1 /* 2 * Copyright (c) 2006 Oracle. 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 43 #include "rds.h" 44 #include "ib.h" 45 46 unsigned int rds_ib_fmr_1m_pool_size = RDS_FMR_1M_POOL_SIZE; 47 unsigned int rds_ib_fmr_8k_pool_size = RDS_FMR_8K_POOL_SIZE; 48 unsigned int rds_ib_retry_count = RDS_IB_DEFAULT_RETRY_COUNT; 49 50 module_param(rds_ib_fmr_1m_pool_size, int, 0444); 51 MODULE_PARM_DESC(rds_ib_fmr_1m_pool_size, " Max number of 1M fmr per HCA"); 52 module_param(rds_ib_fmr_8k_pool_size, int, 0444); 53 MODULE_PARM_DESC(rds_ib_fmr_8k_pool_size, " Max number of 8K fmr per HCA"); 54 module_param(rds_ib_retry_count, int, 0444); 55 MODULE_PARM_DESC(rds_ib_retry_count, " Number of hw retries before reporting an error"); 56 57 /* 58 * we have a clumsy combination of RCU and a rwsem protecting this list 59 * because it is used both in the get_mr fast path and while blocking in 60 * the FMR flushing path. 61 */ 62 DECLARE_RWSEM(rds_ib_devices_lock); 63 struct list_head rds_ib_devices; 64 65 /* NOTE: if also grabbing ibdev lock, grab this first */ 66 DEFINE_SPINLOCK(ib_nodev_conns_lock); 67 LIST_HEAD(ib_nodev_conns); 68 69 static void rds_ib_nodev_connect(void) 70 { 71 struct rds_ib_connection *ic; 72 73 spin_lock(&ib_nodev_conns_lock); 74 list_for_each_entry(ic, &ib_nodev_conns, ib_node) 75 rds_conn_connect_if_down(ic->conn); 76 spin_unlock(&ib_nodev_conns_lock); 77 } 78 79 static void rds_ib_dev_shutdown(struct rds_ib_device *rds_ibdev) 80 { 81 struct rds_ib_connection *ic; 82 unsigned long flags; 83 84 spin_lock_irqsave(&rds_ibdev->spinlock, flags); 85 list_for_each_entry(ic, &rds_ibdev->conn_list, ib_node) 86 rds_conn_drop(ic->conn); 87 spin_unlock_irqrestore(&rds_ibdev->spinlock, flags); 88 } 89 90 /* 91 * rds_ib_destroy_mr_pool() blocks on a few things and mrs drop references 92 * from interrupt context so we push freing off into a work struct in krdsd. 93 */ 94 static void rds_ib_dev_free(struct work_struct *work) 95 { 96 struct rds_ib_ipaddr *i_ipaddr, *i_next; 97 struct rds_ib_device *rds_ibdev = container_of(work, 98 struct rds_ib_device, free_work); 99 100 if (rds_ibdev->mr_8k_pool) 101 rds_ib_destroy_mr_pool(rds_ibdev->mr_8k_pool); 102 if (rds_ibdev->mr_1m_pool) 103 rds_ib_destroy_mr_pool(rds_ibdev->mr_1m_pool); 104 if (rds_ibdev->pd) 105 ib_dealloc_pd(rds_ibdev->pd); 106 107 list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) { 108 list_del(&i_ipaddr->list); 109 kfree(i_ipaddr); 110 } 111 112 kfree(rds_ibdev); 113 } 114 115 void rds_ib_dev_put(struct rds_ib_device *rds_ibdev) 116 { 117 BUG_ON(atomic_read(&rds_ibdev->refcount) <= 0); 118 if (atomic_dec_and_test(&rds_ibdev->refcount)) 119 queue_work(rds_wq, &rds_ibdev->free_work); 120 } 121 122 static void rds_ib_add_one(struct ib_device *device) 123 { 124 struct rds_ib_device *rds_ibdev; 125 struct ib_device_attr *dev_attr; 126 127 /* Only handle IB (no iWARP) devices */ 128 if (device->node_type != RDMA_NODE_IB_CA) 129 return; 130 131 dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL); 132 if (!dev_attr) 133 return; 134 135 if (ib_query_device(device, dev_attr)) { 136 rdsdebug("Query device failed for %s\n", device->name); 137 goto free_attr; 138 } 139 140 