xref: /openbmc/linux/net/rds/ib.c (revision 4800cd83)
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 
42 #include "rds.h"
43 #include "ib.h"
44 
45 static unsigned int fmr_pool_size = RDS_FMR_POOL_SIZE;
46 unsigned int fmr_message_size = RDS_FMR_SIZE + 1; /* +1 allows for unaligned MRs */
47 unsigned int rds_ib_retry_count = RDS_IB_DEFAULT_RETRY_COUNT;
48 
49 module_param(fmr_pool_size, int, 0444);
50 MODULE_PARM_DESC(fmr_pool_size, " Max number of fmr per HCA");
51 module_param(fmr_message_size, int, 0444);
52 MODULE_PARM_DESC(fmr_message_size, " Max size of a RDMA transfer");
53 module_param(rds_ib_retry_count, int, 0444);
54 MODULE_PARM_DESC(rds_ib_retry_count, " Number of hw retries before reporting an error");
55 
56 /*
57  * we have a clumsy combination of RCU and a rwsem protecting this list
58  * because it is used both in the get_mr fast path and while blocking in
59  * the FMR flushing path.
60  */
61 DECLARE_RWSEM(rds_ib_devices_lock);
62 struct list_head rds_ib_devices;
63 
64 /* NOTE: if also grabbing ibdev lock, grab this first */
65 DEFINE_SPINLOCK(ib_nodev_conns_lock);
66 LIST_HEAD(ib_nodev_conns);
67 
68 static void rds_ib_nodev_connect(void)
69 {
70 	struct rds_ib_connection *ic;
71 
72 	spin_lock(&ib_nodev_conns_lock);
73 	list_for_each_entry(ic, &ib_nodev_conns, ib_node)
74 		rds_conn_connect_if_down(ic->conn);
75 	spin_unlock(&ib_nodev_conns_lock);
76 }
77 
78 static void rds_ib_dev_shutdown(struct rds_ib_device *rds_ibdev)
79 {
80 	struct rds_ib_connection *ic;
81 	unsigned long flags;
82 
83 	spin_lock_irqsave(&rds_ibdev->spinlock, flags);
84 	list_for_each_entry(ic, &rds_ibdev->conn_list, ib_node)
85 		rds_conn_drop(ic->conn);
86 	spin_unlock_irqrestore(&rds_ibdev->spinlock, flags);
87 }
88 
89 /*
90  * rds_ib_destroy_mr_pool() blocks on a few things and mrs drop references
91  * from interrupt context so we push freing off into a work struct in krdsd.
92  */
93 static void rds_ib_dev_free(struct work_struct *work)
94 {
95 	struct rds_ib_ipaddr *i_ipaddr, *i_next;
96 	struct rds_ib_device *rds_ibdev = container_of(work,
97 					struct rds_ib_device, free_work);
98 
99 	if (rds_ibdev->mr_pool)
100 		rds_ib_destroy_mr_pool(rds_ibdev->mr_pool);
101 	if (rds_ibdev->mr)
102 		ib_dereg_mr(rds_ibdev->mr);
103 	if (rds_ibdev->pd)
104 		ib_dealloc_pd(rds_ibdev->pd);
105 
106 	list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) {
107 		list_del(&i_ipaddr->list);
108 		kfree(i_ipaddr);
109 	}
110 
111 	kfree(rds_ibdev);
112 }
113 
114 void rds_ib_dev_put(struct rds_ib_device *rds_ibdev)
115 {
116 	BUG_ON(atomic_read(&rds_ibdev->refcount) <= 0);
117 	if (atomic_dec_and_test(&rds_ibdev->refcount))
118 		queue_work(rds_wq, &rds_ibdev->free_work);
119 }
120 
121 static void rds_ib_add_one(struct ib_device *device)
122 {
123 	struct rds_ib_device *rds_ibdev;
124 	struct ib_device_attr *dev_attr;
125 
126 	/* Only handle IB (no iWARP) devices */
127 	if (device->node_type != RDMA_NODE_IB_CA)
128 		return;
129 
130 	dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
131 	if (!dev_attr)
132 		return;
133 
134 	if (ib_query_device(device, dev_attr)) {
135 		rdsdebug("Query device failed for %s\n", device->name);
136 		goto free_attr;
137 	}
138 
139 	rds_ibdev = kzalloc_node(sizeof(struct rds_ib_device), GFP_KERNEL,
140 				 ibdev_to_node(device));
141 	if (!rds_ibdev)
142 		goto free_attr;
143 
144 	spin_lock_init(&rds_ibdev->spinlock);
145 	atomic_set(&rds_ibdev->refcount, 1);
146 	INIT_WORK(&rds_ibdev->free_work, rds_ib_dev_free);
147 
148 	rds_ibdev->max_wrs = dev_attr->max_qp_wr;
149 	rds_ibdev->max_sge = min(dev_attr->max_sge, RDS_IB_MAX_SGE);
150 
151 	rds_ibdev->fmr_max_remaps = dev_attr->max_map_per_fmr?: 32;
152 	rds_ibdev->max_fmrs = dev_attr->max_fmr ?
