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_FMR_SIZE 256 13 #define RDS_FMR_POOL_SIZE 8192 14 15 #define RDS_IB_MAX_SGE 8 16 #define RDS_IB_RECV_SGE 2 17 18 #define RDS_IB_DEFAULT_RECV_WR 1024 19 #define RDS_IB_DEFAULT_SEND_WR 256 20 21 #define RDS_IB_DEFAULT_RETRY_COUNT 2 22 23 #define RDS_IB_SUPPORTED_PROTOCOLS 0x00000003 /* minor versions supported */ 24 25 #define RDS_IB_RECYCLE_BATCH_COUNT 32 26 27 extern struct rw_semaphore rds_ib_devices_lock; 28 extern struct list_head rds_ib_devices; 29 30 /* 31 * IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to 32 * try and minimize the amount of memory tied up both the device and 33 * socket receive queues. 34 */ 35 struct rds_page_frag { 36 struct list_head f_item; 37 struct list_head f_cache_entry; 38 struct scatterlist f_sg; 39 }; 40 41 struct rds_ib_incoming { 42 struct list_head ii_frags; 43 struct list_head ii_cache_entry; 44 struct rds_incoming ii_inc; 45 }; 46 47 struct rds_ib_cache_head { 48 struct list_head *first; 49 unsigned long count; 50 }; 51 52 struct rds_ib_refill_cache { 53 struct rds_ib_cache_head __percpu *percpu; 54 struct list_head *xfer; 55 struct list_head *ready; 56 }; 57 58 struct rds_ib_connect_private { 59 /* Add new fields at the end, and don't permute existing fields. */ 60 __be32 dp_saddr; 61 __be32 dp_daddr; 62 u8 dp_protocol_major; 63 u8 dp_protocol_minor; 64 __be16 dp_protocol_minor_mask; /* bitmask */ 65 __be32 dp_reserved1; 66 __be64 dp_ack_seq; 67 __be32 dp_credit; /* non-zero enables flow ctl */ 68 }; 69 70 struct rds_ib_send_work { 71 void *s_op; 72 struct ib_send_wr s_wr; 73 struct ib_sge s_sge[RDS_IB_MAX_SGE]; 74 unsigned long s_queued; 75 }; 76 77 struct rds_ib_recv_work { 78 struct rds_ib_incoming *r_ibinc; 79 struct rds_page_frag *r_frag; 80 struct ib_recv_wr r_wr; 81 struct ib_sge r_sge[2]; 82 }; 83 84 struct rds_ib_work_ring { 85 u32 w_nr; 86 u32 w_alloc_ptr; 87 u32 w_alloc_ctr; 88 u32 w_free_ptr; 89 atomic_t w_free_ctr; 90 }; 91 92 struct rds_ib_device; 93 94 struct rds_ib_connection { 95 96 struct list_head ib_node; 97 struct rds_ib_device *rds_ibdev; 98 struct rds_connection *conn; 99 100 /* alphabet soup, IBTA style */ 101 struct rdma_cm_id *i_cm_id; 102 struct ib_pd *i_pd; 103 struct ib_cq *i_send_cq; 104 struct ib_cq *i_recv_cq; 105 106 /* tx */ 107 struct rds_ib_work_ring i_send_ring; 108 struct rm_data_op *i_data_op; 109 struct rds_header *i_send_hdrs; 110 u64 i_send_hdrs_dma; 111 struct rds_ib_send_work *i_sends; 112 atomic_t i_signaled_sends; 113 114 /* rx */ 115 struct tasklet_struct i_recv_tasklet; 116 struct mutex i_recv_mutex; 117 struct rds_ib_work_ring i_recv_ring; 118 struct rds_ib_incoming *i_ibinc; 119 u32 i_recv_data_rem; 120 struct rds_header *i_recv_hdrs; 121 u64 i_recv_hdrs_dma; 122 struct rds_ib_recv_work *i_recvs; 123 u64 i_ack_recv; /* last ACK received */ 124 struct rds_ib_refill_cache i_cache_incs; 125 struct rds_ib_refill_cache i_cache_frags; 126 127 /* sending acks */ 128 unsigned long i_ack_flags; 129 #ifdef KERNEL_HAS_ATOMIC64 130 atomic64_t i_ack_next; /* next ACK to send */ 131 #else 132 spinlock_t i_ack_lock; /* protect i_ack_next */ 133 u64 i_ack_next; /* next ACK to send */ 134 #endif 135 struct rds_header *i_ack; 136 struct ib_send_wr i_ack_wr; 137 struct ib_sge i_ack_sge; 138 u64 i_ack_dma; 139 unsigned long i_ack_queued; 140 141 /* Flow control related information 142 * 143 * Our algorithm uses a pair variables that we need to access 144 * atomically - one for the send credits, and one posted 145 * recv credits we need to transfer to remote. 