Loading...
1/*
2 * Copyright (c) 2006, 2019 Oracle and/or its affiliates. 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#include <net/addrconf.h>
43
44#include "rds_single_path.h"
45#include "rds.h"
46#include "ib.h"
47#include "ib_mr.h"
48
49static unsigned int rds_ib_mr_1m_pool_size = RDS_MR_1M_POOL_SIZE;
50static unsigned int rds_ib_mr_8k_pool_size = RDS_MR_8K_POOL_SIZE;
51unsigned int rds_ib_retry_count = RDS_IB_DEFAULT_RETRY_COUNT;
52static atomic_t rds_ib_unloading;
53
54module_param(rds_ib_mr_1m_pool_size, int, 0444);
55MODULE_PARM_DESC(rds_ib_mr_1m_pool_size, " Max number of 1M mr per HCA");
56module_param(rds_ib_mr_8k_pool_size, int, 0444);
57MODULE_PARM_DESC(rds_ib_mr_8k_pool_size, " Max number of 8K mr per HCA");
58module_param(rds_ib_retry_count, int, 0444);
59MODULE_PARM_DESC(rds_ib_retry_count, " Number of hw retries before reporting an error");
60
61/*
62 * we have a clumsy combination of RCU and a rwsem protecting this list
63 * because it is used both in the get_mr fast path and while blocking in
64 * the FMR flushing path.
65 */
66DECLARE_RWSEM(rds_ib_devices_lock);
67struct list_head rds_ib_devices;
68
69/* NOTE: if also grabbing ibdev lock, grab this first */
70DEFINE_SPINLOCK(ib_nodev_conns_lock);
71LIST_HEAD(ib_nodev_conns);
72
73static void rds_ib_nodev_connect(void)
74{
75 struct rds_ib_connection *ic;
76
77 spin_lock(&ib_nodev_conns_lock);
78 list_for_each_entry(ic, &ib_nodev_conns, ib_node)
79 rds_conn_connect_if_down(ic->conn);
80 spin_unlock(&ib_nodev_conns_lock);
81}
82
83static void rds_ib_dev_shutdown(struct rds_ib_device *rds_ibdev)
84{
85 struct rds_ib_connection *ic;
86 unsigned long flags;
87
88 spin_lock_irqsave(&rds_ibdev->spinlock, flags);
89 list_for_each_entry(ic, &rds_ibdev->conn_list, ib_node)
90 rds_conn_path_drop(&ic->conn->c_path[0], true);
91 spin_unlock_irqrestore(&rds_ibdev->spinlock, flags);
92}
93
94/*
95 * rds_ib_destroy_mr_pool() blocks on a few things and mrs drop references
96 * from interrupt context so we push freing off into a work struct in krdsd.
97 */
98static void rds_ib_dev_free(struct work_struct *work)
99{
100 struct rds_ib_ipaddr *i_ipaddr, *i_next;
101 struct rds_ib_device *rds_ibdev = container_of(work,
102 struct rds_ib_device, free_work);
103
104 if (rds_ibdev->mr_8k_pool)
105 rds_ib_destroy_mr_pool(rds_ibdev->mr_8k_pool);
106 if (rds_ibdev->mr_1m_pool)
107 rds_ib_destroy_mr_pool(rds_ibdev->mr_1m_pool);
108 if (rds_ibdev->pd)
109 ib_dealloc_pd(rds_ibdev->pd);
110
111 list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) {
112 list_del(&i_ipaddr->list);
113 kfree(i_ipaddr);
114 }
115
116 kfree(rds_ibdev->vector_load);
117
118 kfree(rds_ibdev);
119}
120
121void rds_ib_dev_put(struct rds_ib_device *rds_ibdev)
122{
123 BUG_ON(refcount_read(&rds_ibdev->refcount) == 0);
124 if (refcount_dec_and_test(&rds_ibdev->refcount))
125 queue_work(rds_wq, &rds_ibdev->free_work);
126}
127
128static int rds_ib_add_one(struct ib_device *device)
129{
130 struct rds_ib_device *rds_ibdev;
131 int ret;
132
133 /* Only handle IB (no iWARP) devices */
134 if (device->node_type != RDMA_NODE_IB_CA)
135 return -EOPNOTSUPP;
136
137 /* Device must support FRWR */
138 if (!(device->attrs.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS))
139 return -EOPNOTSUPP;
140
141 rds_ibdev = kzalloc_node(sizeof(struct rds_ib_device), GFP_KERNEL,
142 ibdev_to_node(device));
143 if (!rds_ibdev)
144 return -ENOMEM;
145
146 spin_lock_init(&rds_ibdev->spinlock);
147 refcount_set(&rds_ibdev->refcount, 1);
148 INIT_WORK(&rds_ibdev->free_work, rds_ib_dev_free);
149
150 INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
151 INIT_LIST_HEAD(&rds_ibdev->conn_list);
152
153 rds_ibdev->max_wrs = device->attrs.max_qp_wr;
154 rds_ibdev->max_sge = min(device->attrs.max_send_sge, RDS_IB_MAX_SGE);
155
156 rds_ibdev->odp_capable =
157 !!(device->attrs.kernel_cap_flags &
158 IBK_ON_DEMAND_PAGING) &&
159 !!(device->attrs.odp_caps.per_transport_caps.rc_odp_caps &
160 IB_ODP_SUPPORT_WRITE) &&
161 !!(device->attrs.odp_caps.per_transport_caps.rc_odp_caps &
162 IB_ODP_SUPPORT_READ);
163
164 rds_ibdev->max_1m_mrs = device->attrs.max_mr ?
