net: implements ip_route_input_noref()
[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_add_one(struct ib_device *device)
63 {
64         struct rds_ib_device *rds_ibdev;
65         struct ib_device_attr *dev_attr;
66
67         /* Only handle IB (no iWARP) devices */
68         if (device->node_type != RDMA_NODE_IB_CA)
69                 return;
70
71         dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
72         if (!dev_attr)
73                 return;
74
75         if (ib_query_device(device, dev_attr)) {
76                 rdsdebug("Query device failed for %s\n", device->name);
77                 goto free_attr;
78         }
79
80         rds_ibdev = kmalloc(sizeof *rds_ibdev, GFP_KERNEL);
81         if (!rds_ibdev)
82                 goto free_attr;
83
84         spin_lock_init(&rds_ibdev->spinlock);
85
86         rds_ibdev->max_wrs = dev_attr->max_qp_wr;
87         rds_ibdev->max_sge = min(dev_attr->max_sge, RDS_IB_MAX_SGE);
88
89         rds_ibdev->fmr_max_remaps = dev_attr->max_map_per_fmr?: 32;
90         rds_ibdev->max_fmrs = dev_attr->max_fmr ?
91                         min_t(unsigned int, dev_attr->max_fmr, fmr_pool_size) :
92                         fmr_pool_size;
93
94         rds_ibdev->dev = device;
95         rds_ibdev->pd = ib_alloc_pd(device);
96         if (IS_ERR(rds_ibdev->pd))
97                 goto free_dev;
98
99         rds_ibdev->mr = ib_get_dma_mr(rds_ibdev->pd,
100                                       IB_ACCESS_LOCAL_WRITE);
101         if (IS_ERR(rds_ibdev->mr))
102                 goto err_pd;
103
104         rds_ibdev->mr_pool = rds_ib_create_mr_pool(rds_ibdev);
105         if (IS_ERR(rds_ibdev->mr_pool)) {
106                 rds_ibdev->mr_pool = NULL;
107                 goto err_mr;
108         }
109
110         INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
111         INIT_LIST_HEAD(&rds_ibdev->conn_list);
112         list_add_tail(&rds_ibdev->list, &rds_ib_devices);
113
114         ib_set_client_data(device, &rds_ib_client, rds_ibdev);
115
116         goto free_attr;
117
118 err_mr:
119         ib_dereg_mr(rds_ibdev->mr);
120 err_pd:
121         ib_dealloc_pd(rds_ibdev->pd);
122 free_dev:
123         kfree(rds_ibdev);
124 free_attr:
125         kfree(dev_attr);
126 }
127
128 void rds_ib_remove_one(struct ib_device *device)
129 {
130         struct rds_ib_device *rds_ibdev;
131         struct rds_ib_ipaddr *i_ipaddr, *i_next;
132
133         rds_ibdev = ib_get_client_data(device, &rds_ib_client);
134         if (!rds_ibdev)
135                 return;
136
137         list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) {
138                 list_del(&i_ipaddr->list);
139                 kfree(i_ipaddr);
140         }
141
142         rds_ib_destroy_conns(rds_ibdev);
143
144         if (rds_ibdev->mr_pool)
145                 rds_ib_destroy_mr_pool(rds_ibdev->mr_pool);
146
147         ib_dereg_mr(rds_ibdev->mr);
148
149         while (ib_dealloc_pd(rds_ibdev->pd)) {
150                 rdsdebug("Failed to dealloc pd %p\n", rds_ibdev->pd);
151                 msleep(1);
152         }
153
154         list_del(&rds_ibdev->list);
155         kfree(rds_ibdev);
156 }
157
158 struct ib_client rds_ib_client = {
159         .name   = "rds_ib",
160         .add    = rds_ib_add_one,
161         .remove = rds_ib_remove_one
162 };
163
164 static int rds_ib_conn_info_visitor(struct rds_connection *conn,
165                                     void *buffer)
166 {
167         struct rds_info_rdma_connection *iinfo = buffer;
168         struct rds_ib_connection *ic;
169
170         /* We will only ever look at IB transports */
171         if (conn->c_trans != &rds_ib_transport)
172                 return 0;
173
174         iinfo->src_addr = conn->c_laddr;
175         iinfo->dst_addr = conn->c_faddr;
176
177         memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid));
178         memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid));
179         if (rds_conn_state(conn) == RDS_CONN_UP) {
180                 struct rds_ib_device *rds_ibdev;
181                 struct rdma_dev_addr *dev_addr;
182
183                 ic = conn->c_transport_data;
184                 dev_addr = &ic->i_cm_id->route.addr.dev_addr;
185
186                 rdma_addr_get_sgid(dev_addr, (union ib_gid *) &iinfo->src_gid);
187                 rdma_addr_get_dgid(dev_addr, (union ib_gid *) &iinfo->dst_gid);
188
189                 rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
190                 iinfo->max_send_wr = ic->i_send_ring.w_nr;
191                 iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
192                 iinfo->max_send_sge = rds_ibdev->max_sge;
193                 rds_ib_get_mr_info(rds_ibdev, iinfo);
194         }
195         return 1;
196 }
197
198 static void rds_ib_ic_info(struct socket *sock, unsigned int len,
199                            struct rds_info_iterator *iter,
200                            struct rds_info_lengths *lens)
201 {
202         rds_for_each_conn_info(sock, len, iter, lens,
203                                 rds_ib_conn_info_visitor,
204                                 sizeof(struct rds_info_rdma_connection));
205 }
206
207
208 /*
209  * Early RDS/IB was built to only bind to an address if there is an IPoIB
210  * device with that address set.
