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/dmapool.h>
34#include <linux/kernel.h>
35#include <linux/in.h>
36#include <linux/if.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>
44
45#include "rds_single_path.h"
46#include "rds.h"
47#include "ib.h"
48#include "ib_mr.h"
49
50static unsigned int rds_ib_mr_1m_pool_size = RDS_MR_1M_POOL_SIZE;
51static unsigned int rds_ib_mr_8k_pool_size = RDS_MR_8K_POOL_SIZE;
52unsigned int rds_ib_retry_count = RDS_IB_DEFAULT_RETRY_COUNT;
53static atomic_t rds_ib_unloading;
54
55module_param(rds_ib_mr_1m_pool_size, int, 0444);
56MODULE_PARM_DESC(rds_ib_mr_1m_pool_size, " Max number of 1M mr per HCA");
57module_param(rds_ib_mr_8k_pool_size, int, 0444);
58MODULE_PARM_DESC(rds_ib_mr_8k_pool_size, " Max number of 8K mr per HCA");
59module_param(rds_ib_retry_count, int, 0444);
60MODULE_PARM_DESC(rds_ib_retry_count, " Number of hw retries before reporting an error");
61
62/*
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.
66 */
67DECLARE_RWSEM(rds_ib_devices_lock);
68struct list_head rds_ib_devices;
69
70/* NOTE: if also grabbing ibdev lock, grab this first */
71DEFINE_SPINLOCK(ib_nodev_conns_lock);
72LIST_HEAD(ib_nodev_conns);
73
74static void rds_ib_nodev_connect(void)
75{
76 struct rds_ib_connection *ic;
77
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);
82}
83
84static void rds_ib_dev_shutdown(struct rds_ib_device *rds_ibdev)
85{
86 struct rds_ib_connection *ic;
87 unsigned long flags;
88
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);
93}
94
95/*
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.
98 */
99static void rds_ib_dev_free(struct work_struct *work)
100{
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);
104
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);
109 if (rds_ibdev->pd)
110 ib_dealloc_pd(rds_ibdev->pd);
111 dma_pool_destroy(rds_ibdev->rid_hdrs_pool);
112
113 list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) {
114 list_del(&i_ipaddr->list);
115 kfree(i_ipaddr);
116 }
117
118 kfree(rds_ibdev->vector_load);
119
120 kfree(rds_ibdev);
121}
122
123void rds_ib_dev_put(struct rds_ib_device *rds_ibdev)
124{
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);
128}
129
130static int rds_ib_add_one(struct ib_device *device)
131{
132 struct rds_ib_device *rds_ibdev;
133 int ret;
134
135 /* Only handle IB (no iWARP) devices */
136 if (device->node_type != RDMA_NODE_IB_CA)
137 return -EOPNOTSUPP;
138
139 /* Device must support FRWR */
140 if (!(device->attrs.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS))
141 return -EOPNOTSUPP;
142
143 rds_ibdev = kzalloc_node(sizeof(struct rds_ib_device), GFP_KERNEL,
144 ibdev_to_node(device));
145 if (!rds_ibdev)
146 return -ENOMEM;
147
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);
151
152 INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
153 INIT_LIST_HEAD(&rds_ibdev->conn_list);
154
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);
157
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);
165
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;
169
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;
173
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;
176
177 rds_ibdev->vector_load = kcalloc(device->num_comp_vectors,
178 sizeof(int),
179 GFP_KERNEL);
180 if (!rds_ibdev->vector_load) {
181 pr_err("RDS/IB: %s failed to allocate vector memory\n",
182 __func__);
183 ret = -ENOMEM;
184 goto put_dev;
185 }
186
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;
192 goto put_dev;
193 }
194 rds_ibdev->rid_hdrs_pool = dma_pool_create(device->name,
195 device->dma_device,
196 sizeof(struct rds_header),
197 L1_CACHE_BYTES, 0);
198 if (!rds_ibdev->rid_hdrs_pool) {
199 ret = -ENOMEM;
200 goto put_dev;
201 }
202
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;
208 goto put_dev;
209 }
210
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;
216 goto put_dev;
217 }
218
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);
222
223 pr_info("RDS/IB: %s: added\n", device->name);
224
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);
229
230 ib_set_client_data(device, &rds_ib_client, rds_ibdev);
231
232 rds_ib_nodev_connect();
233 return 0;
234
235put_dev:
236 rds_ib_dev_put(rds_ibdev);
237 return ret;
238}
239
240/*
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
245 * clients.)
