1 /* 2 * Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved. 3 * 4 * This software is available to you under a choice of one of two 5 * licenses. You may choose to be licensed under the terms of the GNU 6 * General Public License (GPL) Version 2, available from the file 7 * COPYING in the main directory of this source tree, or the 8 * OpenIB.org BSD license below: 9 * 10 * Redistribution and use in source and binary forms, with or 11 * without modification, are permitted provided that the following 12 * conditions are met: 13 * 14 * - Redistributions of source code must retain the above 15 * copyright notice, this list of conditions and the following 16 * disclaimer. 17 * 18 * - Redistributions in binary form must reproduce the above 19 * copyright notice, this list of conditions and the following 20 * disclaimer in the documentation and/or other materials 21 * provided with the distribution. 22 * 23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 30 * SOFTWARE. 31 * 32 */ 33 #include <linux/kernel.h> 34 #include <linux/in.h> 35 #include <linux/if.h> 36 #include <linux/netdevice.h> 37 #include <linux/inetdevice.h> 38 #include <linux/if_arp.h> 39 #include <linux/delay.h> 40 #include <linux/slab.h> 41 #include <linux/module.h> 42 #include <net/addrconf.h> 43 44 #include "rds_single_path.h" 45 #include "rds.h" 46 #include "ib.h" 47 #include "ib_mr.h" 48 49 static unsigned int rds_ib_mr_1m_pool_size = RDS_MR_1M_POOL_SIZE; 50 static unsigned int rds_ib_mr_8k_pool_size = RDS_MR_8K_POOL_SIZE; 51 unsigned int rds_ib_retry_count = RDS_IB_DEFAULT_RETRY_COUNT; 52 static atomic_t rds_ib_unloading; 53 54 module_param(rds_ib_mr_1m_pool_size, int, 0444); 55 MODULE_PARM_DESC(rds_ib_mr_1m_pool_size, " Max number of 1M mr per HCA"); 56 module_param(rds_ib_mr_8k_pool_size, int, 0444); 57 MODULE_PARM_DESC(rds_ib_mr_8k_pool_size, " Max number of 8K mr per HCA"); 58 module_param(rds_ib_retry_count, int, 0444); 59 MODULE_PARM_DESC(rds_ib_retry_count, " Number of hw retries before reporting an error"); 60 61 /* 62 * we have a clumsy combination of RCU and a rwsem protecting this list 63 * because it is used both in the get_mr fast path and while blocking in 64 * the FMR flushing path. 65 */ 66 DECLARE_RWSEM(rds_ib_devices_lock); 67 struct list_head rds_ib_devices; 68 69 /* NOTE: if also grabbing ibdev lock, grab this first */ 70 DEFINE_SPINLOCK(ib_nodev_conns_lock); 71 LIST_HEAD(ib_nodev_conns); 72 73 static void rds_ib_nodev_connect(void) 74 { 75 struct rds_ib_connection *ic; 76 77 spin_lock(&ib_nodev_conns_lock); 78 list_for_each_entry(ic, &ib_nodev_conns, ib_node) 79 rds_conn_connect_if_down(ic->conn); 80 spin_unlock(&ib_nodev_conns_lock); 81 } 82 83 static void rds_ib_dev_shutdown(struct rds_ib_device *rds_ibdev) 84 { 85 struct rds_ib_connection *ic; 86 unsigned long flags; 87 88 spin_lock_irqsave(&rds_ibdev->spinlock, flags); 89 list_for_each_entry(ic, &rds_ibdev->conn_list, ib_node) 90 rds_conn_drop(ic->conn); 91 spin_unlock_irqrestore(&rds_ibdev->spinlock, flags); 92 } 93 94 /* 95 * rds_ib_destroy_mr_pool() blocks on a few things and mrs drop references 96 * from interrupt context so we push freing off into a work struct in krdsd. 