1 /* SPDX-License-Identifier: GPL-2.0 */ 2 #ifndef _RDS_IB_H 3 #define _RDS_IB_H 4 5 #include <rdma/ib_verbs.h> 6 #include <rdma/rdma_cm.h> 7 #include <linux/interrupt.h> 8 #include <linux/pci.h> 9 #include <linux/slab.h> 10 #include "rds.h" 11 #include "rdma_transport.h" 12 13 #define RDS_IB_MAX_SGE 8 14 #define RDS_IB_RECV_SGE 2 15 16 #define RDS_IB_DEFAULT_RECV_WR 1024 17 #define RDS_IB_DEFAULT_SEND_WR 256 18 #define RDS_IB_DEFAULT_FR_WR 256 19 #define RDS_IB_DEFAULT_FR_INV_WR 256 20 21 #define RDS_IB_DEFAULT_RETRY_COUNT 1 22 23 #define RDS_IB_SUPPORTED_PROTOCOLS 0x00000003 /* minor versions supported */ 24 25 #define RDS_IB_RECYCLE_BATCH_COUNT 32 26 27 #define RDS_IB_WC_MAX 32 28 29 extern struct rw_semaphore rds_ib_devices_lock; 30 extern struct list_head rds_ib_devices; 31 32 /* 33 * IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to 34 * try and minimize the amount of memory tied up both the device and 35 * socket receive queues. 36 */ 37 struct rds_page_frag { 38 struct list_head f_item; 39 struct list_head f_cache_entry; 40 struct scatterlist f_sg; 41 }; 42 43 struct rds_ib_incoming { 44 struct list_head ii_frags; 45 struct list_head ii_cache_entry; 46 struct rds_incoming ii_inc; 47 }; 48 49 struct rds_ib_cache_head { 50 struct list_head *first; 51 unsigned long count; 52 }; 53 54 struct rds_ib_refill_cache { 55 struct rds_ib_cache_head __percpu *percpu; 56 struct list_head *xfer; 57 struct list_head *ready; 58 }; 59 60 struct rds_ib_connect_private { 61 /* Add new fields at the end, and don't permute existing fields. */ 62 __be32 dp_saddr; 63 __be32 dp_daddr; 64 u8 dp_protocol_major; 65 u8 dp_protocol_minor; 66 __be16 dp_protocol_minor_mask; /* bitmask */ 67 __be32 dp_reserved1; 68 __be64 dp_ack_seq; 69 __be32 dp_credit; /* non-zero enables flow ctl */ 70 }; 71 72 struct rds_ib_send_work { 73 void *s_op; 74 union { 75 struct ib_send_wr s_wr; 76 struct ib_rdma_wr s_rdma_wr; 77 struct ib_atomic_wr s_atomic_wr; 78 }; 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 /* To control the number of wrs from fastreg */ 129 atomic_t i_fastreg_wrs; 130 atomic_t i_fastunreg_wrs; 131 132 /* interrupt handling */ 133 struct tasklet_struct i_send_tasklet; 134 struct tasklet_struct i_recv_tasklet; 135 136 /* tx */ 137 struct rds_ib_work_ring i_send_ring; 138 struct rm_data_op *i_data_op; 139 struct rds_header *i_send_hdrs; 140 dma_addr_t i_send_hdrs_dma; 141 struct rds_ib_send_work *i_sends; 142 atomic_t i_signaled_sends; 143 144 /* rx */ 145 struct mutex i_recv_mutex; 146 struct rds_ib_work_ring i_recv_ring; 147 struct rds_ib_incoming *i_ibinc; 148 u32 i_recv_data_rem; 149 struct rds_header *i_recv_hdrs; 150 dma_addr_t i_recv_hdrs_dma; 151 struct rds_ib_recv_work *i_recvs; 152 u64 i_ack_recv; /* last ACK received */ 153 struct rds_ib_refill_cache i_cache_incs; 154 struct rds_ib_refill_cache i_cache_frags; 155 atomic_t i_cache_allocs; 156 157 /* sending acks */ 158 unsigned long i_ack_flags; 159 #ifdef KERNEL_HAS_ATOMIC64 160 atomic64_t i_ack_next; /* next ACK to send */ 161 #else 162 spinlock_t i_ack_lock; /* protect i_ack_next */ 163 u64 i_ack_next; /* next ACK to send */ 164 #endif 165 struct rds_header *i_ack; 166 struct ib_send_wr i_ack_wr; 167 struct ib_sge i_ack_sge; 168 dma_addr_t i_ack_dma; 169 unsigned long i_ack_queued; 170 171 /* Flow control related information 172 * 173 * Our algorithm uses a pair variables that we need to access 174 * atomically - one for the send credits, and one posted 175 * recv credits we need to transfer to remote. 