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 <linux/interrupt.h> 7 #include <linux/pci.h> 8 #include <linux/slab.h> 9 #include "rds.h" 10 #include "rdma_transport.h" 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 #define RDS_IB_DEFAULT_FR_WR 256 18 #define RDS_IB_DEFAULT_FR_INV_WR 256 19 20 #define RDS_IB_DEFAULT_RETRY_COUNT 1 21 22 #define RDS_IB_SUPPORTED_PROTOCOLS 0x00000003 /* minor versions supported */ 23 24 #define RDS_IB_RECYCLE_BATCH_COUNT 32 25 26 #define RDS_IB_WC_MAX 32 27 28 extern struct rw_semaphore rds_ib_devices_lock; 29 extern struct list_head rds_ib_devices; 30 31 /* 32 * IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to 33 * try and minimize the amount of memory tied up both the device and 34 * socket receive queues. 35 */ 36 struct rds_page_frag { 37 struct list_head f_item; 38 struct list_head f_cache_entry; 39 struct scatterlist f_sg; 40 }; 41 42 struct rds_ib_incoming { 43 struct list_head ii_frags; 44 struct list_head ii_cache_entry; 45 struct rds_incoming ii_inc; 46 }; 47 48 struct rds_ib_cache_head { 49 struct list_head *first; 50 unsigned long count; 51 }; 52 53 struct rds_ib_refill_cache { 54 struct rds_ib_cache_head __percpu *percpu; 55 struct list_head *xfer; 56 struct list_head *ready; 57 }; 58 59 struct rds_ib_connect_private { 60 /* Add new fields at the end, and don't permute existing fields. */ 61 __be32 dp_saddr; 62 __be32 dp_daddr; 63 u8 dp_protocol_major; 64 u8 dp_protocol_minor; 65 __be16 dp_protocol_minor_mask; /* bitmask */ 66 __be32 dp_reserved1; 67 __be64 dp_ack_seq; 68 __be32 dp_credit; /* non-zero enables flow ctl */ 69 }; 70 71 struct rds_ib_send_work { 72 void *s_op; 73 union { 74 struct ib_send_wr s_wr; 75 struct ib_rdma_wr s_rdma_wr; 76 struct ib_atomic_wr s_atomic_wr; 77 }; 78 struct ib_sge s_sge[RDS_IB_MAX_SGE]; 79 unsigned long s_queued; 80 }; 81 82 struct rds_ib_recv_work { 83 struct rds_ib_incoming *r_ibinc; 84 struct rds_page_frag *r_frag; 85 struct ib_recv_wr r_wr; 86 struct ib_sge r_sge[2]; 87 }; 88 89 struct rds_ib_work_ring { 90 u32 w_nr; 91 u32 w_alloc_ptr; 92 u32 w_alloc_ctr; 93 u32 w_free_ptr; 94 atomic_t w_free_ctr; 95 }; 96 97 /* Rings are posted with all the allocations they'll need to queue the 98 * incoming message to the receiving socket so this can't fail. 99 * All fragments start with a header, so we can make sure we're not receiving 100 * garbage, and we can tell a small 8 byte fragment from an ACK frame. 101 */ 102 struct rds_ib_ack_state { 103 u64 ack_next; 104 u64 ack_recv; 105 unsigned int ack_required:1; 106 unsigned int ack_next_valid:1; 107 unsigned int ack_recv_valid:1; 108 }; 109 110 111 struct rds_ib_device; 112 113 struct rds_ib_connection { 114 115 struct list_head ib_node; 116 struct rds_ib_device *rds_ibdev; 117 struct rds_connection *conn; 118 119 /* alphabet soup, IBTA style */ 120 struct rdma_cm_id *i_cm_id; 121 struct ib_pd *i_pd; 122 struct ib_cq *i_send_cq; 123 struct ib_cq *i_recv_cq; 124 struct ib_wc i_send_wc[RDS_IB_WC_MAX]; 125 struct ib_wc i_recv_wc[RDS_IB_WC_MAX]; 126 127 /* To control the number of wrs from fastreg */ 128 atomic_t i_fastreg_wrs; 129 atomic_t i_fastunreg_wrs; 130 131 /* interrupt handling */ 132 struct tasklet_struct i_send_tasklet; 133 struct tasklet_struct i_recv_tasklet; 134 135 /* tx */ 136 struct rds_ib_work_ring i_send_ring; 137 struct rm_data_op *i_data_op; 138 struct rds_header *i_send_hdrs; 139 dma_addr_t i_send_hdrs_dma; 140 struct rds_ib_send_work *i_sends; 141 