2 * Copyright (c) 2006 Oracle. All rights reserved.
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:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
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.
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
33 #include <linux/kernel.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>
43 #include "rds_single_path.h"
48 static unsigned int rds_ib_mr_1m_pool_size = RDS_MR_1M_POOL_SIZE;
49 static unsigned int rds_ib_mr_8k_pool_size = RDS_MR_8K_POOL_SIZE;
50 unsigned int rds_ib_retry_count = RDS_IB_DEFAULT_RETRY_COUNT;
51 static atomic_t rds_ib_unloading;
53 module_param(rds_ib_mr_1m_pool_size, int, 0444);
54 MODULE_PARM_DESC(rds_ib_mr_1m_pool_size, " Max number of 1M mr per HCA");
55 module_param(rds_ib_mr_8k_pool_size, int, 0444);
56 MODULE_PARM_DESC(rds_ib_mr_8k_pool_size, " Max number of 8K mr per HCA");
57 module_param(rds_ib_retry_count, int, 0444);
58 MODULE_PARM_DESC(rds_ib_retry_count, " Number of hw retries before reporting an error");
61 * we have a clumsy combination of RCU and a rwsem protecting this list
62 * because it is used both in the get_mr fast path and while blocking in
63 * the FMR flushing path.
65 DECLARE_RWSEM(rds_ib_devices_lock);
66 struct list_head rds_ib_devices;
68 /* NOTE: if also grabbing ibdev lock, grab this first */
69 DEFINE_SPINLOCK(ib_nodev_conns_lock);
70 LIST_HEAD(ib_nodev_conns);
72 static void rds_ib_nodev_connect(void)
74 struct rds_ib_connection *ic;
76 spin_lock(&ib_nodev_conns_lock);
77 list_for_each_entry(ic, &ib_nodev_conns, ib_node)
78 rds_conn_connect_if_down(ic->conn);
79 spin_unlock(&ib_nodev_conns_lock);
82 static void rds_ib_dev_shutdown(struct rds_ib_device *rds_ibdev)
84 struct rds_ib_connection *ic;
87 spin_lock_irqsave(&rds_ibdev->spinlock, flags);
88 list_for_each_entry(ic, &rds_ibdev->conn_list, ib_node)
89 rds_conn_drop(ic->conn);
90 spin_unlock_irqrestore(&rds_ibdev->spinlock, flags);
94 * rds_ib_destroy_mr_pool() blocks on a few things and mrs drop references
95 * from interrupt context so we push freing off into a work struct in krdsd.
97 static void rds_ib_dev_free(struct work_struct *work)
99 struct rds_ib_ipaddr *i_ipaddr, *i_next;
100 struct rds_ib_device *rds_ibdev = container_of(work,
101 struct rds_ib_device, free_work);
103 if (rds_ibdev->mr_8k_pool)
104 rds_ib_destroy_mr_pool(rds_ibdev->mr_8k_pool);
105 if (rds_ibdev->mr_1m_pool)
106 rds_ib_destroy_mr_pool(rds_ibdev->mr_1m_pool);
108 ib_dealloc_pd(rds_ibdev->pd);
110 list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) {
111 list_del(&i_ipaddr->list);
115 kfree(rds_ibdev->vector_load);
120 void rds_ib_dev_put(struct rds_ib_device *rds_ibdev)
122 BUG_ON(refcount_read(&rds_ibdev->refcount) == 0);
123 if (refcount_dec_and_test(&rds_ibdev->refcount))
124 queue_work(rds_wq, &rds_ibdev->free_work);
127 static void rds_ib_add_one(struct ib_device *device)
129 struct rds_ib_device *rds_ibdev;
130 bool has_fr, has_fmr;
132 /* Only handle IB (no iWARP) devices */
133 if (device->node_type != RDMA_NODE_IB_CA)
136 rds_ibdev = kzalloc_node(sizeof(struct rds_ib_device), GFP_KERNEL,
137 ibdev_to_node(device));
141 spin_lock_init(&rds_ibdev->spinlock);
142 refcount_set(&rds_ibdev->refcount, 1);
143 INIT_WORK(&rds_ibdev->free_work, rds_ib_dev_free);
145 rds_ibdev->max_wrs = device->attrs.max_qp_wr;
146 rds_ibdev->max_sge = min(device->attrs.max_sge, RDS_IB_MAX_SGE);
148 has_fr = (device->attrs.device_cap_flags &
149 IB_DEVICE_MEM_MGT_EXTENSIONS);
150 has_fmr = (device->alloc_fmr && device->dealloc_fmr &&
151 device->map_phys_fmr && device->unmap_fmr);
152 rds_ibdev->use_fastreg = (has_fr && !has_fmr);
154 rds_ibdev->fmr_max_remaps = device->attrs.max_map_per_fmr?: 32;
155 rds_ibdev->max_1m_mrs = device->attrs.max_mr ?
