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 void rds_ib_add_one(struct ib_device *device)
132 struct rds_ib_device *rds_ibdev;
133 bool has_fr, has_fmr;
135 /* Only handle IB (no iWARP) devices */
136 if (device->node_type != RDMA_NODE_IB_CA)
139 rds_ibdev = kzalloc_node(sizeof(struct rds_ib_device), GFP_KERNEL,
140 ibdev_to_node(device));
144 spin_lock_init(&rds_ibdev->spinlock);
145 refcount_set(&rds_ibdev->refcount, 1);
146 INIT_WORK(&rds_ibdev->free_work, rds_ib_dev_free);
148 INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
149 INIT_LIST_HEAD(&rds_ibdev->conn_list);
151 rds_ibdev->max_wrs = device->attrs.max_qp_wr;
152 rds_ibdev->max_sge = min(device->attrs.max_send_sge, RDS_IB_MAX_SGE);
154 has_fr = (device->attrs.device_cap_flags &
155 IB_DEVICE_MEM_MGT_EXTENSIONS);
156 has_fmr = (device->ops.alloc_fmr && device->ops.dealloc_fmr &&
157 device->ops.map_phys_fmr && device->ops.unmap_fmr);
158 rds_ibdev->use_fastreg = (has_fr && !has_fmr);
160 rds_ibdev->fmr_max_remaps = device->attrs.max_map_per_fmr?: 32;
161 rds_ibdev->max_1m_mrs = device->attrs.max_mr ?
162 min_t(unsigned int, (device->attrs.max_mr / 2),
163 rds_ib_mr_1m_pool_size) : rds_ib_mr_1m_pool_size;
165 rds_ibdev->max_8k_mrs = device->attrs.max_mr ?
166 min_t(unsigned int, ((device->attrs.max_mr / 2) * RDS_MR_8K_SCALE),
167 rds_ib_mr_8k_pool_size) : rds_ib_mr_8k_pool_size;
169 rds_ibdev->max_initiator_depth = device->attrs.max_qp_init_rd_atom;
170 rds_ibdev->max_responder_resources = device->attrs.max_qp_rd_atom;
172 rds_ibdev->vector_load = kcalloc(device->num_comp_vectors,
175 if (!rds_ibdev->vector_load) {
176 pr_err("RDS/IB: %s failed to allocate vector memory\n",
181 rds_ibdev->dev = device;
182 rds_ibdev->pd = ib_alloc_pd(device, 0);
183 if (IS_ERR(rds_ibdev->pd)) {
184 rds_ibdev->pd = NULL;
187 rds_ibdev->rid_hdrs_pool = dma_pool_create(device->name,
189 sizeof(struct rds_header),
191 if (!rds_ibdev->rid_hdrs_pool)
194 rds_ibdev->mr_1m_pool =
195 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_1M_POOL);
196 if (IS_ERR(rds_ibdev->mr_1m_pool)) {
197 rds_ibdev->mr_1m_pool = NULL;
201 rds_ibdev->mr_8k_pool =
202 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_8K_POOL);
203 if (IS_ERR(rds_ibdev->mr_8k_pool)) {
204 rds_ibdev->mr_8k_pool = NULL;
208 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",
209 device->attrs.max_fmr, rds_ibdev->max_wrs, rds_ibdev->max_sge,
210 rds_ibdev->fmr_max_remaps, rds_ibdev->max_1m_mrs,
211 rds_ibdev->max_8k_mrs);
213 pr_info("RDS/IB: %s: %s supported and preferred\n",
215 rds_ibdev->use_fastreg ? "FRMR" : "FMR");
217 down_write(&rds_ib_devices_lock);
218 list_add_tail_rcu(&rds_ibdev->list, &rds_ib_devices);
219 up_write(&rds_ib_devices_lock);
220 refcount_inc(&rds_ibdev->refcount);
222 ib_set_client_data(device, &rds_ib_client, rds_ibdev);
223 refcount_inc(&rds_ibdev->refcount);
225 rds_ib_nodev_connect();
228 rds_ib_dev_put(rds_ibdev);
232 * New connections use this to find the device to associate with the
233 * connection. It's not in the fast path so we're not concerned about the
234 * performance of the IB call. (As of this writing, it uses an interrupt
235 * blocking spinlock to serialize walking a per-device list of all registered
238 * RCU is used to handle incoming connections racing with device teardown.