165 min_t(unsigned int, (device->attrs.max_mr / 2),
166 rds_ib_mr_1m_pool_size) : rds_ib_mr_1m_pool_size;
167
168 rds_ibdev->max_8k_mrs = device->attrs.max_mr ?
169 min_t(unsigned int, ((device->attrs.max_mr / 2) * RDS_MR_8K_SCALE),
170 rds_ib_mr_8k_pool_size) : rds_ib_mr_8k_pool_size;
171
172 rds_ibdev->max_initiator_depth = device->attrs.max_qp_init_rd_atom;
173 rds_ibdev->max_responder_resources = device->attrs.max_qp_rd_atom;
174
175 rds_ibdev->vector_load = kcalloc(device->num_comp_vectors,
176 sizeof(int),
177 GFP_KERNEL);
178 if (!rds_ibdev->vector_load) {
179 pr_err("RDS/IB: %s failed to allocate vector memory\n",
180 __func__);
181 ret = -ENOMEM;
182 goto put_dev;
183 }
184
185 rds_ibdev->dev = device;
186 rds_ibdev->pd = ib_alloc_pd(device, 0);
187 if (IS_ERR(rds_ibdev->pd)) {
188 ret = PTR_ERR(rds_ibdev->pd);
189 rds_ibdev->pd = NULL;
190 goto put_dev;
191 }
192
193 rds_ibdev->mr_1m_pool =
194 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_1M_POOL);
195 if (IS_ERR(rds_ibdev->mr_1m_pool)) {
196 ret = PTR_ERR(rds_ibdev->mr_1m_pool);
197 rds_ibdev->mr_1m_pool = NULL;
198 goto put_dev;
199 }
200
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 ret = PTR_ERR(rds_ibdev->mr_8k_pool);
205 rds_ibdev->mr_8k_pool = NULL;
206 goto put_dev;
207 }
208
209 rdsdebug("RDS/IB: max_mr = %d, max_wrs = %d, max_sge = %d, max_1m_mrs = %d, max_8k_mrs = %d\n",
210 device->attrs.max_mr, rds_ibdev->max_wrs, rds_ibdev->max_sge,
211 rds_ibdev->max_1m_mrs, rds_ibdev->max_8k_mrs);
212
213 pr_info("RDS/IB: %s: added\n", device->name);
214
215 down_write(&rds_ib_devices_lock);
216 list_add_tail_rcu(&rds_ibdev->list, &rds_ib_devices);
217 up_write(&rds_ib_devices_lock);
218 refcount_inc(&rds_ibdev->refcount);
219
220 ib_set_client_data(device, &rds_ib_client, rds_ibdev);
221
222 rds_ib_nodev_connect();
223 return 0;
224
225put_dev:
226 rds_ib_dev_put(rds_ibdev);
227 return ret;
228}
229
230/*
231 * New connections use this to find the device to associate with the
232 * connection. It's not in the fast path so we're not concerned about the
233 * performance of the IB call. (As of this writing, it uses an interrupt
234 * blocking spinlock to serialize walking a per-device list of all registered
235 * clients.)
236 *
237 * RCU is used to handle incoming connections racing with device teardown.
238 * Rather than use a lock to serialize removal from the client_data and
239 * getting a new reference, we use an RCU grace period. The destruction
240 * path removes the device from client_data and then waits for all RCU
241 * readers to finish.
242 *
243 * A new connection can get NULL from this if its arriving on a
244 * device that is in the process of being removed.