246 *
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
251 * readers to finish.
252 *
253 * A new connection can get NULL from this if its arriving on a
254 * device that is in the process of being removed.
255 */
256struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device)
257{
258 struct rds_ib_device *rds_ibdev;
259
260 rcu_read_lock();
261 rds_ibdev = ib_get_client_data(device, &rds_ib_client);
262 if (rds_ibdev)
263 refcount_inc(&rds_ibdev->refcount);
264 rcu_read_unlock();
265 return rds_ibdev;
266}
267
268/*
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.
272 *
273 * This can be called at any time and can be racing with any other RDS path.
274 */
275static void rds_ib_remove_one(struct ib_device *device, void *client_data)
276{
277 struct rds_ib_device *rds_ibdev = client_data;
278
279 rds_ib_dev_shutdown(rds_ibdev);
280
281 /* stop connection attempts from getting a reference to this device. */
282 ib_set_client_data(device, &rds_ib_client, NULL);
283
284 down_write(&rds_ib_devices_lock);
285 list_del_rcu(&rds_ibdev->list);
286 up_write(&rds_ib_devices_lock);
287
288 /*
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.
292 */
293 synchronize_rcu();
294 rds_ib_dev_put(rds_ibdev);
295 rds_ib_dev_put(rds_ibdev);
296}
297
298struct ib_client rds_ib_client = {
299 .name = "rds_ib",
300 .add = rds_ib_add_one,
301 .remove = rds_ib_remove_one
302};
303
304static int rds_ib_conn_info_visitor(struct rds_connection *conn,
305 void *buffer)
306{
307 struct rds_info_rdma_connection *iinfo = buffer;
308 struct rds_ib_connection *ic = conn->c_transport_data;
309
310 /* We will only ever look at IB transports */
311 if (conn->c_trans != &rds_ib_transport)
312 return 0;
313 if (conn->c_isv6)
314 return 0;
315
316 iinfo->src_addr = conn->c_laddr.s6_addr32[3];
317 iinfo->dst_addr = conn->c_faddr.s6_addr32[3];
318 if (ic) {
319 iinfo->tos = conn->c_tos;
320 iinfo->sl = ic->i_sl;
321 }
322
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;
327
328 rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo->src_gid,
329 (union ib_gid *)&iinfo->dst_gid);
330
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);
337 }
338 return 1;
339}
340
341#if IS_ENABLED(CONFIG_IPV6)
342/* IPv6 version of rds_ib_conn_info_visitor(). */
343static int rds6_ib_conn_info_visitor(struct rds_connection *conn,
344 void *buffer)
345{
346 struct rds6_info_rdma_connection *iinfo6 = buffer;
347 struct rds_ib_connection *ic = conn->c_transport_data;
348
349 /* We will only ever look at IB transports */
350 if (conn->c_trans != &rds_ib_transport)
351 return 0;
352
353 iinfo6->src_addr = conn->c_laddr;
354 iinfo6->dst_addr = conn->c_faddr;
355 if (ic) {
356 iinfo6->tos = conn->c_tos;
357 iinfo6->sl = ic->i_sl;
358 }
359
360 memset(&iinfo6->src_gid, 0, sizeof(iinfo6->src_gid));
361 memset(&iinfo6->dst_gid, 0, sizeof(iinfo6->dst_gid));
362
363 if (rds_conn_state(conn) == RDS_CONN_UP) {
364 struct rds_ib_device *rds_ibdev;
365
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);
374 }
375 return 1;
376}
377#endif
378
379static void rds_ib_ic_info(struct socket *sock, unsigned int len,
380 struct rds_info_iterator *iter,
381 struct rds_info_lengths *lens)
382{
383 u64 buffer[(sizeof(struct rds_info_rdma_connection) + 7) / 8];
384
385 rds_for_each_conn_info(sock, len, iter, lens,
386 rds_ib_conn_info_visitor,
387 buffer,
388 sizeof(struct rds_info_rdma_connection));
389}
390
391#if IS_ENABLED(CONFIG_IPV6)
392/* IPv6 version of rds_ib_ic_info(). */
393static void rds6_ib_ic_info(struct socket *sock, unsigned int len,
394 struct rds_info_iterator *iter,
395 struct rds_info_lengths *lens)
396{
397 u64 buffer[(sizeof(struct rds6_info_rdma_connection) + 7) / 8];
398
399 rds_for_each_conn_info(sock, len, iter, lens,
400 rds6_ib_conn_info_visitor,
401 buffer,
402 sizeof(struct rds6_info_rdma_connection));
403}
404#endif
405
406/*
407 * Early RDS/IB was built to only bind to an address if there is an IPoIB
408 * device with that address set.