153 			min_t(unsigned int, dev_attr->max_fmr, fmr_pool_size) :
154 			fmr_pool_size;
155 
156 	rds_ibdev->max_initiator_depth = dev_attr->max_qp_init_rd_atom;
157 	rds_ibdev->max_responder_resources = dev_attr->max_qp_rd_atom;
158 
159 	rds_ibdev->dev = device;
160 	rds_ibdev->pd = ib_alloc_pd(device);
161 	if (IS_ERR(rds_ibdev->pd)) {
162 		rds_ibdev->pd = NULL;
163 		goto put_dev;
164 	}
165 
166 	rds_ibdev->mr = ib_get_dma_mr(rds_ibdev->pd, IB_ACCESS_LOCAL_WRITE);
167 	if (IS_ERR(rds_ibdev->mr)) {
168 		rds_ibdev->mr = NULL;
169 		goto put_dev;
170 	}
171 
172 	rds_ibdev->mr_pool = rds_ib_create_mr_pool(rds_ibdev);
173 	if (IS_ERR(rds_ibdev->mr_pool)) {
174 		rds_ibdev->mr_pool = NULL;
175 		goto put_dev;
176 	}
177 
178 	INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
179 	INIT_LIST_HEAD(&rds_ibdev->conn_list);
180 
181 	down_write(&rds_ib_devices_lock);
182 	list_add_tail_rcu(&rds_ibdev->list, &rds_ib_devices);
183 	up_write(&rds_ib_devices_lock);
184 	atomic_inc(&rds_ibdev->refcount);
185 
186 	ib_set_client_data(device, &rds_ib_client, rds_ibdev);
187 	atomic_inc(&rds_ibdev->refcount);
188 
189 	rds_ib_nodev_connect();
190 
191 put_dev:
192 	rds_ib_dev_put(rds_ibdev);
193 free_attr:
194 	kfree(dev_attr);
195 }
196 
197 /*
198  * New connections use this to find the device to associate with the
199  * connection.  It's not in the fast path so we're not concerned about the
200  * performance of the IB call.  (As of this writing, it uses an interrupt
201  * blocking spinlock to serialize walking a per-device list of all registered
202  * clients.)
203  *
204  * RCU is used to handle incoming connections racing with device teardown.
205  * Rather than use a lock to serialize removal from the client_data and
206  * getting a new reference, we use an RCU grace period.  The destruction
207  * path removes the device from client_data and then waits for all RCU
208  * readers to finish.
209  *
210  * A new connection can get NULL from this if its arriving on a
211  * device that is in the process of being removed.
212  */
213 struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device)
214 {
215 	struct rds_ib_device *rds_ibdev;
216 
217 	rcu_read_lock();
218 	rds_ibdev = ib_get_client_data(device, &rds_ib_client);
219 	if (rds_ibdev)
220 		atomic_inc(&rds_ibdev->refcount);
221 	rcu_read_unlock();
222 	return rds_ibdev;
223 }
224 
225 /*
226  * The IB stack is letting us know that a device is going away.  This can
227  * happen if the underlying HCA driver is removed or if PCI hotplug is removing
228  * the pci function, for example.
229  *
230  * This can be called at any time and can be racing with any other RDS path.
231  */
232 static void rds_ib_remove_one(struct ib_device *device)
233 {
234 	struct rds_ib_device *rds_ibdev;
235 
236 	rds_ibdev = ib_get_client_data(device, &rds_ib_client);
237 	if (!rds_ibdev)
238 		return;
239 
240 	rds_ib_dev_shutdown(rds_ibdev);
241 
242 	/* stop connection attempts from getting a reference to this device. */
243 	ib_set_client_data(device, &rds_ib_client, NULL);
244 
245 	down_write(&rds_ib_devices_lock);
246 	list_del_rcu(&rds_ibdev->list);
247 	up_write(&rds_ib_devices_lock);
248 
249 	/*
250 	 * This synchronize rcu is waiting for readers of both the ib
251 	 * client data and the devices list to finish before we drop
252 	 * both of those references.