245 */
246struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device)
247{
248 struct rds_ib_device *rds_ibdev;
249
250 rcu_read_lock();
251 rds_ibdev = ib_get_client_data(device, &rds_ib_client);
252 if (rds_ibdev)
253 refcount_inc(&rds_ibdev->refcount);
254 rcu_read_unlock();
255 return rds_ibdev;
256}
257
258/*
259 * The IB stack is letting us know that a device is going away. This can
260 * happen if the underlying HCA driver is removed or if PCI hotplug is removing
261 * the pci function, for example.
262 *
263 * This can be called at any time and can be racing with any other RDS path.
264 */
265static void rds_ib_remove_one(struct ib_device *device, void *client_data)
266{
267 struct rds_ib_device *rds_ibdev = client_data;
268
269 rds_ib_dev_shutdown(rds_ibdev);
270
271 /* stop connection attempts from getting a reference to this device. */
272 ib_set_client_data(device, &rds_ib_client, NULL);
273
274 down_write(&rds_ib_devices_lock);
275 list_del_rcu(&rds_ibdev->list);
276 up_write(&rds_ib_devices_lock);
277
278 /*
279 * This synchronize rcu is waiting for readers of both the ib
280 * client data and the devices list to finish before we drop
281 * both of those references.
282 */
283 synchronize_rcu();
284 rds_ib_dev_put(rds_ibdev);
285 rds_ib_dev_put(rds_ibdev);
286}
287
288struct ib_client rds_ib_client = {
289 .name = "rds_ib",
290 .add = rds_ib_add_one,
291 .remove = rds_ib_remove_one
292};
293
294static int rds_ib_conn_info_visitor(struct rds_connection *conn,
295 void *buffer)
296{
297 struct rds_info_rdma_connection *iinfo = buffer;
298 struct rds_ib_connection *ic = conn->c_transport_data;
299
300 /* We will only ever look at IB transports */
301 if (conn->c_trans != &rds_ib_transport)
302 return 0;
303 if (conn->c_isv6)
304 return 0;
305
306 iinfo->src_addr = conn->c_laddr.s6_addr32[3];
307 iinfo->dst_addr = conn->c_faddr.s6_addr32[3];
308 if (ic) {
309 iinfo->tos = conn->c_tos;
310 iinfo->sl = ic->i_sl;
311 }
312
313 memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid));
314 memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid));
315 if (rds_conn_state(conn) == RDS_CONN_UP) {
316 struct rds_ib_device *rds_ibdev;
317
318 rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo->src_gid,
319 (union ib_gid *)&iinfo->dst_gid);
320
321 rds_ibdev = ic->rds_ibdev;
322 iinfo->max_send_wr = ic->i_send_ring.w_nr;
323 iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
324 iinfo->max_send_sge = rds_ibdev->max_sge;
325 rds_ib_get_mr_info(rds_ibdev, iinfo);
326 iinfo->cache_allocs = atomic_read(&ic->i_cache_allocs);
327 }
328 return 1;
329}
330
331#if IS_ENABLED(CONFIG_IPV6)
332/* IPv6 version of rds_ib_conn_info_visitor(). */
333static int rds6_ib_conn_info_visitor(struct rds_connection *conn,
334 void *buffer)
335{
336 struct rds6_info_rdma_connection *iinfo6 = buffer;
337 struct rds_ib_connection *ic = conn->c_transport_data;
338
339 /* We will only ever look at IB transports */
340 if (conn->c_trans != &rds_ib_transport)
341 return 0;
342
343 iinfo6->src_addr = conn->c_laddr;
344 iinfo6->dst_addr = conn->c_faddr;
345 if (ic) {
346 iinfo6->tos = conn->c_tos;
347 iinfo6->sl = ic->i_sl;
348 }
349
350 memset(&iinfo6->src_gid, 0, sizeof(iinfo6->src_gid));
351 memset(&iinfo6->dst_gid, 0, sizeof(iinfo6->dst_gid));
352
353 if (rds_conn_state(conn) == RDS_CONN_UP) {
354 struct rds_ib_device *rds_ibdev;
355
356 rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo6->src_gid,
357 (union ib_gid *)&iinfo6->dst_gid);
358 rds_ibdev = ic->rds_ibdev;
359 iinfo6->max_send_wr = ic->i_send_ring.w_nr;
360 iinfo6->max_recv_wr = ic->i_recv_ring.w_nr;
361 iinfo6->max_send_sge = rds_ibdev->max_sge;
362 rds6_ib_get_mr_info(rds_ibdev, iinfo6);
363 iinfo6->cache_allocs = atomic_read(&ic->i_cache_allocs);
364 }
365 return 1;
366}
367#endif
368
369static void rds_ib_ic_info(struct socket *sock, unsigned int len,
370 struct rds_info_iterator *iter,
371 struct rds_info_lengths *lens)
372{
373 u64 buffer[(sizeof(struct rds_info_rdma_connection) + 7) / 8];
374
375 rds_for_each_conn_info(sock, len, iter, lens,
376 rds_ib_conn_info_visitor,
377 buffer,
378 sizeof(struct rds_info_rdma_connection));
379}
380
381#if IS_ENABLED(CONFIG_IPV6)
382/* IPv6 version of rds_ib_ic_info(). */
383static void rds6_ib_ic_info(struct socket *sock, unsigned int len,
384 struct rds_info_iterator *iter,
385 struct rds_info_lengths *lens)
386{
387 u64 buffer[(sizeof(struct rds6_info_rdma_connection) + 7) / 8];
388
389 rds_for_each_conn_info(sock, len, iter, lens,
390 rds6_ib_conn_info_visitor,
391 buffer,
392 sizeof(struct rds6_info_rdma_connection));
393}
394#endif
395
396/*
397 * Early RDS/IB was built to only bind to an address if there is an IPoIB
398 * device with that address set.
