1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause 2 3 /* Authors: Bernard Metzler <bmt@zurich.ibm.com> */ 4 /* Copyright (c) 2008-2019, IBM Corporation */ 5 6 #include <linux/init.h> 7 #include <linux/errno.h> 8 #include <linux/netdevice.h> 9 #include <linux/inetdevice.h> 10 #include <net/net_namespace.h> 11 #include <linux/rtnetlink.h> 12 #include <linux/if_arp.h> 13 #include <linux/list.h> 14 #include <linux/kernel.h> 15 #include <linux/sched.h> 16 #include <linux/module.h> 17 #include <linux/dma-mapping.h> 18 19 #include <net/addrconf.h> 20 #include <rdma/ib_verbs.h> 21 #include <rdma/ib_user_verbs.h> 22 #include <rdma/rdma_netlink.h> 23 #include <linux/kthread.h> 24 25 #include "siw.h" 26 #include "siw_verbs.h" 27 28 MODULE_AUTHOR("Bernard Metzler"); 29 MODULE_DESCRIPTION("Software iWARP Driver"); 30 MODULE_LICENSE("Dual BSD/GPL"); 31 32 /* transmit from user buffer, if possible */ 33 const bool zcopy_tx = true; 34 35 /* Restrict usage of GSO, if hardware peer iwarp is unable to process 36 * large packets. try_gso = true lets siw try to use local GSO, 37 * if peer agrees. Not using GSO severly limits siw maximum tx bandwidth. 38 */ 39 const bool try_gso; 40 41 /* Attach siw also with loopback devices */ 42 const bool loopback_enabled = true; 43 44 /* We try to negotiate CRC on, if true */ 45 const bool mpa_crc_required; 46 47 /* MPA CRC on/off enforced */ 48 const bool mpa_crc_strict; 49 50 /* Control TCP_NODELAY socket option */ 51 const bool siw_tcp_nagle; 52 53 /* Select MPA version to be used during connection setup */ 54 u_char mpa_version = MPA_REVISION_2; 55 56 /* Selects MPA P2P mode (additional handshake during connection 57 * setup, if true. 58 */ 59 const bool peer_to_peer; 60 61 struct task_struct *siw_tx_thread[NR_CPUS]; 62 struct crypto_shash *siw_crypto_shash; 63 64 static int siw_device_register(struct siw_device *sdev, const char *name) 65 { 66 struct ib_device *base_dev = &sdev->base_dev; 67 static int dev_id = 1; 68 int rv; 69 70 sdev->vendor_part_id = dev_id++; 71 72 rv = ib_register_device(base_dev, name, NULL); 73 if (rv) { 74 pr_warn("siw: device registration error %d\n", rv); 75 return rv; 76 } 77 78 siw_dbg(base_dev, "HWaddr=%pM\n", sdev->raw_gid); 79 return 0; 80 } 81 82 static void siw_device_cleanup(struct ib_device *base_dev) 83 { 84 struct siw_device *sdev = to_siw_dev(base_dev); 85 86 xa_destroy(&sdev->qp_xa); 87 xa_destroy(&sdev->mem_xa); 88 } 89 90 static int siw_dev_qualified(struct net_device *netdev) 91 { 92 /* 93 * Additional hardware support can be added here 94 * (e.g. ARPHRD_FDDI, ARPHRD_ATM, ...) - see 95 * <linux/if_arp.h> for type identifiers. 