atomic_t i_signaled_sends; 142 143 /* rx */ 144 struct mutex i_recv_mutex; 145 struct rds_ib_work_ring i_recv_ring; 146 struct rds_ib_incoming *i_ibinc; 147 u32 i_recv_data_rem; 148 struct rds_header *i_recv_hdrs; 149 dma_addr_t i_recv_hdrs_dma; 150 struct rds_ib_recv_work *i_recvs; 151 u64 i_ack_recv; /* last ACK received */ 152 struct rds_ib_refill_cache i_cache_incs; 153 struct rds_ib_refill_cache i_cache_frags; 154 atomic_t i_cache_allocs; 155 156 /* sending acks */ 157 unsigned long i_ack_flags; 158 #ifdef KERNEL_HAS_ATOMIC64 159 atomic64_t i_ack_next; /* next ACK to send */ 160 #else 161 spinlock_t i_ack_lock; /* protect i_ack_next */ 162 u64 i_ack_next; /* next ACK to send */ 163 #endif 164 struct rds_header *i_ack; 165 struct ib_send_wr i_ack_wr; 166 struct ib_sge i_ack_sge; 167 dma_addr_t i_ack_dma; 168 unsigned long i_ack_queued; 169 170 /* Flow control related information 171 * 172 * Our algorithm uses a pair variables that we need to access 173 * atomically - one for the send credits, and one posted 174 * recv credits we need to transfer to remote. 175 * Rather than protect them using a slow spinlock, we put both into 176 * a single atomic_t and update it using cmpxchg 177 */ 178 atomic_t i_credits; 179 180 /* Protocol version specific information */ 181 unsigned int i_flowctl:1; /* enable/disable flow ctl */ 182 183 /* Batched completions */ 184 unsigned int i_unsignaled_wrs; 185 186 /* Endpoint role in connection */ 187 bool i_active_side; 188 atomic_t i_cq_quiesce; 189 190 /* Send/Recv vectors */ 191 int i_scq_vector; 192 int i_rcq_vector; 193 }; 194 195 /* This assumes that atomic_t is at least 32 bits */ 196 #define IB_GET_SEND_CREDITS(v) ((v) & 0xffff) 197 #define IB_GET_POST_CREDITS(v) ((v) >> 16) 198 #define IB_SET_SEND_CREDITS(v) ((v) & 0xffff) 199 #define IB_SET_POST_CREDITS(v) ((v) << 16) 200 201 struct rds_ib_ipaddr { 202 struct list_head list; 203 __be32 ipaddr; 204 struct rcu_head rcu; 205 }; 206 207 enum { 208 RDS_IB_MR_8K_POOL, 209 RDS_IB_MR_1M_POOL, 210 }; 211 212 struct rds_ib_device { 213 struct list_head list; 214 struct list_head ipaddr_list; 215 struct list_head conn_list; 216 struct ib_device *dev; 217 struct ib_pd *pd; 218 bool has_fmr; 219 bool has_fr; 220 bool use_fastreg; 221 222 unsigned int max_mrs; 223 struct rds_ib_mr_pool *mr_1m_pool; 224 struct rds_ib_mr_pool *mr_8k_pool; 225 unsigned int fmr_max_remaps; 226 unsigned int max_8k_mrs; 227 unsigned int max_1m_mrs; 228 int max_sge; 229 unsigned int max_wrs; 230 unsigned int max_initiator_depth; 231 unsigned int max_responder_resources; 232 spinlock_t spinlock; /* protect the above */ 233 atomic_t refcount; 234 struct work_struct free_work; 235 int *vector_load; 236 }; 237 238 #define ibdev_to_node(ibdev) dev_to_node((ibdev)->dev.parent) 239 #define rdsibdev_to_node(rdsibdev) ibdev_to_node(rdsibdev->dev) 240 241 /* bits for i_ack_flags */ 242 #define IB_ACK_IN_FLIGHT 0 243 #define IB_ACK_REQUESTED 1 244 245 /* Magic WR_ID for ACKs */ 246 #define RDS_IB_ACK_WR_ID (~(u64) 0) 247 248 struct rds_ib_statistics { 249 uint64_t s_ib_connect_raced; 250 uint64_t s_ib_listen_closed_stale; 251 uint64_t s_ib_evt_handler_call; 252 uint64_t s_ib_tasklet_call; 253 uint64_t s_ib_tx_cq_event; 254 uint64_t s_ib_tx_ring_full; 255 uint64_t s_ib_tx_throttle; 256 uint64_t s_ib_tx_sg_mapping_failure; 257 uint64_t s_ib_tx_stalled; 258 uint64_t s_ib_tx_credit_updates; 259 uint64_t s_ib_rx_cq_event; 260 uint64_t s_ib_rx_ring_empty; 261 uint64_t