mlx4.h revision 275109
1/* 2 * Copyright (c) 2004, 2005 Topspin Communications. All rights reserved. 3 * Copyright (c) 2005, 2006, 2007 Cisco Systems. All rights reserved. 4 * 5 * This software is available to you under a choice of one of two 6 * licenses. You may choose to be licensed under the terms of the GNU 7 * General Public License (GPL) Version 2, available from the file 8 * COPYING in the main directory of this source tree, or the 9 * OpenIB.org BSD license below: 10 * 11 * Redistribution and use in source and binary forms, with or 12 * without modification, are permitted provided that the following 13 * conditions are met: 14 * 15 * - Redistributions of source code must retain the above 16 * copyright notice, this list of conditions and the following 17 * disclaimer. 18 * 19 * - Redistributions in binary form must reproduce the above 20 * copyright notice, this list of conditions and the following 21 * disclaimer in the documentation and/or other materials 22 * provided with the distribution. 23 * 24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 31 * SOFTWARE. 32 */ 33 34#ifndef MLX4_H 35#define MLX4_H 36 37#include <stddef.h> 38 39#include <infiniband/driver.h> 40#include <infiniband/arch.h> 41 42#ifdef HAVE_VALGRIND_MEMCHECK_H 43 44# include <valgrind/memcheck.h> 45 46# if !defined(VALGRIND_MAKE_MEM_DEFINED) || !defined(VALGRIND_MAKE_MEM_UNDEFINED) 47# warning "Valgrind support requested, but VALGRIND_MAKE_MEM_(UN)DEFINED not available" 48# endif 49 50#endif /* HAVE_VALGRIND_MEMCHECK_H */ 51 52#ifndef VALGRIND_MAKE_MEM_DEFINED 53# define VALGRIND_MAKE_MEM_DEFINED(addr,len) 54#endif 55 56#ifndef VALGRIND_MAKE_MEM_UNDEFINED 57# define VALGRIND_MAKE_MEM_UNDEFINED(addr,len) 58#endif 59 60#ifndef rmb 61# define rmb() mb() 62#endif 63 64#ifndef wmb 65# define wmb() mb() 66#endif 67 68#ifndef wc_wmb 69 70#if defined(__i386__) 71#define wc_wmb() asm volatile("lock; addl $0,0(%%esp) " ::: "memory") 72#elif defined(__x86_64__) 73#define wc_wmb() asm volatile("sfence" ::: "memory") 74#elif defined(__ia64__) 75#define wc_wmb() asm volatile("fwb" ::: "memory") 76#else 77#define wc_wmb() wmb() 78#endif 79 80#endif 81 82#ifndef HAVE_IBV_MORE_OPS 83#undef HAVE_IBV_XRC_OPS 84#undef HAVE_IBV_CREATE_QP_EXP 85#endif 86 87#define HIDDEN __attribute__((visibility ("hidden"))) 88 89#define PFX "mlx4: " 90 91#ifndef max 92#define max(a,b) \ 93 ({ typeof (a) _a = (a); \ 94 typeof (b) _b = (b); \ 95 _a > _b ? _a : _b; }) 96#endif 97 98#ifndef min 99#define min(a,b) \ 100 ({ typeof (a) _a = (a); \ 101 typeof (b) _b = (b); \ 102 _a < _b ? _a : _b; }) 103#endif 104 105enum { 106 MLX4_STAT_RATE_OFFSET = 5 107}; 108 109enum { 110 MLX4_QP_TABLE_BITS = 8, 111 MLX4_QP_TABLE_SIZE = 1 << MLX4_QP_TABLE_BITS, 112 MLX4_QP_TABLE_MASK = MLX4_QP_TABLE_SIZE - 1 113}; 114 115enum { 116 MLX4_XRC_SRQ_TABLE_BITS = 8, 117 MLX4_XRC_SRQ_TABLE_SIZE = 1 << MLX4_XRC_SRQ_TABLE_BITS, 118 MLX4_XRC_SRQ_TABLE_MASK = MLX4_XRC_SRQ_TABLE_SIZE - 1 119}; 120 121enum { 122 MLX4_XRC_QPN_BIT = (1 << 23) 123}; 124 125enum mlx4_db_type { 126 MLX4_DB_TYPE_CQ, 127 MLX4_DB_TYPE_RQ, 128 MLX4_NUM_DB_TYPE 129}; 130 131enum { 132 MLX4_OPCODE_NOP = 0x00, 133 