239 * Rather than use a lock to serialize removal from the client_data and
240 * getting a new reference, we use an RCU grace period. The destruction
241 * path removes the device from client_data and then waits for all RCU
244 * A new connection can get NULL from this if its arriving on a
245 * device that is in the process of being removed.
247 struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device)
249 struct rds_ib_device *rds_ibdev;
252 rds_ibdev = ib_get_client_data(device, &rds_ib_client);
254 refcount_inc(&rds_ibdev->refcount);
260 * The IB stack is letting us know that a device is going away. This can
261 * happen if the underlying HCA driver is removed or if PCI hotplug is removing
262 * the pci function, for example.
264 * This can be called at any time and can be racing with any other RDS path.
266 static void rds_ib_remove_one(struct ib_device *device, void *client_data)
268 struct rds_ib_device *rds_ibdev = client_data;
273 rds_ib_dev_shutdown(rds_ibdev);
275 /* stop connection attempts from getting a reference to this device. */
276 ib_set_client_data(device, &rds_ib_client, NULL);
278 down_write(&rds_ib_devices_lock);
279 list_del_rcu(&rds_ibdev->list);
280 up_write(&rds_ib_devices_lock);
283 * This synchronize rcu is waiting for readers of both the ib
284 * client data and the devices list to finish before we drop
285 * both of those references.
288 rds_ib_dev_put(rds_ibdev);
289 rds_ib_dev_put(rds_ibdev);
292 struct ib_client rds_ib_client = {
294 .add = rds_ib_add_one,
295 .remove = rds_ib_remove_one
298 static int rds_ib_conn_info_visitor(struct rds_connection *conn,
301 struct rds_info_rdma_connection *iinfo = buffer;
302 struct rds_ib_connection *ic = conn->c_transport_data;
304 /* We will only ever look at IB transports */
305 if (conn->c_trans != &rds_ib_transport)
310 iinfo->src_addr = conn->c_laddr.s6_addr32[3];
311 iinfo->dst_addr = conn->c_faddr.s6_addr32[3];
313 iinfo->tos = conn->c_tos;
314 iinfo->sl = ic->i_sl;
317 memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid));
318 memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid));
319 if (rds_conn_state(conn) == RDS_CONN_UP) {
320 struct rds_ib_device *rds_ibdev;
322 rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo->src_gid,
323 (union ib_gid *)&iinfo->dst_gid);
325 rds_ibdev = ic->rds_ibdev;
326 iinfo->max_send_wr = ic->i_send_ring.w_nr;
327 iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
328 iinfo->max_send_sge = rds_ibdev->max_sge;
329 rds_ib_get_mr_info(rds_ibdev, iinfo);
330 iinfo->cache_allocs = atomic_read(&ic->i_cache_allocs);
335 #if IS_ENABLED(CONFIG_IPV6)
336 /* IPv6 version of rds_ib_conn_info_visitor(). */
337 static int rds6_ib_conn_info_visitor(struct rds_connection *conn,
340 struct rds6_info_rdma_connection *iinfo6 = buffer;
341 struct rds_ib_connection *ic = conn->c_transport_data;
343 /* We will only ever look at IB transports */
344 if (conn->c_trans != &rds_ib_transport)
347 iinfo6->src_addr = conn->c_laddr;
348 iinfo6->dst_addr = conn->c_faddr;