409 *
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".
415 */
416static int rds_ib_laddr_check(struct net *net, const struct in6_addr *addr,
417 __u32 scope_id)
418{
419 int ret;
420 struct rdma_cm_id *cm_id;
421#if IS_ENABLED(CONFIG_IPV6)
422 struct sockaddr_in6 sin6;
423#endif
424 struct sockaddr_in sin;
425 struct sockaddr *sa;
426 bool isv4;
427
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.
431 */
432 cm_id = rdma_create_id(&init_net, rds_rdma_cm_event_handler,
433 NULL, RDMA_PS_TCP, IB_QPT_RC);
434 if (IS_ERR(cm_id))
435 return PTR_ERR(cm_id);
436
437 if (isv4) {
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;
442 } else {
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;
449
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
453 * system.
454 */
455 if (ipv6_addr_type(addr) & IPV6_ADDR_LINKLOCAL) {
456 struct net_device *dev;
457
458 if (scope_id == 0) {
459 ret = -EADDRNOTAVAIL;
460 goto out;
461 }
462
463 /* Use init_net for now as RDS is not network
464 * name space aware.
465 */
466 dev = dev_get_by_index(&init_net, scope_id);
467 if (!dev) {
468 ret = -EADDRNOTAVAIL;
469 goto out;
470 }
471 if (!ipv6_chk_addr(&init_net, addr, dev, 1)) {
472 dev_put(dev);
473 ret = -EADDRNOTAVAIL;
474 goto out;
475 }
476 dev_put(dev);
477 }
478#else
479 ret = -EADDRNOTAVAIL;
480 goto out;
481#endif
482 }
483
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
487 check node_type. */
488 if (ret || !cm_id->device ||
489 cm_id->device->node_type != RDMA_NODE_IB_CA)
490 ret = -EADDRNOTAVAIL;
491
492 rdsdebug("addr %pI6c%%%u ret %d node type %d\n",
493 addr, scope_id, ret,
494 cm_id->device ? cm_id->device->node_type : -1);
495
496out:
497 rdma_destroy_id(cm_id);
498
499 return ret;
500}
501
502static void rds_ib_unregister_client(void)
503{
504 ib_unregister_client(&rds_ib_client);
505 /* wait for rds_ib_dev_free() to complete */
506 flush_workqueue(rds_wq);
507}
508
509static void rds_ib_set_unloading(void)
510{
511 atomic_set(&rds_ib_unloading, 1);
512}
513
514static bool rds_ib_is_unloading(struct rds_connection *conn)
515{
516 struct rds_conn_path *cp = &conn->c_path[0];
517
518 return (test_bit(RDS_DESTROY_PENDING, &cp->cp_flags) ||
519 atomic_read(&rds_ib_unloading) != 0);
520}
521
522void rds_ib_exit(void)
523{
524 rds_ib_set_unloading();
525 synchronize_rcu();
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);
529#endif
530 rds_ib_unregister_client();
531 rds_ib_destroy_nodev_conns();
532 rds_ib_sysctl_exit();
533 rds_ib_recv_exit();
534 rds_trans_unregister(&rds_ib_transport);
535 rds_ib_mr_exit();
536}
537
538static u8 rds_ib_get_tos_map(u8 tos)
539{
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.