399 *
400 * If it were me, I'd advocate for something more flexible. Sending and
401 * receiving should be device-agnostic. Transports would try and maintain
402 * connections between peers who have messages queued. Userspace would be
403 * allowed to influence which paths have priority. We could call userspace
404 * asserting this policy "routing".
405 */
406static int rds_ib_laddr_check(struct net *net, const struct in6_addr *addr,
407 __u32 scope_id)
408{
409 int ret;
410 struct rdma_cm_id *cm_id;
411#if IS_ENABLED(CONFIG_IPV6)
412 struct sockaddr_in6 sin6;
413#endif
414 struct sockaddr_in sin;
415 struct sockaddr *sa;
416 bool isv4;
417
418 isv4 = ipv6_addr_v4mapped(addr);
419 /* Create a CMA ID and try to bind it. This catches both
420 * IB and iWARP capable NICs.
421 */
422 cm_id = rdma_create_id(&init_net, rds_rdma_cm_event_handler,
423 NULL, RDMA_PS_TCP, IB_QPT_RC);
424 if (IS_ERR(cm_id))
425 return PTR_ERR(cm_id);
426
427 if (isv4) {
428 memset(&sin, 0, sizeof(sin));
429 sin.sin_family = AF_INET;
430 sin.sin_addr.s_addr = addr->s6_addr32[3];
431 sa = (struct sockaddr *)&sin;
432 } else {
433#if IS_ENABLED(CONFIG_IPV6)
434 memset(&sin6, 0, sizeof(sin6));
435 sin6.sin6_family = AF_INET6;
436 sin6.sin6_addr = *addr;
437 sin6.sin6_scope_id = scope_id;
438 sa = (struct sockaddr *)&sin6;
439
440 /* XXX Do a special IPv6 link local address check here. The
441 * reason is that rdma_bind_addr() always succeeds with IPv6
442 * link local address regardless it is indeed configured in a
443 * system.
444 */
445 if (ipv6_addr_type(addr) & IPV6_ADDR_LINKLOCAL) {
446 struct net_device *dev;
447
448 if (scope_id == 0) {
449 ret = -EADDRNOTAVAIL;
450 goto out;
451 }
452
453 /* Use init_net for now as RDS is not network
454 * name space aware.
