2 * Copyright (c) 2006, 2019 Oracle and/or its affiliates. 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/dmapool.h>
34 #include <linux/kernel.h>
37 #include <linux/netdevice.h>
38 #include <linux/inetdevice.h>
39 #include <linux/if_arp.h>
40 #include <linux/delay.h>
41 #include <linux/slab.h>
42 #include <linux/module.h>
43 #include <net/addrconf.h>
45 #include "rds_single_path.h"
50 static unsigned int rds_ib_mr_1m_pool_size = RDS_MR_1M_POOL_SIZE;
51 static unsigned int rds_ib_mr_8k_pool_size = RDS_MR_8K_POOL_SIZE;
52 unsigned int rds_ib_retry_count = RDS_IB_DEFAULT_RETRY_COUNT;
53 static atomic_t rds_ib_unloading;
55 module_param(rds_ib_mr_1m_pool_size, int, 0444);
56 MODULE_PARM_DESC(rds_ib_mr_1m_pool_size, " Max number of 1M mr per HCA");
57 module_param(rds_ib_mr_8k_pool_size, int, 0444);
58 MODULE_PARM_DESC(rds_ib_mr_8k_pool_size, " Max number of 8K mr per HCA");
59 module_param(rds_ib_retry_count, int, 0444);
60 MODULE_PARM_DESC(rds_ib_retry_count, " Number of hw retries before reporting an error");
63 * we have a clumsy combination of RCU and a rwsem protecting this list
64 * because it is used both in the get_mr fast path and while blocking in
65 * the FMR flushing path.
67 DECLARE_RWSEM(rds_ib_devices_lock);
68 struct list_head rds_ib_devices;
70 /* NOTE: if also grabbing ibdev lock, grab this first */
71 DEFINE_SPINLOCK(ib_nodev_conns_lock);
72 LIST_HEAD(ib_nodev_conns);
74 static void rds_ib_nodev_connect(void)
76 struct rds_ib_connection *ic;
78 spin_lock(&ib_nodev_conns_lock);
79 list_for_each_entry(ic, &ib_nodev_conns, ib_node)
80 rds_conn_connect_if_down(ic->conn);
81 spin_unlock(&ib_nodev_conns_lock);
84 static void rds_ib_dev_shutdown(struct rds_ib_device *rds_ibdev)
86 struct rds_ib_connection *ic;
89 spin_lock_irqsave(&rds_ibdev->spinlock, flags);
90 list_for_each_entry(ic, &rds_ibdev->conn_list, ib_node)
91 rds_conn_path_drop(&ic->conn->c_path[0], true);
92 spin_unlock_irqrestore(&rds_ibdev->spinlock, flags);
96 * rds_ib_destroy_mr_pool() blocks on a few things and mrs drop references
97 * from interrupt context so we push freing off into a work struct in krdsd.