543 */
544 return tos;
545}
546
547struct rds_transport rds_ib_transport = {
548 .laddr_check = rds_ib_laddr_check,
549 .xmit_path_complete = rds_ib_xmit_path_complete,
550 .xmit = rds_ib_xmit,
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,
564 .exit = rds_ib_exit,
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
574};
575
576int rds_ib_init(void)
577{
578 int ret;
579
580 INIT_LIST_HEAD(&rds_ib_devices);
581
582 ret = rds_ib_mr_init();
583 if (ret)
584 goto out;
585
586 ret = ib_register_client(&rds_ib_client);
587 if (ret)
588 goto out_mr_exit;
589
590 ret = rds_ib_sysctl_init();
591 if (ret)
592 goto out_ibreg;
593
594 ret = rds_ib_recv_init();
595 if (ret)
596 goto out_sysctl;
597
598 rds_trans_register(&rds_ib_transport);
599
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);
603#endif
604
605 goto out;
606
607out_sysctl:
608 rds_ib_sysctl_exit();
609out_ibreg:
610 rds_ib_unregister_client();
611out_mr_exit:
612 rds_ib_mr_exit();
613out:
614 return ret;
615}
616
617MODULE_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.h"
44#include "ib.h"
45
46static unsigned int fmr_pool_size = RDS_FMR_POOL_SIZE;
47unsigned int fmr_message_size = RDS_FMR_SIZE + 1; /* +1 allows for unaligned MRs */
48unsigned int rds_ib_retry_count = RDS_IB_DEFAULT_RETRY_COUNT;
49
50module_param(fmr_pool_size, int, 0444);
51MODULE_PARM_DESC(fmr_pool_size, " Max number of fmr per HCA");
52module_param(fmr_message_size, int, 0444);
53MODULE_PARM_DESC(fmr_message_size, " Max size of a RDMA transfer");
54module_param(rds_ib_retry_count, int, 0444);
55MODULE_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 */
62DECLARE_RWSEM(rds_ib_devices_lock);
63struct list_head rds_ib_devices;
64
65/* NOTE: if also grabbing ibdev lock, grab this first */
66DEFINE_SPINLOCK(ib_nodev_conns_lock);
67LIST_HEAD(ib_nodev_conns);
68
69static 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
79static 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 */
94static 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->mr)
103 ib_dereg_mr(rds_ibdev->mr);
104 if (rds_ibdev->pd)
105 ib_dealloc_pd(rds_ibdev->pd);
106
107 list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) {
108 list_del(&i_ipaddr->list);
109 kfree(i_ipaddr);
110 }
111
112 kfree(rds_ibdev);
113}
114
115void rds_ib_dev_put(struct rds_ib_device *rds_ibdev)
116{
117 BUG_ON(atomic_read(&rds_ibdev->refcount) <= 0);
118 if (atomic_dec_and_test(&rds_ibdev->refcount))
119 queue_work(rds_wq, &rds_ibdev->free_work);
120}
121
122static void rds_ib_add_one(struct ib_device *device)
123{
124 struct rds_ib_device *rds_ibdev;
125 struct ib_device_attr *dev_attr;
126
127 /* Only handle IB (no iWARP) devices */
128 if (device->node_type != RDMA_NODE_IB_CA)
129 return;
130
131 dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
132 if (!dev_attr)
133 return;
134
135 if (ib_query_device(device, dev_attr)) {
136 rdsdebug("Query device failed for %s\n", device->name);
137 goto free_attr;
138 }
139
140 rds_ibdev = kzalloc_node(sizeof(struct rds_ib_device), GFP_KERNEL,
141 ibdev_to_node(device));
142 if (!rds_ibdev)
143 goto free_attr;
144
145 spin_lock_init(&rds_ibdev->spinlock);
146 atomic_set(&rds_ibdev->refcount, 1);
147 INIT_WORK(&rds_ibdev->free_work, rds_ib_dev_free);
148
149 rds_ibdev->max_wrs = dev_attr->max_qp_wr;
150 rds_ibdev->max_sge = min(dev_attr->max_sge, RDS_IB_MAX_SGE);
151
152 rds_ibdev->fmr_max_remaps = dev_attr->max_map_per_fmr?: 32;
153 rds_ibdev->max_fmrs = dev_attr->max_fmr ?
154 min_t(unsigned int, dev_attr->max_fmr, fmr_pool_size) :
155 fmr_pool_size;
156
157 rds_ibdev->max_initiator_depth = dev_attr->max_qp_init_rd_atom;
158 rds_ibdev->max_responder_resources = dev_attr->max_qp_rd_atom;
159
160 rds_ibdev->dev = device;
161 rds_ibdev->pd = ib_alloc_pd(device);
162 if (IS_ERR(rds_ibdev->pd)) {
163 rds_ibdev->pd = NULL;
164 goto put_dev;
165 }
166
167 rds_ibdev->mr = ib_get_dma_mr(rds_ibdev->pd, IB_ACCESS_LOCAL_WRITE);
168 if (IS_ERR(rds_ibdev->mr)) {
169 rds_ibdev->mr = NULL;
170 goto put_dev;
171 }
172
173 rds_ibdev->mr_pool = rds_ib_create_mr_pool(rds_ibdev);
174 if (IS_ERR(rds_ibdev->mr_pool)) {
175 rds_ibdev->mr_pool = NULL;
176 goto put_dev;
177 }
178
179 INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
180 INIT_LIST_HEAD(&rds_ibdev->conn_list);
181
182 down_write(&rds_ib_devices_lock);
183 list_add_tail_rcu(&rds_ibdev->list, &rds_ib_devices);
184 up_write(&rds_ib_devices_lock);
185 atomic_inc(&rds_ibdev->refcount);
186
187 ib_set_client_data(device, &rds_ib_client, rds_ibdev);
188 atomic_inc(&rds_ibdev->refcount);
189
190 rds_ib_nodev_connect();
191
192put_dev:
193 rds_ib_dev_put(rds_ibdev);
194free_attr:
195 kfree(dev_attr);
196}
197
198/*
199 * New connections use this to find the device to associate with the
200 * connection. It's not in the fast path so we're not concerned about the
201 * performance of the IB call. (As of this writing, it uses an interrupt
202 * blocking spinlock to serialize walking a per-device list of all registered
203 * clients.)