96 */ 97 if (netdev->type == ARPHRD_ETHER || netdev->type == ARPHRD_IEEE802 || 98 netdev->type == ARPHRD_NONE || 99 (netdev->type == ARPHRD_LOOPBACK && loopback_enabled)) 100 return 1; 101 102 return 0; 103 } 104 105 static DEFINE_PER_CPU(atomic_t, siw_use_cnt); 106 107 static struct { 108 struct cpumask **tx_valid_cpus; 109 int num_nodes; 110 } siw_cpu_info; 111 112 static void siw_destroy_cpulist(int number) 113 { 114 int i = 0; 115 116 while (i < number) 117 kfree(siw_cpu_info.tx_valid_cpus[i++]); 118 119 kfree(siw_cpu_info.tx_valid_cpus); 120 siw_cpu_info.tx_valid_cpus = NULL; 121 } 122 123 static int siw_init_cpulist(void) 124 { 125 int i, num_nodes = nr_node_ids; 126 127 memset(siw_tx_thread, 0, sizeof(siw_tx_thread)); 128 129 siw_cpu_info.num_nodes = num_nodes; 130 131 siw_cpu_info.tx_valid_cpus = 132 kcalloc(num_nodes, sizeof(struct cpumask *), GFP_KERNEL); 133 if (!siw_cpu_info.tx_valid_cpus) { 134 siw_cpu_info.num_nodes = 0; 135 return -ENOMEM; 136 } 137 for (i = 0; i < siw_cpu_info.num_nodes; i++) { 138 siw_cpu_info.tx_valid_cpus[i] = 139 kzalloc(sizeof(struct cpumask), GFP_KERNEL); 140 if (!siw_cpu_info.tx_valid_cpus[i]) 141 goto out_err; 142 143 cpumask_clear(siw_cpu_info.tx_valid_cpus[i]); 144 } 145 for_each_possible_cpu(i) 146 cpumask_set_cpu(i, siw_cpu_info.tx_valid_cpus[cpu_to_node(i)]); 147 148 return 0; 149 150 out_err: 151 siw_cpu_info.num_nodes = 0; 152 siw_destroy_cpulist(i); 153 154 return -ENOMEM; 155 } 156 157 /* 158 * Choose CPU with least number of active QP's from NUMA node of 159 * TX interface. 160 */ 161 int siw_get_tx_cpu(struct siw_device *sdev) 162 { 163 const struct cpumask *tx_cpumask; 164 int i, num_cpus, cpu, min_use, node = sdev->numa_node, tx_cpu = -1; 165 166 if (node < 0) 167 tx_cpumask = cpu_online_mask; 168 else 169 tx_cpumask = siw_cpu_info.tx_valid_cpus[node]; 170 171 num_cpus = cpumask_weight(tx_cpumask); 172 if (!num_cpus) { 173 /* no CPU on this NUMA node */ 174 tx_cpumask = cpu_online_mask; 175 num_cpus = cpumask_weight(tx_cpumask); 176 } 177 if (!num_cpus) 178 goto out; 179 180 cpu = cpumask_first(tx_cpumask); 181 182 for (i = 0, min_use = SIW_MAX_QP; i < num_cpus; 183 i++, cpu = cpumask_next(cpu, tx_cpumask)) { 184 int usage; 185 186 /* Skip any cores which have no TX thread */ 187 if (!siw_tx_thread[cpu]) 188 continue; 189 190 usage = atomic_read(&per_cpu(siw_use_cnt, cpu)); 191 if (usage <= min_use) { 192 tx_cpu = cpu; 193 min_use = usage; 194 } 195 } 196 siw_dbg(&sdev->base_dev, 197 "tx cpu %d, node %d, %d qp's\n", tx_cpu, node, min_use); 198 199 out: 200 if (tx_cpu >= 0) 201 atomic_inc(&per_cpu(siw_use_cnt, tx_cpu)); 202 else 203 pr_warn("siw: no tx cpu found\n"); 204 205 return tx_cpu; 206 } 207 208 void siw_put_tx_cpu(int cpu) 209 { 210 atomic_dec(&per_cpu(siw_use_cnt, cpu)); 211 } 212 213 static struct ib_qp *siw_get_base_qp(struct ib_device *base_dev, int id) 214 { 215 struct siw_qp *qp = siw_qp_id2obj(to_siw_dev(base_dev), id); 216 217 if (qp) { 218 /* 219 * siw_qp_id2obj() increments object reference count 220 */ 221 siw_qp_put(qp); 222 return &qp->base_qp; 223 } 224 return NULL; 225 } 226 227 static const struct ib_device_ops siw_device_ops = { 228 .owner = THIS_MODULE, 229 .uverbs_abi_ver = SIW_ABI_VERSION, 230 .driver_id = RDMA_DRIVER_SIW, 231 232 .alloc_mr = siw_alloc_mr, 233 .alloc_pd = siw_alloc_pd, 234 .alloc_ucontext = siw_alloc_ucontext, 235 .create_cq = siw_create_cq, 236 .create_qp = siw_create_qp, 237 .create_srq = siw_create_srq, 238 .dealloc_driver = siw_device_cleanup, 239 .dealloc_pd = siw_dealloc_pd, 