s_ib_rx_refill_from_cq; 262 uint64_t s_ib_rx_refill_from_thread; 263 uint64_t s_ib_rx_alloc_limit; 264 uint64_t s_ib_rx_total_frags; 265 uint64_t s_ib_rx_total_incs; 266 uint64_t s_ib_rx_credit_updates; 267 uint64_t s_ib_ack_sent; 268 uint64_t s_ib_ack_send_failure; 269 uint64_t s_ib_ack_send_delayed; 270 uint64_t s_ib_ack_send_piggybacked; 271 uint64_t s_ib_ack_received; 272 uint64_t s_ib_rdma_mr_8k_alloc; 273 uint64_t s_ib_rdma_mr_8k_free; 274 uint64_t s_ib_rdma_mr_8k_used; 275 uint64_t s_ib_rdma_mr_8k_pool_flush; 276 uint64_t s_ib_rdma_mr_8k_pool_wait; 277 uint64_t s_ib_rdma_mr_8k_pool_depleted; 278 uint64_t s_ib_rdma_mr_1m_alloc; 279 uint64_t s_ib_rdma_mr_1m_free; 280 uint64_t s_ib_rdma_mr_1m_used; 281 uint64_t s_ib_rdma_mr_1m_pool_flush; 282 uint64_t s_ib_rdma_mr_1m_pool_wait; 283 uint64_t s_ib_rdma_mr_1m_pool_depleted; 284 uint64_t s_ib_rdma_mr_8k_reused; 285 uint64_t s_ib_rdma_mr_1m_reused; 286 uint64_t s_ib_atomic_cswp; 287 uint64_t s_ib_atomic_fadd; 288 uint64_t s_ib_recv_added_to_cache; 289 uint64_t s_ib_recv_removed_from_cache; 290 }; 291 292 extern struct workqueue_struct *rds_ib_wq; 293 294 /* 295 * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h 296 * doesn't define it. 297 */ 298 static inline void rds_ib_dma_sync_sg_for_cpu(struct ib_device *dev, 299 struct scatterlist *sglist, 300 unsigned int sg_dma_len, 301 int direction) 302 { 303 struct scatterlist *sg; 304 unsigned int i; 305 306 for_each_sg(sglist, sg, sg_dma_len, i) { 307 ib_dma_sync_single_for_cpu(dev, 308 ib_sg_dma_address(dev, sg), 309 ib_sg_dma_len(dev, sg), 310 direction); 311 } 312 } 313 #define ib_dma_sync_sg_for_cpu rds_ib_dma_sync_sg_for_cpu 314 315 static inline void rds_ib_dma_sync_sg_for_device(struct ib_device *dev, 316 struct scatterlist *sglist, 317 unsigned int sg_dma_len, 318 int direction) 319 { 320 struct scatterlist *sg; 321 unsigned int i; 322 323 for_each_sg(sglist, sg, sg_dma_len, i) { 324 ib_dma_sync_single_for_device(dev, 325 ib_sg_dma_address(dev, sg), 326 ib_sg_dma_len(dev, sg), 327 direction); 328 } 329 } 330 #define ib_dma_sync_sg_for_device rds_ib_dma_sync_sg_for_device 331 332 333 /* ib.c */ 334 extern struct rds_transport rds_ib_transport; 335 struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device); 336 void rds_ib_dev_put(struct rds_ib_device *rds_ibdev); 337 extern struct ib_client rds_ib_client; 338 339 extern unsigned int rds_ib_retry_count; 340 341 extern spinlock_t ib_nodev_conns_lock; 342 extern struct list_head ib_nodev_conns; 343 344 /* ib_cm.c */ 345 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp); 346 void rds_ib_conn_free(void *arg); 347 int rds_ib_conn_path_connect(struct rds_conn_path *cp); 348 void rds_ib_conn_path_shutdown(struct rds_conn_path *cp); 349 void rds_ib_state_change(struct sock *sk); 350 int rds_ib_listen_init(void); 351 void rds_ib_listen_stop(void); 352 __printf(2, 3) 353 void __rds_ib_conn_error(struct rds_connection *conn, const char *, ...); 354 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id, 355 struct rdma_cm_event *event); 356 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id); 357 void rds_ib_cm_connect_complete(struct rds_connection *conn, 358 struct rdma_cm_event *event); 359 360 361 #define rds_ib_conn_error(conn, fmt...) \ 362 __rds_ib_conn_error(conn, KERN_WARNING "RDS/IB: " fmt) 363 364 /* ib_rdma.c */ 365 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr); 366 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn); 