97 */ 98 static void rds_ib_dev_free(struct work_struct *work) 99 { 100 struct rds_ib_ipaddr *i_ipaddr, *i_next; 101 struct rds_ib_device *rds_ibdev = container_of(work, 102 struct rds_ib_device, free_work); 103 104 if (rds_ibdev->mr_8k_pool) 105 rds_ib_destroy_mr_pool(rds_ibdev->mr_8k_pool); 106 if (rds_ibdev->mr_1m_pool) 107 rds_ib_destroy_mr_pool(rds_ibdev->mr_1m_pool); 108 if (rds_ibdev->pd) 109 ib_dealloc_pd(rds_ibdev->pd); 110 111 list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) { 112 list_del(&i_ipaddr->list); 113 kfree(i_ipaddr); 114 } 115 116 kfree(rds_ibdev->vector_load); 117 118 kfree(rds_ibdev); 119 } 120 121 void rds_ib_dev_put(struct rds_ib_device *rds_ibdev) 122 { 123 BUG_ON(refcount_read(&rds_ibdev->refcount) == 0); 124 if (refcount_dec_and_test(&rds_ibdev->refcount)) 125 queue_work(rds_wq, &rds_ibdev->free_work); 126 } 127 128 static void rds_ib_add_one(struct ib_device *device) 129 { 130 struct rds_ib_device *rds_ibdev; 131 bool has_fr, has_fmr; 132 133 /* Only handle IB (no iWARP) devices */ 134 if (device->node_type != RDMA_NODE_IB_CA) 135 return; 136 137 rds_ibdev = kzalloc_node(sizeof(struct rds_ib_device), GFP_KERNEL, 138 ibdev_to_node(device)); 139 if (!rds_ibdev) 140 return; 141 142 spin_lock_init(&rds_ibdev->spinlock); 143 refcount_set(&rds_ibdev->refcount, 1); 144 INIT_WORK(&rds_ibdev->free_work, rds_ib_dev_free); 145 146 rds_ibdev->max_wrs = device->attrs.max_qp_wr; 147 rds_ibdev->max_sge = min(device->attrs.max_send_sge, RDS_IB_MAX_SGE); 148 149 has_fr = (device->attrs.device_cap_flags & 150 IB_DEVICE_MEM_MGT_EXTENSIONS); 151 has_fmr = (device->alloc_fmr && device->dealloc_fmr && 152 device->map_phys_fmr && device->unmap_fmr); 153 rds_ibdev->use_fastreg = (has_fr && !has_fmr); 154 155 rds_ibdev->fmr_max_remaps = device->attrs.max_map_per_fmr?: 32; 156 rds_ibdev->max_1m_mrs = device->attrs.max_mr ? 157 min_t(unsigned int, (device->attrs.max_mr / 2), 158 rds_ib_mr_1m_pool_size) : rds_ib_mr_1m_pool_size; 159 160 rds_ibdev->max_8k_mrs = device->attrs.max_mr ? 161 min_t(unsigned int, ((device->attrs.max_mr / 2) * RDS_MR_8K_SCALE), 162 rds_ib_mr_8k_pool_size) : rds_ib_mr_8k_pool_size; 163 164 rds_ibdev->max_initiator_depth = device->attrs.max_qp_init_rd_atom; 165 rds_ibdev->max_responder_resources = device->attrs.max_qp_rd_atom; 166 167 rds_ibdev->vector_load = kcalloc(device->num_comp_vectors, 168 sizeof(int), 169 GFP_KERNEL); 170 if (!rds_ibdev->vector_load) { 171 pr_err("RDS/IB: %s failed to allocate vector memory\n", 172 __func__); 173 goto put_dev; 174 } 175 176 rds_ibdev->dev = device; 177 rds_ibdev->pd = ib_alloc_pd(device, 0); 178 if (IS_ERR(rds_ibdev->pd)) { 179 rds_ibdev->pd = NULL; 180 goto put_dev; 181 } 182 183 rds_ibdev->mr_1m_pool = 184 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_1M_POOL); 