blob: a792d8a3872aca5b835d7bca66cf12a6859cc7eb [file] [log] [blame]
Andy Groverec162272009-02-24 15:30:30 +00001/*
Ka-Cheong Poon9b17f582019-10-02 21:11:08 -07002 * Copyright (c) 2006, 2019 Oracle and/or its affiliates. All rights reserved.
Andy Groverec162272009-02-24 15:30:30 +00003 *
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 */
YueHaibingd131c5b2019-10-06 15:08:32 +080033#include <linux/dmapool.h>
Andy Groverec162272009-02-24 15:30:30 +000034#include <linux/kernel.h>
35#include <linux/in.h>
36#include <linux/if.h>
37#include <linux/netdevice.h>
38#include <linux/inetdevice.h>
39#include <linux/if_arp.h>
40#include <linux/delay.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090041#include <linux/slab.h>
Paul Gortmaker3a9a2312011-05-27 09:12:25 -040042#include <linux/module.h>
Ka-Cheong Poon1e2b44e2018-07-23 20:51:22 -070043#include <net/addrconf.h>
Andy Groverec162272009-02-24 15:30:30 +000044
Sowmini Varadhan0cb43962016-06-13 09:44:26 -070045#include "rds_single_path.h"
Andy Groverec162272009-02-24 15:30:30 +000046#include "rds.h"
47#include "ib.h"
santosh.shilimkar@oracle.comf6df6832016-03-01 15:20:46 -080048#include "ib_mr.h"
Andy Groverec162272009-02-24 15:30:30 +000049
Zhu Yanjun4f7bfb32017-02-28 01:45:40 -050050static unsigned int rds_ib_mr_1m_pool_size = RDS_MR_1M_POOL_SIZE;
51static unsigned int rds_ib_mr_8k_pool_size = RDS_MR_8K_POOL_SIZE;
Andy Grover3ba23ad2009-07-17 13:13:22 +000052unsigned int rds_ib_retry_count = RDS_IB_DEFAULT_RETRY_COUNT;
Sowmini Varadhanebeeb1a2018-02-03 04:26:51 -080053static atomic_t rds_ib_unloading;
Andy Groverec162272009-02-24 15:30:30 +000054
santosh.shilimkar@oracle.comf6df6832016-03-01 15:20:46 -080055module_param(rds_ib_mr_1m_pool_size, int, 0444);
56MODULE_PARM_DESC(rds_ib_mr_1m_pool_size, " Max number of 1M mr per HCA");
57module_param(rds_ib_mr_8k_pool_size, int, 0444);
58MODULE_PARM_DESC(rds_ib_mr_8k_pool_size, " Max number of 8K mr per HCA");
Andy Grover3ba23ad2009-07-17 13:13:22 +000059module_param(rds_ib_retry_count, int, 0444);
60MODULE_PARM_DESC(rds_ib_retry_count, " Number of hw retries before reporting an error");
Andy Groverec162272009-02-24 15:30:30 +000061
Zach Brownea819862010-07-15 12:34:33 -070062/*
63 * we have a clumsy combination of RCU and a rwsem protecting this list
64 * because it is used both in the get_mr fast path and while blocking in
65 * the FMR flushing path.
