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