99 static void rds_ib_dev_free(struct work_struct *work)
101 struct rds_ib_ipaddr *i_ipaddr, *i_next;
102 struct rds_ib_device *rds_ibdev = container_of(work,
103 struct rds_ib_device, free_work);
105 if (rds_ibdev->mr_8k_pool)
106 rds_ib_destroy_mr_pool(rds_ibdev->mr_8k_pool);
107 if (rds_ibdev->mr_1m_pool)
108 rds_ib_destroy_mr_pool(rds_ibdev->mr_1m_pool);
110 ib_dealloc_pd(rds_ibdev->pd);
111 dma_pool_destroy(rds_ibdev->rid_hdrs_pool);
113 list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) {
114 list_del(&i_ipaddr->list);
118 kfree(rds_ibdev->vector_load);
123 void rds_ib_dev_put(struct rds_ib_device *rds_ibdev)
125 BUG_ON(refcount_read(&rds_ibdev->refcount) == 0);
126 if (refcount_dec_and_test(&rds_ibdev->refcount))
127 queue_work(rds_wq, &rds_ibdev->free_work);
130 static int rds_ib_add_one(struct ib_device *device)
132 struct rds_ib_device *rds_ibdev;
135 /* Only handle IB (no iWARP) devices */
136 if (device->node_type != RDMA_NODE_IB_CA)
139 /* Device must support FRWR */
140 if (!(device->attrs.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS))
143 rds_ibdev = kzalloc_node(sizeof(struct rds_ib_device), GFP_KERNEL,
144 ibdev_to_node(device));
148 spin_lock_init(&rds_ibdev->spinlock);
149 refcount_set(&rds_ibdev->refcount, 1);
150 INIT_WORK(&rds_ibdev->free_work, rds_ib_dev_free);
152 INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
153 INIT_LIST_HEAD(&rds_ibdev->conn_list);
155 rds_ibdev->max_wrs = device->attrs.max_qp_wr;
156 rds_ibdev->max_sge = min(device->attrs.max_send_sge, RDS_IB_MAX_SGE);
158 rds_ibdev->odp_capable =
159 !!(device->attrs.device_cap_flags &
160 IB_DEVICE_ON_DEMAND_PAGING) &&
161 !!(device->attrs.odp_caps.per_transport_caps.rc_odp_caps &
162 IB_ODP_SUPPORT_WRITE) &&
163 !!(device->attrs.odp_caps.per_transport_caps.rc_odp_caps &
164 IB_ODP_SUPPORT_READ);
166 rds_ibdev->max_1m_mrs = device->attrs.max_mr ?
167 min_t(unsigned int, (device->attrs.max_mr / 2),
168 rds_ib_mr_1m_pool_size) : rds_ib_mr_1m_pool_size;
170 rds_ibdev->max_8k_mrs = device->attrs.max_mr ?
171 min_t(unsigned int, ((device->attrs.max_mr / 2) * RDS_MR_8K_SCALE),
172 rds_ib_mr_8k_pool_size) : rds_ib_mr_8k_pool_size;
174 rds_ibdev->max_initiator_depth = device->attrs.max_qp_init_rd_atom;
175 rds_ibdev->max_responder_resources = device->attrs.max_qp_rd_atom;
177 rds_ibdev->vector_load = kcalloc(device->num_comp_vectors,
180 if (!rds_ibdev->vector_load) {
181 pr_err("RDS/IB: %s failed to allocate vector memory\n",
187 rds_ibdev->dev = device;
188 rds_ibdev->pd = ib_alloc_pd(device, 0);
189 if (IS_ERR(rds_ibdev->pd)) {
190 ret = PTR_ERR(rds_ibdev->pd);
191 rds_ibdev->pd = NULL;
194 rds_ibdev->rid_hdrs_pool = dma_pool_create(device->name,
196 sizeof(struct rds_header),
198 if (!rds_ibdev->rid_hdrs_pool) {
203 rds_ibdev->mr_1m_pool =
204 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_1M_POOL);
205 if (IS_ERR(rds_ibdev->mr_1m_pool)) {
206 ret = PTR_ERR(rds_ibdev->mr_1m_pool);
207 rds_ibdev->mr_1m_pool = NULL;
211 rds_ibdev->mr_8k_pool =
212 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_8K_POOL);
213 if (IS_ERR(rds_ibdev->mr_8k_pool)) {
214 ret = PTR_ERR(rds_ibdev->mr_8k_pool);
215 rds_ibdev->mr_8k_pool = NULL;
219 rdsdebug("RDS/IB: max_mr = %d, max_wrs = %d, max_sge = %d, max_1m_mrs = %d, max_8k_mrs = %d\n",
220 device->attrs.max_mr, rds_ibdev->max_wrs, rds_ibdev->max_sge,
221 rds_ibdev->max_1m_mrs, rds_ibdev->max_8k_mrs);
223 pr_info("RDS/IB: %s: added\n", device->name);
225 down_write(&rds_ib_devices_lock);
226 list_add_tail_rcu(&rds_ibdev->list, &rds_ib_devices);
227 up_write(&rds_ib_devices_lock);
228 refcount_inc(&rds_ibdev->refcount);
230 ib_set_client_data(device, &rds_ib_client, rds_ibdev);
232 rds_ib_nodev_connect();
236 rds_ib_dev_put(rds_ibdev);
241 * New connections use this to find the device to associate with the
242 * connection. It's not in the fast path so we're not concerned about the
243 * performance of the IB call. (As of this writing, it uses an interrupt
244 * blocking spinlock to serialize walking a per-device list of all registered
247 * RCU is used to handle incoming connections racing with device teardown.