66 */
67DECLARE_RWSEM(rds_ib_devices_lock);
Andy Groverec162272009-02-24 15:30:30 +000068struct list_head rds_ib_devices;
69
Andy Grover745cbcc2009-04-01 08:20:19 +000070/* NOTE: if also grabbing ibdev lock, grab this first */
Andy Groverec162272009-02-24 15:30:30 +000071DEFINE_SPINLOCK(ib_nodev_conns_lock);
72LIST_HEAD(ib_nodev_conns);
73
stephen hemmingerff51bf82010-10-19 08:08:33 +000074static void rds_ib_nodev_connect(void)
Zach Brownfc19de32010-05-24 13:16:57 -070075{
76 struct rds_ib_connection *ic;
77
78 spin_lock(&ib_nodev_conns_lock);
79 list_for_each_entry(ic, &ib_nodev_conns, ib_node)
80 rds_conn_connect_if_down(ic->conn);
81 spin_unlock(&ib_nodev_conns_lock);
82}
83
stephen hemmingerff51bf82010-10-19 08:08:33 +000084static void rds_ib_dev_shutdown(struct rds_ib_device *rds_ibdev)
Zach Brownfc19de32010-05-24 13:16:57 -070085{
86 struct rds_ib_connection *ic;
87 unsigned long flags;
88
89 spin_lock_irqsave(&rds_ibdev->spinlock, flags);
90 list_for_each_entry(ic, &rds_ibdev->conn_list, ib_node)
Zhu Yanjunb50e0582019-06-03 08:48:19 -040091 rds_conn_path_drop(&ic->conn->c_path[0], true);
Zach Brownfc19de32010-05-24 13:16:57 -070092 spin_unlock_irqrestore(&rds_ibdev->spinlock, flags);
93}
94
Zach Brown3e0249f2010-05-18 15:48:51 -070095/*
96 * rds_ib_destroy_mr_pool() blocks on a few things and mrs drop references
97 * from interrupt context so we push freing off into a work struct in krdsd.
98 */
99static void rds_ib_dev_free(struct work_struct *work)
100{
101 struct rds_ib_ipaddr *i_ipaddr, *i_next;
102 struct rds_ib_device *rds_ibdev = container_of(work,
103 struct rds_ib_device, free_work);
104
Santosh Shilimkar06766512015-09-10 21:20:57 -0700105 if (rds_ibdev->mr_8k_pool)
106 rds_ib_destroy_mr_pool(rds_ibdev->mr_8k_pool);
107 if (rds_ibdev->mr_1m_pool)
108 rds_ib_destroy_mr_pool(rds_ibdev->mr_1m_pool);
Zach Brown3e0249f2010-05-18 15:48:51 -0700109 if (rds_ibdev->pd)
110 ib_dealloc_pd(rds_ibdev->pd);
YueHaibingce753e62019-10-15 19:47:36 +0800111 dma_pool_destroy(rds_ibdev->rid_hdrs_pool);
Zach Brown3e0249f2010-05-18 15:48:51 -0700112
113 list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) {
114 list_del(&i_ipaddr->list);
115 kfree(i_ipaddr);
116 }
117
Wu Fengguang77cc7ae2017-02-23 20:47:01 +0800118 kfree(rds_ibdev->vector_load);
Santosh Shilimkarbe2f76e2016-07-04 16:16:36 -0700119
Zach Brown3e0249f2010-05-18 15:48:51 -0700120 kfree(rds_ibdev);
121}
122
123void rds_ib_dev_put(struct rds_ib_device *rds_ibdev)
124{
Reshetova, Elena50d61ff2017-07-04 15:53:15 +0300125 BUG_ON(refcount_read(&rds_ibdev->refcount) == 0);
126 if (refcount_dec_and_test(&rds_ibdev->refcount))
Zach Brown3e0249f2010-05-18 15:48:51 -0700127 queue_work(rds_wq, &rds_ibdev->free_work);
128}
129