rds_ibdev = kzalloc_node(sizeof(struct rds_ib_device), GFP_KERNEL, 141 ibdev_to_node(device)); 142 if (!rds_ibdev) 143 goto free_attr; 144 145 spin_lock_init(&rds_ibdev->spinlock); 146 atomic_set(&rds_ibdev->refcount, 1); 147 INIT_WORK(&rds_ibdev->free_work, rds_ib_dev_free); 148 149 rds_ibdev->max_wrs = dev_attr->max_qp_wr; 150 rds_ibdev->max_sge = min(dev_attr->max_sge, RDS_IB_MAX_SGE); 151 152 rds_ibdev->fmr_max_remaps = dev_attr->max_map_per_fmr?: 32; 153 rds_ibdev->max_1m_fmrs = dev_attr->max_mr ? 154 min_t(unsigned int, (dev_attr->max_mr / 2), 155 rds_ib_fmr_1m_pool_size) : rds_ib_fmr_1m_pool_size; 156 157 rds_ibdev->max_8k_fmrs = dev_attr->max_mr ? 158 min_t(unsigned int, ((dev_attr->max_mr / 2) * RDS_MR_8K_SCALE), 159 rds_ib_fmr_8k_pool_size) : rds_ib_fmr_8k_pool_size; 160 161 rds_ibdev->max_initiator_depth = dev_attr->max_qp_init_rd_atom; 162 rds_ibdev->max_responder_resources = dev_attr->max_qp_rd_atom; 163 164 rds_ibdev->dev = device; 165 rds_ibdev->pd = ib_alloc_pd(device); 166 if (IS_ERR(rds_ibdev->pd)) { 167 rds_ibdev->pd = NULL; 168 goto put_dev; 169 } 170 171 rds_ibdev->mr_1m_pool = 172 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_1M_POOL); 173 if (IS_ERR(rds_ibdev->mr_1m_pool)) { 174 rds_ibdev->mr_1m_pool = NULL; 175 goto put_dev; 176 } 177 178 rds_ibdev->mr_8k_pool = 179 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_8K_POOL); 180 if (IS_ERR(rds_ibdev->mr_8k_pool)) { 181 rds_ibdev->mr_8k_pool = NULL; 182 goto put_dev; 183 } 184 185 rdsdebug("RDS/IB: max_mr = %d, max_wrs = %d, max_sge = %d, fmr_max_remaps = %d, max_1m_fmrs = %d, max_8k_fmrs = %d\n", 186 dev_attr->max_fmr, rds_ibdev->max_wrs, rds_ibdev->max_sge, 187 rds_ibdev->fmr_max_remaps, rds_ibdev->max_1m_fmrs, 188 rds_ibdev->max_8k_fmrs); 189 190 INIT_LIST_HEAD(&rds_ibdev->ipaddr_list); 191 INIT_LIST_HEAD(&rds_ibdev->conn_list); 192 193 down_write(&rds_ib_devices_lock); 194 list_add_tail_rcu(&rds_ibdev->list, &rds_ib_devices); 195 up_write(&rds_ib_devices_lock); 196 atomic_inc(&rds_ibdev->refcount); 197 198 ib_set_client_data(device, &rds_ib_client, rds_ibdev); 199 atomic_inc(&rds_ibdev->refcount); 200 201 rds_ib_nodev_connect(); 202 203 put_dev: 204 rds_ib_dev_put(rds_ibdev); 205 free_attr: 206 kfree(dev_attr); 207 } 208 209 /* 210 * New connections use this to find the device to associate with the 211 * connection. It's not in the fast path so we're not concerned about the 212 * performance of the IB call. (As of this writing, it uses an interrupt 213 * blocking spinlock to serialize walking a per-device list of all registered 214 * clients.) 215 * 216 * RCU is used to handle incoming connections racing with device teardown. 217 * Rather than use a lock to serialize removal from the client_data and 218 * getting a new reference, we use an RCU grace period. The destruction 219 * path removes the device from client_data and then waits for all RCU 220 * readers to finish. 221 * 222 * A new connection can get NULL from this if its arriving on a 223 * device that is in the process of being removed. 