Merge branch 'for-linus' of git://neil.brown.name/md
[linux-2.6] / 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
41 #include "rds.h"
42 #include "ib.h"
43
44 unsigned int fmr_pool_size = RDS_FMR_POOL_SIZE;
45 unsigned int fmr_message_size = RDS_FMR_SIZE + 1; /* +1 allows for unaligned MRs */
46
47 module_param(fmr_pool_size, int, 0444);
48 MODULE_PARM_DESC(fmr_pool_size, " Max number of fmr per HCA");
49 module_param(fmr_message_size, int, 0444);
50 MODULE_PARM_DESC(fmr_message_size, " Max size of a RDMA transfer");
51
52 struct list_head rds_ib_devices;
53
54 /* NOTE: if also grabbing ibdev lock, grab this first */
55 DEFINE_SPINLOCK(ib_nodev_conns_lock);
56 LIST_HEAD(ib_nodev_conns);
57
58 void rds_ib_add_one(struct ib_device *device)
59 {
60         struct rds_ib_device *rds_ibdev;
61         struct ib_device_attr *dev_attr;
62
63         /* Only handle IB (no iWARP) devices */
64         if (device->node_type != RDMA_NODE_IB_CA)
65                 return;
66
67         dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
68         if (!dev_attr)
69                 return;
70
71         if (ib_query_device(device, dev_attr)) {
72                 rdsdebug("Query device failed for %s\n", device->name);
73                 goto free_attr;
74         }
75
76         rds_ibdev = kmalloc(sizeof *rds_ibdev, GFP_KERNEL);
77         if (!rds_ibdev)
78                 goto free_attr;
79
80         spin_lock_init(&rds_ibdev->spinlock);
81
82         rds_ibdev->max_wrs = dev_attr->max_qp_wr;
83         rds_ibdev->max_sge = min(dev_attr->max_sge, RDS_IB_MAX_SGE);
84
85         rds_ibdev->fmr_page_shift = max(9, ffs(dev_attr->page_size_cap) - 1);
86         rds_ibdev->fmr_page_size  = 1 << rds_ibdev->fmr_page_shift;
87         rds_ibdev->fmr_page_mask  = ~((u64) rds_ibdev->fmr_page_size - 1);
88         rds_ibdev->fmr_max_remaps = dev_attr->max_map_per_fmr?: 32;
89         rds_ibdev->max_fmrs = dev_attr->max_fmr ?
90                         min_t(unsigned int, dev_attr->max_fmr, fmr_pool_size) :
91                         fmr_pool_size;
92
93         rds_ibdev->dev = device;
94         rds_ibdev->pd = ib_alloc_pd(device);
95         if (IS_ERR(rds_ibdev->pd))
96                 goto free_dev;
97
98         rds_ibdev->mr = ib_get_dma_mr(rds_ibdev->pd,
99                                       IB_ACCESS_LOCAL_WRITE);
100         if (IS_ERR(rds_ibdev->mr))
101                 goto err_pd;
102
103         rds_ibdev->mr_pool = rds_ib_create_mr_pool(rds_ibdev);
104         if (IS_ERR(rds_ibdev->mr_pool)) {
105                 rds_ibdev->mr_pool = NULL;
106                 goto err_mr;
107         }
108
109         INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
110         INIT_LIST_HEAD(&rds_ibdev->conn_list);
111         list_add_tail(&rds_ibdev->list, &rds_ib_devices);
112
113         ib_set_client_data(device, &rds_ib_client, rds_ibdev);
114
115         goto free_attr;
116
117 err_mr:
118         ib_dereg_mr(rds_ibdev->mr);
119 err_pd:
120         ib_dealloc_pd(rds_ibdev->pd);
121 free_dev:
122         kfree(rds_ibdev);
123 free_attr:
124         kfree(dev_attr);
125 }
126
127 void rds_ib_remove_one(struct ib_device *device)
128 {
129         struct rds_ib_device *rds_ibdev;
130         struct rds_ib_ipaddr *i_ipaddr, *i_next;
131
132         rds_ibdev = ib_get_client_data(device, &rds_ib_client);
133         if (!rds_ibdev)
134                 return;
135
136         list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) {
137                 list_del(&i_ipaddr->list);
138                 kfree(i_ipaddr);
139         }
140
141         rds_ib_destroy_conns(rds_ibdev);
142
143         if (rds_ibdev->mr_pool)
144                 rds_ib_destroy_mr_pool(rds_ibdev->mr_pool);
145
146         ib_dereg_mr(rds_ibdev->mr);
147
148         while (ib_dealloc_pd(rds_ibdev->pd)) {
149                 rdsdebug("Failed to dealloc pd %p\n", rds_ibdev->pd);
150                 msleep(1);
151         }
152
153         list_del(&rds_ibdev->list);
154         kfree(rds_ibdev);
155 }
156
157 struct ib_client rds_ib_client = {
158         .name   = "rds_ib",
159         .add    = rds_ib_add_one,
160         .remove = rds_ib_remove_one
161 };
162
163 static int rds_ib_conn_info_visitor(struct rds_connection *conn,
164                                     void *buffer)
165 {
166         struct rds_info_rdma_connection *iinfo = buffer;
167         struct rds_ib_connection *ic;
168
169         /* We will only ever look at IB transports */
170         if (conn->c_trans != &rds_ib_transport)
171                 return 0;
172
173         iinfo->src_addr = conn->c_laddr;
174         iinfo->dst_addr = conn->c_faddr;
175
176         memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid));
177         memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid));
178         if (rds_conn_state(conn) == RDS_CONN_UP) {
179                 struct rds_ib_device *rds_ibdev;
180                 struct rdma_dev_addr *dev_addr;
181
182                 ic = conn->c_transport_data;
183                 dev_addr = &ic->i_cm_id->route.addr.dev_addr;
184
185                 ib_addr_get_sgid(dev_addr, (union ib_gid *) &iinfo->src_gid);
186                 ib_addr_get_dgid(dev_addr, (union ib_gid *) &iinfo->dst_gid);
187
188                 rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
189                 iinfo->max_send_wr = ic->i_send_ring.w_nr;
190                 iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
191                 iinfo->max_send_sge = rds_ibdev->max_sge;
192                 rds_ib_get_mr_info(rds_ibdev, iinfo);
193         }
194         return 1;
195 }
196
197 static void rds_ib_ic_info(struct socket *sock, unsigned int len,
198                            struct rds_info_iterator *iter,
199                            struct rds_info_lengths *lens)
200 {
201         rds_for_each_conn_info(sock, len, iter, lens,
202                                 rds_ib_conn_info_visitor,
203                                 sizeof(struct rds_info_rdma_connection));
204 }
205
206
207 /*
208  * Early RDS/IB was built to only bind to an address if there is an IPoIB
209  * device with that address set.