176 * Rather than protect them using a slow spinlock, we put both into 177 * a single atomic_t and update it using cmpxchg 178 */ 179 atomic_t i_credits; 180 181 /* Protocol version specific information */ 182 unsigned int i_flowctl:1; /* enable/disable flow ctl */ 183 184 /* Batched completions */ 185 unsigned int i_unsignaled_wrs; 186 187 /* Endpoint role in connection */ 188 bool i_active_side; 189 atomic_t i_cq_quiesce; 190 191 /* Send/Recv vectors */ 192 int i_scq_vector; 193 int i_rcq_vector; 194 }; 195 196 /* This assumes that atomic_t is at least 32 bits */ 197 #define IB_GET_SEND_CREDITS(v) ((v) & 0xffff) 198 #define IB_GET_POST_CREDITS(v) ((v) >> 16) 199 #define IB_SET_SEND_CREDITS(v) ((v) & 0xffff) 200 #define IB_SET_POST_CREDITS(v) ((v) << 16) 201 202 struct rds_ib_ipaddr { 203 struct list_head list; 204 __be32 ipaddr; 205 struct rcu_head rcu; 206 }; 207 208 enum { 209 RDS_IB_MR_8K_POOL, 210 RDS_IB_MR_1M_POOL, 211 }; 212 213 struct rds_ib_device { 214 struct list_head list; 215 struct list_head ipaddr_list; 216 struct list_head conn_list; 217 struct ib_device *dev; 218 struct ib_pd *pd; 219 bool use_fastreg; 220 221 unsigned int max_mrs; 222 struct rds_ib_mr_pool *mr_1m_pool; 223 struct rds_ib_mr_pool *mr_8k_pool; 224 unsigned int fmr_max_remaps; 225 unsigned int max_8k_mrs; 226 unsigned int max_1m_mrs; 227 int max_sge; 228 unsigned int max_wrs; 229 unsigned int max_initiator_depth; 230 unsigned int max_responder_resources; 231 spinlock_t spinlock; /* protect the above */ 232 refcount_t refcount; 233 struct work_struct free_work; 234 int *vector_load; 235 }; 236 237 #define ibdev_to_node(ibdev) dev_to_node((ibdev)->dev.parent) 238 #define rdsibdev_to_node(rdsibdev) ibdev_to_node(rdsibdev->dev) 239 240 /* bits for i_ack_flags */ 241 #define IB_ACK_IN_FLIGHT 0 242 #define IB_ACK_REQUESTED 1 243 244 /* Magic WR_ID for ACKs */ 245 #define RDS_IB_ACK_WR_ID (~(u64) 0) 246 247 struct rds_ib_statistics { 248 uint64_t s_ib_connect_raced; 249 uint64_t s_ib_listen_closed_stale; 250 uint64_t s_ib_evt_handler_call; 251 uint64_t s_ib_tasklet_call; 252 uint64_t s_ib_tx_cq_event; 253 uint64_t s_ib_tx_ring_full; 254 uint64_t s_ib_tx_throttle; 255 uint64_t s_ib_tx_sg_mapping_failure; 256 uint64_t s_ib_tx_stalled; 257 uint64_t s_ib_tx_credit_updates; 258 uint64_t s_ib_rx_cq_event; 259 uint64_t s_ib_rx_ring_empty; 260 uint64_t s_ib_rx_refill_from_cq; 261 uint64_t s_ib_rx_refill_from_thread; 262 uint64_t s_ib_rx_alloc_limit; 263 uint64_t s_ib_rx_total_frags; 264 uint64_t s_ib_rx_total_incs; 265 uint64_t s_ib_rx_credit_updates; 266 uint64_t s_ib_ack_sent; 267 uint64_t s_ib_ack_send_failure; 268 uint64_t s_ib_ack_send_delayed; 269 uint64_t s_ib_ack_send_piggybacked; 270 uint64_t s_ib_ack_received; 271 uint64_t s_ib_rdma_mr_8k_alloc; 272 uint64_t s_ib_rdma_mr_8k_free; 273 uint64_t s_ib_rdma_mr_8k_used; 274 uint64_t s_ib_rdma_mr_8k_pool_flush; 275 