240 .dealloc_ucontext = siw_dealloc_ucontext, 241 .dereg_mr = siw_dereg_mr, 242 .destroy_cq = siw_destroy_cq, 243 .destroy_qp = siw_destroy_qp, 244 .destroy_srq = siw_destroy_srq, 245 .get_dma_mr = siw_get_dma_mr, 246 .get_port_immutable = siw_get_port_immutable, 247 .iw_accept = siw_accept, 248 .iw_add_ref = siw_qp_get_ref, 249 .iw_connect = siw_connect, 250 .iw_create_listen = siw_create_listen, 251 .iw_destroy_listen = siw_destroy_listen, 252 .iw_get_qp = siw_get_base_qp, 253 .iw_reject = siw_reject, 254 .iw_rem_ref = siw_qp_put_ref, 255 .map_mr_sg = siw_map_mr_sg, 256 .mmap = siw_mmap, 257 .mmap_free = siw_mmap_free, 258 .modify_qp = siw_verbs_modify_qp, 259 .modify_srq = siw_modify_srq, 260 .poll_cq = siw_poll_cq, 261 .post_recv = siw_post_receive, 262 .post_send = siw_post_send, 263 .post_srq_recv = siw_post_srq_recv, 264 .query_device = siw_query_device, 265 .query_gid = siw_query_gid, 266 .query_port = siw_query_port, 267 .query_qp = siw_query_qp, 268 .query_srq = siw_query_srq, 269 .req_notify_cq = siw_req_notify_cq, 270 .reg_user_mr = siw_reg_user_mr, 271 272 INIT_RDMA_OBJ_SIZE(ib_cq, siw_cq, base_cq), 273 INIT_RDMA_OBJ_SIZE(ib_pd, siw_pd, base_pd), 274 INIT_RDMA_OBJ_SIZE(ib_qp, siw_qp, base_qp), 275 INIT_RDMA_OBJ_SIZE(ib_srq, siw_srq, base_srq), 276 INIT_RDMA_OBJ_SIZE(ib_ucontext, siw_ucontext, base_ucontext), 277 }; 278 279 static struct siw_device *siw_device_create(struct net_device *netdev) 280 { 281 struct siw_device *sdev = NULL; 282 struct ib_device *base_dev; 283 int rv; 284 285 sdev = ib_alloc_device(siw_device, base_dev); 286 if (!sdev) 287 return NULL; 288 289 base_dev = &sdev->base_dev; 290 sdev->netdev = netdev; 291 292 if (netdev->addr_len) { 293 memcpy(sdev->raw_gid, netdev->dev_addr, 294 min_t(unsigned int, netdev->addr_len, ETH_ALEN)); 295 } else { 296 /* 297 * This device does not have a HW address, but 298 * connection mangagement requires a unique gid. 299 */ 300 eth_random_addr(sdev->raw_gid); 301 } 302 addrconf_addr_eui48((u8 *)&base_dev->node_guid, sdev->raw_gid); 303 304 base_dev->uverbs_cmd_mask |= BIT_ULL(IB_USER_VERBS_CMD_POST_SEND); 305 306 base_dev->node_type = RDMA_NODE_RNIC; 307 memcpy(base_dev->node_desc, SIW_NODE_DESC_COMMON, 308 sizeof(SIW_NODE_DESC_COMMON)); 309 310 /* 311 * Current model (one-to-one device association): 312 * One Softiwarp device per net_device or, equivalently, 313 * per physical port. 314 */ 315 base_dev->phys_port_cnt = 1; 316 base_dev->num_comp_vectors = num_possible_cpus(); 317 318 xa_init_flags(&sdev->qp_xa, XA_FLAGS_ALLOC1); 319 xa_init_flags(&sdev->mem_xa, XA_FLAGS_ALLOC1); 320 321 ib_set_device_ops(base_dev, &siw_device_ops); 322 rv = ib_device_set_netdev(base_dev, netdev, 1); 323 if (rv) 324 goto error; 325 326 memcpy(base_dev->iw_ifname, netdev->name, 327 sizeof(base_dev->iw_ifname)); 328 329 /* Disable TCP port mapping */ 330 base_dev->iw_driver_flags = IW_F_NO_PORT_MAP; 331 332 sdev->attrs.max_qp = SIW_MAX_QP; 333 sdev->attrs.max_qp_wr = SIW_MAX_QP_WR; 334 sdev->attrs.max_ord = SIW_MAX_ORD_QP; 335 sdev->attrs.max_ird = SIW_MAX_IRD_QP; 336 sdev->attrs.max_sge = SIW_MAX_SGE; 337 sdev->attrs.max_sge_rd = SIW_MAX_SGE_RD; 338 sdev->attrs.max_cq = SIW_MAX_CQ; 339 sdev->attrs.max_cqe = SIW_MAX_CQE; 340 sdev->attrs.max_mr = SIW_MAX_MR; 341 sdev->attrs.max_pd = SIW_MAX_PD; 342 sdev->attrs.max_mw = SIW_MAX_MW; 343 sdev->attrs.max_srq = SIW_MAX_SRQ; 344 sdev->attrs.max_srq_wr = SIW_MAX_SRQ_WR; 345 sdev->attrs.max_srq_sge = SIW_MAX_SGE; 346 347 INIT_LIST_HEAD(&sdev->cep_list); 348 INIT_LIST_HEAD(&sdev->qp_list); 349 350 atomic_set(&sdev->num_ctx, 0); 351 atomic_set(&sdev->num_srq, 0); 352 atomic_set(&sdev->num_qp, 0); 353 atomic_set(&sdev->num_cq, 0); 354 atomic_set(&sdev->num_mr, 0); 355 atomic_set(&sdev->num_pd, 0); 356 357 sdev->numa_node = dev_to_node(&netdev->dev); 358 spin_lock_init(&sdev->lock); 359 360 return sdev; 361 error: 362 ib_dealloc_device(base_dev); 363 364 return NULL; 365 } 366 367 /* 368 * Network link becomes unavailable. Mark all 369 * affected QP's accordingly. 370 */ 371 static void siw_netdev_down(struct work_struct *work) 372 { 373 struct siw_device *sdev = 374 container_of(work, struct siw_device, netdev_down); 375 376 struct siw_qp_attrs qp_attrs; 377 struct list_head *pos, *tmp; 378 379 memset(&qp_attrs, 0, sizeof(qp_attrs)); 380 qp_attrs.state = SIW_QP_STATE_ERROR; 381 382 list_for_each_safe(pos, tmp, &sdev->qp_list) { 383 struct siw_qp *qp = list_entry(pos, struct siw_qp, devq); 384 385 down_write(&qp->state_lock); 386 WARN_ON(siw_qp_modify(qp, &qp_attrs, SIW_QP_ATTR_STATE)); 387 up_write(&qp->state_lock); 388 } 389 ib_device_put(&sdev->base_dev); 390 } 391 392 static void siw_device_goes_down(struct siw_device *sdev) 393 { 394 if (ib_device_try_get(&sdev->base_dev)) { 395 INIT_WORK(&sdev->netdev_down, siw_netdev_down); 396 schedule_work(&sdev->netdev_down); 397 } 398 } 399 400 static int siw_netdev_event(struct notifier_block *nb, unsigned long event, 401 void *arg) 402 { 403 struct net_device *netdev = netdev_notifier_info_to_dev(arg); 404 struct ib_device *base_dev; 405 struct siw_device *sdev; 406 407 dev_dbg(&netdev->dev, "siw: event %lu\n", event); 408 409 base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW); 410 if (!base_dev) 411 return NOTIFY_OK; 412 413 sdev = to_siw_dev(base_dev); 414 415 switch (event) { 416 case NETDEV_UP: 417 sdev->state = IB_PORT_ACTIVE; 418 siw_port_event(sdev, 1, IB_EVENT_PORT_ACTIVE); 419 break; 420 421 case NETDEV_GOING_DOWN: 422 siw_device_goes_down(sdev); 423 break; 424 425 case NETDEV_DOWN: 426 sdev->state = IB_PORT_DOWN; 427 siw_port_event(sdev, 1, IB_EVENT_PORT_ERR); 428 break; 429 430 case NETDEV_REGISTER: 431 /* 432 * Device registration now handled only by 433 * rdma netlink commands. So it shall be impossible 434 * to end up here with a valid siw device. 435 */ 436 siw_dbg(base_dev, "unexpected NETDEV_REGISTER event\n"); 437 break; 438 439 case NETDEV_UNREGISTER: 440 ib_unregister_device_queued(&sdev->base_dev); 441 break; 442 443 case NETDEV_CHANGEADDR: 444 siw_port_event(sdev, 1, IB_EVENT_LID_CHANGE); 445 break; 446 /* 447 * Todo: Below netdev events are currently not handled. 