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_1M_POOL_SIZE (8192 / 2) 13 #define RDS_FMR_1M_MSG_SIZE 256 14 #define RDS_FMR_8K_MSG_SIZE 2 15 #define RDS_MR_8K_SCALE (256 / (RDS_FMR_8K_MSG_SIZE + 1)) 16 #define RDS_FMR_8K_POOL_SIZE (RDS_MR_8K_SCALE * (8192 / 2)) 17 18 #define RDS_IB_MAX_SGE 8 19 #define RDS_IB_RECV_SGE 2 20 21 #define RDS_IB_DEFAULT_RECV_WR 1024 22 #define RDS_IB_DEFAULT_SEND_WR 256 23 24 #define RDS_IB_DEFAULT_RETRY_COUNT 2 25 26 #define RDS_IB_SUPPORTED_PROTOCOLS 0x00000003 /* minor versions supported */ 27 28 #define RDS_IB_RECYCLE_BATCH_COUNT 32 29 30 #define RDS_IB_WC_MAX 32 31 #define RDS_IB_SEND_OP BIT_ULL(63) 32 33 extern struct rw_semaphore rds_ib_devices_lock; 34 extern struct list_head rds_ib_devices; 35 36 /* 37 * IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to 38 * try and minimize the amount of memory tied up both the device and 39 * socket receive queues. 40 */ 41 struct rds_page_frag { 42 struct list_head f_item; 43 struct list_head f_cache_entry; 44 struct scatterlist f_sg; 45 }; 46 47 struct rds_ib_incoming { 48 struct list_head ii_frags; 49 struct list_head ii_cache_entry; 50 struct rds_incoming ii_inc; 51 }; 52 53 struct rds_ib_cache_head { 54 struct list_head *first; 55 unsigned long count; 56 }; 57 58 struct rds_ib_refill_cache { 59 struct rds_ib_cache_head __percpu *percpu; 60 struct list_head *xfer; 61 struct list_head *ready; 62 }; 63 64 struct rds_ib_connect_private { 65 /* Add new fields at the end, and don't permute existing fields. */ 66 __be32 dp_saddr; 67 __be32 dp_daddr; 68 u8 dp_protocol_major; 69 u8 dp_protocol_minor; 70 __be16 dp_protocol_minor_mask; /* bitmask */ 71 __be32 dp_reserved1; 72 __be64 dp_ack_seq; 73 __be32 dp_credit; /* non-zero enables flow ctl */ 74 }; 75 76 struct rds_ib_send_work { 77 void *s_op; 78 struct ib_send_wr s_wr; 79 struct ib_sge s_sge[RDS_IB_MAX_SGE]; 80 unsigned long s_queued; 81 }; 82 83 struct rds_ib_recv_work { 84 struct rds_ib_incoming *r_ibinc; 85 struct rds_page_frag *r_frag; 86 struct ib_recv_wr r_wr; 87 struct ib_sge r_sge[2]; 88 }; 89 90 struct rds_ib_work_ring { 91 u32 w_nr; 92 u32 w_alloc_ptr; 93 u32 w_alloc_ctr; 94 u32 w_free_ptr; 95 atomic_t w_free_ctr; 96 }; 97 98 /* Rings are posted with all the allocations they'll need to queue the 99 * incoming message to the receiving socket so this can't fail. 100 * All fragments start with a header, so we can make sure we're not receiving 101 * garbage, and we can tell a small 8 byte fragment from an ACK frame. 102 */ 103 struct rds_ib_ack_state { 104 u64 ack_next; 105 u64 ack_recv; 106 unsigned int ack_required:1; 107 unsigned int ack_next_valid:1; 108 unsigned int ack_recv_valid:1; 109 }; 110 111 112 struct rds_ib_device; 113 114 struct rds_ib_connection { 115 116 struct list_head ib_node; 117 struct rds_ib_device *rds_ibdev; 118 struct rds_connection *conn; 119 120 /* alphabet soup, IBTA style */ 121 struct rdma_cm_id *i_cm_id; 122 struct ib_pd *i_pd; 123 struct ib_cq *i_send_cq; 124 struct ib_cq *i_recv_cq; 125 struct ib_wc i_send_wc[RDS_IB_WC_MAX]; 126 struct ib_wc i_recv_wc[RDS_IB_WC_MAX]; 127 128 /* interrupt handling */ 129 struct tasklet_struct i_send_tasklet; 130 struct tasklet_struct i_recv_tasklet; 131 132 /* tx */ 133 struct rds_ib_work_ring i_send_ring; 134 struct rm_data_op *i_data_op; 135 struct rds_header *i_send_hdrs; 136 u64 i_send_hdrs_dma; 137 struct rds_ib_send_work *i_sends; 138 atomic_t i_signaled_sends; 