stephen hemmingerff51bf82010-10-19 08:08:33 +0000130static void rds_ib_add_one(struct ib_device *device)
Andy Groverec162272009-02-24 15:30:30 +0000131{
132 struct rds_ib_device *rds_ibdev;
Avinash Repaka9dff9932017-10-04 12:10:43 -0700133 bool has_fr, has_fmr;
Andy Groverec162272009-02-24 15:30:30 +0000134
135 /* Only handle IB (no iWARP) devices */
136 if (device->node_type != RDMA_NODE_IB_CA)
137 return;
138
Zach Brown3e0249f2010-05-18 15:48:51 -0700139 rds_ibdev = kzalloc_node(sizeof(struct rds_ib_device), GFP_KERNEL,
140 ibdev_to_node(device));
Andy Groverec162272009-02-24 15:30:30 +0000141 if (!rds_ibdev)
Or Gerlitz03532612015-12-18 10:59:47 +0200142 return;
Andy Groverec162272009-02-24 15:30:30 +0000143
144 spin_lock_init(&rds_ibdev->spinlock);
Reshetova, Elena50d61ff2017-07-04 15:53:15 +0300145 refcount_set(&rds_ibdev->refcount, 1);
Zach Brown3e0249f2010-05-18 15:48:51 -0700146 INIT_WORK(&rds_ibdev->free_work, rds_ib_dev_free);
Andy Groverec162272009-02-24 15:30:30 +0000147
Dotan Barakd64bf892019-10-01 10:21:02 -0700148 INIT_LIST_HEAD(&rds_ibdev->ipaddr_list);
149 INIT_LIST_HEAD(&rds_ibdev->conn_list);
150
Or Gerlitz03532612015-12-18 10:59:47 +0200151 rds_ibdev->max_wrs = device->attrs.max_qp_wr;
Steve Wise33023fb2018-06-18 08:05:26 -0700152 rds_ibdev->max_sge = min(device->attrs.max_send_sge, RDS_IB_MAX_SGE);
Andy Groverec162272009-02-24 15:30:30 +0000153
Avinash Repaka9dff9932017-10-04 12:10:43 -0700154 has_fr = (device->attrs.device_cap_flags &
155 IB_DEVICE_MEM_MGT_EXTENSIONS);
Kamal Heib3023a1e2018-12-10 21:09:48 +0200156 has_fmr = (device->ops.alloc_fmr && device->ops.dealloc_fmr &&
157 device->ops.map_phys_fmr && device->ops.unmap_fmr);
Avinash Repaka9dff9932017-10-04 12:10:43 -0700158 rds_ibdev->use_fastreg = (has_fr && !has_fmr);
Hans Westgaard Ry2eafa172020-01-15 14:43:39 +0200159 rds_ibdev->odp_capable =
160 !!(device->attrs.device_cap_flags &
161 IB_DEVICE_ON_DEMAND_PAGING) &&
162 !!(device->attrs.odp_caps.per_transport_caps.rc_odp_caps &
163 IB_ODP_SUPPORT_WRITE) &&
164 !!(device->attrs.odp_caps.per_transport_caps.rc_odp_caps &
165 IB_ODP_SUPPORT_READ);
santosh.shilimkar@oracle.com2cb2912d2016-03-01 15:20:52 -0800166
Or Gerlitz03532612015-12-18 10:59:47 +0200167 rds_ibdev->fmr_max_remaps = device->attrs.max_map_per_fmr?: 32;
santosh.shilimkar@oracle.comf6df6832016-03-01 15:20:46 -0800168 rds_ibdev->max_1m_mrs = device->attrs.max_mr ?
Or Gerlitz03532612015-12-18 10:59:47 +0200169 min_t(unsigned int, (device->attrs.max_mr / 2),
santosh.shilimkar@oracle.comf6df6832016-03-01 15:20:46 -0800170 rds_ib_mr_1m_pool_size) : rds_ib_mr_1m_pool_size;
Santosh Shilimkar06766512015-09-10 21:20:57 -0700171
santosh.shilimkar@oracle.comf6df6832016-03-01 15:20:46 -0800172 rds_ibdev->max_8k_mrs = device->attrs.max_mr ?