uint64_t s_ib_rdma_mr_8k_pool_wait; 276 uint64_t s_ib_rdma_mr_8k_pool_depleted; 277 uint64_t s_ib_rdma_mr_1m_alloc; 278 uint64_t s_ib_rdma_mr_1m_free; 279 uint64_t s_ib_rdma_mr_1m_used; 280 uint64_t s_ib_rdma_mr_1m_pool_flush; 281 uint64_t s_ib_rdma_mr_1m_pool_wait; 282 uint64_t s_ib_rdma_mr_1m_pool_depleted; 283 uint64_t s_ib_rdma_mr_8k_reused; 284 uint64_t s_ib_rdma_mr_1m_reused; 285 uint64_t s_ib_atomic_cswp; 286 uint64_t s_ib_atomic_fadd; 287 uint64_t s_ib_recv_added_to_cache; 288 uint64_t s_ib_recv_removed_from_cache; 289 }; 290 291 extern struct workqueue_struct *rds_ib_wq; 292 293 /* 294 * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h 295 * doesn't define it. 296 */ 297 static inline void rds_ib_dma_sync_sg_for_cpu(struct ib_device *dev, 298 struct scatterlist *sglist, 299 unsigned int sg_dma_len, 300 int direction) 301 { 302 struct scatterlist *sg; 303 unsigned int i; 304 305 for_each_sg(sglist, sg, sg_dma_len, i) { 306 ib_dma_sync_single_for_cpu(dev, 307 ib_sg_dma_address(dev, sg), 308 ib_sg_dma_len(dev, sg), 309 direction); 310 } 311 } 312 #define ib_dma_sync_sg_for_cpu rds_ib_dma_sync_sg_for_cpu 313 314 static inline void rds_ib_dma_sync_sg_for_device(struct ib_device *dev, 315 struct scatterlist *sglist, 316 unsigned int sg_dma_len, 317 int direction) 318 { 319 struct scatterlist *sg; 320 unsigned int i; 321 322 for_each_sg(sglist, sg, sg_dma_len, i) { 323 ib_dma_sync_single_for_device(dev, 324 ib_sg_dma_address(dev, sg), 325 ib_sg_dma_len(dev, sg), 326 direction); 327 } 328 } 329 #define ib_dma_sync_sg_for_device rds_ib_dma_sync_sg_for_device 330 331 332 /* ib.c */ 333 extern struct rds_transport rds_ib_transport; 334 struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device); 335 void rds_ib_dev_put(struct rds_ib_device *rds_ibdev); 336 extern struct ib_client rds_ib_client; 337 338 extern unsigned int rds_ib_retry_count; 339 340 extern spinlock_t ib_nodev_conns_lock; 341 extern struct list_head ib_nodev_conns; 342 343 /* ib_cm.c */ 344 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp); 345 void rds_ib_conn_free(void *arg); 346 int rds_ib_conn_path_connect(struct rds_conn_path *cp); 347 void rds_ib_conn_path_shutdown(struct rds_conn_path *cp); 348 void rds_ib_state_change(struct sock *sk); 349 int rds_ib_listen_init(void); 350 void rds_ib_listen_stop(void); 351 __printf(2, 3) 352 void __rds_ib_conn_error(struct rds_connection *conn, const char *, ...); 353 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id, 354 struct rdma_cm_event *event); 355 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id); 356 void rds_ib_cm_connect_complete(struct rds_connection *conn, 357 struct rdma_cm_event *event); 358 359 360 #define rds_ib_conn_error(conn, fmt...) \ 361 __rds_ib_conn_error(conn, KERN_WARNING "RDS/IB: " fmt) 362 363 /* ib_rdma.c */ 364 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr); 365 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn); 366 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn); 367 void rds_ib_destroy_nodev_conns(void); 368 void rds_ib_mr_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc); 369 370 /* ib_recv.c */ 371 int rds_ib_recv_init(void); 372 void rds_ib_recv_exit(void); 