455 */
456 dev = dev_get_by_index(&init_net, scope_id);
457 if (!dev) {
458 ret = -EADDRNOTAVAIL;
459 goto out;
460 }
461 if (!ipv6_chk_addr(&init_net, addr, dev, 1)) {
462 dev_put(dev);
463 ret = -EADDRNOTAVAIL;
464 goto out;
465 }
466 dev_put(dev);
467 }
468#else
469 ret = -EADDRNOTAVAIL;
470 goto out;
471#endif
472 }
473
474 /* rdma_bind_addr will only succeed for IB & iWARP devices */
475 ret = rdma_bind_addr(cm_id, sa);
476 /* due to this, we will claim to support iWARP devices unless we
477 check node_type. */
478 if (ret || !cm_id->device ||
479 cm_id->device->node_type != RDMA_NODE_IB_CA)
480 ret = -EADDRNOTAVAIL;
481
482 rdsdebug("addr %pI6c%%%u ret %d node type %d\n",
483 addr, scope_id, ret,
484 cm_id->device ? cm_id->device->node_type : -1);
485
486out:
487 rdma_destroy_id(cm_id);
488
489 return ret;
490}
491
492static void rds_ib_unregister_client(void)
493{
494 ib_unregister_client(&rds_ib_client);
495 /* wait for rds_ib_dev_free() to complete */
496 flush_workqueue(rds_wq);
497}
498
499static void rds_ib_set_unloading(void)
500{
501 atomic_set(&rds_ib_unloading, 1);
502}
503
504static bool rds_ib_is_unloading(struct rds_connection *conn)
505{
506 struct rds_conn_path *cp = &conn->c_path[0];
507
508 return (test_bit(RDS_DESTROY_PENDING, &cp->cp_flags) ||
509 atomic_read(&rds_ib_unloading) != 0);
510}
511
512void rds_ib_exit(void)
513{
514 rds_ib_set_unloading();
515 synchronize_rcu();
516 rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
517#if IS_ENABLED(CONFIG_IPV6)
518 rds_info_deregister_func(RDS6_INFO_IB_CONNECTIONS, rds6_ib_ic_info);
519#endif
520 rds_ib_unregister_client();
521 rds_ib_destroy_nodev_conns();
522 rds_ib_sysctl_exit();
523 rds_ib_recv_exit();
524 rds_trans_unregister(&rds_ib_transport);
525 rds_ib_mr_exit();
526}
527
528static u8 rds_ib_get_tos_map(u8 tos)
529{
530 /* 1:1 user to transport map for RDMA transport.
531 * In future, if custom map is desired, hook can export
532 * user configurable map.
533 */
534 return tos;
535}
536
537struct rds_transport rds_ib_transport = {
538 .laddr_check = rds_ib_laddr_check,
539 .xmit_path_complete = rds_ib_xmit_path_complete,
540 .xmit = rds_ib_xmit,
541 .xmit_rdma = rds_ib_xmit_rdma,
542 .xmit_atomic = rds_ib_xmit_atomic,
543 .recv_path = rds_ib_recv_path,
544 .conn_alloc = rds_ib_conn_alloc,
545 .conn_free = rds_ib_conn_free,
546 .conn_path_connect = rds_ib_conn_path_connect,
547 .conn_path_shutdown = rds_ib_conn_path_shutdown,
548 .inc_copy_to_user = rds_ib_inc_copy_to_user,
549 .inc_free = rds_ib_inc_free,
550 .cm_initiate_connect = rds_ib_cm_initiate_connect,
551 .cm_handle_connect = rds_ib_cm_handle_connect,
552 .cm_connect_complete = rds_ib_cm_connect_complete,
553 .stats_info_copy = rds_ib_stats_info_copy,
554 .exit = rds_ib_exit,
555 .get_mr = rds_ib_get_mr,
556 .sync_mr = rds_ib_sync_mr,
557 .free_mr = rds_ib_free_mr,
558 .flush_mrs = rds_ib_flush_mrs,
559 .get_tos_map = rds_ib_get_tos_map,
560 .t_owner = THIS_MODULE,
561 .t_name = "infiniband",
562 .t_unloading = rds_ib_is_unloading,
563 .t_type = RDS_TRANS_IB
564};
565
566int rds_ib_init(void)
567{
568 int ret;
569
570 INIT_LIST_HEAD(&rds_ib_devices);
571
572 ret = rds_ib_mr_init();
573 if (ret)
574 goto out;
575
576 ret = ib_register_client(&rds_ib_client);
577 if (ret)
578 goto out_mr_exit;
579
580 ret = rds_ib_sysctl_init();
581 if (ret)
582 goto out_ibreg;
583
584 ret = rds_ib_recv_init();
585 if (ret)
586 goto out_sysctl;
587
588 rds_trans_register(&rds_ib_transport);
589
590 rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
591#if IS_ENABLED(CONFIG_IPV6)
592 rds_info_register_func(RDS6_INFO_IB_CONNECTIONS, rds6_ib_ic_info);
593#endif
594
595 goto out;
596
597out_sysctl:
598 rds_ib_sysctl_exit();
599out_ibreg:
600 rds_ib_unregister_client();
601out_mr_exit:
602 rds_ib_mr_exit();
603out:
604 return ret;
605}
606
607MODULE_LICENSE("GPL");
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_single_path.h"
44#include "rds.h"
45#include "ib.h"
46#include "ib_mr.h"
47
48static unsigned int rds_ib_mr_1m_pool_size = RDS_MR_1M_POOL_SIZE;
49static unsigned int rds_ib_mr_8k_pool_size = RDS_MR_8K_POOL_SIZE;
50unsigned int rds_ib_retry_count = RDS_IB_DEFAULT_RETRY_COUNT;
51static atomic_t rds_ib_unloading;
52
53module_param(rds_ib_mr_1m_pool_size, int, 0444);
54MODULE_PARM_DESC(rds_ib_mr_1m_pool_size, " Max number of 1M mr per HCA");
55module_param(rds_ib_mr_8k_pool_size, int, 0444);
56MODULE_PARM_DESC(rds_ib_mr_8k_pool_size, " Max number of 8K mr per HCA");
57module_param(rds_ib_retry_count, int, 0444);
58MODULE_PARM_DESC(rds_ib_retry_count, " Number of hw retries before reporting an error");
59
60/*
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.
