d2007b9316160e3c6bf9df526af316c4772c86dc
[linux-2.6.git] / net / rds / ib.c
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 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 struct list_head rds_ib_devices;
57
58 /* NOTE: if also grabbing ibdev lock, grab this first */
59 DEFINE_SPINLOCK(ib_nodev_conns_lock);
60 LIST_HEAD(ib_nodev_conns);
61
62 void rds_ib_nodev_connect(void)
63 {
64         struct rds_ib_connection *ic;
65
66         spin_lock(&ib_nodev_conns_lock);
67         list_for_each_entry(ic, &ib_nodev_conns, ib_node)
68                 rds_conn_connect_if_down(ic->conn);
69         spin_unlock(&ib_nodev_conns_lock);
70 }
71
72 void rds_ib_dev_shutdown(struct rds_ib_device *rds_ibdev)
73 {
74         struct rds_ib_connection *ic;
75         unsigned long flags;
76
77         spin_lock_irqsave(&rds_ibdev->spinlock, flags);
78         list_for_each_entry(ic, &rds_ibdev->conn_list, ib_node)
79                 rds_conn_drop(ic->conn);
80         spin_unlock_irqrestore(&rds_ibdev->spinlock, flags);
81 }
82
83 /*
84  * rds_ib_destroy_mr_pool() blocks on a few things and mrs drop references
85  * from interrupt context so we push freing off into a work struct in krdsd.
86  */
87 static void rds_ib_dev_free(struct work_struct *work)
88 {
89         struct rds_ib_ipaddr *i_ipaddr, *i_next;
90         struct rds_ib_device *rds_ibdev = container_of(work,
91                                         struct rds_ib_device, free_work);
92
93         if (rds_ibdev->mr_pool)
94                 rds_ib_destroy_mr_pool(rds_ibdev->mr_pool);
95         if (rds_ibdev->mr)
96                 ib_dereg_mr(rds_ibdev->mr);
97         if (rds_ibdev->pd)
98                 ib_dealloc_pd(rds_ibdev->pd);
99
100         list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) {
101                 list_del(&i_ipaddr->list);
102                 kfree(i_ipaddr);
103         }
104
105         kfree(rds_ibdev);
106 }
107
108 void rds_ib_dev_put(struct rds_ib_device *rds_ibdev)
109 {
110         BUG_ON(atomic_read(&rds_ibdev->refcount) <= 0);
111         if (atomic_dec_and_test(&rds_ibdev->refcount))
112                 queue_work(rds_wq, &rds_ibdev->free_work);
113 }
114
115 void rds_ib_add_one(struct ib_device *device)
116 {
117         struct rds_ib_device *rds_ibdev;
118         struct ib_device_attr *dev_attr;
119
120         /* Only handle IB (no iWARP) devices */
121         if (device->node_type != RDMA_NODE_IB_CA)
122                 return;
123
124         dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
125         if (!dev_attr)
126                 return;
127
128         if (ib_query_device(device, dev_attr)) {
129                 rdsdebug("Query device failed for %s\n", device->name);
130                 goto free_attr;
131         }
132
133         rds_ibdev = kzalloc_node(sizeof(struct rds_ib_device), GFP_KERNEL,
134                                  ibdev_to_node(device));
135         if (!rds_ibdev)
136                 goto free_attr;
137
138         spin_lock_init(&rds_ibdev->spinlock);
139         atomic_set(&rds_ibdev->refcount, 1);
140         INIT_WORK(&rds_ibdev->free_work, rds_ib_dev_free);
141
142         rds_ibdev->max_wrs = dev_attr->max_qp_wr;
143         rds_ibdev->max_sge = min(dev_attr->max_sge, RDS_IB_MAX_SGE);
144
145         rds_ibdev->fmr_max_remaps = dev_attr->max_map_per_fmr?: 32;
146         rds_ibdev->max_fmrs = dev_attr->max_fmr ?