185 if (IS_ERR(rds_ibdev->mr_1m_pool)) { 186 rds_ibdev->mr_1m_pool = NULL; 187 goto put_dev; 188 } 189 190 rds_ibdev->mr_8k_pool = 191 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_8K_POOL); 192 if (IS_ERR(rds_ibdev->mr_8k_pool)) { 193 rds_ibdev->mr_8k_pool = NULL; 194 goto put_dev; 195 } 196 197 rdsdebug("RDS/IB: max_mr = %d, max_wrs = %d, max_sge = %d, fmr_max_remaps = %d, max_1m_mrs = %d, max_8k_mrs = %d\n", 198 device->attrs.max_fmr, rds_ibdev->max_wrs, rds_ibdev->max_sge, 199 rds_ibdev->fmr_max_remaps, rds_ibdev->max_1m_mrs, 200 rds_ibdev->max_8k_mrs); 201 202 pr_info("RDS/IB: %s: %s supported and preferred\n", 203 device->name, 204 rds_ibdev->use_fastreg ? "FRMR" : "FMR"); 205 206 INIT_LIST_HEAD(&rds_ibdev->ipaddr_list); 207 INIT_LIST_HEAD(&rds_ibdev->conn_list); 208 209 down_write(&rds_ib_devices_lock); 210 list_add_tail_rcu(&rds_ibdev->list, &rds_ib_devices); 211 up_write(&rds_ib_devices_lock); 212 refcount_inc(&rds_ibdev->refcount); 213 214 ib_set_client_data(device, &rds_ib_client, rds_ibdev); 215 refcount_inc(&rds_ibdev->refcount); 216 217 rds_ib_nodev_connect(); 218 219 put_dev: 220 rds_ib_dev_put(rds_ibdev); 221 } 222 223 /* 224 * New connections use this to find the device to associate with the 225 * connection. It's not in the fast path so we're not concerned about the 226 * performance of the IB call. (As of this writing, it uses an interrupt 227 * blocking spinlock to serialize walking a per-device list of all registered 228 * clients.) 229 * 230 * RCU is used to handle incoming connections racing with device teardown. 231 * Rather than use a lock to serialize removal from the client_data and 232 * getting a new reference, we use an RCU grace period. The destruction 233 * path removes the device from client_data and then waits for all RCU 234 * readers to finish. 235 * 236 * A new connection can get NULL from this if its arriving on a 237 * device that is in the process of being removed. 238 */ 239 struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device) 240 { 241 struct rds_ib_device *rds_ibdev; 242 243 rcu_read_lock(); 244 rds_ibdev = ib_get_client_data(device, &rds_ib_client); 245 if (rds_ibdev) 246 refcount_inc(&rds_ibdev->refcount); 247 rcu_read_unlock(); 248 return rds_ibdev; 249 } 250 251 /* 252 * The IB stack is letting us know that a device is going away. This can 253 * happen if the underlying HCA driver is removed or if PCI hotplug is removing 254 * the pci function, for example. 255 * 256 * This can be called at any time and can be racing with any other RDS path. 257 */ 258 static void rds_ib_remove_one(struct ib_device *device, void *client_data) 259 { 260 struct rds_ib_device *rds_ibdev = client_data; 261 262 if (!rds_ibdev) 263 return; 264 265 rds_ib_dev_shutdown(rds_ibdev); 266 267 /* stop connection attempts from getting a reference to this device. */ 268 ib_set_client_data(device, &rds_ib_client, NULL); 269 270 down_write(&rds_ib_devices_lock); 271 list_del_rcu(&rds_ibdev->list); 272 up_write(&rds_ib_devices_lock); 273 274 /* 275 * This synchronize rcu is waiting for readers of both the ib 276 * client data and the devices list to finish before we drop 277 * both of those references. 