64 */
65DECLARE_RWSEM(rds_ib_devices_lock);
66struct list_head rds_ib_devices;
67
68/* NOTE: if also grabbing ibdev lock, grab this first */
69DEFINE_SPINLOCK(ib_nodev_conns_lock);
70LIST_HEAD(ib_nodev_conns);
71
72static void rds_ib_nodev_connect(void)
73{
74 struct rds_ib_connection *ic;
75
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);
80}
81
82static void rds_ib_dev_shutdown(struct rds_ib_device *rds_ibdev)
83{
84 struct rds_ib_connection *ic;
85 unsigned long flags;
86
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);
91}
92
93/*
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.
96 */
97static void rds_ib_dev_free(struct work_struct *work)
98{
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);
102
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);
107 if (rds_ibdev->pd)
108 ib_dealloc_pd(rds_ibdev->pd);
109
110 list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) {
111 list_del(&i_ipaddr->list);
112 kfree(i_ipaddr);
113 }
114
115 kfree(rds_ibdev->vector_load);
116
117 kfree(rds_ibdev);
118}
119
120void rds_ib_dev_put(struct rds_ib_device *rds_ibdev)
121{
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);
125}
126
127static void rds_ib_add_one(struct ib_device *device)
128{
129 struct rds_ib_device *rds_ibdev;
130 bool has_fr, has_fmr;
131
132 /* Only handle IB (no iWARP) devices */
133 if (device->node_type != RDMA_NODE_IB_CA)
134 return;
135
136 rds_ibdev = kzalloc_node(sizeof(struct rds_ib_device), GFP_KERNEL,
137 ibdev_to_node(device));
138 if (!rds_ibdev)
139 return;
140
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);
144
145 rds_ibdev->max_wrs = device->attrs.max_qp_wr;
146 rds_ibdev->max_sge = min(device->attrs.max_sge, RDS_IB_MAX_SGE);
147
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);
153
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;
158
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;
162
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;
165
166 rds_ibdev->vector_load = kzalloc(sizeof(int) * device->num_comp_vectors,
167 GFP_KERNEL);
168 if (!rds_ibdev->vector_load) {
169 pr_err("RDS/IB: %s failed to allocate vector memory\n",
170 __func__);
171 goto put_dev;
172 }
173
174 rds_ibdev->dev = device;
175 rds_ibdev->pd = ib_alloc_pd(device, 0);
176 if (IS_ERR(rds_ibdev->pd)) {
177 rds_ibdev->pd = NULL;
178 goto put_dev;
179 }
180
181 rds_ibdev->mr_1m_pool =
182 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_1M_POOL);
183 if (IS_ERR(rds_ibdev->mr_1m_pool)) {
184 rds_ibdev->mr_1m_pool = NULL;
185 goto put_dev;
186 }
187
188 rds_ibdev->mr_8k_pool =
189 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_8K_POOL);
190 if (IS_ERR(rds_ibdev->mr_8k_pool)) {
191 rds_ibdev->mr_8k_pool = NULL;
192 goto put_dev;
193 }
194
195 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",
196 device->attrs.max_fmr, rds_ibdev->max_wrs, rds_ibdev->max_sge,
197 rds_ibdev->fmr_max_remaps, rds_ibdev->max_1m_mrs,
198 rds_ibdev->max_8k_mrs);
199
200 pr_info("RDS/IB: %s: %s supported and preferred\n",
201 device->name,
202 rds_ibdev->use_fastreg ? "FRMR" : "FMR");
203
204 INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
205 INIT_LIST_HEAD(&rds_ibdev->conn_list);
206
207 down_write(&rds_ib_devices_lock);
208 list_add_tail_rcu(&rds_ibdev->list, &rds_ib_devices);
209 up_write(&rds_ib_devices_lock);
210 refcount_inc(&rds_ibdev->refcount);
211
212 ib_set_client_data(device, &rds_ib_client, rds_ibdev);
213 refcount_inc(&rds_ibdev->refcount);
214
215 rds_ib_nodev_connect();
216
217put_dev:
218 rds_ib_dev_put(rds_ibdev);
219}
220
221/*
222 * New connections use this to find the device to associate with the
223 * connection. It's not in the fast path so we're not concerned about the
224 * performance of the IB call. (As of this writing, it uses an interrupt
225 * blocking spinlock to serialize walking a per-device list of all registered
226 * clients.)