248 * Rather than use a lock to serialize removal from the client_data and
249 * getting a new reference, we use an RCU grace period. The destruction
250 * path removes the device from client_data and then waits for all RCU
253 * A new connection can get NULL from this if its arriving on a
254 * device that is in the process of being removed.
256 struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device)
258 struct rds_ib_device *rds_ibdev;
261 rds_ibdev = ib_get_client_data(device, &rds_ib_client);
263 refcount_inc(&rds_ibdev->refcount);
269 * The IB stack is letting us know that a device is going away. This can
270 * happen if the underlying HCA driver is removed or if PCI hotplug is removing
271 * the pci function, for example.
273 * This can be called at any time and can be racing with any other RDS path.
275 static void rds_ib_remove_one(struct ib_device *device, void *client_data)
277 struct rds_ib_device *rds_ibdev = client_data;
279 rds_ib_dev_shutdown(rds_ibdev);
281 /* stop connection attempts from getting a reference to this device. */
282 ib_set_client_data(device, &rds_ib_client, NULL);
284 down_write(&rds_ib_devices_lock);
285 list_del_rcu(&rds_ibdev->list);
286 up_write(&rds_ib_devices_lock);
289 * This synchronize rcu is waiting for readers of both the ib
290 * client data and the devices list to finish before we drop
291 * both of those references.
294 rds_ib_dev_put(rds_ibdev);
295 rds_ib_dev_put(rds_ibdev);
298 struct ib_client rds_ib_client = {
300 .add = rds_ib_add_one,
301 .remove = rds_ib_remove_one
304 static int rds_ib_conn_info_visitor(struct rds_connection *conn,
307 struct rds_info_rdma_connection *iinfo = buffer;
308 struct rds_ib_connection *ic = conn->c_transport_data;
310 /* We will only ever look at IB transports */
311 if (conn->c_trans != &rds_ib_transport)
316 iinfo->src_addr = conn->c_laddr.s6_addr32[3];
317 iinfo->dst_addr = conn->c_faddr.s6_addr32[3];
319 iinfo->tos = conn->c_tos;
320 iinfo->sl = ic->i_sl;
323 memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid));
324 memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid));
325 if (rds_conn_state(conn) == RDS_CONN_UP) {
326 struct rds_ib_device *rds_ibdev;
328 rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo->src_gid,
329 (union ib_gid *)&iinfo->dst_gid);
331 rds_ibdev = ic->rds_ibdev;
332 iinfo->max_send_wr = ic->i_send_ring.w_nr;
333 iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
334 iinfo->max_send_sge = rds_ibdev->max_sge;
335 rds_ib_get_mr_info(rds_ibdev, iinfo);
336 iinfo->cache_allocs = atomic_read(&ic->i_cache_allocs);
341 #if IS_ENABLED(CONFIG_IPV6)
342 /* IPv6 version of rds_ib_conn_info_visitor(). */
343 static int rds6_ib_conn_info_visitor(struct rds_connection *conn,
346 struct rds6_info_rdma_connection *iinfo6 = buffer;
347 struct rds_ib_connection *ic = conn->c_transport_data;
349 /* We will only ever look at IB transports */
350 if (conn->c_trans != &rds_ib_transport)
353 iinfo6->src_addr = conn->c_laddr;
354 iinfo6->dst_addr = conn->c_faddr;