350 iinfo6->tos = conn->c_tos;
351 iinfo6->sl = ic->i_sl;
354 memset(&iinfo6->src_gid, 0, sizeof(iinfo6->src_gid));
355 memset(&iinfo6->dst_gid, 0, sizeof(iinfo6->dst_gid));
357 if (rds_conn_state(conn) == RDS_CONN_UP) {
358 struct rds_ib_device *rds_ibdev;
360 rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo6->src_gid,
361 (union ib_gid *)&iinfo6->dst_gid);
362 rds_ibdev = ic->rds_ibdev;
363 iinfo6->max_send_wr = ic->i_send_ring.w_nr;
364 iinfo6->max_recv_wr = ic->i_recv_ring.w_nr;
365 iinfo6->max_send_sge = rds_ibdev->max_sge;
366 rds6_ib_get_mr_info(rds_ibdev, iinfo6);
367 iinfo6->cache_allocs = atomic_read(&ic->i_cache_allocs);
373 static void rds_ib_ic_info(struct socket *sock, unsigned int len,
374 struct rds_info_iterator *iter,
375 struct rds_info_lengths *lens)
377 u64 buffer[(sizeof(struct rds_info_rdma_connection) + 7) / 8];
379 rds_for_each_conn_info(sock, len, iter, lens,
380 rds_ib_conn_info_visitor,
382 sizeof(struct rds_info_rdma_connection));
385 #if IS_ENABLED(CONFIG_IPV6)
386 /* IPv6 version of rds_ib_ic_info(). */
387 static void rds6_ib_ic_info(struct socket *sock, unsigned int len,
388 struct rds_info_iterator *iter,
389 struct rds_info_lengths *lens)
391 u64 buffer[(sizeof(struct rds6_info_rdma_connection) + 7) / 8];
393 rds_for_each_conn_info(sock, len, iter, lens,
394 rds6_ib_conn_info_visitor,
396 sizeof(struct rds6_info_rdma_connection));
401 * Early RDS/IB was built to only bind to an address if there is an IPoIB
402 * device with that address set.
404 * If it were me, I'd advocate for something more flexible. Sending and
405 * receiving should be device-agnostic. Transports would try and maintain
406 * connections between peers who have messages queued. Userspace would be
407 * allowed to influence which paths have priority. We could call userspace
408 * asserting this policy "routing".
410 static int rds_ib_laddr_check(struct net *net, const struct in6_addr *addr,
414 struct rdma_cm_id *cm_id;
415 #if IS_ENABLED(CONFIG_IPV6)
416 struct sockaddr_in6 sin6;
418 struct sockaddr_in sin;
422 isv4 = ipv6_addr_v4mapped(addr);
423 /* Create a CMA ID and try to bind it. This catches both
424 * IB and iWARP capable NICs.
426 cm_id = rdma_create_id(&init_net, rds_rdma_cm_event_handler,
427 NULL, RDMA_PS_TCP, IB_QPT_RC);
429 return PTR_ERR(cm_id);
432 memset(&sin, 0, sizeof(sin));
433 sin.sin_family = AF_INET;
434 sin.sin_addr.s_addr = addr->s6_addr32[3];
435 sa = (struct sockaddr *)&sin;
437 #if IS_ENABLED(CONFIG_IPV6)
438 memset(&sin6, 0, sizeof(sin6));
439 sin6.sin6_family = AF_INET6;
440 sin6.sin6_addr = *addr;
441 sin6.sin6_scope_id = scope_id;
442 sa = (struct sockaddr *)&sin6;
444 /* XXX Do a special IPv6 link local address check here. The
445 * reason is that rdma_bind_addr() always succeeds with IPv6
446 * link local address regardless it is indeed configured in a
449 if (ipv6_addr_type(addr) & IPV6_ADDR_LINKLOCAL) {
450 struct net_device *dev;
453 ret = -EADDRNOTAVAIL;
457 /* Use init_net for now as RDS is not network
460 dev = dev_get_by_index(&init_net, scope_id);