139 140 /* rx */ 141 struct mutex i_recv_mutex; 142 struct rds_ib_work_ring i_recv_ring; 143 struct rds_ib_incoming *i_ibinc; 144 u32 i_recv_data_rem; 145 struct rds_header *i_recv_hdrs; 146 u64 i_recv_hdrs_dma; 147 struct rds_ib_recv_work *i_recvs; 148 u64 i_ack_recv; /* last ACK received */ 149 struct rds_ib_refill_cache i_cache_incs; 150 struct rds_ib_refill_cache i_cache_frags; 151 152 /* sending acks */ 153 unsigned long i_ack_flags; 154 #ifdef KERNEL_HAS_ATOMIC64 155 atomic64_t i_ack_next; /* next ACK to send */ 156 #else 157 spinlock_t i_ack_lock; /* protect i_ack_next */ 158 u64 i_ack_next; /* next ACK to send */ 159 #endif 160 struct rds_header *i_ack; 161 struct ib_send_wr i_ack_wr; 162 struct ib_sge i_ack_sge; 163 u64 i_ack_dma; 164 unsigned long i_ack_queued; 165 166 /* Flow control related information 167 * 168 * Our algorithm uses a pair variables that we need to access 169 * atomically - one for the send credits, and one posted 170 * recv credits we need to transfer to remote. 171 * Rather than protect them using a slow spinlock, we put both into 172 * a single atomic_t and update it using cmpxchg 173 */ 174 atomic_t i_credits; 175 176 /* Protocol version specific information */ 177 unsigned int i_flowctl:1; /* enable/disable flow ctl */ 178 179 /* Batched completions */ 180 unsigned int i_unsignaled_wrs; 181 }; 182 183 /* This assumes that atomic_t is at least 32 bits */ 184 #define IB_GET_SEND_CREDITS(v) ((v) & 0xffff) 185 #define IB_GET_POST_CREDITS(v) ((v) >> 16) 186 #define IB_SET_SEND_CREDITS(v) ((v) & 0xffff) 187 #define IB_SET_POST_CREDITS(v) ((v) << 16) 188 189 struct rds_ib_ipaddr { 190 struct list_head list; 191 __be32 ipaddr; 192 struct rcu_head rcu; 193 }; 194 195 enum { 196 RDS_IB_MR_8K_POOL, 197 RDS_IB_MR_1M_POOL, 198 }; 199 200 struct rds_ib_device { 201 struct list_head list; 202 struct list_head ipaddr_list; 203 struct list_head conn_list; 204 struct ib_device *dev; 205 struct ib_pd *pd; 206 unsigned int max_fmrs; 207 struct rds_ib_mr_pool *mr_1m_pool; 208 struct rds_ib_mr_pool *mr_8k_pool; 209 unsigned int fmr_max_remaps; 210 unsigned int max_8k_fmrs; 211 unsigned int max_1m_fmrs; 212 int max_sge; 213 unsigned int max_wrs; 214 unsigned int max_initiator_depth; 215 unsigned int max_responder_resources; 216 spinlock_t spinlock; /* protect the above */ 217 atomic_t refcount; 218 struct work_struct free_work; 219 }; 220 221 #define ibdev_to_node(ibdev) dev_to_node(ibdev->dma_device) 222 #define rdsibdev_to_node(rdsibdev) ibdev_to_node(rdsibdev->dev) 223 224 /* bits for i_ack_flags */ 225 #define IB_ACK_IN_FLIGHT 0 226 #define IB_ACK_REQUESTED 1 227 228 /* Magic WR_ID for ACKs */ 229 #define RDS_IB_ACK_WR_ID (~(u64) 0) 230 231 struct rds_ib_statistics { 232 uint64_t s_ib_connect_raced; 233 uint64_t s_ib_listen_closed_stale; 234 uint64_t s_ib_evt_handler_call; 235 uint64_t s_ib_tasklet_call; 236 uint64_t s_ib_tx_cq_event; 237 uint64_t s_ib_tx_ring_full; 238 uint64_t s_ib_tx_throttle; 239 uint64_t s_ib_tx_sg_mapping_failure; 240 uint64_t s_ib_tx_stalled; 241 uint64_t s_ib_tx_credit_updates; 242 uint64_t s_ib_rx_cq_event; 243 uint64_t s_ib_rx_ring_empty; 244 uint64_t s_ib_rx_refill_from_cq; 245 uint64_t s_ib_rx_refill_from_thread; 246 uint64_t s_ib_rx_alloc_limit; 247 uint64_t s_ib_rx_credit_updates; 248 uint64_t s_ib_ack_sent; 249 uint64_t s_ib_ack_send_failure; 250 uint64_t s_ib_ack_send_delayed; 251 uint64_t s_ib_ack_send_piggybacked; 