356 iinfo6->tos = conn->c_tos;
357 iinfo6->sl = ic->i_sl;
360 memset(&iinfo6->src_gid, 0, sizeof(iinfo6->src_gid));
361 memset(&iinfo6->dst_gid, 0, sizeof(iinfo6->dst_gid));
363 if (rds_conn_state(conn) == RDS_CONN_UP) {
364 struct rds_ib_device *rds_ibdev;
366 rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo6->src_gid,
367 (union ib_gid *)&iinfo6->dst_gid);
368 rds_ibdev = ic->rds_ibdev;
369 iinfo6->max_send_wr = ic->i_send_ring.w_nr;
370 iinfo6->max_recv_wr = ic->i_recv_ring.w_nr;
371 iinfo6->max_send_sge = rds_ibdev->max_sge;
372 rds6_ib_get_mr_info(rds_ibdev, iinfo6);
373 iinfo6->cache_allocs = atomic_read(&ic->i_cache_allocs);
379 static void rds_ib_ic_info(struct socket *sock, unsigned int len,
380 struct rds_info_iterator *iter,
381 struct rds_info_lengths *lens)
383 u64 buffer[(sizeof(struct rds_info_rdma_connection) + 7) / 8];
385 rds_for_each_conn_info(sock, len, iter, lens,
386 rds_ib_conn_info_visitor,
388 sizeof(struct rds_info_rdma_connection));
391 #if IS_ENABLED(CONFIG_IPV6)
392 /* IPv6 version of rds_ib_ic_info(). */
393 static void rds6_ib_ic_info(struct socket *sock, unsigned int len,
394 struct rds_info_iterator *iter,
395 struct rds_info_lengths *lens)
397 u64 buffer[(sizeof(struct rds6_info_rdma_connection) + 7) / 8];
399 rds_for_each_conn_info(sock, len, iter, lens,
400 rds6_ib_conn_info_visitor,
402 sizeof(struct rds6_info_rdma_connection));
407 * Early RDS/IB was built to only bind to an address if there is an IPoIB
408 * device with that address set.
410 * If it were me, I'd advocate for something more flexible. Sending and
411 * receiving should be device-agnostic. Transports would try and maintain
412 * connections between peers who have messages queued. Userspace would be
413 * allowed to influence which paths have priority. We could call userspace
414 * asserting this policy "routing".
416 static int rds_ib_laddr_check(struct net *net, const struct in6_addr *addr,
420 struct rdma_cm_id *cm_id;
421 #if IS_ENABLED(CONFIG_IPV6)
422 struct sockaddr_in6 sin6;
424 struct sockaddr_in sin;
428 isv4 = ipv6_addr_v4mapped(addr);
429 /* Create a CMA ID and try to bind it. This catches both
430 * IB and iWARP capable NICs.
432 cm_id = rdma_create_id(&init_net, rds_rdma_cm_event_handler,
433 NULL, RDMA_PS_TCP, IB_QPT_RC);
435 return PTR_ERR(cm_id);
438 memset(&sin, 0, sizeof(sin));
439 sin.sin_family = AF_INET;
440 sin.sin_addr.s_addr = addr->s6_addr32[3];
441 sa = (struct sockaddr *)&sin;
443 #if IS_ENABLED(CONFIG_IPV6)
444 memset(&sin6, 0, sizeof(sin6));
445 sin6.sin6_family = AF_INET6;
446 sin6.sin6_addr = *addr;
447 sin6.sin6_scope_id = scope_id;
448 sa = (struct sockaddr *)&sin6;
450 /* XXX Do a special IPv6 link local address check here. The
451 * reason is that rdma_bind_addr() always succeeds with IPv6
452 * link local address regardless it is indeed configured in a
455 if (ipv6_addr_type(addr) & IPV6_ADDR_LINKLOCAL) {
456 struct net_device *dev;
459 ret = -EADDRNOTAVAIL;
463 /* Use init_net for now as RDS is not network
466 dev = dev_get_by_index(&init_net, scope_id);
