1 #ifndef DEF_RDMA_VT_H 2 #define DEF_RDMA_VT_H 3 4 /* 5 * Copyright(c) 2016 Intel Corporation. 6 * 7 * This file is provided under a dual BSD/GPLv2 license. When using or 8 * redistributing this file, you may do so under either license. 9 * 10 * GPL LICENSE SUMMARY 11 * 12 * This program is free software; you can redistribute it and/or modify 13 * it under the terms of version 2 of the GNU General Public License as 14 * published by the Free Software Foundation. 15 * 16 * This program is distributed in the hope that it will be useful, but 17 * WITHOUT ANY WARRANTY; without even the implied warranty of 18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 19 * General Public License for more details. 20 * 21 * BSD LICENSE 22 * 23 * Redistribution and use in source and binary forms, with or without 24 * modification, are permitted provided that the following conditions 25 * are met: 26 * 27 * - Redistributions of source code must retain the above copyright 28 * notice, this list of conditions and the following disclaimer. 29 * - Redistributions in binary form must reproduce the above copyright 30 * notice, this list of conditions and the following disclaimer in 31 * the documentation and/or other materials provided with the 32 * distribution. 33 * - Neither the name of Intel Corporation nor the names of its 34 * contributors may be used to endorse or promote products derived 35 * from this software without specific prior written permission. 36 * 37 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 38 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 39 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 40 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 41 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 42 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 43 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 44 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 45 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 46 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 47 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 48 * 49 */ 50 51 /* 52 * Structure that low level drivers will populate in order to register with the 53 * rdmavt layer. 54 */ 55 56 #include <linux/spinlock.h> 57 #include <linux/list.h> 58 #include <linux/hash.h> 59 #include <rdma/ib_verbs.h> 60 #include <rdma/rdmavt_mr.h> 61 #include <rdma/rdmavt_qp.h> 62 63 #define RVT_MAX_PKEY_VALUES 16 64 65 struct rvt_ibport { 66 struct rvt_qp __rcu *qp[2]; 67 struct ib_mad_agent *send_agent; /* agent for SMI (traps) */ 68 struct rb_root mcast_tree; 69 spinlock_t lock; /* protect changes in this struct */ 70 71 /* non-zero when timer is set */ 72 unsigned long mkey_lease_timeout; 73 unsigned long trap_timeout; 74 __be64 gid_prefix; /* in network order */ 75 __be64 mkey; 76 u64 tid; 77 u32 port_cap_flags; 78 u32 pma_sample_start; 79 u32 pma_sample_interval; 80 __be16 pma_counter_select[5]; 81 u16 pma_tag; 82 u16 mkey_lease_period; 83 u16 sm_lid; 84 u8 sm_sl; 85 u8 mkeyprot; 86 u8 subnet_timeout; 87 u8 vl_high_limit; 88 89 /* 90 * Driver is expected to keep these up to date. These 91 * counters are informational only and not required to be 92 * completely accurate. 93 */ 94 u64 n_rc_resends; 95 u64 n_seq_naks; 96 u64 n_rdma_seq; 97 u64 n_rnr_naks; 98 u64 n_other_naks; 99 u64 n_loop_pkts; 100 u64 n_pkt_drops; 101 u64 n_vl15_dropped; 102 u64 n_rc_timeouts; 103 u64 n_dmawait; 104 u64 n_unaligned; 105 u64 n_rc_dupreq; 106 u64 n_rc_seqnak; 107 u16 pkey_violations; 108 u16 qkey_violations; 109 u16 mkey_violations; 110 111 /* Hot-path per CPU counters to avoid cacheline trading to update */ 112 u64 z_rc_acks; 113 u64 z_rc_qacks; 114 u64 z_rc_delayed_comp; 115 u64 __percpu *rc_acks; 116 u64 __percpu *rc_qacks; 117 u64 __percpu *rc_delayed_comp; 118 119 void *priv; /* driver private data */ 120 121 /* 122 * The pkey table is