1 #ifndef DEF_RDMA_VT_H 2 #define DEF_RDMA_VT_H 3 4 /* 5 * Copyright(c) 2016 - 2019 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/ib_mad.h> 61 #include <rdma/rdmavt_mr.h> 62 63 #define RVT_MAX_PKEY_VALUES 16 64 65 #define RVT_MAX_TRAP_LEN 100 /* Limit pending trap list */ 66 #define RVT_MAX_TRAP_LISTS 5 /*((IB_NOTICE_TYPE_INFO & 0x0F) + 1)*/ 67 #define RVT_TRAP_TIMEOUT 4096 /* 4.096 usec */ 68 69 struct trap_list { 70 u32 list_len; 71 struct list_head list; 72 }; 73 74 struct rvt_qp; 75 struct rvt_qpn_table; 76 struct rvt_ibport { 77 struct rvt_qp __rcu *qp[2]; 78 struct ib_mad_agent *send_agent; /* agent for SMI (traps) */ 79 struct rb_root mcast_tree; 80 spinlock_t lock; /* protect changes in this struct */ 81 82 /* non-zero when timer is set */ 83 unsigned long mkey_lease_timeout; 84 unsigned long trap_timeout; 85 __be64 gid_prefix; /* in network order */ 86 __be64 mkey; 87 u64 tid; 88 u32 port_cap_flags; 89 u16 port_cap3_flags; 90 u32 pma_sample_start; 91 u32 pma_sample_interval; 92 __be16 pma_counter_select[5]; 93 u16 pma_tag; 94 u16 mkey_lease_period; 95 u32 sm_lid; 96 u8 sm_sl; 97 u8 mkeyprot; 98 u8 subnet_timeout; 99 u8 vl_high_limit; 100 101 /* 102 * Driver is expected to keep these up to date. These 103 * counters are informational only and not required to be 104 * completely accurate. 105 */ 106 u64 n_rc_resends; 107 u64 n_seq_naks; 108 u64 n_rdma_seq; 109 u64 n_rnr_naks; 110 u64 n_other_naks; 111 u64 n_loop_pkts; 112 u64 n_pkt_drops; 113 u64 n_vl15_dropped; 114 u64 n_rc_timeouts; 115 u64 n_dmawait; 116 u64 n_unaligned; 117 u64 n_rc_dupreq; 118 u64 n_rc_seqnak; 119 u16 pkey_violations; 120 u16 qkey_violations; 121 u16 mkey_violations; 122 123 /* Hot-path per CPU counters to avoid cacheline trading to update */ 124 u64 z_rc_acks; 125 u64 z_rc_qacks; 126 u64 z_rc_delayed_comp; 127 u64 __percpu *rc_acks; 128 u64 __percpu *rc_qacks; 129 u64 __percpu *rc_delayed_comp; 130 131 void *priv; /* driver private data */ 132 133 /* 134 * The pkey table is allocated and maintained by the driver. Drivers 135 * need to have access to this before registering with rdmav. However 136 * rdmavt will need access to it so drivers need to proviee this during 137 * the attach port API call. 138 */ 139 u16 *pkey_table; 140 141 struct rvt_ah *sm_ah; 142 143 /* 144 * Keep a list of traps that have not been repressed. They will be 145 * resent based on trap_timer. 146 */ 147 struct trap_list trap_lists[RVT_MAX_TRAP_LISTS]; 148 struct timer_list trap_timer; 149 }; 150 151 #define RVT_CQN_MAX 16 /* maximum length of cq name */ 152 153 #define RVT_SGE_COPY_MEMCPY 0 154 #define RVT_SGE_COPY_CACHELESS 1 155 #define RVT_SGE_COPY_ADAPTIVE 2 156 157 /* 158 * Things that are driver specific, module parameters in hfi1 and qib 159 */ 160 struct rvt_driver_params { 161 struct ib_device_attr props; 162 163 /* 164 * Anything driver specific that is not covered by props 165 * For instance special module parameters. Goes here. 166 */ 167 unsigned int lkey_table_size; 168 unsigned int qp_table_size; 169 unsigned int sge_copy_mode; 170 unsigned int wss_threshold; 171 unsigned int wss_clean_period; 172 int qpn_start; 173 int qpn_inc; 174 int qpn_res_start; 175 int qpn_res_end; 176 int nports; 177 int npkeys; 178 int node; 179 int psn_mask; 180 int psn_shift; 181 int psn_modify_mask; 182 u32 core_cap_flags; 183 u32 max_mad_size; 184 u8 qos_shift; 185 u8 max_rdma_atomic; 186 u8 extra_rdma_atomic; 187 u8 reserved_operations; 188 }; 189 190 /* User context */ 191 struct rvt_ucontext { 192 struct ib_ucontext ibucontext; 193 }; 194 195 /* Protection domain */ 196 struct rvt_pd { 197 struct ib_pd ibpd; 198 bool user; 199 }; 200 201 /* Address handle */ 202 struct rvt_ah { 203 struct ib_ah ibah; 204 struct rdma_ah_attr attr; 205 u8 vl; 206 u8 log_pmtu; 207 }; 208 209 /* 210 * This structure is used by rvt_mmap() to validate an offset 211 * when an mmap() request is made. The