xref: /openbmc/linux/net/rds/ib.c (revision 93d90ad7)
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 
46 static unsigned int fmr_pool_size = RDS_FMR_POOL_SIZE;
47 unsigned int fmr_message_size = RDS_FMR_SIZE + 1; /* +1 allows for unaligned MRs */
48 unsigned int rds_ib_retry_count = RDS_IB_DEFAULT_RETRY_COUNT;
49 
50 module_param(fmr_pool_size, int, 0444);
51 MODULE_PARM_DESC(fmr_pool_size, " Max number of fmr per HCA");
52 module_param(fmr_message_size, int, 0444);
53 MODULE_PARM_DESC(fmr_message_size, " Max size of a RDMA transfer");
54 module_param(rds_ib_retry_count, int, 0444);
55 MODULE_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  */
62 DECLARE_RWSEM(rds_ib_devices_lock);
63 struct list_head rds_ib_devices;
64 
65 /* NOTE: if also grabbing ibdev lock, grab this first */
66 DEFINE_SPINLOCK(ib_nodev_conns_lock);
67 LIST_HEAD(ib_nodev_conns);
68 
69 static 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 
79 static 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  */
94 static 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 
115 void 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 
122 static 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 
192 put_dev:
193 	rds_ib_dev_put(rds_ibdev);
194 free_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  */
214 struct 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  */
233 static 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 
260 struct ib_client rds_ib_client = {
261 	.name   = "rds_ib",
262 	.add    = rds_ib_add_one,
263 	.remove = rds_ib_remove_one
264 };
265 
266 static 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 
300 static 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  */
320 static 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 ||
342 	    cm_id->device->node_type != RDMA_NODE_IB_CA)
343 		ret = -EADDRNOTAVAIL;
344 
345 	rdsdebug("addr %pI4 ret %d node type %d\n",
346 		&addr, ret,
347 		cm_id->device ? cm_id->device->node_type : -1);
348 
349 	rdma_destroy_id(cm_id);
350 
351 	return ret;
352 }
353 
354 static void rds_ib_unregister_client(void)
355 {
356 	ib_unregister_client(&rds_ib_client);
357 	/* wait for rds_ib_dev_free() to complete */
358 	flush_workqueue(rds_wq);
359 }
360 
361 void rds_ib_exit(void)
362 {
363 	rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
364 	rds_ib_unregister_client();
365 	rds_ib_destroy_nodev_conns();
366 	rds_ib_sysctl_exit();
367 	rds_ib_recv_exit();
368 	rds_trans_unregister(&rds_ib_transport);
369 }
370 
371 struct rds_transport rds_ib_transport = {
372 	.laddr_check		= rds_ib_laddr_check,
373 	.xmit_complete		= rds_ib_xmit_complete,
374 	.xmit			= rds_ib_xmit,
375 	.xmit_rdma		= rds_ib_xmit_rdma,
376 	.xmit_atomic		= rds_ib_xmit_atomic,
377 	.recv			= rds_ib_recv,
378 	.conn_alloc		= rds_ib_conn_alloc,
379 	.conn_free		= rds_ib_conn_free,
380 	.conn_connect		= rds_ib_conn_connect,
381 	.conn_shutdown		= rds_ib_conn_shutdown,
382 	.inc_copy_to_user	= rds_ib_inc_copy_to_user,
383 	.inc_free		= rds_ib_inc_free,
384 	.cm_initiate_connect	= rds_ib_cm_initiate_connect,
385 	.cm_handle_connect	= rds_ib_cm_handle_connect,
386 	.cm_connect_complete	= rds_ib_cm_connect_complete,
387 	.stats_info_copy	= rds_ib_stats_info_copy,
388 	.exit			= rds_ib_exit,
389 	.get_mr			= rds_ib_get_mr,
390 	.sync_mr		= rds_ib_sync_mr,
391 	.free_mr		= rds_ib_free_mr,
392 	.flush_mrs		= rds_ib_flush_mrs,
393 	.t_owner		= THIS_MODULE,
394 	.t_name			= "infiniband",
395 	.t_type			= RDS_TRANS_IB
396 };
397 
398 int rds_ib_init(void)
399 {
400 	int ret;
401 
402 	INIT_LIST_HEAD(&rds_ib_devices);
403 
404 	ret = ib_register_client(&rds_ib_client);
405 	if (ret)
406 		goto out;
407 
408 	ret = rds_ib_sysctl_init();
409 	if (ret)
410 		goto out_ibreg;
411 
412 	ret = rds_ib_recv_init();
413 	if (ret)
414 		goto out_sysctl;
415 
416 	ret = rds_trans_register(&rds_ib_transport);
417 	if (ret)
418 		goto out_recv;
419 
420 	rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
421 
422 	goto out;
423 
424 out_recv:
425 	rds_ib_recv_exit();
426 out_sysctl:
427 	rds_ib_sysctl_exit();
428 out_ibreg:
429 	rds_ib_unregister_client();
430 out:
431 	return ret;
432 }
433 
434 MODULE_LICENSE("GPL");
435 
436