MLX4_OPCODE_SEND_INVAL = 0x01, 134 MLX4_OPCODE_RDMA_WRITE = 0x08, 135 MLX4_OPCODE_RDMA_WRITE_IMM = 0x09, 136 MLX4_OPCODE_SEND = 0x0a, 137 MLX4_OPCODE_SEND_IMM = 0x0b, 138 MLX4_OPCODE_LSO = 0x0e, 139 MLX4_OPCODE_RDMA_READ = 0x10, 140 MLX4_OPCODE_ATOMIC_CS = 0x11, 141 MLX4_OPCODE_ATOMIC_FA = 0x12, 142 MLX4_OPCODE_ATOMIC_MASK_CS = 0x14, 143 MLX4_OPCODE_ATOMIC_MASK_FA = 0x15, 144 MLX4_OPCODE_BIND_MW = 0x18, 145 MLX4_OPCODE_FMR = 0x19, 146 MLX4_OPCODE_LOCAL_INVAL = 0x1b, 147 MLX4_OPCODE_CONFIG_CMD = 0x1f, 148 149 MLX4_RECV_OPCODE_RDMA_WRITE_IMM = 0x00, 150 MLX4_RECV_OPCODE_SEND = 0x01, 151 MLX4_RECV_OPCODE_SEND_IMM = 0x02, 152 MLX4_RECV_OPCODE_SEND_INVAL = 0x03, 153 154 MLX4_CQE_OPCODE_ERROR = 0x1e, 155 MLX4_CQE_OPCODE_RESIZE = 0x16, 156}; 157 158enum { 159 MLX4_MAX_WQE_SIZE = 1008 160}; 161 162struct mlx4_device { 163 struct ibv_device ibv_dev; 164 int page_size; 165}; 166 167struct mlx4_db_page; 168 169struct mlx4_context { 170 struct ibv_context ibv_ctx; 171 172 void *uar; 173 pthread_spinlock_t uar_lock; 174 175 void *bf_page; 176 int bf_buf_size; 177 int bf_offset; 178 pthread_spinlock_t bf_lock; 179 180 struct { 181 struct mlx4_qp **table; 182 int refcnt; 183 } qp_table[MLX4_QP_TABLE_SIZE]; 184 pthread_mutex_t qp_table_mutex; 185 int num_qps; 186 int qp_table_shift; 187 int qp_table_mask; 188 int max_qp_wr; 189 int max_sge; 190 int max_cqe; 191 int cqe_size; 192 193 struct { 194 struct mlx4_srq **table; 195 int refcnt; 196 } xrc_srq_table[MLX4_XRC_SRQ_TABLE_SIZE]; 197 pthread_mutex_t xrc_srq_table_mutex; 198 int num_xrc_srqs; 199 int xrc_srq_table_shift; 200 int xrc_srq_table_mask; 201 202 struct mlx4_db_page *db_list[MLX4_NUM_DB_TYPE]; 203 pthread_mutex_t db_list_mutex; 204}; 205 206struct mlx4_buf { 207 void *buf; 208 size_t length; 209}; 210 211struct mlx4_pd { 212 struct ibv_pd ibv_pd; 213 uint32_t pdn; 214}; 215 216struct mlx4_cq { 217 struct ibv_cq ibv_cq; 218 struct mlx4_buf buf; 219 struct mlx4_buf resize_buf; 220 pthread_spinlock_t lock; 221 uint32_t cqn; 222 uint32_t cons_index; 223 uint32_t *set_ci_db; 224 uint32_t *arm_db; 225 int arm_sn; 226 int cqe_size; 227}; 228 229struct mlx4_srq { 230 struct ibv_srq ibv_srq; 231 struct mlx4_buf buf; 232 pthread_spinlock_t lock; 233 uint64_t *wrid; 234 uint32_t srqn; 235 int max; 236 int max_gs; 237 int wqe_shift; 238 int head; 239 int tail; 240 uint32_t *db; 241 uint16_t counter; 242}; 243 244struct mlx4_wq { 245 uint64_t *wrid; 246 pthread_spinlock_t lock; 247 int wqe_cnt; 248 int max_post; 249 unsigned head; 250 unsigned tail; 251 int max_gs; 252 int wqe_shift; 253 int offset; 254}; 255 256struct mlx4_qp { 257 struct ibv_qp ibv_qp; 258 struct mlx4_buf buf; 259 int max_inline_data; 260 int buf_size; 261 262 uint32_t doorbell_qpn; 263 uint32_t sq_signal_bits; 264 int sq_spare_wqes; 265 struct mlx4_wq sq; 266 267 uint32_t *db; 268 struct mlx4_wq rq; 269}; 270 271struct mlx4_av { 272 uint32_t port_pd; 273 uint8_t reserved1; 274 uint8_t g_slid; 275 uint16_t dlid; 276 uint8_t reserved2; 277 uint8_t gid_index; 278 uint8_t stat_rate; 279 uint8_t hop_limit; 280 uint32_t sl_tclass_flowlabel; 281 uint8_t dgid[16]; 282 uint8_t mac[8]; 283}; 284 285struct mlx4_ah { 