156 min_t(unsigned int, (device->attrs.max_mr / 2),
157 rds_ib_mr_1m_pool_size) : rds_ib_mr_1m_pool_size;
159 rds_ibdev->max_8k_mrs = device->attrs.max_mr ?
160 min_t(unsigned int, ((device->attrs.max_mr / 2) * RDS_MR_8K_SCALE),
161 rds_ib_mr_8k_pool_size) : rds_ib_mr_8k_pool_size;
163 rds_ibdev->max_initiator_depth = device->attrs.max_qp_init_rd_atom;
164 rds_ibdev->max_responder_resources = device->attrs.max_qp_rd_atom;
166 rds_ibdev->vector_load = kcalloc(device->num_comp_vectors,
169 if (!rds_ibdev->vector_load) {
170 pr_err("RDS/IB: %s failed to allocate vector memory\n",
175 rds_ibdev->dev = device;
176 rds_ibdev->pd = ib_alloc_pd(device, 0);
177 if (IS_ERR(rds_ibdev->pd)) {
178 rds_ibdev->pd = NULL;
182 rds_ibdev->mr_1m_pool =
183 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_1M_POOL);
184 if (IS_ERR(rds_ibdev->mr_1m_pool)) {
185 rds_ibdev->mr_1m_pool = NULL;
189 rds_ibdev->mr_8k_pool =
190 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_8K_POOL);
191 if (IS_ERR(rds_ibdev->mr_8k_pool)) {
192 rds_ibdev->mr_8k_pool = NULL;
196 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",
197 device->attrs.max_fmr, rds_ibdev->max_wrs, rds_ibdev->max_sge,
198 rds_ibdev->fmr_max_remaps, rds_ibdev->max_1m_mrs,
199 rds_ibdev->max_8k_mrs);
201 pr_info("RDS/IB: %s: %s supported and preferred\n",
203 rds_ibdev->use_fastreg ? "FRMR" : "FMR");
205 INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
206 INIT_LIST_HEAD(&rds_ibdev->conn_list);
208 down_write(&rds_ib_devices_lock);
209 list_add_tail_rcu(&rds_ibdev->list, &rds_ib_devices);
210 up_write(&rds_ib_devices_lock);
211 refcount_inc(&rds_ibdev->refcount);
213 ib_set_client_data(device, &rds_ib_client, rds_ibdev);
214 refcount_inc(&rds_ibdev->refcount);
216 rds_ib_nodev_connect();
219 rds_ib_dev_put(rds_ibdev);
223 * New connections use this to find the device to associate with the
224 * connection. It's not in the fast path so we're not concerned about the
225 * performance of the IB call. (As of this writing, it uses an interrupt
226 * blocking spinlock to serialize walking a per-device list of all registered
229 * RCU is used to handle incoming connections racing with device teardown.
230 * Rather than use a lock to serialize removal from the client_data and
231 * getting a new reference, we use an RCU grace period. The destruction
232 * path removes the device from client_data and then waits for all RCU
235 * A new connection can get NULL from this if its arriving on a
236 * device that is in the process of being removed.
238 struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device)
240 struct rds_ib_device *rds_ibdev;
243 rds_ibdev = ib_get_client_data(device, &rds_ib_client);
245 refcount_inc(&rds_ibdev->refcount);
251 * The IB stack is letting us know that a device is going away. This can
252 * happen if the underlying HCA driver is removed or if PCI hotplug is removing
253 * the pci function, for example.
255 * This can be called at any time and can be racing with any other RDS path.
257 static void rds_ib_remove_one(struct ib_device *device, void *client_data)
259 struct rds_ib_device *rds_ibdev = client_data;
264 rds_ib_dev_shutdown(rds_ibdev);
266 /* stop connection attempts from getting a reference to this device. */
267 ib_set_client_data(device, &rds_ib_client, NULL);
269 down_write(&rds_ib_devices_lock);
270 list_del_rcu(&rds_ibdev->list);
271 up_write(&rds_ib_devices_lock);
274 * This synchronize rcu is waiting for readers of both the ib
275 * client data and the devices list to finish before we drop
276 * both of those references.
279 rds_ib_dev_put(rds_ibdev);
280 rds_ib_dev_put(rds_ibdev);
283 struct ib_client rds_ib_client = {
285 .add = rds_ib_add_one,
286 .remove = rds_ib_remove_one
289 static int rds_ib_conn_info_visitor(struct rds_connection *conn,
292 struct rds_info_rdma_connection *iinfo = buffer;
293 struct rds_ib_connection *ic;
295 /* We will only ever look at IB transports */
296 if (conn->c_trans != &rds_ib_transport)
299 iinfo->src_addr = conn->c_laddr;
300 iinfo->dst_addr = conn->c_faddr;
302 memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid));
303 memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid));
304 if (rds_conn_state(conn) == RDS_CONN_UP) {
305 struct rds_ib_device *rds_ibdev;
307 ic = conn->c_transport_data;
309 rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo->src_gid,
310 (union ib_gid *)&iinfo->dst_gid);
312 rds_ibdev = ic->rds_ibdev;
313 iinfo->max_send_wr = ic->i_send_ring.w_nr;
314 iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
315 iinfo->max_send_sge = rds_ibdev->max_sge;
316 rds_ib_get_mr_info(rds_ibdev, iinfo);
321 static void rds_ib_ic_info(struct socket *sock, unsigned int len,
322 struct rds_info_iterator *iter,
323 struct rds_info_lengths *lens)
325 u64 buffer[(sizeof(struct rds_info_rdma_connection) + 7) / 8];
327 rds_for_each_conn_info(sock, len, iter, lens,
328 rds_ib_conn_info_visitor,
330 sizeof(struct rds_info_rdma_connection));
335 * Early RDS/IB was built to only bind to an address if there is an IPoIB
336 * device with that address set.