Or Gerlitz03532612015-12-18 10:59:47 +0200173 min_t(unsigned int, ((device->attrs.max_mr / 2) * RDS_MR_8K_SCALE),
santosh.shilimkar@oracle.comf6df6832016-03-01 15:20:46 -0800174 rds_ib_mr_8k_pool_size) : rds_ib_mr_8k_pool_size;
Andy Groverec162272009-02-24 15:30:30 +0000175
Or Gerlitz03532612015-12-18 10:59:47 +0200176 rds_ibdev->max_initiator_depth = device->attrs.max_qp_init_rd_atom;
177 rds_ibdev->max_responder_resources = device->attrs.max_qp_rd_atom;
Andy Grover40589e72010-01-12 10:50:48 -0800178
Kees Cook6396bb22018-06-12 14:03:40 -0700179 rds_ibdev->vector_load = kcalloc(device->num_comp_vectors,
180 sizeof(int),
Santosh Shilimkarbe2f76e2016-07-04 16:16:36 -0700181 GFP_KERNEL);
182 if (!rds_ibdev->vector_load) {
183 pr_err("RDS/IB: %s failed to allocate vector memory\n",
184 __func__);
185 goto put_dev;
186 }
187
Andy Groverec162272009-02-24 15:30:30 +0000188 rds_ibdev->dev = device;
Christoph Hellwiged082d32016-09-05 12:56:17 +0200189 rds_ibdev->pd = ib_alloc_pd(device, 0);
Zach Brown3e0249f2010-05-18 15:48:51 -0700190 if (IS_ERR(rds_ibdev->pd)) {
191 rds_ibdev->pd = NULL;
192 goto put_dev;
193 }
Ka-Cheong Poon9b17f582019-10-02 21:11:08 -0700194 rds_ibdev->rid_hdrs_pool = dma_pool_create(device->name,
195 device->dma_device,
196 sizeof(struct rds_header),
197 L1_CACHE_BYTES, 0);
198 if (!rds_ibdev->rid_hdrs_pool)
199 goto put_dev;
Andy Groverec162272009-02-24 15:30:30 +0000200
Santosh Shilimkar06766512015-09-10 21:20:57 -0700201 rds_ibdev->mr_1m_pool =
202 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_1M_POOL);
203 if (IS_ERR(rds_ibdev->mr_1m_pool)) {
204 rds_ibdev->mr_1m_pool = NULL;
Zach Brown3e0249f2010-05-18 15:48:51 -0700205 goto put_dev;
Andy Groverec162272009-02-24 15:30:30 +0000206 }
207
Santosh Shilimkar06766512015-09-10 21:20:57 -0700208 rds_ibdev->mr_8k_pool =
209 rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_8K_POOL);
210 if (IS_ERR(rds_ibdev->mr_8k_pool)) {
211 rds_ibdev->mr_8k_pool = NULL;
212 goto put_dev;
213 }
214
santosh.shilimkar@oracle.comf6df6832016-03-01 15:20:46 -0800215 rdsdebug("RDS/IB: max_mr = %d, max_wrs = %d, max_sge = %d, fmr_max_remaps = %d, max_1m_mrs = %d, max_8k_mrs = %d\n",
Or Gerlitz03532612015-12-18 10:59:47 +0200216 device->attrs.max_fmr, rds_ibdev->max_wrs, rds_ibdev->max_sge,
santosh.shilimkar@oracle.comf6df6832016-03-01 15:20:46 -0800217 rds_ibdev->fmr_max_remaps, rds_ibdev->max_1m_mrs,
218 rds_ibdev->max_8k_mrs);
Santosh Shilimkar06766512015-09-10 21:20:57 -0700219
santosh.shilimkar@oracle.com2cb2912d2016-03-01 15:20:52 -0800220 pr_info("RDS/IB: %s: %s supported and preferred\n",
221 device->name,
222 rds_ibdev->use_fastreg ? "FRMR" : "FMR");
223
Zach Brownea819862010-07-15 12:34:33 -0700224 down_write(&rds_ib_devices_lock);
225 list_add_tail_rcu(&rds_ibdev->list, &rds_ib_devices);
226 up_write(&rds_ib_devices_lock);
Reshetova, Elena50d61ff2017-07-04 15:53:15 +0300227 refcount_inc(&rds_ibdev->refcount);
Andy Groverec162272009-02-24 15:30:30 +0000228
229 ib_set_client_data(device, &rds_ib_client, rds_ibdev);
Reshetova, Elena50d61ff2017-07-04 15:53:15 +0300230 refcount_inc(&rds_ibdev->refcount);
Andy Groverec162272009-02-24 15:30:30 +0000231
Zach Brownfc19de32010-05-24 13:16:57 -0700232 rds_ib_nodev_connect();
233
Zach Brown3e0249f2010-05-18 15:48:51 -0700234put_dev:
235 rds_ib_dev_put(rds_ibdev);
Andy Groverec162272009-02-24 15:30:30 +0000236}
237
Zach Brown3e0249f2010-05-18 15:48:51 -0700238/*
239 * New connections use this to find the device to associate with the
240 * connection. It's not in the fast path so we're not concerned about the
241 * performance of the IB call. (As of this writing, it uses an interrupt
242 * blocking spinlock to serialize walking a per-device list of all registered
243 * clients.)