448 */ 449 case NETDEV_CHANGEMTU: 450 case NETDEV_CHANGE: 451 break; 452 453 default: 454 break; 455 } 456 ib_device_put(&sdev->base_dev); 457 458 return NOTIFY_OK; 459 } 460 461 static struct notifier_block siw_netdev_nb = { 462 .notifier_call = siw_netdev_event, 463 }; 464 465 static int siw_newlink(const char *basedev_name, struct net_device *netdev) 466 { 467 struct ib_device *base_dev; 468 struct siw_device *sdev = NULL; 469 int rv = -ENOMEM; 470 471 if (!siw_dev_qualified(netdev)) 472 return -EINVAL; 473 474 base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW); 475 if (base_dev) { 476 ib_device_put(base_dev); 477 return -EEXIST; 478 } 479 sdev = siw_device_create(netdev); 480 if (sdev) { 481 dev_dbg(&netdev->dev, "siw: new device\n"); 482 483 if (netif_running(netdev) && netif_carrier_ok(netdev)) 484 sdev->state = IB_PORT_ACTIVE; 485 else 486 sdev->state = IB_PORT_DOWN; 487 488 rv = siw_device_register(sdev, basedev_name); 489 if (rv) 490 ib_dealloc_device(&sdev->base_dev); 491 } 492 return rv; 493 } 494 495 static struct rdma_link_ops siw_link_ops = { 496 .type = "siw", 497 .newlink = siw_newlink, 498 }; 499 500 /* 501 * siw_init_module - Initialize Softiwarp module and register with netdev 502 * subsystem. 503 */ 504 static __init int siw_init_module(void) 505 { 506 int rv; 507 508 if (SENDPAGE_THRESH < SIW_MAX_INLINE) { 509 pr_info("siw: sendpage threshold too small: %u\n", 510 (int)SENDPAGE_THRESH); 511 rv = -EINVAL; 512 goto out_error; 513 } 514 rv = siw_init_cpulist(); 515 if (rv) 516 goto out_error; 517 518 rv = siw_cm_init(); 519 if (rv) 520 goto out_error; 521 522 if (!siw_create_tx_threads()) { 523 pr_info("siw: Could not start any TX thread\n"); 524 rv = -ENOMEM; 525 goto out_error; 526 } 527 /* 528 * Locate CRC32 algorithm. If unsuccessful, fail 529 * loading siw only, if CRC is required. 530 */ 531 siw_crypto_shash = crypto_alloc_shash("crc32c", 0, 0); 532 if (IS_ERR(siw_crypto_shash)) { 533 pr_info("siw: Loading CRC32c failed: %ld\n", 534 PTR_ERR(siw_crypto_shash)); 535 siw_crypto_shash = NULL; 536 if (mpa_crc_required) { 537 rv = -EOPNOTSUPP; 538 goto out_error; 539 } 540 } 541 rv = register_netdevice_notifier(&siw_netdev_nb); 542 if (rv) 543 goto out_error; 544 545 rdma_link_register(&siw_link_ops); 546 547 pr_info("SoftiWARP attached\n"); 548 return 0; 549 550 out_error: 551 siw_stop_tx_threads(); 552 553 if (siw_crypto_shash) 554 crypto_free_shash(siw_crypto_shash); 555 556 pr_info("SoftIWARP attach failed. Error: %d\n", rv); 557 558 siw_cm_exit(); 559 siw_destroy_cpulist(siw_cpu_info.num_nodes); 560 561 return rv; 562 } 563 564 static void __exit siw_exit_module(void) 565 { 566 siw_stop_tx_threads(); 567 568 unregister_netdevice_notifier(&siw_netdev_nb); 569 rdma_link_unregister(&siw_link_ops); 570 ib_unregister_driver(RDMA_DRIVER_SIW); 571 572 siw_cm_exit(); 573 574 siw_destroy_cpulist(siw_cpu_info.num_nodes); 575 576 if (siw_crypto_shash) 577 crypto_free_shash(siw_crypto_shash); 578 579 pr_info("SoftiWARP detached\n"); 580 } 581 582 module_init(siw_init_module); 583 module_exit(siw_exit_module); 584 585 MODULE_ALIAS_RDMA_LINK("siw"); 586