RDS: Set retry_count to 2 and make modifiable via modparam
[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
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 unsigned int rds_ib_retry_count = RDS_IB_DEFAULT_RETRY_COUNT;
47
48 module_param(fmr_pool_size, int, 0444);
49 MODULE_PARM_DESC(fmr_pool_size, " Max number of fmr per HCA");
50 module_param(fmr_message_size, int, 0444);
51 MODULE_PARM_DESC(fmr_message_size, " Max size of a RDMA transfer");
52 module_param(rds_ib_retry_count, int, 0444);
53 MODULE_PARM_DESC(rds_ib_retry_count, " Number of hw retries before reporting an error");
54
55 struct list_head rds_ib_devices;
56
57 /* NOTE: if also grabbing ibdev lock, grab this first */
58 DEFINE_SPINLOCK(ib_nodev_conns_lock);
59 LIST_HEAD(ib_nodev_conns);
60
61 void rds_ib_add_one(struct ib_device *device)
62 {
63         struct rds_ib_device *rds_ibdev;
64         struct ib_device_attr *dev_attr;
65
66         /* Only handle IB (no iWARP) devices */
67         if (device->node_type != RDMA_NODE_IB_CA)
68                 return;
69
70         dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
71         if (!dev_attr)
72                 return;
73
74         if (ib_query_device(device, dev_attr)) {
75                 rdsdebug("Query device failed for %s\n", device->name);
76                 goto free_attr;
77         }
78
79         rds_ibdev = kmalloc(sizeof *rds_ibdev, GFP_KERNEL);
80         if (!rds_ibdev)
81                 goto free_attr;
82
83         spin_lock_init(&rds_ibdev->spinlock);
84
85         rds_ibdev->max_wrs = dev_attr->max_qp_wr;
86         rds_ibdev->max_sge = min(dev_attr->max_sge, RDS_IB_MAX_SGE);
87
88         rds_ibdev->fmr_page_shift = max(9, ffs(dev_attr->page_size_cap) - 1);
89         rds_ibdev->fmr_page_size  = 1 << rds_ibdev->fmr_page_shift;
90         rds_ibdev->fmr_page_mask  = ~((u64) rds_ibdev->fmr_page_size - 1);
91         rds_ibdev->fmr_max_remaps = dev_attr->max_map_per_fmr?: 32;
92         rds_ibdev->max_fmrs = dev_attr->max_fmr ?
93                         min_t(unsigned int, dev_attr->max_fmr, fmr_pool_size) :
94                         fmr_pool_size;
95
96         rds_ibdev->dev = device;
97         rds_ibdev->pd = ib_alloc_pd(device);
98         if (IS_ERR(rds_ibdev->pd))
99                 goto free_dev;
100
101         rds_ibdev->mr = ib_get_dma_mr(rds_ibdev->pd,
102                                       IB_ACCESS_LOCAL_WRITE);
103         if (IS_ERR(rds_ibdev->mr))
104                 goto err_pd;
105
106         rds_ibdev->mr_pool = rds_ib_create_mr_pool(rds_ibdev);
107         if (IS_ERR(rds_ibdev->mr_pool)) {
108                 rds_ibdev->mr_pool = NULL;
109                 goto err_mr;
110         }
111
112         INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
113         INIT_LIST_HEAD(&rds_ibdev->conn_list);
114         list_add_tail(&rds_ibdev->list, &rds_ib_devices);
115
116         ib_set_client_data(device, &rds_ib_client, rds_ibdev);
117
118         goto free_attr;
119
120 err_mr:
121         ib_dereg_mr(rds_ibdev->mr);
122 err_pd:
123         ib_dealloc_pd(rds_ibdev->pd);
124 free_dev:
125         kfree(rds_ibdev);
126 free_attr:
127         kfree(dev_attr);
128 }
129
130 void rds_ib_remove_one(struct ib_device *device)
131 {
132         struct rds_ib_device *rds_ibdev;
133         struct rds_ib_ipaddr *i_ipaddr, *i_next;
134
135         rds_ibdev = ib_get_client_data(device, &rds_ib_client);
136         if (!rds_ibdev)
137                 return;
138
139         list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) {
140                 list_del(&i_ipaddr->list);
141                 kfree(i_ipaddr);
142         }
143
144         rds_ib_destroy_conns(rds_ibdev);
145
146         if (rds_ibdev->mr_pool)
147                 rds_ib_destroy_mr_pool(rds_ibdev->mr_pool);
148
149         ib_dereg_mr(rds_ibdev->mr);
150
151         while (ib_dealloc_pd(rds_ibdev->pd)) {
152                 rdsdebug("Failed to dealloc pd %p\n", rds_ibdev->pd);
153                 msleep(1);
154         }
155
156         list_del(&rds_ibdev->list);
157         kfree(rds_ibdev);
158 }
159
160 struct ib_client rds_ib_client = {
161         .name   = "rds_ib",
162         .add    = rds_ib_add_one,
163         .remove = rds_ib_remove_one
164 };
165
166 static int rds_ib_conn_info_visitor(struct rds_connection *conn,
167                                     void *buffer)
168 {
169         struct rds_info_rdma_connection *iinfo = buffer;
170         struct rds_ib_connection *ic;
171
172         /* We will only ever look at IB transports */
173         if (conn->c_trans != &rds_ib_transport)
174                 return 0;
175
176         iinfo->src_addr = conn->c_laddr;
177         iinfo->dst_addr = conn->c_faddr;
178
179         memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid));
180         memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid));
181         if (rds_conn_state(conn) == RDS_CONN_UP) {
182                 struct rds_ib_device *rds_ibdev;
183                 struct rdma_dev_addr *dev_addr;
184
185                 ic = conn->c_transport_data;
186                 dev_addr = &ic->i_cm_id->route.addr.dev_addr;
187
188                 ib_addr_get_sgid(dev_addr, (union ib_gid *) &iinfo->src_gid);
189                 ib_addr_get_dgid(dev_addr, (union ib_gid *) &iinfo->dst_gid);
190
191                 rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
192                 iinfo->max_send_wr = ic->i_send_ring.w_nr;
193                 iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
194                 iinfo->max_send_sge = rds_ibdev->max_sge;
195                 rds_ib_get_mr_info(rds_ibdev, iinfo);
196         }
197         return 1;
198 }
199
200 static void rds_ib_ic_info(struct socket *sock, unsigned int len,
201                            struct rds_info_iterator *iter,
202                            struct rds_info_lengths *lens)
203 {
204         rds_for_each_conn_info(sock, len, iter, lens,
205                                 rds_ib_conn_info_visitor,
206                                 sizeof(struct rds_info_rdma_connection));
207 }
208
209
210 /*
211  * Early RDS/IB was built to only bind to an address if there is an IPoIB
212  * device with that address set.