462 ret = -EADDRNOTAVAIL;
465 if (!ipv6_chk_addr(&init_net, addr, dev, 1)) {
467 ret = -EADDRNOTAVAIL;
473 ret = -EADDRNOTAVAIL;
478 /* rdma_bind_addr will only succeed for IB & iWARP devices */
479 ret = rdma_bind_addr(cm_id, sa);
480 /* due to this, we will claim to support iWARP devices unless we
482 if (ret || !cm_id->device ||
483 cm_id->device->node_type != RDMA_NODE_IB_CA)
484 ret = -EADDRNOTAVAIL;
486 rdsdebug("addr %pI6c%%%u ret %d node type %d\n",
488 cm_id->device ? cm_id->device->node_type : -1);
491 rdma_destroy_id(cm_id);
496 static void rds_ib_unregister_client(void)
498 ib_unregister_client(&rds_ib_client);
499 /* wait for rds_ib_dev_free() to complete */
500 flush_workqueue(rds_wq);
503 static void rds_ib_set_unloading(void)
505 atomic_set(&rds_ib_unloading, 1);
508 static bool rds_ib_is_unloading(struct rds_connection *conn)
510 struct rds_conn_path *cp = &conn->c_path[0];
512 return (test_bit(RDS_DESTROY_PENDING, &cp->cp_flags) ||
513 atomic_read(&rds_ib_unloading) != 0);
516 void rds_ib_exit(void)
518 rds_ib_set_unloading();
520 rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
521 #if IS_ENABLED(CONFIG_IPV6)
522 rds_info_deregister_func(RDS6_INFO_IB_CONNECTIONS, rds6_ib_ic_info);
524 rds_ib_unregister_client();
525 rds_ib_destroy_nodev_conns();
526 rds_ib_sysctl_exit();
528 rds_trans_unregister(&rds_ib_transport);
532 static u8 rds_ib_get_tos_map(u8 tos)
534 /* 1:1 user to transport map for RDMA transport.
535 * In future, if custom map is desired, hook can export
536 * user configurable map.
541 struct rds_transport rds_ib_transport = {
542 .laddr_check = rds_ib_laddr_check,
543 .xmit_path_complete = rds_ib_xmit_path_complete,
545 .xmit_rdma = rds_ib_xmit_rdma,
546 .xmit_atomic = rds_ib_xmit_atomic,
547 .recv_path = rds_ib_recv_path,
548 .conn_alloc = rds_ib_conn_alloc,
549 .conn_free = rds_ib_conn_free,
550 .conn_path_connect = rds_ib_conn_path_connect,
551 .conn_path_shutdown = rds_ib_conn_path_shutdown,
552 .inc_copy_to_user = rds_ib_inc_copy_to_user,
553 .inc_free = rds_ib_inc_free,
554 .cm_initiate_connect = rds_ib_cm_initiate_connect,
555 .cm_handle_connect = rds_ib_cm_handle_connect,
556 .cm_connect_complete = rds_ib_cm_connect_complete,
557 .stats_info_copy = rds_ib_stats_info_copy,
559 .get_mr = rds_ib_get_mr,
560 .sync_mr = rds_ib_sync_mr,
561 .free_mr = rds_ib_free_mr,
562 .flush_mrs = rds_ib_flush_mrs,
563 .get_tos_map = rds_ib_get_tos_map,
564 .t_owner = THIS_MODULE,
565 .t_name = "infiniband",
566 .t_unloading = rds_ib_is_unloading,
567 .t_type = RDS_TRANS_IB
570 int rds_ib_init(void)
574 INIT_LIST_HEAD(&rds_ib_devices);
576 ret = rds_ib_mr_init();
580 ret = ib_register_client(&rds_ib_client);
584 ret = rds_ib_sysctl_init();
588 ret = rds_ib_recv_init();
592 rds_trans_register(&rds_ib_transport);
594 rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
595 #if IS_ENABLED(CONFIG_IPV6)
596 rds_info_register_func(RDS6_INFO_IB_CONNECTIONS, rds6_ib_ic_info);
602 rds_ib_sysctl_exit();
604 rds_ib_unregister_client();
611 MODULE_LICENSE("GPL");