286 struct ibv_ah ibv_ah; 287 struct mlx4_av av; 288 uint16_t vlan; 289 uint8_t mac[6]; 290 uint8_t tagged; 291}; 292 293struct mlx4_xrc_domain { 294 struct ibv_xrc_domain ibv_xrcd; 295 uint32_t xrcdn; 296}; 297 298static inline unsigned long align(unsigned long val, unsigned long align) 299{ 300 return (val + align - 1) & ~(align - 1); 301} 302 303#define to_mxxx(xxx, type) \ 304 ((struct mlx4_##type *) \ 305 ((void *) ib##xxx - offsetof(struct mlx4_##type, ibv_##xxx))) 306 307static inline struct mlx4_device *to_mdev(struct ibv_device *ibdev) 308{ 309 return to_mxxx(dev, device); 310} 311 312static inline struct mlx4_context *to_mctx(struct ibv_context *ibctx) 313{ 314 return to_mxxx(ctx, context); 315} 316 317static inline struct mlx4_pd *to_mpd(struct ibv_pd *ibpd) 318{ 319 return to_mxxx(pd, pd); 320} 321 322static inline struct mlx4_cq *to_mcq(struct ibv_cq *ibcq) 323{ 324 return to_mxxx(cq, cq); 325} 326 327static inline struct mlx4_srq *to_msrq(struct ibv_srq *ibsrq) 328{ 329 return to_mxxx(srq, srq); 330} 331 332static inline struct mlx4_qp *to_mqp(struct ibv_qp *ibqp) 333{ 334 return to_mxxx(qp, qp); 335} 336 337static inline struct mlx4_ah *to_mah(struct ibv_ah *ibah) 338{ 339 return to_mxxx(ah, ah); 340} 341 342#ifdef HAVE_IBV_XRC_OPS 343static inline struct mlx4_xrc_domain *to_mxrcd(struct ibv_xrc_domain *ibxrcd) 344{ 345 return to_mxxx(xrcd, xrc_domain); 346} 347#endif 348 349int mlx4_alloc_buf(struct mlx4_buf *buf, size_t size, int page_size); 350void mlx4_free_buf(struct mlx4_buf *buf); 351 352uint32_t *mlx4_alloc_db(struct mlx4_context *context, enum mlx4_db_type type); 353void mlx4_free_db(struct mlx4_context *context, enum mlx4_db_type type, uint32_t *db); 354 355int mlx4_query_device(struct ibv_context *context, 356 struct ibv_device_attr *attr); 357int mlx4_query_port(struct ibv_context *context, uint8_t port, 358 struct ibv_port_attr *attr); 359 360struct ibv_pd *mlx4_alloc_pd(struct ibv_context *context); 361int mlx4_free_pd(struct ibv_pd *pd); 362 363struct ibv_mr *mlx4_reg_mr(struct ibv_pd *pd, void *addr, 364 size_t length, enum ibv_access_flags access); 365int mlx4_dereg_mr(struct ibv_mr *mr); 366 367struct ibv_cq *mlx4_create_cq(struct ibv_context *context, int cqe, 368 struct ibv_comp_channel *channel, 369 int comp_vector); 370int mlx4_alloc_cq_buf(struct mlx4_device *dev, struct mlx4_buf *buf, int nent, 371 int entry_size); 372int mlx4_resize_cq(struct ibv_cq *cq, int cqe); 373int mlx4_destroy_cq(struct ibv_cq *cq); 374int mlx4_poll_cq(struct ibv_cq *cq, int ne, struct ibv_wc *wc); 375int mlx4_arm_cq(struct ibv_cq *cq, int solicited); 376void mlx4_cq_event(struct ibv_cq *cq); 377void __mlx4_cq_clean(struct mlx4_cq *cq, uint32_t qpn, struct mlx4_srq *srq); 378void mlx4_cq_clean(struct mlx4_cq *cq, uint32_t qpn, struct mlx4_srq *srq); 379int mlx4_get_outstanding_cqes(struct mlx4_cq *cq); 380void mlx4_cq_resize_copy_cqes(struct mlx4_cq *cq, void *buf, int new_cqe); 381 382struct ibv_srq *mlx4_create_srq(struct ibv_pd *pd, 383 struct ibv_srq_init_attr *attr); 384int mlx4_modify_srq(struct ibv_srq *srq, 385 struct ibv_srq_attr *attr, 386 enum ibv_srq_attr_mask mask); 387int