244 *
245 * RCU is used to handle incoming connections racing with device teardown.
246 * Rather than use a lock to serialize removal from the client_data and
247 * getting a new reference, we use an RCU grace period. The destruction
248 * path removes the device from client_data and then waits for all RCU
249 * readers to finish.
250 *
251 * A new connection can get NULL from this if its arriving on a
252 * device that is in the process of being removed.
253 */
254struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device)
255{
256 struct rds_ib_device *rds_ibdev;
257
258 rcu_read_lock();
259 rds_ibdev = ib_get_client_data(device, &rds_ib_client);
260 if (rds_ibdev)
Reshetova, Elena50d61ff2017-07-04 15:53:15 +0300261 refcount_inc(&rds_ibdev->refcount);
Zach Brown3e0249f2010-05-18 15:48:51 -0700262 rcu_read_unlock();
263 return rds_ibdev;
264}
265
266/*
267 * The IB stack is letting us know that a device is going away. This can
268 * happen if the underlying HCA driver is removed or if PCI hotplug is removing
269 * the pci function, for example.
270 *
271 * This can be called at any time and can be racing with any other RDS path.
272 */
Haggai Eran7c1eb452015-07-30 17:50:14 +0300273static void rds_ib_remove_one(struct ib_device *device, void *client_data)
Andy Groverec162272009-02-24 15:30:30 +0000274{
Haggai Eran7c1eb452015-07-30 17:50:14 +0300275 struct rds_ib_device *rds_ibdev = client_data;
Andy Groverec162272009-02-24 15:30:30 +0000276
Andy Groverec162272009-02-24 15:30:30 +0000277 if (!rds_ibdev)
278 return;
279
Zach Brownfc19de32010-05-24 13:16:57 -0700280 rds_ib_dev_shutdown(rds_ibdev);
Andy Groverec162272009-02-24 15:30:30 +0000281
Zach Brownea819862010-07-15 12:34:33 -0700282 /* stop connection attempts from getting a reference to this device. */
Zach Brown3e0249f2010-05-18 15:48:51 -0700283 ib_set_client_data(device, &rds_ib_client, NULL);
Zach Brownea819862010-07-15 12:34:33 -0700284
285 down_write(&rds_ib_devices_lock);
286 list_del_rcu(&rds_ibdev->list);
287 up_write(&rds_ib_devices_lock);
288
289 /*
290 * This synchronize rcu is waiting for readers of both the ib
291 * client data and the devices list to finish before we drop
292 * both of those references.