227 *
228 * RCU is used to handle incoming connections racing with device teardown.
229 * Rather than use a lock to serialize removal from the client_data and
230 * getting a new reference, we use an RCU grace period. The destruction
231 * path removes the device from client_data and then waits for all RCU
232 * readers to finish.
233 *
234 * A new connection can get NULL from this if its arriving on a
235 * device that is in the process of being removed.
236 */
237struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device)
238{
239 struct rds_ib_device *rds_ibdev;
240
241 rcu_read_lock();
242 rds_ibdev = ib_get_client_data(device, &rds_ib_client);
243 if (rds_ibdev)
244 refcount_inc(&rds_ibdev->refcount);
245 rcu_read_unlock();
246 return rds_ibdev;
247}
248
249/*
250 * The IB stack is letting us know that a device is going away. This can
251 * happen if the underlying HCA driver is removed or if PCI hotplug is removing
252 * the pci function, for example.
253 *
254 * This can be called at any time and can be racing with any other RDS path.
255 */
256static void rds_ib_remove_one(struct ib_device *device, void *client_data)
257{
258 struct rds_ib_device *rds_ibdev = client_data;
259
260 if (!rds_ibdev)
261 return;
262
263 rds_ib_dev_shutdown(rds_ibdev);
264
265 /* stop connection attempts from getting a reference to this device. */
266 ib_set_client_data(device, &rds_ib_client, NULL);
267
268 down_write(&rds_ib_devices_lock);
269 list_del_rcu(&rds_ibdev->list);
270 up_write(&rds_ib_devices_lock);
271
272 /*
273 * This synchronize rcu is waiting for readers of both the ib
274 * client data and the devices list to finish before we drop
275 * both of those references.
276 */
277 synchronize_rcu();
278 rds_ib_dev_put(rds_ibdev);
279 rds_ib_dev_put(rds_ibdev);
280}
281
282struct ib_client rds_ib_client = {
283 .name = "rds_ib",
284 .add = rds_ib_add_one,
285 .remove = rds_ib_remove_one
286};
287
288static int rds_ib_conn_info_visitor(struct rds_connection *conn,
289 void *buffer)
290{
291 struct rds_info_rdma_connection *iinfo = buffer;
292 struct rds_ib_connection *ic;
293
294 /* We will only ever look at IB transports */
295 if (conn->c_trans != &rds_ib_transport)
296 return 0;
297
298 iinfo->src_addr = conn->c_laddr;
299 iinfo->dst_addr = conn->c_faddr;
300
301 memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid));
302 memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid));
303 if (rds_conn_state(conn) == RDS_CONN_UP) {
304 struct rds_ib_device *rds_ibdev;
305
306 ic = conn->c_transport_data;
307
308 rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo->src_gid,
309 (union ib_gid *)&iinfo->dst_gid);
310
311 rds_ibdev = ic->rds_ibdev;
312 iinfo->max_send_wr = ic->i_send_ring.w_nr;
313 iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
314 iinfo->max_send_sge = rds_ibdev->max_sge;
315 rds_ib_get_mr_info(rds_ibdev, iinfo);
316 }
317 return 1;
318}
319
320static void rds_ib_ic_info(struct socket *sock, unsigned int len,
321 struct rds_info_iterator *iter,
322 struct rds_info_lengths *lens)
323{
324 u64 buffer[(sizeof(struct rds_info_rdma_connection) + 7) / 8];
325
326 rds_for_each_conn_info(sock, len, iter, lens,
327 rds_ib_conn_info_visitor,
328 buffer,
329 sizeof(struct rds_info_rdma_connection));
330}
331
332
333/*
334 * Early RDS/IB was built to only bind to an address if there is an IPoIB
335 * device with that address set.
336 *
337 * If it were me, I'd advocate for something more flexible. Sending and
338 * receiving should be device-agnostic. Transports would try and maintain
339 * connections between peers who have messages queued. Userspace would be
340 * allowed to influence which paths have priority. We could call userspace
341 * asserting this policy "routing".
342 */
343static int rds_ib_laddr_check(struct net *net, __be32 addr)
344{
345 int ret;
346 struct rdma_cm_id *cm_id;
347 struct sockaddr_in sin;
348
349 /* Create a CMA ID and try to bind it. This catches both
350 * IB and iWARP capable NICs.