vm_area_struct then uses 212 * this as its vm_private_data. 213 */ 214 struct rvt_mmap_info { 215 struct list_head pending_mmaps; 216 struct ib_ucontext *context; 217 void *obj; 218 __u64 offset; 219 struct kref ref; 220 u32 size; 221 }; 222 223 /* memory working set size */ 224 struct rvt_wss { 225 unsigned long *entries; 226 atomic_t total_count; 227 atomic_t clean_counter; 228 atomic_t clean_entry; 229 230 int threshold; 231 int num_entries; 232 long pages_mask; 233 unsigned int clean_period; 234 }; 235 236 struct rvt_dev_info; 237 struct rvt_swqe; 238 struct rvt_driver_provided { 239 /* 240 * Which functions are required depends on which verbs rdmavt is 241 * providing and which verbs the driver is overriding. See 242 * check_support() for details. 243 */ 244 245 /* hot path calldowns in a single cacheline */ 246 247 /* 248 * Give the driver a notice that there is send work to do. It is up to 249 * the driver to generally push the packets out, this just queues the 250 * work with the driver. There are two variants here. The no_lock 251 * version requires the s_lock not to be held. The other assumes the 252 * s_lock is held. 253 */ 254 bool (*schedule_send)(struct rvt_qp *qp); 255 bool (*schedule_send_no_lock)(struct rvt_qp *qp); 256 257 /* 258 * Driver specific work request setup and checking. 259 * This function is allowed to perform any setup, checks, or 260 * adjustments required to the SWQE in order to be usable by 261 * underlying protocols. This includes private data structure 262 * allocations. 263 */ 264 int (*setup_wqe)(struct rvt_qp *qp, struct rvt_swqe *wqe, 265 bool *call_send); 266 267 /* 268 * Sometimes rdmavt needs to kick the driver's send progress. That is 269 * done by this call back. 270 */ 271 void (*do_send)(struct rvt_qp *qp); 272 273 /* 274 * Returns a pointer to the undelying hardware's PCI device. This is 275 * used to display information as to what hardware is being referenced 276 * in an output message 277 */ 278 struct pci_dev * (*get_pci_dev)(struct rvt_dev_info *rdi); 279 280 /* 281 * Allocate a private queue pair data structure for driver specific 282 * information which is opaque to rdmavt. Errors are returned via 283 * ERR_PTR(err). The driver is free to return NULL or a valid 284 * pointer. 285 */ 286 void * (*qp_priv_alloc)(struct rvt_dev_info *rdi, struct rvt_qp *qp); 287 288 /* 289 * Init a struture allocated with qp_priv_alloc(). This should be 290 * called after all qp fields have been initialized in rdmavt. 291 */ 292 int (*qp_priv_init)(struct rvt_dev_info *rdi, struct rvt_qp *qp, 293 struct ib_qp_init_attr *init_attr); 294 295 /* 296 * Free the driver's private qp structure. 297 */ 298 void (*qp_priv_free)(struct rvt_dev_info *rdi, struct rvt_qp *qp); 299 300 /* 301 * Inform the driver the particular qp in quesiton has been reset so 302 * that it can clean up anything it needs to. 303 */ 304 void (*notify_qp_reset)(struct rvt_qp *qp); 305 306 /* 307 * Get a path mtu from the driver based on qp attributes. 308 */ 309 int (*get_pmtu_from_attr)(struct rvt_dev_info *rdi, struct rvt_qp *qp, 310 struct ib_qp_attr *attr); 311 312 /* 313 * Notify driver that it needs to flush any outstanding IO requests that 314 * are waiting on a qp. 315 */ 316 void (*flush_qp_waiters)(struct rvt_qp *qp); 317 318 /* 319 * Notify driver to stop its queue of sending packets. Nothing else 320 * should be posted to the queue pair after this has been called. 321 */ 322 void (*stop_send_queue)(struct rvt_qp *qp); 323 324 /* 325 * Have the drivr drain any in progress operations 326 */ 327 void (*quiesce_qp)(struct rvt_qp *qp); 328 329 /* 330 * Inform the driver a qp has went to error state. 