252 uint64_t s_ib_ack_received; 253 uint64_t s_ib_rdma_mr_8k_alloc; 254 uint64_t s_ib_rdma_mr_8k_free; 255 uint64_t s_ib_rdma_mr_8k_used; 256 uint64_t s_ib_rdma_mr_8k_pool_flush; 257 uint64_t s_ib_rdma_mr_8k_pool_wait; 258 uint64_t s_ib_rdma_mr_8k_pool_depleted; 259 uint64_t s_ib_rdma_mr_1m_alloc; 260 uint64_t s_ib_rdma_mr_1m_free; 261 uint64_t s_ib_rdma_mr_1m_used; 262 uint64_t s_ib_rdma_mr_1m_pool_flush; 263 uint64_t s_ib_rdma_mr_1m_pool_wait; 264 uint64_t s_ib_rdma_mr_1m_pool_depleted; 265 uint64_t s_ib_atomic_cswp; 266 uint64_t s_ib_atomic_fadd; 267 }; 268 269 extern struct workqueue_struct *rds_ib_wq; 270 271 /* 272 * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h 273 * doesn't define it. 274 */ 275 static inline void rds_ib_dma_sync_sg_for_cpu(struct ib_device *dev, 276 struct scatterlist *sglist, 277 unsigned int sg_dma_len, 278 int direction) 279 { 280 struct scatterlist *sg; 281 unsigned int i; 282 283 for_each_sg(sglist, sg, sg_dma_len, i) { 284 ib_dma_sync_single_for_cpu(dev, 285 ib_sg_dma_address(dev, sg), 286 ib_sg_dma_len(dev, sg), 287 direction); 288 } 289 } 290 #define ib_dma_sync_sg_for_cpu rds_ib_dma_sync_sg_for_cpu 291 292 static inline void rds_ib_dma_sync_sg_for_device(struct ib_device *dev, 293 struct scatterlist *sglist, 294 unsigned int sg_dma_len, 295 int direction) 296 { 297 struct scatterlist *sg; 298 unsigned int i; 299 300 for_each_sg(sglist, sg, sg_dma_len, i) { 301 ib_dma_sync_single_for_device(dev, 302 ib_sg_dma_address(dev, sg), 303 ib_sg_dma_len(dev, sg), 304 direction); 305 } 306 } 307 #define ib_dma_sync_sg_for_device rds_ib_dma_sync_sg_for_device 308 309 310 /* ib.c */ 311 extern struct rds_transport rds_ib_transport; 312 struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device); 313 void rds_ib_dev_put(struct rds_ib_device *rds_ibdev); 314 extern struct ib_client rds_ib_client; 315 316 extern unsigned int rds_ib_fmr_1m_pool_size; 317 extern unsigned int rds_ib_fmr_8k_pool_size; 318 extern unsigned int rds_ib_retry_count; 319 320 extern spinlock_t ib_nodev_conns_lock; 321 extern struct list_head ib_nodev_conns; 322 323 /* ib_cm.c */ 324 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp); 325 void rds_ib_conn_free(void *arg); 326 int rds_ib_conn_connect(struct rds_connection *conn); 327 void rds_ib_conn_shutdown(struct rds_connection *conn); 328 void rds_ib_state_change(struct sock *sk); 329 int rds_ib_listen_init(void); 330 void rds_ib_listen_stop(void); 331 void __rds_ib_conn_error(struct rds_connection *conn, const char *, ...); 332 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id, 333 struct rdma_cm_event *event); 334 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id); 335 void rds_ib_cm_connect_complete(struct rds_connection *conn, 336 struct rdma_cm_event *event); 337 338 339 #define rds_ib_conn_error(conn, fmt...) \ 340 __rds_ib_conn_error(conn, KERN_WARNING "RDS/IB: " fmt) 341 342 /* ib_rdma.c */ 343 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr); 344 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn); 345 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn); 346 void rds_ib_destroy_nodev_conns(void); 347 struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_dev, 348 int npages); 349 void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo); 350 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *); 351 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents, 352 