224 */ 225 struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device) 226 { 227 struct rds_ib_device *rds_ibdev; 228 229 rcu_read_lock(); 230 rds_ibdev = ib_get_client_data(device, &rds_ib_client); 231 if (rds_ibdev) 232 atomic_inc(&rds_ibdev->refcount); 233 rcu_read_unlock(); 234 return rds_ibdev; 235 } 236 237 /* 238 * The IB stack is letting us know that a device is going away. This can 239 * happen if the underlying HCA driver is removed or if PCI hotplug is removing 240 * the pci function, for example. 241 * 242 * This can be called at any time and can be racing with any other RDS path. 243 */ 244 static void rds_ib_remove_one(struct ib_device *device, void *client_data) 245 { 246 struct rds_ib_device *rds_ibdev = client_data; 247 248 if (!rds_ibdev) 249 return; 250 251 rds_ib_dev_shutdown(rds_ibdev); 252 253 /* stop connection attempts from getting a reference to this device. */ 254 ib_set_client_data(device, &rds_ib_client, NULL); 255 256 down_write(&rds_ib_devices_lock); 257 list_del_rcu(&rds_ibdev->list); 258 up_write(&rds_ib_devices_lock); 259 260 /* 261 * This synchronize rcu is waiting for readers of both the ib 262 * client data and the devices list to finish before we drop 263 * both of those references. 264 */ 265 synchronize_rcu(); 266 rds_ib_dev_put(rds_ibdev); 267 rds_ib_dev_put(rds_ibdev); 268 } 269 270 struct ib_client rds_ib_client = { 271 .name = "rds_ib", 272 .add = rds_ib_add_one, 273 .remove = rds_ib_remove_one 274 }; 275 276 static int rds_ib_conn_info_visitor(struct rds_connection *conn, 277 void *buffer) 278 { 279 struct rds_info_rdma_connection *iinfo = buffer; 280 struct rds_ib_connection *ic; 281 282 /* We will only ever look at IB transports */ 283 if (conn->c_trans != &rds_ib_transport) 284 return 0; 285 286 iinfo->src_addr = conn->c_laddr; 287 iinfo->dst_addr = conn->c_faddr; 288 289 memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid)); 290 memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid)); 291 if (rds_conn_state(conn) == RDS_CONN_UP) { 292 struct rds_ib_device *rds_ibdev; 293 struct rdma_dev_addr *dev_addr; 294 295 ic = conn->c_transport_data; 296 dev_addr = &ic->i_cm_id->route.addr.dev_addr; 297 298 rdma_addr_get_sgid(dev_addr, (union ib_gid *) &iinfo->src_gid); 299 rdma_addr_get_dgid(dev_addr, (union ib_gid *) &iinfo->dst_gid); 300 301 rds_ibdev = ic->rds_ibdev; 302 iinfo->max_send_wr = ic->i_send_ring.w_nr; 303 iinfo->max_recv_wr = ic->i_recv_ring.w_nr; 304 iinfo->max_send_sge = rds_ibdev->max_sge; 305 rds_ib_get_mr_info(rds_ibdev, iinfo); 306 } 307 return 1; 308 } 309 310 static void rds_ib_ic_info(struct socket *sock, unsigned int len, 311 struct rds_info_iterator *iter, 312 struct rds_info_lengths *lens) 313 { 314 rds_for_each_conn_info(sock, len, iter, lens, 315 rds_ib_conn_info_visitor, 316 sizeof(struct rds_info_rdma_connection)); 317 } 318 319 320 /* 321 * Early RDS/IB was built to only bind to an address if there is an IPoIB 322 * device with that address set. 323 * 324 * If it were me, I'd advocate for something more flexible. Sending and 325 * receiving should be device-agnostic. Transports would try and maintain 326 * connections between peers who have messages queued. Userspace would be 327 * allowed to influence which paths have priority. We could call userspace 328 * asserting this policy "routing". 