331 */ 332 void (*notify_error_qp)(struct rvt_qp *qp); 333 334 /* 335 * Get an MTU for a qp. 336 */ 337 u32 (*mtu_from_qp)(struct rvt_dev_info *rdi, struct rvt_qp *qp, 338 u32 pmtu); 339 /* 340 * Convert an mtu to a path mtu 341 */ 342 int (*mtu_to_path_mtu)(u32 mtu); 343 344 /* 345 * Get the guid of a port in big endian byte order 346 */ 347 int (*get_guid_be)(struct rvt_dev_info *rdi, struct rvt_ibport *rvp, 348 int guid_index, __be64 *guid); 349 350 /* 351 * Query driver for the state of the port. 352 */ 353 int (*query_port_state)(struct rvt_dev_info *rdi, u8 port_num, 354 struct ib_port_attr *props); 355 356 /* 357 * Tell driver to shutdown a port 358 */ 359 int (*shut_down_port)(struct rvt_dev_info *rdi, u8 port_num); 360 361 /* Tell driver to send a trap for changed port capabilities */ 362 void (*cap_mask_chg)(struct rvt_dev_info *rdi, u8 port_num); 363 364 /* 365 * The following functions can be safely ignored completely. Any use of 366 * these is checked for NULL before blindly calling. Rdmavt should also 367 * be functional if drivers omit these. 368 */ 369 370 /* Called to inform the driver that all qps should now be freed. */ 371 unsigned (*free_all_qps)(struct rvt_dev_info *rdi); 372 373 /* Driver specific AH validation */ 374 int (*check_ah)(struct ib_device *, struct rdma_ah_attr *); 375 376 /* Inform the driver a new AH has been created */ 377 void (*notify_new_ah)(struct ib_device *, struct rdma_ah_attr *, 378 struct rvt_ah *); 379 380 /* Let the driver pick the next queue pair number*/ 381 int (*alloc_qpn)(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt, 382 enum ib_qp_type type, u8 port_num); 383 384 /* Determine if its safe or allowed to modify the qp */ 385 int (*check_modify_qp)(struct rvt_qp *qp, struct ib_qp_attr *attr, 386 int attr_mask, struct ib_udata *udata); 387 388 /* Driver specific QP modification/notification-of */ 389 void (*modify_qp)(struct rvt_qp *qp, struct ib_qp_attr *attr, 390 int attr_mask, struct ib_udata *udata); 391 392 /* Notify driver a mad agent has been created */ 393 void (*notify_create_mad_agent)(struct rvt_dev_info *rdi, int port_idx); 394 395 /* Notify driver a mad agent has been removed */ 396 void (*notify_free_mad_agent)(struct rvt_dev_info *rdi, int port_idx); 397 398 /* Notify driver to restart rc */ 399 void (*notify_restart_rc)(struct rvt_qp *qp, u32 psn, int wait); 400 401 /* Get and return CPU to pin CQ processing thread */ 402 int (*comp_vect_cpu_lookup)(struct rvt_dev_info *rdi, int comp_vect); 403 }; 404 405 struct rvt_dev_info { 406 struct ib_device ibdev; /* Keep this first. Nothing above here */ 407 408 /* 409 * Prior to calling for registration the driver will be responsible for 410 * allocating space for this structure. 411 * 412 * The driver will also be responsible for filling in certain members of 413 * dparms.props. The driver needs to fill in dparms exactly as it would 414 * want values reported to a ULP. This will be returned to the caller 415 * in rdmavt's device. The driver should also therefore refrain from 416 * modifying this directly after registration with rdmavt. 417 */ 418 419 /* Driver specific properties */ 420 struct rvt_driver_params dparms; 421 422 /* post send table */ 423 const struct rvt_operation_params *post_parms; 424 425 /* opcode translation table */ 426 const enum ib_wc_opcode *wc_opcode; 427 428 /* Driver specific helper functions */ 429 struct rvt_driver_provided driver_f; 430 431 struct rvt_mregion __rcu *dma_mr; 432 struct rvt_lkey_table lkey_table; 433 434 /* Internal use */ 435 int n_pds_allocated; 436 spinlock_t n_pds_lock; /* Protect pd allocated count */ 437 438 int n_ahs_allocated; 439 spinlock_t n_ahs_lock; /* Protect ah allocated count */ 440 441 u32 n_srqs_allocated; 442 spinlock_t n_srqs_lock; /* Protect srqs allocated count */ 443 444 int flags; 445 struct rvt_ibport **ports; 446 447 /* QP */ 448 struct rvt_qp_ibdev *qp_dev; 449 u32 n_qps_allocated; /* number of QPs allocated for device */ 450 u32 n_rc_qps; /* number of RC QPs allocated for device */ 451 u32 busy_jiffies; /* timeout scaling based on RC QP count */ 452 spinlock_t n_qps_lock; /* protect qps, rc