147                         min_t(unsigned int, dev_attr->max_fmr, fmr_pool_size) :
148                         fmr_pool_size;
149
150         rds_ibdev->max_initiator_depth = dev_attr->max_qp_init_rd_atom;
151         rds_ibdev->max_responder_resources = dev_attr->max_qp_rd_atom;
152
153         rds_ibdev->dev = device;
154         rds_ibdev->pd = ib_alloc_pd(device);
155         if (IS_ERR(rds_ibdev->pd)) {
156                 rds_ibdev->pd = NULL;
157                 goto put_dev;
158         }
159
160         rds_ibdev->mr = ib_get_dma_mr(rds_ibdev->pd, IB_ACCESS_LOCAL_WRITE);
161         if (IS_ERR(rds_ibdev->mr)) {
162                 rds_ibdev->mr = NULL;
163                 goto put_dev;
164         }
165
166         rds_ibdev->mr_pool = rds_ib_create_mr_pool(rds_ibdev);
167         if (IS_ERR(rds_ibdev->mr_pool)) {
168                 rds_ibdev->mr_pool = NULL;
169                 goto put_dev;
170         }
171
172         INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
173         INIT_LIST_HEAD(&rds_ibdev->conn_list);
174         list_add_tail(&rds_ibdev->list, &rds_ib_devices);
175         atomic_inc(&rds_ibdev->refcount);
176
177         ib_set_client_data(device, &rds_ib_client, rds_ibdev);
178         atomic_inc(&rds_ibdev->refcount);
179
180         rds_ib_nodev_connect();
181
182 put_dev:
183         rds_ib_dev_put(rds_ibdev);
184 free_attr:
185         kfree(dev_attr);
186 }
187
188 /*
189  * New connections use this to find the device to associate with the
190  * connection.  It's not in the fast path so we're not concerned about the
191  * performance of the IB call.  (As of this writing, it uses an interrupt
192  * blocking spinlock to serialize walking a per-device list of all registered
193  * clients.)
194  *
195  * RCU is used to handle incoming connections racing with device teardown.
196  * Rather than use a lock to serialize removal from the client_data and
197  * getting a new reference, we use an RCU grace period.  The destruction
198  * path removes the device from client_data and then waits for all RCU
199  * readers to finish.
200  *
201  * A new connection can get NULL from this if its arriving on a
202  * device that is in the process of being removed.
203  */
204 struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device)
205 {
206         struct rds_ib_device *rds_ibdev;
207
208         rcu_read_lock();
209         rds_ibdev = ib_get_client_data(device, &rds_ib_client);
210         if (rds_ibdev)
211                 atomic_inc(&rds_ibdev->refcount);
212         rcu_read_unlock();
213         return rds_ibdev;
214 }
215
216 /*
217  * The IB stack is letting us know that a device is going away.  This can
218  * happen if the underlying HCA driver is removed or if PCI hotplug is removing
219  * the pci function, for example.
220  *
221  * This can be called at any time and can be racing with any other RDS path.
222  */
223 void rds_ib_remove_one(struct ib_device *device)
224 {
225         struct rds_ib_device *rds_ibdev;
226
227         rds_ibdev = ib_get_client_data(device, &rds_ib_client);
228         if (!rds_ibdev)
229                 return;
230
231         rds_ib_dev_shutdown(rds_ibdev);
232
233         /*
234          * prevent future connection attempts from getting a reference to this
235          * device and wait for currently racing connection attempts to finish
236          * getting their reference
237          */
238         ib_set_client_data(device, &rds_ib_client, NULL);
239         synchronize_rcu();
240         rds_ib_dev_put(rds_ibdev);
241
242         list_del(&rds_ibdev->list);
243         rds_ib_dev_put(rds_ibdev);
244 }
245
246 struct ib_client rds_ib_client = {
247         .name   = "rds_ib",
248         .add    = rds_ib_add_one,
249         .remove = rds_ib_remove_one
250 };
251
252 static int rds_ib_conn_info_visitor(struct rds_connection *conn,
253                                     void *buffer)
254 {
255         struct rds_info_rdma_connection *iinfo = buffer;
256         struct rds_ib_connection *ic;
257
258         /* We will only ever look at IB transports */
259         if (conn->c_trans != &rds_ib_transport)
260                 return 0;
261
262         iinfo->src_addr = conn->c_laddr;
263         iinfo->dst_addr = conn->c_faddr;
264
265         memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid));
266         memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid));
267         if (rds_conn_state(conn) == RDS_CONN_UP) {
268                 struct rds_ib_device *rds_ibdev;
269                 struct rdma_dev_addr *dev_addr;
270
271                 ic = conn->c_transport_data;
272                 dev_addr = &ic->i_cm_id->route.addr.dev_addr;
273
274                 rdma_addr_get_sgid(dev_addr, (union ib_gid *) &iinfo->src_gid);
275                 rdma_addr_get_dgid(dev_addr, (union ib_gid *) &iinfo->dst_gid);
276
277                 rds_ibdev = ic->rds_ibdev;
278                 iinfo->max_send_wr = ic->i_send_ring.w_nr;
279                 iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
280                 iinfo->max_send_sge = rds_ibdev->max_sge;
281                 rds_ib_get_mr_info(rds_ibdev, iinfo);
282         }
283         return 1;
284 }
285
286 static void rds_ib_ic_info(struct socket *sock, unsigned int len,
287                            struct rds_info_iterator *iter,
288                            struct rds_info_lengths *lens)
289 {
290         rds_for_each_conn_info(sock, len, iter, lens,
291                                 rds_ib_conn_info_visitor,
292                                 sizeof(struct rds_info_rdma_connection));
293 }
294
295
296 /*
297  * Early RDS/IB was built to only bind to an address if there is an IPoIB
298  * device with that address set.