146 * Rather than protect them using a slow spinlock, we put both into 147 * a single atomic_t and update it using cmpxchg 148 */ 149 atomic_t i_credits; 150 151 /* Protocol version specific information */ 152 unsigned int i_flowctl:1; /* enable/disable flow ctl */ 153 154 /* Batched completions */ 155 unsigned int i_unsignaled_wrs; 156 }; 157 158 /* This assumes that atomic_t is at least 32 bits */ 159 #define IB_GET_SEND_CREDITS(v) ((v) & 0xffff) 160 #define IB_GET_POST_CREDITS(v) ((v) >> 16) 161 #define IB_SET_SEND_CREDITS(v) ((v) & 0xffff) 162 #define IB_SET_POST_CREDITS(v) ((v) << 16) 163 164 struct rds_ib_ipaddr { 165 struct list_head list; 166 __be32 ipaddr; 167 }; 168 169 struct rds_ib_device { 170 struct list_head list; 171 struct list_head ipaddr_list; 172 struct list_head conn_list; 173 struct ib_device *dev; 174 struct ib_pd *pd; 175 struct rds_ib_mr_pool *mr_pool; 176 unsigned int fmr_max_remaps; 177 unsigned int max_fmrs; 178 int max_sge; 179 unsigned int max_wrs; 180 unsigned int max_initiator_depth; 181 unsigned int max_responder_resources; 182 spinlock_t spinlock; /* protect the above */ 183 atomic_t refcount; 184 struct work_struct free_work; 185 }; 186 187 #define ibdev_to_node(ibdev) dev_to_node(ibdev->dma_device) 188 #define rdsibdev_to_node(rdsibdev) ibdev_to_node(rdsibdev->dev) 189 190 /* bits for i_ack_flags */ 191 #define IB_ACK_IN_FLIGHT 0 192 #define IB_ACK_REQUESTED 1 193 194 /* Magic WR_ID for ACKs */ 195 #define RDS_IB_ACK_WR_ID (~(u64) 0) 196 197 struct rds_ib_statistics { 198 uint64_t s_ib_connect_raced; 199 uint64_t s_ib_listen_closed_stale; 200 uint64_t s_ib_tx_cq_call; 201 uint64_t s_ib_tx_cq_event; 202 uint64_t s_ib_tx_ring_full; 203 uint64_t s_ib_tx_throttle; 204 uint64_t s_ib_tx_sg_mapping_failure; 205 uint64_t s_ib_tx_stalled; 206 uint64_t s_ib_tx_credit_updates; 207 uint64_t s_ib_rx_cq_call; 208 uint64_t s_ib_rx_cq_event; 209 uint64_t s_ib_rx_ring_empty; 210 uint64_t s_ib_rx_refill_from_cq; 211 uint64_t s_ib_rx_refill_from_thread; 212 uint64_t s_ib_rx_alloc_limit; 213 uint64_t s_ib_rx_credit_updates; 214 uint64_t s_ib_ack_sent; 215 uint64_t s_ib_ack_send_failure; 216 uint64_t s_ib_ack_send_delayed; 217 uint64_t s_ib_ack_send_piggybacked; 218 uint64_t s_ib_ack_received; 219 uint64_t s_ib_rdma_mr_alloc; 220 uint64_t s_ib_rdma_mr_free; 221 uint64_t s_ib_rdma_mr_used; 222 uint64_t s_ib_rdma_mr_pool_flush; 223 uint64_t s_ib_rdma_mr_pool_wait; 224 uint64_t s_ib_rdma_mr_pool_depleted; 225 uint64_t s_ib_atomic_cswp; 226 uint64_t s_ib_atomic_fadd; 227 }; 228 229 extern struct workqueue_struct *rds_ib_wq; 230 231 /* 232 * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h 233 * doesn't define it. 234 */ 235 static inline void rds_ib_dma_sync_sg_for_cpu(struct ib_device *dev, 236 struct scatterlist *sglist, 237 unsigned int sg_dma_len, 238 int direction) 239 { 240 struct scatterlist *sg; 241 unsigned int i; 242 243 for_each_sg(sglist, sg, sg_dma_len, i) { 244 ib_dma_sync_single_for_cpu(dev, 245 ib_sg_dma_address(dev, sg), 246 ib_sg_dma_len(dev, sg), 247 direction); 248 } 249 } 250 #define ib_dma_sync_sg_for_cpu rds_ib_dma_sync_sg_for_cpu 251 252 static inline void rds_ib_dma_sync_sg_for_device(struct ib_device *dev, 253 struct scatterlist *sglist, 254 unsigned int sg_dma_len, 255 int direction) 256 { 257 struct scatterlist *sg; 258 unsigned int i; 259 260 for_each_sg(sglist, sg, sg_dma_len, i) { 261 ib_dma_sync_single_for_device(dev, 262 ib_sg_dma_address(dev, sg), 263 ib_sg_dma_len(dev, sg), 264 direction); 265 } 266 } 267 #define ib_dma_sync_sg_for_device rds_ib_dma_sync_sg_for_device 268 269 270 /* ib.c */ 271 extern struct rds_transport rds_ib_transport; 272 struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device); 273 void rds_ib_dev_put(struct rds_ib_device *rds_ibdev); 274 extern struct ib_client rds_ib_client; 275 276 extern unsigned int fmr_message_size; 277 extern unsigned int rds_ib_retry_count; 278 279 extern spinlock_t ib_nodev_conns_lock; 280 extern struct list_head ib_nodev_conns; 281 282 /* ib_cm.c */ 283 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp); 284 void rds_ib_conn_free(void *arg); 285 int rds_ib_conn_connect(struct rds_connection *conn); 286 void rds_ib_conn_shutdown(struct rds_connection *conn); 287 void rds_ib_state_change(struct sock *sk); 288 int rds_ib_listen_init(void); 289 void rds_ib_listen_stop(void); 290 void __rds_ib_conn_error(struct rds_connection *conn, const char *, ...); 291 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id, 292 struct rdma_cm_event *event); 293 