210  *
211  * If it were me, I'd advocate for something more flexible.  Sending and
212  * receiving should be device-agnostic.  Transports would try and maintain
213  * connections between peers who have messages queued.  Userspace would be
214  * allowed to influence which paths have priority.  We could call userspace
215  * asserting this policy "routing".
216  */
217 static int rds_ib_laddr_check(__be32 addr)
218 {
219         int ret;
220         struct rdma_cm_id *cm_id;
221         struct sockaddr_in sin;
222
223         /* Create a CMA ID and try to bind it. This catches both
224          * IB and iWARP capable NICs.
225          */
226         cm_id = rdma_create_id(NULL, NULL, RDMA_PS_TCP);
227         if (!cm_id)
228                 return -EADDRNOTAVAIL;
229
230         memset(&sin, 0, sizeof(sin));
231         sin.sin_family = AF_INET;
232         sin.sin_addr.s_addr = addr;
233
234         /* rdma_bind_addr will only succeed for IB & iWARP devices */
235         ret = rdma_bind_addr(cm_id, (struct sockaddr *)&sin);
236         /* due to this, we will claim to support iWARP devices unless we
237            check node_type. */
238         if (ret || cm_id->device->node_type != RDMA_NODE_IB_CA)
239                 ret = -EADDRNOTAVAIL;
240
241         rdsdebug("addr %pI4 ret %d node type %d\n",
242                 &addr, ret,
243                 cm_id->device ? cm_id->device->node_type : -1);
244
245         rdma_destroy_id(cm_id);
246
247         return ret;
248 }
249
250 void rds_ib_exit(void)
251 {
252         rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
253         rds_ib_destroy_nodev_conns();
254         ib_unregister_client(&rds_ib_client);
255         rds_ib_sysctl_exit();
256         rds_ib_recv_exit();
257         rds_trans_unregister(&rds_ib_transport);
258 }
259
260 struct rds_transport rds_ib_transport = {
261         .laddr_check            = rds_ib_laddr_check,
262         .xmit_complete          = rds_ib_xmit_complete,
263         .xmit                   = rds_ib_xmit,
264         .xmit_cong_map          = NULL,
265         .xmit_rdma              = rds_ib_xmit_rdma,
266         .recv                   = rds_ib_recv,
267         .conn_alloc             = rds_ib_conn_alloc,
268         .conn_free              = rds_ib_conn_free,
269         .conn_connect           = rds_ib_conn_connect,
270         .conn_shutdown          = rds_ib_conn_shutdown,
271         .inc_copy_to_user       = rds_ib_inc_copy_to_user,
272         .inc_purge              = rds_ib_inc_purge,
273         .inc_free               = rds_ib_inc_free,
274         .cm_initiate_connect    = rds_ib_cm_initiate_connect,
275         .cm_handle_connect      = rds_ib_cm_handle_connect,
276         .cm_connect_complete    = rds_ib_cm_connect_complete,
277         .stats_info_copy        = rds_ib_stats_info_copy,
278         .exit                   = rds_ib_exit,
279         .get_mr                 = rds_ib_get_mr,
280         .sync_mr                = rds_ib_sync_mr,
281         .free_mr                = rds_ib_free_mr,
282         .flush_mrs              = rds_ib_flush_mrs,
283         .t_owner                = THIS_MODULE,
284         .t_name                 = "infiniband",
285 };
286
287 int __init rds_ib_init(void)
288 {
289         int ret;
290
291         INIT_LIST_HEAD(&rds_ib_devices);
292
293         ret = ib_register_client(&rds_ib_client);
294         if (ret)
295                 goto out;
296
297         ret = rds_ib_sysctl_init();
298         if (ret)
299                 goto out_ibreg;
300
301         ret = rds_ib_recv_init();
302         if (ret)
303                 goto out_sysctl;
304
305         ret = rds_trans_register(&rds_ib_transport);
306         if (ret)
307                 goto out_recv;
308
309         rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
310
311         goto out;
312
313 out_recv:
314         rds_ib_recv_exit();
315 out_sysctl:
316         rds_ib_sysctl_exit();
317 out_ibreg:
318         ib_unregister_client(&rds_ib_client);
319 out:
320         return ret;
321 }
322
323 MODULE_LICENSE("GPL");
324