213  *
214  * If it were me, I'd advocate for something more flexible.  Sending and
215  * receiving should be device-agnostic.  Transports would try and maintain
216  * connections between peers who have messages queued.  Userspace would be
217  * allowed to influence which paths have priority.  We could call userspace
218  * asserting this policy "routing".
219  */
220 static int rds_ib_laddr_check(__be32 addr)
221 {
222         int ret;
223         struct rdma_cm_id *cm_id;
224         struct sockaddr_in sin;
225
226         /* Create a CMA ID and try to bind it. This catches both
227          * IB and iWARP capable NICs.
228          */
229         cm_id = rdma_create_id(NULL, NULL, RDMA_PS_TCP);
230         if (IS_ERR(cm_id))
231                 return PTR_ERR(cm_id);
232
233         memset(&sin, 0, sizeof(sin));
234         sin.sin_family = AF_INET;
235         sin.sin_addr.s_addr = addr;
236
237         /* rdma_bind_addr will only succeed for IB & iWARP devices */
238         ret = rdma_bind_addr(cm_id, (struct sockaddr *)&sin);
239         /* due to this, we will claim to support iWARP devices unless we
240            check node_type. */
241         if (ret || cm_id->device->node_type != RDMA_NODE_IB_CA)
242                 ret = -EADDRNOTAVAIL;
243
244         rdsdebug("addr %pI4 ret %d node type %d\n",
245                 &addr, ret,
246                 cm_id->device ? cm_id->device->node_type : -1);
247
248         rdma_destroy_id(cm_id);
249
250         return ret;
251 }
252
253 void rds_ib_exit(void)
254 {
255         rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
256         rds_ib_destroy_nodev_conns();
257         ib_unregister_client(&rds_ib_client);
258         rds_ib_sysctl_exit();
259         rds_ib_recv_exit();
260         rds_trans_unregister(&rds_ib_transport);
261 }
262
263 struct rds_transport rds_ib_transport = {
264         .laddr_check            = rds_ib_laddr_check,
265         .xmit_complete          = rds_ib_xmit_complete,
266         .xmit                   = rds_ib_xmit,
267         .xmit_cong_map          = NULL,
268         .xmit_rdma              = rds_ib_xmit_rdma,
269         .recv                   = rds_ib_recv,
270         .conn_alloc             = rds_ib_conn_alloc,
271         .conn_free              = rds_ib_conn_free,
272         .conn_connect           = rds_ib_conn_connect,
273         .conn_shutdown          = rds_ib_conn_shutdown,
274         .inc_copy_to_user       = rds_ib_inc_copy_to_user,
275         .inc_purge              = rds_ib_inc_purge,
276         .inc_free               = rds_ib_inc_free,
277         .cm_initiate_connect    = rds_ib_cm_initiate_connect,
278         .cm_handle_connect      = rds_ib_cm_handle_connect,
279         .cm_connect_complete    = rds_ib_cm_connect_complete,
280         .stats_info_copy        = rds_ib_stats_info_copy,
281         .exit                   = rds_ib_exit,
282         .get_mr                 = rds_ib_get_mr,
283         .sync_mr                = rds_ib_sync_mr,
284         .free_mr                = rds_ib_free_mr,
285         .flush_mrs              = rds_ib_flush_mrs,
286         .t_owner                = THIS_MODULE,
287         .t_name                 = "infiniband",
288 };
289
290 int __init rds_ib_init(void)
291 {
292         int ret;
293
294         INIT_LIST_HEAD(&rds_ib_devices);
295
296         ret = ib_register_client(&rds_ib_client);
297         if (ret)
298                 goto out;
299
300         ret = rds_ib_sysctl_init();
301         if (ret)
302                 goto out_ibreg;
303
304         ret = rds_ib_recv_init();
305         if (ret)
306                 goto out_sysctl;
307
308         ret = rds_trans_register(&rds_ib_transport);
309         if (ret)
310                 goto out_recv;
311
312         rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
313
314         goto out;
315
316 out_recv:
317         rds_ib_recv_exit();
318 out_sysctl:
319         rds_ib_sysctl_exit();
320 out_ibreg:
321         ib_unregister_client(&rds_ib_client);
322 out:
323         return ret;
324 }
325
326 MODULE_LICENSE("GPL");
327