278 */ 279 synchronize_rcu(); 280 rds_ib_dev_put(rds_ibdev); 281 rds_ib_dev_put(rds_ibdev); 282 } 283 284 struct ib_client rds_ib_client = { 285 .name = "rds_ib", 286 .add = rds_ib_add_one, 287 .remove = rds_ib_remove_one 288 }; 289 290 static int rds_ib_conn_info_visitor(struct rds_connection *conn, 291 void *buffer) 292 { 293 struct rds_info_rdma_connection *iinfo = buffer; 294 struct rds_ib_connection *ic; 295 296 /* We will only ever look at IB transports */ 297 if (conn->c_trans != &rds_ib_transport) 298 return 0; 299 if (conn->c_isv6) 300 return 0; 301 302 iinfo->src_addr = conn->c_laddr.s6_addr32[3]; 303 iinfo->dst_addr = conn->c_faddr.s6_addr32[3]; 304 305 memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid)); 306 memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid)); 307 if (rds_conn_state(conn) == RDS_CONN_UP) { 308 struct rds_ib_device *rds_ibdev; 309 310 ic = conn->c_transport_data; 311 312 rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo->src_gid, 313 (union ib_gid *)&iinfo->dst_gid); 314 315 rds_ibdev = ic->rds_ibdev; 316 iinfo->max_send_wr = ic->i_send_ring.w_nr; 317 iinfo->max_recv_wr = ic->i_recv_ring.w_nr; 318 iinfo->max_send_sge = rds_ibdev->max_sge; 319 rds_ib_get_mr_info(rds_ibdev, iinfo); 320 } 321 return 1; 322 } 323 324 #if IS_ENABLED(CONFIG_IPV6) 325 /* IPv6 version of rds_ib_conn_info_visitor(). */ 326 static int rds6_ib_conn_info_visitor(struct rds_connection *conn, 327 void *buffer) 328 { 329 struct rds6_info_rdma_connection *iinfo6 = buffer; 330 struct rds_ib_connection *ic; 331 332 /* We will only ever look at IB transports */ 333 if (conn->c_trans != &rds_ib_transport) 334 return 0; 335 336 iinfo6->src_addr = conn->c_laddr; 337 iinfo6->dst_addr = conn->c_faddr; 338 339 memset(&iinfo6->src_gid, 0, sizeof(iinfo6->src_gid)); 340 memset(&iinfo6->dst_gid, 0, sizeof(iinfo6->dst_gid)); 341 342 if (rds_conn_state(conn) == RDS_CONN_UP) { 343 struct rds_ib_device *rds_ibdev; 344 345 ic = conn->c_transport_data; 346 rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo6->src_gid, 347 (union ib_gid *)&iinfo6->dst_gid); 348 rds_ibdev = ic->rds_ibdev; 349 iinfo6->max_send_wr = ic->i_send_ring.w_nr; 350 iinfo6->max_recv_wr = ic->i_recv_ring.w_nr; 351 iinfo6->max_send_sge = rds_ibdev->max_sge; 352 rds6_ib_get_mr_info(rds_ibdev, iinfo6); 353 } 354 return 1; 355 } 356 #endif 357 358 static void rds_ib_ic_info(struct socket *sock, unsigned int len, 359 struct rds_info_iterator *iter, 360 struct rds_info_lengths *lens) 361 { 362 u64 buffer[(sizeof(struct rds_info_rdma_connection) + 7) / 8]; 363 364 rds_for_each_conn_info(sock, len, iter, lens, 365 rds_ib_conn_info_visitor, 366 buffer, 367 sizeof(struct rds_info_rdma_connection)); 368 } 369 370 #if IS_ENABLED(CONFIG_IPV6) 