_dma_len; ++i) { |
| 233 | ib_dma_sync_single_for_device(dev, |
| 234 | ib_sg_dma_address(dev, &sg[i]), |
| 235 | ib_sg_dma_len(dev, &sg[i]), |
| 236 | direction); |
| 237 | } |
| 238 | } |
| 239 | #define ib_dma_sync_sg_for_device rds_ib_dma_sync_sg_for_device |
| 240 | |
| 241 | |
| 242 | /* ib.c */ |
| 243 | extern struct rds_transport rds_ib_transport; |
| 244 | extern void rds_ib_add_one(struct ib_device *device); |
| 245 | extern void rds_ib_remove_one(struct ib_device *device); |
| 246 | extern struct ib_client rds_ib_client; |
| 247 | |
| 248 | extern unsigned int fmr_pool_size; |
| 249 | extern unsigned int fmr_message_size; |
| 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); |
Andy Grover | 745cbcc | 2009-04-01 08:20:19 +0000 | [diff] [blame] | 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 | } |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 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, |
| 336 | u32 *adv_credits, int need_posted); |
| 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) |
| 358 | */ |
| 359 | static inline struct ib_sge * |
| 360 | rds_ib_header_sge(struct rds_ib_connection *ic, struct ib_sge *sge) |
| 361 | { |
| 362 | return &sge[0]; |
| 363 | } |
| 364 | |
| 365 | static inline struct ib_sge * |
| 366 | rds_ib_data_sge(struct rds_ib_connection *ic, struct ib_sge *sge) |
| 367 | { |
| 368 | return &sge[1]; |
| 369 | } |
| 370 | |
Andy Grover | ec16227 | 2009-02-24 15:30:30 +0000 | [diff] [blame] | 371 | #endif |