329 */ 330 static int rds_ib_laddr_check(struct net *net, __be32 addr) 331 { 332 int ret; 333 struct rdma_cm_id *cm_id; 334 struct sockaddr_in sin; 335 336 /* Create a CMA ID and try to bind it. This catches both 337 * IB and iWARP capable NICs. 338 */ 339 cm_id = rdma_create_id(&init_net, NULL, NULL, RDMA_PS_TCP, IB_QPT_RC); 340 if (IS_ERR(cm_id)) 341 return PTR_ERR(cm_id); 342 343 memset(&sin, 0, sizeof(sin)); 344 sin.sin_family = AF_INET; 345 sin.sin_addr.s_addr = addr; 346 347 /* rdma_bind_addr will only succeed for IB & iWARP devices */ 348 ret = rdma_bind_addr(cm_id, (struct sockaddr *)&sin); 349 /* due to this, we will claim to support iWARP devices unless we 350 check node_type. */ 351 if (ret || !cm_id->device || 352 cm_id->device->node_type != RDMA_NODE_IB_CA) 353 ret = -EADDRNOTAVAIL; 354 355 rdsdebug("addr %pI4 ret %d node type %d\n", 356 &addr, ret, 357 cm_id->device ? cm_id->device->node_type : -1); 358 359 rdma_destroy_id(cm_id); 360 361 return ret; 362 } 363 364 static void rds_ib_unregister_client(void) 365 { 366 ib_unregister_client(&rds_ib_client); 367 /* wait for rds_ib_dev_free() to complete */ 368 flush_workqueue(rds_wq); 369 } 370 371 void rds_ib_exit(void) 372 { 373 rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info); 374 rds_ib_unregister_client(); 375 rds_ib_destroy_nodev_conns(); 376 rds_ib_sysctl_exit(); 377 rds_ib_recv_exit(); 378 rds_trans_unregister(&rds_ib_transport); 379 rds_ib_fmr_exit(); 380 } 381 382 struct rds_transport rds_ib_transport = { 383 .laddr_check = rds_ib_laddr_check, 384 .xmit_complete = rds_ib_xmit_complete, 385 .xmit = rds_ib_xmit, 386 .xmit_rdma = rds_ib_xmit_rdma, 387 .xmit_atomic = rds_ib_xmit_atomic, 388 .recv = rds_ib_recv, 389 .conn_alloc = rds_ib_conn_alloc, 390 .conn_free = rds_ib_conn_free, 391 .conn_connect = rds_ib_conn_connect, 392 .conn_shutdown = rds_ib_conn_shutdown, 393 .inc_copy_to_user = rds_ib_inc_copy_to_user, 394 .inc_free = rds_ib_inc_free, 395 .cm_initiate_connect = rds_ib_cm_initiate_connect, 396 .cm_handle_connect = rds_ib_cm_handle_connect, 397 .cm_connect_complete = rds_ib_cm_connect_complete, 398 .stats_info_copy = rds_ib_stats_info_copy, 399 .exit = rds_ib_exit, 400 .get_mr = rds_ib_get_mr, 401 .sync_mr = rds_ib_sync_mr, 402 .free_mr = rds_ib_free_mr, 403 .flush_mrs = rds_ib_flush_mrs, 404 .t_owner = THIS_MODULE, 405 .t_name = "infiniband", 406 .t_type = RDS_TRANS_IB 407 }; 408 409 int rds_ib_init(void) 410 { 411 int ret; 412 413 INIT_LIST_HEAD(&rds_ib_devices); 414 415 ret = rds_ib_fmr_init(); 416 if (ret) 417 goto out; 418 419 ret = ib_register_client(&rds_ib_client); 420 if (ret) 421 goto out_fmr_exit; 422 423 ret = rds_ib_sysctl_init(); 424 if (ret) 425 goto out_ibreg; 426 427 ret = rds_ib_recv_init(); 428 if (ret) 429 goto out_sysctl; 430 431 ret = rds_trans_register(&rds_ib_transport); 432 if (ret) 433 goto out_recv; 434 435 rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info); 436 437 goto out; 438 439 out_recv: 440 rds_ib_recv_exit(); 441 out_sysctl: 442 rds_ib_sysctl_exit(); 443 out_ibreg: 444 rds_ib_unregister_client(); 445 out_fmr_exit: 446 rds_ib_fmr_exit(); 447 out: 448 return ret; 449 } 450 451 MODULE_LICENSE("GPL"); 452 453