351 */
352 cm_id = rdma_create_id(&init_net, rds_rdma_cm_event_handler,
353 NULL, RDMA_PS_TCP, IB_QPT_RC);
354 if (IS_ERR(cm_id))
355 return PTR_ERR(cm_id);
356
357 memset(&sin, 0, sizeof(sin));
358 sin.sin_family = AF_INET;
359 sin.sin_addr.s_addr = addr;
360
361 /* rdma_bind_addr will only succeed for IB & iWARP devices */
362 ret = rdma_bind_addr(cm_id, (struct sockaddr *)&sin);
363 /* due to this, we will claim to support iWARP devices unless we
364 check node_type. */
365 if (ret || !cm_id->device ||
366 cm_id->device->node_type != RDMA_NODE_IB_CA)
367 ret = -EADDRNOTAVAIL;
368
369 rdsdebug("addr %pI4 ret %d node type %d\n",
370 &addr, ret,
371 cm_id->device ? cm_id->device->node_type : -1);
372
373 rdma_destroy_id(cm_id);
374
375 return ret;
376}
377
378static void rds_ib_unregister_client(void)
379{
380 ib_unregister_client(&rds_ib_client);
381 /* wait for rds_ib_dev_free() to complete */
382 flush_workqueue(rds_wq);
383}
384
385static void rds_ib_set_unloading(void)
386{
387 atomic_set(&rds_ib_unloading, 1);
388}
389
390static bool rds_ib_is_unloading(struct rds_connection *conn)
391{
392 struct rds_conn_path *cp = &conn->c_path[0];
393
394 return (test_bit(RDS_DESTROY_PENDING, &cp->cp_flags) ||
395 atomic_read(&rds_ib_unloading) != 0);
396}
397
398void rds_ib_exit(void)
399{
400 rds_ib_set_unloading();
401 synchronize_rcu();
402 rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
403 rds_ib_unregister_client();
404 rds_ib_destroy_nodev_conns();
405 rds_ib_sysctl_exit();
406 rds_ib_recv_exit();
407 rds_trans_unregister(&rds_ib_transport);
408 rds_ib_mr_exit();
409}
410
411struct rds_transport rds_ib_transport = {
412 .laddr_check = rds_ib_laddr_check,
413 .xmit_path_complete = rds_ib_xmit_path_complete,
414 .xmit = rds_ib_xmit,
415 .xmit_rdma = rds_ib_xmit_rdma,
416 .xmit_atomic = rds_ib_xmit_atomic,
417 .recv_path = rds_ib_recv_path,
418 .conn_alloc = rds_ib_conn_alloc,
419 .conn_free = rds_ib_conn_free,
420 .conn_path_connect = rds_ib_conn_path_connect,
421 .conn_path_shutdown = rds_ib_conn_path_shutdown,
422 .inc_copy_to_user = rds_ib_inc_copy_to_user,
423 .inc_free = rds_ib_inc_free,
424 .cm_initiate_connect = rds_ib_cm_initiate_connect,
425 .cm_handle_connect = rds_ib_cm_handle_connect,
426 .cm_connect_complete = rds_ib_cm_connect_complete,
427 .stats_info_copy = rds_ib_stats_info_copy,
428 .exit = rds_ib_exit,
429 .get_mr = rds_ib_get_mr,
430 .sync_mr = rds_ib_sync_mr,
431 .free_mr = rds_ib_free_mr,
432 .flush_mrs = rds_ib_flush_mrs,
433 .t_owner = THIS_MODULE,
434 .t_name = "infiniband",
435 .t_unloading = rds_ib_is_unloading,
436 .t_type = RDS_TRANS_IB
437};
438
439int rds_ib_init(void)
440{
441 int ret;
442
443 INIT_LIST_HEAD(&rds_ib_devices);
444
445 ret = rds_ib_mr_init();
446 if (ret)
447 goto out;
448
449 ret = ib_register_client(&rds_ib_client);
450 if (ret)
451 goto out_mr_exit;
452
453 ret = rds_ib_sysctl_init();
454 if (ret)
455 goto out_ibreg;
456
457 ret = rds_ib_recv_init();
458 if (ret)
459 goto out_sysctl;
460
461 rds_trans_register(&rds_ib_transport);
462
463 rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
464
465 goto out;
466
467out_sysctl:
468 rds_ib_sysctl_exit();
469out_ibreg:
470 rds_ib_unregister_client();
471out_mr_exit:
472 rds_ib_mr_exit();
473out:
474 return ret;
475}
476
477MODULE_LICENSE("GPL");
478