253 	 */
254 	synchronize_rcu();
255 	rds_ib_dev_put(rds_ibdev);
256 	rds_ib_dev_put(rds_ibdev);
257 }
258 
259 struct ib_client rds_ib_client = {
260 	.name   = "rds_ib",
261 	.add    = rds_ib_add_one,
262 	.remove = rds_ib_remove_one
263 };
264 
265 static int rds_ib_conn_info_visitor(struct rds_connection *conn,
266 				    void *buffer)
267 {
268 	struct rds_info_rdma_connection *iinfo = buffer;
269 	struct rds_ib_connection *ic;
270 
271 	/* We will only ever look at IB transports */
272 	if (conn->c_trans != &rds_ib_transport)
273 		return 0;
274 
275 	iinfo->src_addr = conn->c_laddr;
276 	iinfo->dst_addr = conn->c_faddr;
277 
278 	memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid));
279 	memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid));
280 	if (rds_conn_state(conn) == RDS_CONN_UP) {
281 		struct rds_ib_device *rds_ibdev;
282 		struct rdma_dev_addr *dev_addr;
283 
284 		ic = conn->c_transport_data;
285 		dev_addr = &ic->i_cm_id->route.addr.dev_addr;
286 
287 		rdma_addr_get_sgid(dev_addr, (union ib_gid *) &iinfo->src_gid);
288 		rdma_addr_get_dgid(dev_addr, (union ib_gid *) &iinfo->dst_gid);
289 
290 		rds_ibdev = ic->rds_ibdev;
291 		iinfo->max_send_wr = ic->i_send_ring.w_nr;
292 		iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
293 		iinfo->max_send_sge = rds_ibdev->max_sge;
294 		rds_ib_get_mr_info(rds_ibdev, iinfo);
295 	}
296 	return 1;
297 }
298 
299 static void rds_ib_ic_info(struct socket *sock, unsigned int len,
300 			   struct rds_info_iterator *iter,
301 			   struct rds_info_lengths *lens)
302 {
303 	rds_for_each_conn_info(sock, len, iter, lens,
304 				rds_ib_conn_info_visitor,
305 				sizeof(struct rds_info_rdma_connection));
306 }
307 
308 
309 /*
310  * Early RDS/IB was built to only bind to an address if there is an IPoIB
311  * device with that address set.
312  *
313  * If it were me, I'd advocate for something more flexible.  Sending and
314  * receiving should be device-agnostic.  Transports would try and maintain
315  * connections between peers who have messages queued.  Userspace would be
316  * allowed to influence which paths have priority.  We could call userspace
317  * asserting this policy "routing".
318  */
319 static int rds_ib_laddr_check(__be32 addr)
320 {
321 	int ret;
322 	struct rdma_cm_id *cm_id;
323 	struct sockaddr_in sin;
324 
325 	/* Create a CMA ID and try to bind it. This catches both
326 	 * IB and iWARP capable NICs.
327 	 */
328 	cm_id = rdma_create_id(NULL, NULL, RDMA_PS_TCP);
329 	if (IS_ERR(cm_id))
330 		return PTR_ERR(cm_id);
331 
332 	memset(&sin, 0, sizeof(sin));
333 	sin.sin_family = AF_INET;
334 	sin.sin_addr.s_addr = addr;
335 
336 	/* rdma_bind_addr will only succeed for IB & iWARP devices */
337 	ret = rdma_bind_addr(cm_id, (struct sockaddr *)&sin);
338 	/* due to this, we will claim to support iWARP devices unless we
339 	   check node_type. */
340 	if (ret || cm_id->device->node_type != RDMA_NODE_IB_CA)
341 		ret = -EADDRNOTAVAIL;
342 
343 	rdsdebug("addr %pI4 ret %d node type %d\n",
344 		&addr, ret,
345 		cm_id->device ? cm_id->device->node_type : -1);
346 
347 	rdma_destroy_id(cm_id);
348 
349 	return ret;
350 }
351 
352 static void rds_ib_unregister_client(void)
353 {
354 	ib_unregister_client(&rds_ib_client);
355 	/* wait for rds_ib_dev_free() to complete */
356 	flush_workqueue(rds_wq);
357 }
358 
359 void rds_ib_exit(void)
360 {
361 	rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
362 	rds_ib_unregister_client();
363 	rds_ib_destroy_nodev_conns();
364 	rds_ib_sysctl_exit();
365 	rds_ib_recv_exit();
366 	rds_trans_unregister(&rds_ib_transport);
367 	rds_ib_fmr_exit();
368 }
369 
370 struct 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 
397 int rds_ib_init(void)
398 {
399 	int ret;
400 
401 	INIT_LIST_HEAD(&rds_ib_devices);
402 
403 	ret = rds_ib_fmr_init();
404 	if (ret)
405 		goto out;
406 
407 	ret = ib_register_client(&rds_ib_client);
408 	if (ret)
409 		goto out_fmr_exit;
410 
411 	ret = rds_ib_sysctl_init();
412 	if (ret)
413 		goto out_ibreg;
414 
415 	ret = rds_ib_recv_init();
416 	if (ret)
417 		goto out_sysctl;
418 
419 	ret = rds_trans_register(&rds_ib_transport);
420 	if (ret)
421 		goto out_recv;
422 
423 	rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
424 
425 	goto out;
426 
427 out_recv:
428 	rds_ib_recv_exit();
429 out_sysctl:
430 	rds_ib_sysctl_exit();
431 out_ibreg:
432 	rds_ib_unregister_client();
433 out_fmr_exit:
434 	rds_ib_fmr_exit();
435 out:
436 	return ret;
437 }
438 
439 MODULE_LICENSE("GPL");
440 
441