299  *
300  * If it were me, I'd advocate for something more flexible.  Sending and
301  * receiving should be device-agnostic.  Transports would try and maintain
302  * connections between peers who have messages queued.  Userspace would be
303  * allowed to influence which paths have priority.  We could call userspace
304  * asserting this policy "routing".
305  */
306 static int rds_ib_laddr_check(__be32 addr)
307 {
308         int ret;
309         struct rdma_cm_id *cm_id;
310         struct sockaddr_in sin;
311
312         /* Create a CMA ID and try to bind it. This catches both
313          * IB and iWARP capable NICs.
314          */
315         cm_id = rdma_create_id(NULL, NULL, RDMA_PS_TCP);
316         if (IS_ERR(cm_id))
317                 return PTR_ERR(cm_id);
318
319         memset(&sin, 0, sizeof(sin));
320         sin.sin_family = AF_INET;
321         sin.sin_addr.s_addr = addr;
322
323         /* rdma_bind_addr will only succeed for IB & iWARP devices */
324         ret = rdma_bind_addr(cm_id, (struct sockaddr *)&sin);
325         /* due to this, we will claim to support iWARP devices unless we
326            check node_type. */
327         if (ret || cm_id->device->node_type != RDMA_NODE_IB_CA)
328                 ret = -EADDRNOTAVAIL;
329
330         rdsdebug("addr %pI4 ret %d node type %d\n",
331                 &addr, ret,
332                 cm_id->device ? cm_id->device->node_type : -1);
333
334         rdma_destroy_id(cm_id);
335
336         return ret;
337 }
338
339 static void rds_ib_unregister_client(void)
340 {
341         ib_unregister_client(&rds_ib_client);
342         /* wait for rds_ib_dev_free() to complete */
343         flush_workqueue(rds_wq);
344 }
345
346 void rds_ib_exit(void)
347 {
348         rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
349         rds_ib_unregister_client();
350         rds_ib_destroy_nodev_conns();
351         rds_ib_sysctl_exit();
352         rds_ib_recv_exit();
353         rds_trans_unregister(&rds_ib_transport);
354         rds_ib_fmr_exit();
355 }
356
357 struct rds_transport rds_ib_transport = {
358         .laddr_check            = rds_ib_laddr_check,
359         .xmit_complete          = rds_ib_xmit_complete,
360         .xmit                   = rds_ib_xmit,
361         .xmit_rdma              = rds_ib_xmit_rdma,
362         .xmit_atomic            = rds_ib_xmit_atomic,
363         .recv                   = rds_ib_recv,
364         .conn_alloc             = rds_ib_conn_alloc,
365         .conn_free              = rds_ib_conn_free,
366         .conn_connect           = rds_ib_conn_connect,
367         .conn_shutdown          = rds_ib_conn_shutdown,
368         .inc_copy_to_user       = rds_ib_inc_copy_to_user,
369         .inc_free               = rds_ib_inc_free,
370         .cm_initiate_connect    = rds_ib_cm_initiate_connect,
371         .cm_handle_connect      = rds_ib_cm_handle_connect,
372         .cm_connect_complete    = rds_ib_cm_connect_complete,
373         .stats_info_copy        = rds_ib_stats_info_copy,
374         .exit                   = rds_ib_exit,
375         .get_mr                 = rds_ib_get_mr,
376         .sync_mr                = rds_ib_sync_mr,
377         .free_mr                = rds_ib_free_mr,
378         .flush_mrs              = rds_ib_flush_mrs,
379         .t_owner                = THIS_MODULE,
380         .t_name                 = "infiniband",
381         .t_type                 = RDS_TRANS_IB
382 };
383
384 int __init rds_ib_init(void)
385 {
386         int ret;
387
388         INIT_LIST_HEAD(&rds_ib_devices);
389
390         ret = rds_ib_fmr_init();
391         if (ret)
392                 goto out;
393
394         ret = ib_register_client(&rds_ib_client);
395         if (ret)
396                 goto out_fmr_exit;
397
398         ret = rds_ib_sysctl_init();
399         if (ret)
400                 goto out_ibreg;
401
402         ret = rds_ib_recv_init();
403         if (ret)
404                 goto out_sysctl;
405
406         ret = rds_trans_register(&rds_ib_transport);
407         if (ret)
408                 goto out_recv;
409
410         rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
411
412         goto out;
413
414 out_recv:
415         rds_ib_recv_exit();
416 out_sysctl:
417         rds_ib_sysctl_exit();
418 out_ibreg:
419         rds_ib_unregister_client();
420 out_fmr_exit:
421         rds_ib_fmr_exit();
422 out:
423         return ret;
424 }
425
426 MODULE_LICENSE("GPL");
427