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id); 294 void rds_ib_cm_connect_complete(struct rds_connection *conn, 295 struct rdma_cm_event *event); 296 297 298 #define rds_ib_conn_error(conn, fmt...) \ 299 __rds_ib_conn_error(conn, KERN_WARNING "RDS/IB: " fmt) 300 301 /* ib_rdma.c */ 302 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr); 303 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn); 304 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn); 305 void rds_ib_destroy_nodev_conns(void); 306 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *); 307 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo); 308 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *); 309 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents, 310 struct rds_sock *rs, u32 *key_ret); 311 void rds_ib_sync_mr(void *trans_private, int dir); 312 void rds_ib_free_mr(void *trans_private, int invalidate); 313 void rds_ib_flush_mrs(void); 314 int rds_ib_fmr_init(void); 315 void rds_ib_fmr_exit(void); 316 317 /* ib_recv.c */ 318 int rds_ib_recv_init(void); 319 void rds_ib_recv_exit(void); 320 int rds_ib_recv(struct rds_connection *conn); 321 int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic); 322 void rds_ib_recv_free_caches(struct rds_ib_connection *ic); 323 void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp); 324 void rds_ib_inc_free(struct rds_incoming *inc); 325 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to); 326 void rds_ib_recv_cq_comp_handler(struct ib_cq *cq, void *context); 327 void rds_ib_recv_tasklet_fn(unsigned long data); 328 void rds_ib_recv_init_ring(struct rds_ib_connection *ic); 329 void rds_ib_recv_clear_ring(struct rds_ib_connection *ic); 330 void rds_ib_recv_init_ack(struct rds_ib_connection *ic); 331 void rds_ib_attempt_ack(struct rds_ib_connection *ic); 332 void rds_ib_ack_send_complete(struct rds_ib_connection *ic); 333 u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic); 334 335 /* ib_ring.c */ 336 void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr); 337 void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr); 338 u32 rds_ib_ring_alloc(struct rds_ib_work_ring *ring, u32 val, u32 *pos); 339 void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val); 340 void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val); 341 int rds_ib_ring_empty(struct rds_ib_work_ring *ring); 342 int rds_ib_ring_low(struct rds_ib_work_ring *ring); 343 u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring); 344 u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest); 345 extern wait_queue_head_t rds_ib_ring_empty_wait; 346 347 /* ib_send.c */ 348 void rds_ib_xmit_complete(struct rds_connection *conn); 349 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm, 350 unsigned int hdr_off, unsigned int sg, unsigned int off); 351 void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context); 352 void rds_ib_send_init_ring(struct rds_ib_connection *ic); 353 void rds_ib_send_clear_ring(struct rds_ib_connection *ic); 354 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op); 355 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits); 356 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted); 357 int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted, 358 u32 *adv_credits, int need_posted, int max_posted); 359 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op); 360 361 /* ib_stats.c */ 362 DECLARE_PER_CPU(struct rds_ib_statistics, rds_ib_stats); 363 #define rds_ib_stats_inc(member) rds_stats_inc_which(rds_ib_stats, member) 364 unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter, 365 unsigned int avail); 366 367 /* ib_sysctl.c */ 368 int rds_ib_sysctl_init(void); 369 void rds_ib_sysctl_exit(void); 370 extern unsigned long rds_ib_sysctl_max_send_wr; 371 extern unsigned long rds_ib_sysctl_max_recv_wr; 372 extern unsigned long rds_ib_sysctl_max_unsig_wrs; 373 extern unsigned long rds_ib_sysctl_max_unsig_bytes; 374 extern unsigned long rds_ib_sysctl_max_recv_allocation; 375 extern unsigned int rds_ib_sysctl_flow_control; 376 377 #endif 378