338 * If it were me, I'd advocate for something more flexible. Sending and
339 * receiving should be device-agnostic. Transports would try and maintain
340 * connections between peers who have messages queued. Userspace would be
341 * allowed to influence which paths have priority. We could call userspace
342 * asserting this policy "routing".
344 static int rds_ib_laddr_check(struct net *net, __be32 addr)
347 struct rdma_cm_id *cm_id;
348 struct sockaddr_in sin;
350 /* Create a CMA ID and try to bind it. This catches both
351 * IB and iWARP capable NICs.
353 cm_id = rdma_create_id(&init_net, rds_rdma_cm_event_handler,
354 NULL, RDMA_PS_TCP, IB_QPT_RC);
356 return PTR_ERR(cm_id);
358 memset(&sin, 0, sizeof(sin));
359 sin.sin_family = AF_INET;
360 sin.sin_addr.s_addr = addr;
362 /* rdma_bind_addr will only succeed for IB & iWARP devices */
363 ret = rdma_bind_addr(cm_id, (struct sockaddr *)&sin);
364 /* due to this, we will claim to support iWARP devices unless we
366 if (ret || !cm_id->device ||
367 cm_id->device->node_type != RDMA_NODE_IB_CA)
368 ret = -EADDRNOTAVAIL;
370 rdsdebug("addr %pI4 ret %d node type %d\n",
372 cm_id->device ? cm_id->device->node_type : -1);
374 rdma_destroy_id(cm_id);
379 static void rds_ib_unregister_client(void)
381 ib_unregister_client(&rds_ib_client);
382 /* wait for rds_ib_dev_free() to complete */
383 flush_workqueue(rds_wq);
386 static void rds_ib_set_unloading(void)
388 atomic_set(&rds_ib_unloading, 1);
391 static bool rds_ib_is_unloading(struct rds_connection *conn)
393 struct rds_conn_path *cp = &conn->c_path[0];
395 return (test_bit(RDS_DESTROY_PENDING, &cp->cp_flags) ||
396 atomic_read(&rds_ib_unloading) != 0);
399 void rds_ib_exit(void)
401 rds_ib_set_unloading();
403 rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
404 rds_ib_unregister_client();
405 rds_ib_destroy_nodev_conns();
406 rds_ib_sysctl_exit();
408 rds_trans_unregister(&rds_ib_transport);
412 struct rds_transport rds_ib_transport = {
413 .laddr_check = rds_ib_laddr_check,
414 .xmit_path_complete = rds_ib_xmit_path_complete,
416 .xmit_rdma = rds_ib_xmit_rdma,
417 .xmit_atomic = rds_ib_xmit_atomic,
418 .recv_path = rds_ib_recv_path,
419 .conn_alloc = rds_ib_conn_alloc,
420 .conn_free = rds_ib_conn_free,
421 .conn_path_connect = rds_ib_conn_path_connect,
422 .conn_path_shutdown = rds_ib_conn_path_shutdown,
423 .inc_copy_to_user = rds_ib_inc_copy_to_user,
424 .inc_free = rds_ib_inc_free,
425 .cm_initiate_connect = rds_ib_cm_initiate_connect,
426 .cm_handle_connect = rds_ib_cm_handle_connect,
427 .cm_connect_complete = rds_ib_cm_connect_complete,
428 .stats_info_copy = rds_ib_stats_info_copy,
430 .get_mr = rds_ib_get_mr,
431 .sync_mr = rds_ib_sync_mr,
432 .free_mr = rds_ib_free_mr,
433 .flush_mrs = rds_ib_flush_mrs,
434 .t_owner = THIS_MODULE,
435 .t_name = "infiniband",
436 .t_unloading = rds_ib_is_unloading,
437 .t_type = RDS_TRANS_IB
440 int rds_ib_init(void)
444 INIT_LIST_HEAD(&rds_ib_devices);
446 ret = rds_ib_mr_init();
450 ret = ib_register_client(&rds_ib_client);
454 ret = rds_ib_sysctl_init();
458 ret = rds_ib_recv_init();
462 rds_trans_register(&rds_ib_transport);
464 rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
469 rds_ib_sysctl_exit();
471 rds_ib_unregister_client();
478 MODULE_LICENSE("GPL");