mlx4_query_srq(struct ibv_srq *srq, 388 struct ibv_srq_attr *attr); 389int mlx4_destroy_srq(struct ibv_srq *srq); 390int mlx4_alloc_srq_buf(struct ibv_pd *pd, struct ibv_srq_attr *attr, 391 struct mlx4_srq *srq); 392void mlx4_free_srq_wqe(struct mlx4_srq *srq, int ind); 393int mlx4_post_srq_recv(struct ibv_srq *ibsrq, 394 struct ibv_recv_wr *wr, 395 struct ibv_recv_wr **bad_wr); 396struct mlx4_srq *mlx4_find_xrc_srq(struct mlx4_context *ctx, uint32_t xrc_srqn); 397int mlx4_store_xrc_srq(struct mlx4_context *ctx, uint32_t xrc_srqn, 398 struct mlx4_srq *srq); 399void mlx4_clear_xrc_srq(struct mlx4_context *ctx, uint32_t xrc_srqn); 400 401struct ibv_qp *mlx4_create_qp(struct ibv_pd *pd, struct ibv_qp_init_attr *attr); 402int mlx4_query_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr, 403 enum ibv_qp_attr_mask attr_mask, 404 struct ibv_qp_init_attr *init_attr); 405int mlx4_modify_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr, 406 enum ibv_qp_attr_mask attr_mask); 407int mlx4_destroy_qp(struct ibv_qp *qp); 408void mlx4_init_qp_indices(struct mlx4_qp *qp); 409void mlx4_qp_init_sq_ownership(struct mlx4_qp *qp); 410int mlx4_post_send(struct ibv_qp *ibqp, struct ibv_send_wr *wr, 411 struct ibv_send_wr **bad_wr); 412int mlx4_post_recv(struct ibv_qp *ibqp, struct ibv_recv_wr *wr, 413 struct ibv_recv_wr **bad_wr); 414void mlx4_calc_sq_wqe_size(struct ibv_qp_cap *cap, enum ibv_qp_type type, 415 struct mlx4_qp *qp); 416int num_inline_segs(int data, enum ibv_qp_type type); 417int mlx4_alloc_qp_buf(struct ibv_pd *pd, struct ibv_qp_cap *cap, 418 enum ibv_qp_type type, struct mlx4_qp *qp); 419void mlx4_set_sq_sizes(struct mlx4_qp *qp, struct ibv_qp_cap *cap, 420 enum ibv_qp_type type); 421struct mlx4_qp *mlx4_find_qp(struct mlx4_context *ctx, uint32_t qpn); 422int mlx4_store_qp(struct mlx4_context *ctx, uint32_t qpn, struct mlx4_qp *qp); 423void mlx4_clear_qp(struct mlx4_context *ctx, uint32_t qpn); 424struct ibv_ah *mlx4_create_ah(struct ibv_pd *pd, struct ibv_ah_attr *attr); 425int mlx4_destroy_ah(struct ibv_ah *ah); 426int mlx4_alloc_av(struct mlx4_pd *pd, struct ibv_ah_attr *attr, 427 struct mlx4_ah *ah); 428void mlx4_free_av(struct mlx4_ah *ah); 429#ifdef HAVE_IBV_XRC_OPS 430struct ibv_srq *mlx4_create_xrc_srq(struct ibv_pd *pd, 431 struct ibv_xrc_domain *xrc_domain, 432 struct ibv_cq *xrc_cq, 433 struct ibv_srq_init_attr *attr); 434struct ibv_xrc_domain *mlx4_open_xrc_domain(struct ibv_context *context, 435 int fd, int oflag); 436 437int mlx4_close_xrc_domain(struct ibv_xrc_domain *d); 438int mlx4_create_xrc_rcv_qp(struct ibv_qp_init_attr *init_attr, 439 uint32_t *xrc_qp_num); 440int mlx4_modify_xrc_rcv_qp(struct ibv_xrc_domain *xrc_domain, 441 uint32_t xrc_qp_num, 442 struct ibv_qp_attr *attr, 443 int attr_mask); 444int mlx4_query_xrc_rcv_qp(struct ibv_xrc_domain *xrc_domain, 445 uint32_t xrc_qp_num, 446 struct ibv_qp_attr *attr, 447 int attr_mask, 448 struct ibv_qp_init_attr *init_attr); 449int mlx4_reg_xrc_rcv_qp(struct ibv_xrc_domain *xrc_domain, 450 uint32_t xrc_qp_num); 451int mlx4_unreg_xrc_rcv_qp(struct ibv_xrc_domain *xrc_domain, 452 uint32_t xrc_qp_num); 453#endif 454 455 456#endif /* MLX4_H */ 457