204 *
205 * RCU is used to handle incoming connections racing with device teardown.
206 * Rather than use a lock to serialize removal from the client_data and
207 * getting a new reference, we use an RCU grace period. The destruction
208 * path removes the device from client_data and then waits for all RCU
209 * readers to finish.
210 *
211 * A new connection can get NULL from this if its arriving on a
212 * device that is in the process of being removed.
213 */
214struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device)
215{
216 struct rds_ib_device *rds_ibdev;
217
218 rcu_read_lock();
219 rds_ibdev = ib_get_client_data(device, &rds_ib_client);
220 if (rds_ibdev)
221 atomic_inc(&rds_ibdev->refcount);
222 rcu_read_unlock();
223 return rds_ibdev;
224}
225
226/*
227 * The IB stack is letting us know that a device is going away. This can
228 * happen if the underlying HCA driver is removed or if PCI hotplug is removing
229 * the pci function, for example.
230 *
231 * This can be called at any time and can be racing with any other RDS path.
232 */
233static void rds_ib_remove_one(struct ib_device *device)
234{
235 struct rds_ib_device *rds_ibdev;
236
237 rds_ibdev = ib_get_client_data(device, &rds_ib_client);
238 if (!rds_ibdev)
239 return;
240
241 rds_ib_dev_shutdown(rds_ibdev);
242
243 /* stop connection attempts from getting a reference to this device. */
244 ib_set_client_data(device, &rds_ib_client, NULL);
245
246 down_write(&rds_ib_devices_lock);
247 list_del_rcu(&rds_ibdev->list);
248 up_write(&rds_ib_devices_lock);
249
250 /*
251 * This synchronize rcu is waiting for readers of both the ib
252 * client data and the devices list to finish before we drop
253 * both of those references.
254 */
255 synchronize_rcu();
256 rds_ib_dev_put(rds_ibdev);
257 rds_ib_dev_put(rds_ibdev);
258}
259
260struct ib_client rds_ib_client = {
261 .name = "rds_ib",
262 .add = rds_ib_add_one,
263 .remove = rds_ib_remove_one
264};
265
266static int rds_ib_conn_info_visitor(struct rds_connection *conn,
267 void *buffer)
268{
269 struct rds_info_rdma_connection *iinfo = buffer;
270 struct rds_ib_connection *ic;
271
272 /* We will only ever look at IB transports */
273 if (conn->c_trans != &rds_ib_transport)
274 return 0;
275
276 iinfo->src_addr = conn->c_laddr;
277 iinfo->dst_addr = conn->c_faddr;
278
279 memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid));
280 memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid));
281 if (rds_conn_state(conn) == RDS_CONN_UP) {
282 struct rds_ib_device *rds_ibdev;
283 struct rdma_dev_addr *dev_addr;
284
285 ic = conn->c_transport_data;
286 dev_addr = &ic->i_cm_id->route.addr.dev_addr;
287
288 rdma_addr_get_sgid(dev_addr, (union ib_gid *) &iinfo->src_gid);
289 rdma_addr_get_dgid(dev_addr, (union ib_gid *) &iinfo->dst_gid);
290
291 rds_ibdev = ic->rds_ibdev;
292 iinfo->max_send_wr = ic->i_send_ring.w_nr;
293 iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
294 iinfo->max_send_sge = rds_ibdev->max_sge;
295 rds_ib_get_mr_info(rds_ibdev, iinfo);
296 }
297 return 1;
298}
299
300static void rds_ib_ic_info(struct socket *sock, unsigned int len,
301 struct rds_info_iterator *iter,
302 struct rds_info_lengths *lens)
303{
304 rds_for_each_conn_info(sock, len, iter, lens,
305 rds_ib_conn_info_visitor,
306 sizeof(struct rds_info_rdma_connection));
307}
308
309
310/*
311 * Early RDS/IB was built to only bind to an address if there is an IPoIB
312 * device with that address set.