293 */
Zach Brown3e0249f2010-05-18 15:48:51 -0700294 synchronize_rcu();
295 rds_ib_dev_put(rds_ibdev);
Zach Brown3e0249f2010-05-18 15:48:51 -0700296 rds_ib_dev_put(rds_ibdev);
Andy Groverec162272009-02-24 15:30:30 +0000297}
298
299struct ib_client rds_ib_client = {
300 .name = "rds_ib",
301 .add = rds_ib_add_one,
302 .remove = rds_ib_remove_one
303};
304
305static int rds_ib_conn_info_visitor(struct rds_connection *conn,
306 void *buffer)
307{
308 struct rds_info_rdma_connection *iinfo = buffer;
Zhu Yanjune0e6d062019-08-23 21:04:16 -0400309 struct rds_ib_connection *ic = conn->c_transport_data;
Andy Groverec162272009-02-24 15:30:30 +0000310
311 /* We will only ever look at IB transports */
312 if (conn->c_trans != &rds_ib_transport)
313 return 0;
Ka-Cheong Poon1e2b44e2018-07-23 20:51:22 -0700314 if (conn->c_isv6)
315 return 0;
Andy Groverec162272009-02-24 15:30:30 +0000316
Ka-Cheong Pooneee2fa62018-07-23 20:51:21 -0700317 iinfo->src_addr = conn->c_laddr.s6_addr32[3];
318 iinfo->dst_addr = conn->c_faddr.s6_addr32[3];
Zhu Yanjune0e6d062019-08-23 21:04:16 -0400319 if (ic) {
320 iinfo->tos = conn->c_tos;
321 iinfo->sl = ic->i_sl;
322 }
Andy Groverec162272009-02-24 15:30:30 +0000323
324 memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid));
325 memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid));
326 if (rds_conn_state(conn) == RDS_CONN_UP) {
327 struct rds_ib_device *rds_ibdev;
Andy Groverec162272009-02-24 15:30:30 +0000328
Parav Pandita2e812e2018-01-18 10:11:18 +0200329 rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo->src_gid,
330 (union ib_gid *)&iinfo->dst_gid);
Andy Groverec162272009-02-24 15:30:30 +0000331
Zach Brown3e0249f2010-05-18 15:48:51 -0700332 rds_ibdev = ic->rds_ibdev;
Andy Groverec162272009-02-24 15:30:30 +0000333 iinfo->max_send_wr = ic->i_send_ring.w_nr;
334 iinfo->max_recv_wr = ic->i_recv_ring.w_nr;
335 iinfo->max_send_sge = rds_ibdev->max_sge;
336 rds_ib_get_mr_info(rds_ibdev, iinfo);
Zhu Yanjunfe3475a2019-06-03 00:28:01 -0400337 iinfo->cache_allocs = atomic_read(&ic->i_cache_allocs);
Andy Groverec162272009-02-24 15:30:30 +0000338 }
339 return 1;
340}
341
Ka-Cheong Poone65d4d92018-07-30 22:48:42 -0700342#if IS_ENABLED(CONFIG_IPV6)
Ka-Cheong Poonb7ff8b12018-07-23 20:51:23 -0700343/* IPv6 version of rds_ib_conn_info_visitor(). */
344static int rds6_ib_conn_info_visitor(struct rds_connection *conn,
345 void *buffer)
346{
347 struct rds6_info_rdma_connection *iinfo6 = buffer;
Zhu Yanjune0e6d062019-08-23 21:04:16 -0400348 struct rds_ib_connection *ic = conn->c_transport_data;
Ka-Cheong Poonb7ff8b12018-07-23 20:51:23 -0700349
350 /* We will only ever look at IB transports */
351 if (conn->c_trans != &rds_ib_transport)
352 return 0;
353
354 iinfo6->src_addr = conn->c_laddr;
355 iinfo6->dst_addr = conn->c_faddr;
Zhu Yanjune0e6d062019-08-23 21:04:16 -0400356 if (ic) {
357 iinfo6->tos = conn->c_tos;
358 iinfo6->sl = ic->i_sl;
359 }
Ka-Cheong Poonb7ff8b12018-07-23 20:51:23 -0700360
361 memset(&iinfo6->src_gid, 0, sizeof(iinfo6->src_gid));
362 memset(&iinfo6->dst_gid, 0, sizeof(iinfo6->dst_gid));
363
364 if (rds_conn_state(conn) == RDS_CONN_UP) {
365 struct rds_ib_device *rds_ibdev;