allocated and maintained by the driver. Drivers 123 * need to have access to this before registering with rdmav. However 124 * rdmavt will need access to it so drivers need to proviee this during 125 * the attach port API call. 126 */ 127 u16 *pkey_table; 128 129 struct rvt_ah *sm_ah; 130 }; 131 132 #define RVT_CQN_MAX 16 /* maximum length of cq name */ 133 134 /* 135 * Things that are driver specific, module parameters in hfi1 and qib 136 */ 137 struct rvt_driver_params { 138 struct ib_device_attr props; 139 140 /* 141 * Anything driver specific that is not covered by props 142 * For instance special module parameters. Goes here. 143 */ 144 unsigned int lkey_table_size; 145 unsigned int qp_table_size; 146 int qpn_start; 147 int qpn_inc; 148 int qpn_res_start; 149 int qpn_res_end; 150 int nports; 151 int npkeys; 152 char cq_name[RVT_CQN_MAX]; 153 int node; 154 int psn_mask; 155 int psn_shift; 156 int psn_modify_mask; 157 u32 core_cap_flags; 158 u32 max_mad_size; 159 u8 qos_shift; 160 u8 max_rdma_atomic; 161 }; 162 163 /* Protection domain */ 164 struct rvt_pd { 165 struct ib_pd ibpd; 166 int user; /* non-zero if created from user space */ 167 }; 168 169 /* Address handle */ 170 struct rvt_ah { 171 struct ib_ah ibah; 172 struct ib_ah_attr attr; 173 atomic_t refcount; 174 u8 vl; 175 u8 log_pmtu; 176 }; 177 178 struct rvt_dev_info; 179 struct rvt_swqe; 180 struct rvt_driver_provided { 181 /* 182 * Which functions are required depends on which verbs rdmavt is 183 * providing and which verbs the driver is overriding. See 184 * check_support() for details. 185 */ 186 187 /* Passed to ib core registration. Callback to create syfs files */ 188 int (*port_callback)(struct ib_device *, u8, struct kobject *); 189 190 /* 191 * Returns a string to represent the device for which is being 192 * registered. This is primarily used for error and debug messages on 193 * the console. 194 */ 195 const char * (*get_card_name)(struct rvt_dev_info *rdi); 196 197 /* 198 * Returns a pointer to the undelying hardware's PCI device. This is 199 * used to display information as to what hardware is being referenced 200 * in an output message 201 */ 202 struct pci_dev * (*get_pci_dev)(struct rvt_dev_info *rdi); 203 204 /* 205 * Allocate a private queue pair data structure for driver specific 206 * information which is opaque to rdmavt. Errors are returned via 207 * ERR_PTR(err). The driver is free to return NULL or a valid 208 * pointer. 209 */ 210 void * (*qp_priv_alloc)(struct rvt_dev_info *rdi, struct rvt_qp *qp, 211 gfp_t gfp); 212 213 /* 214 * Free the driver's private qp structure. 215 */ 216 void (*qp_priv_free)(struct rvt_dev_info *rdi, struct rvt_qp *qp); 217 218 /* 219 * Inform the driver the particular qp in quesiton has been reset so 220 * that it can clean up anything it needs to. 221 */ 222 void (*notify_qp_reset)(struct rvt_qp *qp); 223 224 /* 225 * Give the driver a notice that there is send work to do. It is up to 226 * the driver to generally push the packets out, this just queues the 227 * work with the driver. There are two variants here. The no_lock 228 * version requires the s_lock not to be held. The other assumes the 229 * s_lock is held. 230 */ 231 void (*schedule_send)(struct rvt_qp *qp); 232 void (*schedule_send_no_lock)(struct rvt_qp *qp); 233 234 /* 235 * Sometimes rdmavt needs to kick the driver's send progress. That is 236 * done by this call back. 237 */ 238 void (*do_send)(struct rvt_qp *qp); 239 240 /* 241 * Get a path mtu from the driver based on qp attributes. 242 */ 243 int (*get_pmtu_from_attr)(struct rvt_dev_info *rdi, struct rvt_qp *qp, 244 struct ib_qp_attr *attr); 245 246 /* 247 * Notify driver that it needs to flush any outstanding IO requests that 248 * are waiting on a qp. 249 */ 250 void (*flush_qp_waiters)(struct rvt_qp *qp); 251 252 /* 253 * Notify driver to stop its queue of sending packets. Nothing else 254 * should be posted to the queue pair after this has been called. 