468 ret = -EADDRNOTAVAIL;
471 if (!ipv6_chk_addr(&init_net, addr, dev, 1)) {
473 ret = -EADDRNOTAVAIL;
479 ret = -EADDRNOTAVAIL;
484 /* rdma_bind_addr will only succeed for IB & iWARP devices */
485 ret = rdma_bind_addr(cm_id, sa);
486 /* due to this, we will claim to support iWARP devices unless we
488 if (ret || !cm_id->device ||
489 cm_id->device->node_type != RDMA_NODE_IB_CA)
490 ret = -EADDRNOTAVAIL;
492 rdsdebug("addr %pI6c%%%u ret %d node type %d\n",
494 cm_id->device ? cm_id->device->node_type : -1);
497 rdma_destroy_id(cm_id);
502 static void rds_ib_unregister_client(void)
504 ib_unregister_client(&rds_ib_client);
505 /* wait for rds_ib_dev_free() to complete */
506 flush_workqueue(rds_wq);
509 static void rds_ib_set_unloading(void)
511 atomic_set(&rds_ib_unloading, 1);
514 static bool rds_ib_is_unloading(struct rds_connection *conn)
516 struct rds_conn_path *cp = &conn->c_path[0];
518 return (test_bit(RDS_DESTROY_PENDING, &cp->cp_flags) ||
519 atomic_read(&rds_ib_unloading) != 0);
522 void rds_ib_exit(void)
524 rds_ib_set_unloading();
526 rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
527 #if IS_ENABLED(CONFIG_IPV6)
528 rds_info_deregister_func(RDS6_INFO_IB_CONNECTIONS, rds6_ib_ic_info);
530 rds_ib_unregister_client();
531 rds_ib_destroy_nodev_conns();
532 rds_ib_sysctl_exit();
534 rds_trans_unregister(&rds_ib_transport);
538 static u8 rds_ib_get_tos_map(u8 tos)
540 /* 1:1 user to transport map for RDMA transport.
541 * In future, if custom map is desired, hook can export
542 * user configurable map.
547 struct rds_transport rds_ib_transport = {
548 .laddr_check = rds_ib_laddr_check,
549 .xmit_path_complete = rds_ib_xmit_path_complete,
551 .xmit_rdma = rds_ib_xmit_rdma,
552 .xmit_atomic = rds_ib_xmit_atomic,
553 .recv_path = rds_ib_recv_path,
554 .conn_alloc = rds_ib_conn_alloc,
555 .conn_free = rds_ib_conn_free,
556 .conn_path_connect = rds_ib_conn_path_connect,
557 .conn_path_shutdown = rds_ib_conn_path_shutdown,
558 .inc_copy_to_user = rds_ib_inc_copy_to_user,
559 .inc_free = rds_ib_inc_free,
560 .cm_initiate_connect = rds_ib_cm_initiate_connect,
561 .cm_handle_connect = rds_ib_cm_handle_connect,
562 .cm_connect_complete = rds_ib_cm_connect_complete,
563 .stats_info_copy = rds_ib_stats_info_copy,
565 .get_mr = rds_ib_get_mr,
566 .sync_mr = rds_ib_sync_mr,
567 .free_mr = rds_ib_free_mr,
568 .flush_mrs = rds_ib_flush_mrs,
569 .get_tos_map = rds_ib_get_tos_map,
570 .t_owner = THIS_MODULE,
571 .t_name = "infiniband",
572 .t_unloading = rds_ib_is_unloading,
573 .t_type = RDS_TRANS_IB
576 int rds_ib_init(void)
580 INIT_LIST_HEAD(&rds_ib_devices);
582 ret = rds_ib_mr_init();
586 ret = ib_register_client(&rds_ib_client);
590 ret = rds_ib_sysctl_init();
594 ret = rds_ib_recv_init();
598 rds_trans_register(&rds_ib_transport);
600 rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
601 #if IS_ENABLED(CONFIG_IPV6)
602 rds_info_register_func(RDS6_INFO_IB_CONNECTIONS, rds6_ib_ic_info);
608 rds_ib_sysctl_exit();
610 rds_ib_unregister_client();
617 MODULE_LICENSE("GPL");