211  *
212  * If it were me, I'd advocate for something more flexible.  Sending and
213  * receiving should be device-agnostic.  Transports would try and maintain
214  * connections between peers who have messages queued.  Userspace would be
215  * allowed to influence which paths have priority.  We could call userspace
216  * asserting this policy "routing".
217  */
218 static int rds_ib_laddr_check(__be32 addr)
219 {
220         int ret;
221         struct rdma_cm_id *cm_id;
222         struct sockaddr_in sin;
223
224         /* Create a CMA ID and try to bind it. This catches both
225          * IB and iWARP capable NICs.
226          */
227         cm_id = rdma_create_id(NULL, NULL, RDMA_PS_TCP);
228         if (IS_ERR(cm_id))
229                 return PTR_ERR(cm_id);
230
231         memset(&sin, 0, sizeof(sin));
232         sin.sin_family = AF_INET;
233         sin.sin_addr.s_addr = addr;
234
235         /* rdma_bind_addr will only succeed for IB & iWARP devices */
236         ret = rdma_bind_addr(cm_id, (struct sockaddr *)&sin);
237         /* due to this, we will claim to support iWARP devices unless we
238            check node_type. */
239         if (ret || cm_id->device->node_type != RDMA_NODE_IB_CA)
240                 ret = -EADDRNOTAVAIL;
241
242         rdsdebug("addr %pI4 ret %d node type %d\n",
243                 &addr, ret,
244                 cm_id->device ? cm_id->device->node_type : -1);
245
246         rdma_destroy_id(cm_id);
247
248         return ret;
249 }
250
251 void rds_ib_exit(void)
252 {
253         rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
254         rds_ib_destroy_nodev_conns();
255         ib_unregister_client(&rds_ib_client);
256         rds_ib_sysctl_exit();
257         rds_ib_recv_exit();
258         rds_trans_unregister(&rds_ib_transport);
259 }
260
261 struct rds_transport rds_ib_transport = {
262         .laddr_check            = rds_ib_laddr_check,
263         .xmit_complete          = rds_ib_xmit_complete,
264         .xmit                   = rds_ib_xmit,
265         .xmit_cong_map          = NULL,
266         .xmit_rdma              = rds_ib_xmit_rdma,
267         .recv                   = rds_ib_recv,
268         .conn_alloc             = rds_ib_conn_alloc,
269         .conn_free              = rds_ib_conn_free,
270         .conn_connect           = rds_ib_conn_connect,
271         .conn_shutdown          = rds_ib_conn_shutdown,
272         .inc_copy_to_user       = rds_ib_inc_copy_to_user,
273         .inc_purge              = rds_ib_inc_purge,
274         .inc_free               = rds_ib_inc_free,
275         .cm_initiate_connect    = rds_ib_cm_initiate_connect,
276         .cm_handle_connect      = rds_ib_cm_handle_connect,
277         .cm_connect_complete    = rds_ib_cm_connect_complete,
278         .stats_info_copy        = rds_ib_stats_info_copy,
279         .exit                   = rds_ib_exit,
280         .get_mr                 = rds_ib_get_mr,
281         .sync_mr                = rds_ib_sync_mr,
282         .free_mr                = rds_ib_free_mr,
283         .flush_mrs              = rds_ib_flush_mrs,
284         .t_owner                = THIS_MODULE,
285         .t_name                 = "infiniband",
286         .t_type                 = RDS_TRANS_IB
287 };
288
289 int __init rds_ib_init(void)
290 {
291         int ret;
292
293         INIT_LIST_HEAD(&rds_ib_devices);
294
295         ret = ib_register_client(&rds_ib_client);
296         if (ret)
297                 goto out;
298
299         ret = rds_ib_sysctl_init();
300         if (ret)
301                 goto out_ibreg;
302
303         ret = rds_ib_recv_init();
304         if (ret)
305                 goto out_sysctl;
306
307         ret = rds_trans_register(&rds_ib_transport);
308         if (ret)
309                 goto out_recv;
310
311         rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
312
313         goto out;
314
315 out_recv:
316         rds_ib_recv_exit();
317 out_sysctl:
318         rds_ib_sysctl_exit();
319 out_ibreg:
320         ib_unregister_client(&rds_ib_client);
321 out:
322         return ret;
323 }
324
325 MODULE_LICENSE("GPL");
326