255 */ 256 void (*stop_send_queue)(struct rvt_qp *qp); 257 258 /* 259 * Have the drivr drain any in progress operations 260 */ 261 void (*quiesce_qp)(struct rvt_qp *qp); 262 263 /* 264 * Inform the driver a qp has went to error state. 265 */ 266 void (*notify_error_qp)(struct rvt_qp *qp); 267 268 /* 269 * Get an MTU for a qp. 270 */ 271 u32 (*mtu_from_qp)(struct rvt_dev_info *rdi, struct rvt_qp *qp, 272 u32 pmtu); 273 /* 274 * Convert an mtu to a path mtu 275 */ 276 int (*mtu_to_path_mtu)(u32 mtu); 277 278 /* 279 * Get the guid of a port in big endian byte order 280 */ 281 int (*get_guid_be)(struct rvt_dev_info *rdi, struct rvt_ibport *rvp, 282 int guid_index, __be64 *guid); 283 284 /* 285 * Query driver for the state of the port. 286 */ 287 int (*query_port_state)(struct rvt_dev_info *rdi, u8 port_num, 288 struct ib_port_attr *props); 289 290 /* 291 * Tell driver to shutdown a port 292 */ 293 int (*shut_down_port)(struct rvt_dev_info *rdi, u8 port_num); 294 295 /* Tell driver to send a trap for changed port capabilities */ 296 void (*cap_mask_chg)(struct rvt_dev_info *rdi, u8 port_num); 297 298 /* 299 * The following functions can be safely ignored completely. Any use of 300 * these is checked for NULL before blindly calling. Rdmavt should also 301 * be functional if drivers omit these. 302 */ 303 304 /* Called to inform the driver that all qps should now be freed. */ 305 unsigned (*free_all_qps)(struct rvt_dev_info *rdi); 306 307 /* Driver specific AH validation */ 308 int (*check_ah)(struct ib_device *, struct ib_ah_attr *); 309 310 /* Inform the driver a new AH has been created */ 311 void (*notify_new_ah)(struct ib_device *, struct ib_ah_attr *, 312 struct rvt_ah *); 313 314 /* Let the driver pick the next queue pair number*/ 315 int (*alloc_qpn)(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt, 316 enum ib_qp_type type, u8 port_num, gfp_t gfp); 317 318 /* Determine if its safe or allowed to modify the qp */ 319 int (*check_modify_qp)(struct rvt_qp *qp, struct ib_qp_attr *attr, 320 int attr_mask, struct ib_udata *udata); 321 322 /* Driver specific QP modification/notification-of */ 323 void (*modify_qp)(struct rvt_qp *qp, struct ib_qp_attr *attr, 324 int attr_mask, struct ib_udata *udata); 325 326 /* Driver specific work request checking */ 327 int (*check_send_wqe)(struct rvt_qp *qp, struct rvt_swqe *wqe); 328 329 /* Notify driver a mad agent has been created */ 330 void (*notify_create_mad_agent)(struct rvt_dev_info *rdi, int port_idx); 331 332 /* Notify driver a mad agent has been removed */ 333 void (*notify_free_mad_agent)(struct rvt_dev_info *rdi, int port_idx); 334 335 }; 336 337 struct rvt_dev_info { 338 struct ib_device ibdev; /* Keep this first. Nothing above here */ 339 340 /* 341 * Prior to calling for registration the driver will be responsible for 342 * allocating space for this structure. 343 * 344 * The driver will also be responsible for filling in certain members of 345 * dparms.props. The driver needs to fill in dparms exactly as it would 346 * want values reported to a ULP. This will be returned to the caller 347 * in rdmavt's device. The driver should also therefore refrain from 348 * modifying this directly after registration with rdmavt. 349 */ 350 351 /* Driver specific properties */ 352 struct rvt_driver_params dparms; 353 354 struct rvt_mregion __rcu *dma_mr; 355 struct rvt_lkey_table lkey_table; 356 357 /* Driver specific helper functions */ 358 struct rvt_driver_provided driver_f; 359 360 /* Internal use */ 361 int n_pds_allocated; 362 spinlock_t n_pds_lock; /* Protect pd allocated count */ 363 364 int n_ahs_allocated; 365 spinlock_t n_ahs_lock; /* Protect ah allocated count */ 366 367 u32 n_srqs_allocated; 368 spinlock_t n_srqs_lock; /* Protect srqs allocated count */ 369 370 int flags; 371 struct rvt_ibport **ports; 372 373 /* QP */ 374 struct rvt_qp_ibdev *qp_dev; 375 u32 n_qps_allocated; /* number of QPs allocated for device */ 376 u32 n_rc_qps; /* number of RC QPs allocated for device */ 377 u32 busy_jiffies; /* timeout scaling based on RC QP count */ 378 spinlock_t n_qps_lock; /* protect qps, rc qps and busy jiffy counts */ 379 380 /* memory maps */ 381 struct list_head pending_mmaps; 382 spinlock_t mmap_offset_lock; /* protect mmap_offset */ 383 u32 mmap_offset; 384 spinlock_t pending_lock; /* protect pending mmap list */ 385 386 /* CQ */ 387 struct kthread_worker *worker; /* per device cq worker */ 388 u32 n_cqs_allocated; /* number of CQs allocated for device */ 389 spinlock_t n_cqs_lock; /* protect count of in use cqs */ 390 391 /* Multicast */ 392 u32 n_mcast_grps_allocated; /* number of mcast groups allocated */ 393 spinlock_t n_mcast_grps_lock; 394 395 }; 396 397 static inline struct rvt_pd *ibpd_to_rvtpd(struct ib_pd *ibpd) 398 { 399 return container_of(ibpd, struct rvt_pd, ibpd); 400 } 401 402 static inline struct rvt_ah *ibah_to_rvtah(struct ib_ah *ibah) 403 { 404 return container_of(ibah, struct rvt_ah, ibah); 405 } 406 407 static inline struct rvt_dev_info *ib_to_rvt(struct ib_device *ibdev) 408 { 409 return container_of(ibdev, struct rvt_dev_info, ibdev); 410 } 411 412 static inline struct rvt_srq *ibsrq_to_rvtsrq(struct ib_srq *ibsrq) 413 { 414 return container_of(ibsrq, struct rvt_srq, ibsrq); 415 } 416 417 static inline struct rvt_qp *ibqp_to_rvtqp(struct ib_qp *ibqp) 418 { 419 return container_of(ibqp, struct rvt_qp, ibqp); 420 } 421 422 static inline unsigned rvt_get_npkeys(struct rvt_dev_info *rdi) 423 { 424 /* 425 * All ports have same number of pkeys. 426 */ 427 return rdi->dparms.npkeys; 428 } 429 430 /* 431 * Return the max atomic suitable for determining 432 * the size of the ack ring buffer in a QP. 433 */ 434 static inline unsigned int rvt_max_atomic(struct rvt_dev_info *rdi) 435 { 436 return rdi->dparms.max_rdma_atomic + 1; 437 } 438 439 /* 440 * Return the indexed PKEY from the port PKEY table. 441 */ 442 static inline u16 rvt_get_pkey(struct rvt_dev_info *rdi, 443 int port_index, 444 unsigned index) 445 { 446 if (index >= rvt_get_npkeys(rdi)) 447 return 0; 448 else 449 return rdi->ports[port_index]->pkey_table[index]; 450 } 451 452 /** 453 * rvt_lookup_qpn - return the QP with the given QPN 454 * @ibp: the ibport 455 * @qpn: the QP number to look up 456 * 457 * The caller must hold the rcu_read_lock(), and keep the lock until 458 * the returned qp is no longer in use. 459 */ 460 /* TODO: Remove this and put in rdmavt/qp.h when no longer needed by drivers */ 461 static inline struct rvt_qp *rvt_lookup_qpn(struct rvt_dev_info *rdi, 462 struct rvt_ibport *rvp, 463 u32 qpn) __must_hold(RCU) 464 { 465 struct rvt_qp *qp = NULL; 466 467 if (unlikely(qpn <= 1)) { 468 qp = rcu_dereference(rvp->qp[qpn]); 469 } else { 470 u32 n = hash_32(qpn, rdi->qp_dev->qp_table_bits); 471 472 for (qp = rcu_dereference(rdi->qp_dev->qp_table[n]); qp; 473 qp = rcu_dereference(qp->next)) 474 if (qp->ibqp.qp_num == qpn) 475 break; 476 } 477 return qp; 478 } 479 480 struct rvt_dev_info *rvt_alloc_device(size_t size, int nports); 481 void rvt_dealloc_device(struct rvt_dev_info *rdi); 482 int rvt_register_device(struct rvt_dev_info *rvd); 483 void rvt_unregister_device(struct rvt_dev_info *rvd); 484 int rvt_check_ah(struct ib_device *ibdev, struct ib_ah_attr *ah_attr); 485 int rvt_init_port(struct rvt_dev_info *rdi, struct rvt_ibport *port, 486 int port_index, u16 *pkey_table); 487 int rvt_rkey_ok(struct rvt_qp *qp, struct rvt_sge *sge, 488 u32 len, u64 vaddr, u32 rkey, int acc); 489 int rvt_lkey_ok(struct rvt_lkey_table *rkt, struct rvt_pd *pd, 490 struct rvt_sge *isge, struct ib_sge *sge, int acc); 491 struct rvt_mcast *rvt_mcast_find(struct rvt_ibport *ibp, union ib_gid *mgid); 492 493 #endif /* DEF_RDMA_VT_H */ 494