313 *
314 * If it were me, I'd advocate for something more flexible. Sending and
315 * receiving should be device-agnostic. Transports would try and maintain
316 * connections between peers who have messages queued. Userspace would be
317 * allowed to influence which paths have priority. We could call userspace
318 * asserting this policy "routing".
319 */
320static int rds_ib_laddr_check(__be32 addr)
321{
322 int ret;
323 struct rdma_cm_id *cm_id;
324 struct sockaddr_in sin;
325
326 /* Create a CMA ID and try to bind it. This catches both
327 * IB and iWARP capable NICs.
328 */
329 cm_id = rdma_create_id(NULL, NULL, RDMA_PS_TCP, IB_QPT_RC);
330 if (IS_ERR(cm_id))
331 return PTR_ERR(cm_id);
332
333 memset(&sin, 0, sizeof(sin));
334 sin.sin_family = AF_INET;
335 sin.sin_addr.s_addr = addr;
336
337 /* rdma_bind_addr will only succeed for IB & iWARP devices */
338 ret = rdma_bind_addr(cm_id, (struct sockaddr *)&sin);
339 /* due to this, we will claim to support iWARP devices unless we
340 check node_type. */
341 if (ret || cm_id->device->node_type != RDMA_NODE_IB_CA)
342 ret = -EADDRNOTAVAIL;
343
344 rdsdebug("addr %pI4 ret %d node type %d\n",
345 &addr, ret,
346 cm_id->device ? cm_id->device->node_type : -1);
347
348 rdma_destroy_id(cm_id);
349
350 return ret;
351}
352
353static void rds_ib_unregister_client(void)
354{
355 ib_unregister_client(&rds_ib_client);
356 /* wait for rds_ib_dev_free() to complete */
357 flush_workqueue(rds_wq);
358}
359
360void rds_ib_exit(void)
361{
362 rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
363 rds_ib_unregister_client();
364 rds_ib_destroy_nodev_conns();
365 rds_ib_sysctl_exit();
366 rds_ib_recv_exit();
367 rds_trans_unregister(&rds_ib_transport);
368}
369
370struct rds_transport rds_ib_transport = {
371 .laddr_check = rds_ib_laddr_check,
372 .xmit_complete = rds_ib_xmit_complete,
373 .xmit = rds_ib_xmit,
374 .xmit_rdma = rds_ib_xmit_rdma,
375 .xmit_atomic = rds_ib_xmit_atomic,
376 .recv = rds_ib_recv,
377 .conn_alloc = rds_ib_conn_alloc,
378 .conn_free = rds_ib_conn_free,
379 .conn_connect = rds_ib_conn_connect,
380 .conn_shutdown = rds_ib_conn_shutdown,
381 .inc_copy_to_user = rds_ib_inc_copy_to_user,
382 .inc_free = rds_ib_inc_free,
383 .cm_initiate_connect = rds_ib_cm_initiate_connect,
384 .cm_handle_connect = rds_ib_cm_handle_connect,
385 .cm_connect_complete = rds_ib_cm_connect_complete,
386 .stats_info_copy = rds_ib_stats_info_copy,
387 .exit = rds_ib_exit,
388 .get_mr = rds_ib_get_mr,
389 .sync_mr = rds_ib_sync_mr,
390 .free_mr = rds_ib_free_mr,
391 .flush_mrs = rds_ib_flush_mrs,
392 .t_owner = THIS_MODULE,
393 .t_name = "infiniband",
394 .t_type = RDS_TRANS_IB
395};
396
397int rds_ib_init(void)
398{
399 int ret;
400
401 INIT_LIST_HEAD(&rds_ib_devices);
402
403 ret = ib_register_client(&rds_ib_client);
404 if (ret)
405 goto out;
406
407 ret = rds_ib_sysctl_init();
408 if (ret)
409 goto out_ibreg;
410
411 ret = rds_ib_recv_init();
412 if (ret)
413 goto out_sysctl;
414
415 ret = rds_trans_register(&rds_ib_transport);
416 if (ret)
417 goto out_recv;
418
419 rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
420
421 goto out;
422
423out_recv:
424 rds_ib_recv_exit();
425out_sysctl:
426 rds_ib_sysctl_exit();
427out_ibreg:
428 rds_ib_unregister_client();
429out:
430 return ret;
431}
432
433MODULE_LICENSE("GPL");
434