367 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn); 368 void rds_ib_destroy_nodev_conns(void); 369 void rds_ib_mr_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc); 370 371 /* ib_recv.c */ 372 int rds_ib_recv_init(void); 373 void rds_ib_recv_exit(void); 374 int rds_ib_recv_path(struct rds_conn_path *conn); 375 int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic); 376 void rds_ib_recv_free_caches(struct rds_ib_connection *ic); 377 void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp); 378 void rds_ib_inc_free(struct rds_incoming *inc); 379 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to); 380 void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc, 381 struct rds_ib_ack_state *state); 382 void rds_ib_recv_tasklet_fn(unsigned long data); 383 void rds_ib_recv_init_ring(struct rds_ib_connection *ic); 384 void rds_ib_recv_clear_ring(struct rds_ib_connection *ic); 385 void rds_ib_recv_init_ack(struct rds_ib_connection *ic); 386 void rds_ib_attempt_ack(struct rds_ib_connection *ic); 387 void rds_ib_ack_send_complete(struct rds_ib_connection *ic); 388 u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic); 389 void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required); 390 391 /* ib_ring.c */ 392 void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr); 393 void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr); 394 u32 rds_ib_ring_alloc(struct rds_ib_work_ring *ring, u32 val, u32 *pos); 395 void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val); 396 void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val); 397 int rds_ib_ring_empty(struct rds_ib_work_ring *ring); 398 int rds_ib_ring_low(struct rds_ib_work_ring *ring); 399 u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring); 400 u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest); 401 extern wait_queue_head_t rds_ib_ring_empty_wait; 402 403 /* ib_send.c */ 404 void rds_ib_xmit_path_complete(struct rds_conn_path *cp); 405 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm, 406 unsigned int hdr_off, unsigned int sg, unsigned int off); 407 void rds_ib_send_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc); 408 void rds_ib_send_init_ring(struct rds_ib_connection *ic); 409 void rds_ib_send_clear_ring(struct rds_ib_connection *ic); 410 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op); 411 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits); 412 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted); 413 int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted, 414 u32 *adv_credits, int need_posted, int max_posted); 415 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op); 416 417 /* ib_stats.c */ 418 DECLARE_PER_CPU(struct rds_ib_statistics, rds_ib_stats); 419 #define rds_ib_stats_inc(member) rds_stats_inc_which(rds_ib_stats, member) 420 #define rds_ib_stats_add(member, count) \ 421 rds_stats_add_which(rds_ib_stats, member, count) 422 unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter, 423 unsigned int avail); 424 425 /* ib_sysctl.c */ 426 int rds_ib_sysctl_init(void); 427 void rds_ib_sysctl_exit(void); 428 extern unsigned long rds_ib_sysctl_max_send_wr; 429 extern unsigned long rds_ib_sysctl_max_recv_wr; 430 extern unsigned long rds_ib_sysctl_max_unsig_wrs; 431 extern unsigned long rds_ib_sysctl_max_unsig_bytes; 432 extern unsigned long rds_ib_sysctl_max_recv_allocation; 433 extern unsigned int rds_ib_sysctl_flow_control; 434 435 #endif 436