373 int rds_ib_recv_path(struct rds_conn_path *conn); 374 int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic); 375 void rds_ib_recv_free_caches(struct rds_ib_connection *ic); 376 void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp); 377 void rds_ib_inc_free(struct rds_incoming *inc); 378 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to); 379 void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc, 380 struct rds_ib_ack_state *state); 381 void rds_ib_recv_tasklet_fn(unsigned long data); 382 void rds_ib_recv_init_ring(struct rds_ib_connection *ic); 383 void rds_ib_recv_clear_ring(struct rds_ib_connection *ic); 384 void rds_ib_recv_init_ack(struct rds_ib_connection *ic); 385 void rds_ib_attempt_ack(struct rds_ib_connection *ic); 386 void rds_ib_ack_send_complete(struct rds_ib_connection *ic); 387 u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic); 388 void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required); 389 390 /* ib_ring.c */ 391 void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr); 392 void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr); 393 u32 rds_ib_ring_alloc(struct rds_ib_work_ring *ring, u32 val, u32 *pos); 394 void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val); 395 void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val); 396 int rds_ib_ring_empty(struct rds_ib_work_ring *ring); 397 int rds_ib_ring_low(struct rds_ib_work_ring *ring); 398 u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring); 399 u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest); 400 extern wait_queue_head_t rds_ib_ring_empty_wait; 401 402 /* ib_send.c */ 403 void rds_ib_xmit_path_complete(struct rds_conn_path *cp); 404 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm, 405 unsigned int hdr_off, unsigned int sg, unsigned int off); 406 void rds_ib_send_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc); 407 void rds_ib_send_init_ring(struct rds_ib_connection *ic); 408 void rds_ib_send_clear_ring(struct rds_ib_connection *ic); 409 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op); 410 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits); 411 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted); 412 int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted, 413 u32 *adv_credits, int need_posted, int max_posted); 414 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op); 415 416 /* ib_stats.c */ 417 DECLARE_PER_CPU(struct rds_ib_statistics, rds_ib_stats); 418 #define rds_ib_stats_inc(member) rds_stats_inc_which(rds_ib_stats, member) 419 #define rds_ib_stats_add(member, count) \ 420 rds_stats_add_which(rds_ib_stats, member, count) 421 unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter, 422 unsigned int avail); 423 424 /* ib_sysctl.c */ 425 int rds_ib_sysctl_init(void); 426 void rds_ib_sysctl_exit(void); 427 extern unsigned long rds_ib_sysctl_max_send_wr; 428 extern unsigned long rds_ib_sysctl_max_recv_wr; 429 extern unsigned long rds_ib_sysctl_max_unsig_wrs; 430 extern unsigned long rds_ib_sysctl_max_unsig_bytes; 431 extern unsigned long rds_ib_sysctl_max_recv_allocation; 432 extern unsigned int rds_ib_sysctl_flow_control; 433 434 #endif 435