qps and busy jiffy counts */ 453 454 /* memory maps */ 455 struct list_head pending_mmaps; 456 spinlock_t mmap_offset_lock; /* protect mmap_offset */ 457 u32 mmap_offset; 458 spinlock_t pending_lock; /* protect pending mmap list */ 459 460 /* CQ */ 461 u32 n_cqs_allocated; /* number of CQs allocated for device */ 462 spinlock_t n_cqs_lock; /* protect count of in use cqs */ 463 464 /* Multicast */ 465 u32 n_mcast_grps_allocated; /* number of mcast groups allocated */ 466 spinlock_t n_mcast_grps_lock; 467 468 /* Memory Working Set Size */ 469 struct rvt_wss *wss; 470 }; 471 472 /** 473 * rvt_set_ibdev_name - Craft an IB device name from client info 474 * @rdi: pointer to the client rvt_dev_info structure 475 * @name: client specific name 476 * @unit: client specific unit number. 477 */ 478 static inline void rvt_set_ibdev_name(struct rvt_dev_info *rdi, 479 const char *fmt, const char *name, 480 const int unit) 481 { 482 /* 483 * FIXME: rvt and its users want to touch the ibdev before 484 * registration and have things like the name work. We don't have the 485 * infrastructure in the core to support this directly today, hack it 486 * to work by setting the name manually here. 487 */ 488 dev_set_name(&rdi->ibdev.dev, fmt, name, unit); 489 strlcpy(rdi->ibdev.name, dev_name(&rdi->ibdev.dev), IB_DEVICE_NAME_MAX); 490 } 491 492 /** 493 * rvt_get_ibdev_name - return the IB name 494 * @rdi: rdmavt device 495 * 496 * Return the registered name of the device. 497 */ 498 static inline const char *rvt_get_ibdev_name(const struct rvt_dev_info *rdi) 499 { 500 return dev_name(&rdi->ibdev.dev); 501 } 502 503 static inline struct rvt_pd *ibpd_to_rvtpd(struct ib_pd *ibpd) 504 { 505 return container_of(ibpd, struct rvt_pd, ibpd); 506 } 507 508 static inline struct rvt_ah *ibah_to_rvtah(struct ib_ah *ibah) 509 { 510 return container_of(ibah, struct rvt_ah, ibah); 511 } 512 513 static inline struct rvt_dev_info *ib_to_rvt(struct ib_device *ibdev) 514 { 515 return container_of(ibdev, struct rvt_dev_info, ibdev); 516 } 517 518 static inline unsigned rvt_get_npkeys(struct rvt_dev_info *rdi) 519 { 520 /* 521 * All ports have same number of pkeys. 522 */ 523 return rdi->dparms.npkeys; 524 } 525 526 /* 527 * Return the max atomic suitable for determining 528 * the size of the ack ring buffer in a QP. 529 */ 530 static inline unsigned int rvt_max_atomic(struct rvt_dev_info *rdi) 531 { 532 return rdi->dparms.max_rdma_atomic + 533 rdi->dparms.extra_rdma_atomic + 1; 534 } 535 536 static inline unsigned int rvt_size_atomic(struct rvt_dev_info *rdi) 537 { 538 return rdi->dparms.max_rdma_atomic + 539 rdi->dparms.extra_rdma_atomic; 540 } 541 542 /* 543 * Return the indexed PKEY from the port PKEY table. 544 */ 545 static inline u16 rvt_get_pkey(struct rvt_dev_info *rdi, 546 int port_index, 547 unsigned index) 548 { 549 if (index >= rvt_get_npkeys(rdi)) 550 return 0; 551 else 552 return rdi->ports[port_index]->pkey_table[index]; 553 } 554 555 struct rvt_dev_info *rvt_alloc_device(size_t size, int nports); 556 void rvt_dealloc_device(struct rvt_dev_info *rdi); 557 int rvt_register_device(struct rvt_dev_info *rvd); 558 void rvt_unregister_device(struct rvt_dev_info *rvd); 559 int rvt_check_ah(struct ib_device *ibdev, struct rdma_ah_attr *ah_attr); 560 int rvt_init_port(struct rvt_dev_info *rdi, struct rvt_ibport *port, 561 int port_index, u16 *pkey_table); 562 int rvt_fast_reg_mr(struct rvt_qp *qp, struct ib_mr *ibmr, u32 key, 563 int access); 564 int rvt_invalidate_rkey(struct rvt_qp *qp, u32 rkey); 565 int rvt_rkey_ok(struct rvt_qp *qp, struct rvt_sge *sge, 566 u32 len, u64 vaddr, u32 rkey, int acc); 567 int rvt_lkey_ok(struct rvt_lkey_table *rkt, struct rvt_pd *pd, 568 struct rvt_sge *isge, struct rvt_sge *last_sge, 569 struct ib_sge *sge, int acc); 570 struct rvt_mcast *rvt_mcast_find(struct rvt_ibport *ibp, union ib_gid *mgid, 571 u16 lid); 572 573 #endif /* DEF_RDMA_VT_H */ 574