371 /* IPv6 version of rds_ib_ic_info(). */ 372 static void rds6_ib_ic_info(struct socket *sock, unsigned int len, 373 struct rds_info_iterator *iter, 374 struct rds_info_lengths *lens) 375 { 376 u64 buffer[(sizeof(struct rds6_info_rdma_connection) + 7) / 8]; 377 378 rds_for_each_conn_info(sock, len, iter, lens, 379 rds6_ib_conn_info_visitor, 380 buffer, 381 sizeof(struct rds6_info_rdma_connection)); 382 } 383 #endif 384 385 /* 386 * Early RDS/IB was built to only bind to an address if there is an IPoIB 387 * device with that address set. 388 * 389 * If it were me, I'd advocate for something more flexible. Sending and 390 * receiving should be device-agnostic. Transports would try and maintain 391 * connections between peers who have messages queued. Userspace would be 392 * allowed to influence which paths have priority. We could call userspace 393 * asserting this policy "routing". 394 */ 395 static int rds_ib_laddr_check(struct net *net, const struct in6_addr *addr, 396 __u32 scope_id) 397 { 398 int ret; 399 struct rdma_cm_id *cm_id; 400 #if IS_ENABLED(CONFIG_IPV6) 401 struct sockaddr_in6 sin6; 402 #endif 403 struct sockaddr_in sin; 404 struct sockaddr *sa; 405 bool isv4; 406 407 isv4 = ipv6_addr_v4mapped(addr); 408 /* Create a CMA ID and try to bind it. This catches both 409 * IB and iWARP capable NICs. 410 */ 411 cm_id = rdma_create_id(&init_net, rds_rdma_cm_event_handler, 412 NULL, RDMA_PS_TCP, IB_QPT_RC); 413 if (IS_ERR(cm_id)) 414 return PTR_ERR(cm_id); 415 416 if (isv4) { 417 memset(&sin, 0, sizeof(sin)); 418 sin.sin_family = AF_INET; 419 sin.sin_addr.s_addr = addr->s6_addr32[3]; 420 sa = (struct sockaddr *)&sin; 421 } else { 422 #if IS_ENABLED(CONFIG_IPV6) 423 memset(&sin6, 0, sizeof(sin6)); 424 sin6.sin6_family = AF_INET6; 425 sin6.sin6_addr = *addr; 426 sin6.sin6_scope_id = scope_id; 427 sa = (struct sockaddr *)&sin6; 428 429 /* XXX Do a special IPv6 link local address check here. The 430 * reason is that rdma_bind_addr() always succeeds with IPv6 431 * link local address regardless it is indeed configured in a 432 * system. 433 */ 434 if (ipv6_addr_type(addr) & IPV6_ADDR_LINKLOCAL) { 435 struct net_device *dev; 436 437 if (scope_id == 0) { 438 ret = -EADDRNOTAVAIL; 439 goto out; 440 } 441 442 /* Use init_net for now as RDS is not network 443 * name space aware. 444 */ 445 dev = dev_get_by_index(&init_net, scope_id); 446 if (!dev) { 447 ret = -EADDRNOTAVAIL; 448 goto out; 449 } 450 if (!ipv6_chk_addr(&init_net, addr, dev, 1)) { 451 dev_put(dev); 452 ret = -EADDRNOTAVAIL; 453 goto out; 454 } 455 dev_put(dev); 456 } 457 #else 458 ret = -EADDRNOTAVAIL; 459 goto out; 460 #endif 461 } 462 463 /* rdma_bind_addr will only succeed for IB & iWARP devices */ 464 ret = rdma_bind_addr(cm_id, sa); 465 /* due to this, we will claim to support iWARP devices unless we 466 check node_type. */ 467 if (ret || !cm_id->device || 468 cm_id->device->node_type != RDMA_NODE_IB_CA) 469 ret = -EADDRNOTAVAIL; 470 471 rdsdebug("addr %pI6c%%%u