struct rds_sock *rs, u32 *key_ret); 353 void rds_ib_sync_mr(void *trans_private, int dir); 354 void rds_ib_free_mr(void *trans_private, int invalidate); 355 void rds_ib_flush_mrs(void); 356 int rds_ib_fmr_init(void); 357 void rds_ib_fmr_exit(void); 358 359 /* ib_recv.c */ 360 int rds_ib_recv_init(void); 361 void rds_ib_recv_exit(void); 362 int rds_ib_recv(struct rds_connection *conn); 363 int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic); 364 void rds_ib_recv_free_caches(struct rds_ib_connection *ic); 365 void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp); 366 void rds_ib_inc_free(struct rds_incoming *inc); 367 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to); 368 void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc, 369 struct rds_ib_ack_state *state); 370 void rds_ib_recv_tasklet_fn(unsigned long data); 371 void rds_ib_recv_init_ring(struct rds_ib_connection *ic); 372 void rds_ib_recv_clear_ring(struct rds_ib_connection *ic); 373 void rds_ib_recv_init_ack(struct rds_ib_connection *ic); 374 void rds_ib_attempt_ack(struct rds_ib_connection *ic); 375 void rds_ib_ack_send_complete(struct rds_ib_connection *ic); 376 u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic); 377 void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required); 378 379 /* ib_ring.c */ 380 void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr); 381 void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr); 382 u32 rds_ib_ring_alloc(struct rds_ib_work_ring *ring, u32 val, u32 *pos); 383 void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val); 384 void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val); 385 int rds_ib_ring_empty(struct rds_ib_work_ring *ring); 386 int rds_ib_ring_low(struct rds_ib_work_ring *ring); 387 u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring); 388 u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest); 389 extern wait_queue_head_t rds_ib_ring_empty_wait; 390 391 /* ib_send.c */ 392 void rds_ib_xmit_complete(struct rds_connection *conn); 393 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm, 394 unsigned int hdr_off, unsigned int sg, unsigned int off); 395 void rds_ib_send_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc); 396 void rds_ib_send_init_ring(struct rds_ib_connection *ic); 397 void rds_ib_send_clear_ring(struct rds_ib_connection *ic); 398 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op); 399 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits); 400 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted); 401 int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted, 402 u32 *adv_credits, int need_posted, int max_posted); 403 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op); 404 405 /* ib_stats.c */ 406 DECLARE_PER_CPU(struct rds_ib_statistics, rds_ib_stats); 407 #define rds_ib_stats_inc(member) rds_stats_inc_which(rds_ib_stats, member) 408 unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter, 409 unsigned int avail); 410 411 /* ib_sysctl.c */ 412 int rds_ib_sysctl_init(void); 413 void rds_ib_sysctl_exit(void); 414 extern unsigned long rds_ib_sysctl_max_send_wr; 415 extern unsigned long rds_ib_sysctl_max_recv_wr; 416 extern unsigned long rds_ib_sysctl_max_unsig_wrs; 417 extern unsigned long rds_ib_sysctl_max_unsig_bytes; 418 extern unsigned long rds_ib_sysctl_max_recv_allocation; 419 extern unsigned int rds_ib_sysctl_flow_control; 420 421 #endif 422