ret %d node type %d\n", 472 addr, scope_id, ret, 473 cm_id->device ? cm_id->device->node_type : -1); 474 475 out: 476 rdma_destroy_id(cm_id); 477 478 return ret; 479 } 480 481 static void rds_ib_unregister_client(void) 482 { 483 ib_unregister_client(&rds_ib_client); 484 /* wait for rds_ib_dev_free() to complete */ 485 flush_workqueue(rds_wq); 486 } 487 488 static void rds_ib_set_unloading(void) 489 { 490 atomic_set(&rds_ib_unloading, 1); 491 } 492 493 static bool rds_ib_is_unloading(struct rds_connection *conn) 494 { 495 struct rds_conn_path *cp = &conn->c_path[0]; 496 497 return (test_bit(RDS_DESTROY_PENDING, &cp->cp_flags) || 498 atomic_read(&rds_ib_unloading) != 0); 499 } 500 501 void rds_ib_exit(void) 502 { 503 rds_ib_set_unloading(); 504 synchronize_rcu(); 505 rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info); 506 #if IS_ENABLED(CONFIG_IPV6) 507 rds_info_deregister_func(RDS6_INFO_IB_CONNECTIONS, rds6_ib_ic_info); 508 #endif 509 rds_ib_unregister_client(); 510 rds_ib_destroy_nodev_conns(); 511 rds_ib_sysctl_exit(); 512 rds_ib_recv_exit(); 513 rds_trans_unregister(&rds_ib_transport); 514 rds_ib_mr_exit(); 515 } 516 517 struct rds_transport rds_ib_transport = { 518 .laddr_check = rds_ib_laddr_check, 519 .xmit_path_complete = rds_ib_xmit_path_complete, 520 .xmit = rds_ib_xmit, 521 .xmit_rdma = rds_ib_xmit_rdma, 522 .xmit_atomic = rds_ib_xmit_atomic, 523 .recv_path = rds_ib_recv_path, 524 .conn_alloc = rds_ib_conn_alloc, 525 .conn_free = rds_ib_conn_free, 526 .conn_path_connect = rds_ib_conn_path_connect, 527 .conn_path_shutdown = rds_ib_conn_path_shutdown, 528 .inc_copy_to_user = rds_ib_inc_copy_to_user, 529 .inc_free = rds_ib_inc_free, 530 .cm_initiate_connect = rds_ib_cm_initiate_connect, 531 .cm_handle_connect = rds_ib_cm_handle_connect, 532 .cm_connect_complete = rds_ib_cm_connect_complete, 533 .stats_info_copy = rds_ib_stats_info_copy, 534 .exit = rds_ib_exit, 535 .get_mr = rds_ib_get_mr, 536 .sync_mr = rds_ib_sync_mr, 537 .free_mr = rds_ib_free_mr, 538 .flush_mrs = rds_ib_flush_mrs, 539 .t_owner = THIS_MODULE, 540 .t_name = "infiniband", 541 .t_unloading = rds_ib_is_unloading, 542 .t_type = RDS_TRANS_IB 543 }; 544 545 int rds_ib_init(void) 546 { 547 int ret; 548 549 INIT_LIST_HEAD(&rds_ib_devices); 550 551 ret = rds_ib_mr_init(); 552 if (ret) 553 goto out; 554 555 ret = ib_register_client(&rds_ib_client); 556 if (ret) 557 goto out_mr_exit; 558 559 ret = rds_ib_sysctl_init(); 560 if (ret) 561 goto out_ibreg; 562 563 ret = rds_ib_recv_init(); 564 if (ret) 565 goto out_sysctl; 566 567 rds_trans_register(&rds_ib_transport); 568 569 rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info); 570 #if IS_ENABLED(CONFIG_IPV6) 571 rds_info_register_func(RDS6_INFO_IB_CONNECTIONS, rds6_ib_ic_info); 572 #endif 573 574 goto out; 575 576 out_sysctl: 577 rds_ib_sysctl_exit(); 578 out_ibreg: 579 rds_ib_unregister_client(); 580 out_mr_exit: 581 rds_ib_mr_exit(); 582 out: 583 return ret; 584 } 585 586 MODULE_LICENSE("GPL"); 587