Ka-Cheong Poonb7ff8b12018-07-23 20:51:23 -0700366
Zhu Yanjun53ae9142018-08-25 15:19:05 +0800367 rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo6->src_gid,
368 (union ib_gid *)&iinfo6->dst_gid);
Ka-Cheong Poonb7ff8b12018-07-23 20:51:23 -0700369 rds_ibdev = ic->rds_ibdev;
370 iinfo6->max_send_wr = ic->i_send_ring.w_nr;
371 iinfo6->max_recv_wr = ic->i_recv_ring.w_nr;
372 iinfo6->max_send_sge = rds_ibdev->max_sge;
373 rds6_ib_get_mr_info(rds_ibdev, iinfo6);
Zhu Yanjunfe3475a2019-06-03 00:28:01 -0400374 iinfo6->cache_allocs = atomic_read(&ic->i_cache_allocs);
Ka-Cheong Poonb7ff8b12018-07-23 20:51:23 -0700375 }
376 return 1;
377}
Ka-Cheong Poone65d4d92018-07-30 22:48:42 -0700378#endif
Ka-Cheong Poonb7ff8b12018-07-23 20:51:23 -0700379
Andy Groverec162272009-02-24 15:30:30 +0000380static void rds_ib_ic_info(struct socket *sock, unsigned int len,
381 struct rds_info_iterator *iter,
382 struct rds_info_lengths *lens)
383{
Salvatore Mesoracaf1cb9d62018-03-11 22:07:49 +0100384 u64 buffer[(sizeof(struct rds_info_rdma_connection) + 7) / 8];
385
Andy Groverec162272009-02-24 15:30:30 +0000386 rds_for_each_conn_info(sock, len, iter, lens,
387 rds_ib_conn_info_visitor,
Salvatore Mesoracaf1cb9d62018-03-11 22:07:49 +0100388 buffer,
Andy Groverec162272009-02-24 15:30:30 +0000389 sizeof(struct rds_info_rdma_connection));
390}
391
Ka-Cheong Poone65d4d92018-07-30 22:48:42 -0700392#if IS_ENABLED(CONFIG_IPV6)
Ka-Cheong Poonb7ff8b12018-07-23 20:51:23 -0700393/* IPv6 version of rds_ib_ic_info(). */
394static void rds6_ib_ic_info(struct socket *sock, unsigned int len,
395 struct rds_info_iterator *iter,
396 struct rds_info_lengths *lens)
397{
398 u64 buffer[(sizeof(struct rds6_info_rdma_connection) + 7) / 8];
399
400 rds_for_each_conn_info(sock, len, iter, lens,
401 rds6_ib_conn_info_visitor,
402 buffer,
403 sizeof(struct rds6_info_rdma_connection));
404}
Ka-Cheong Poone65d4d92018-07-30 22:48:42 -0700405#endif
Andy Groverec162272009-02-24 15:30:30 +0000406
407/*
408 * Early RDS/IB was built to only bind to an address if there is an IPoIB
409 * device with that address set.
410 *
411 * If it were me, I'd advocate for something more flexible. Sending and
412 * receiving should be device-agnostic. Transports would try and maintain
413 * connections between peers who have messages queued. Userspace would be
414 * allowed to influence which paths have priority. We could call userspace
415 * asserting this policy "routing".
416 */
Ka-Cheong Pooneee2fa62018-07-23 20:51:21 -0700417static int rds_ib_laddr_check(struct net *net, const struct in6_addr *addr,
418 __u32 scope_id)
Andy Groverec162272009-02-24 15:30:30 +0000419{
420 int ret;
421 struct rdma_cm_id *cm_id;
Ka-Cheong Poone65d4d92018-07-30 22:48:42 -0700422#if IS_ENABLED(CONFIG_IPV6)
Ka-Cheong Poon1e2b44e2018-07-23 20:51:22 -0700423 struct sockaddr_in6 sin6;
Ka-Cheong Poone65d4d92018-07-30 22:48:42 -0700424#endif
Andy Groverec162272009-02-24 15:30:30 +0000425 struct sockaddr_in sin;
Ka-Cheong Poon1e2b44e2018-07-23 20:51:22 -0700426 struct sockaddr *sa;
427 bool isv4;
Andy Groverec162272009-02-24 15:30:30 +0000428
Ka-Cheong Poon1e2b44e2018-07-23 20:51:22 -0700429 isv4 = ipv6_addr_v4mapped(addr);
Andy Groverec162272009-02-24 15:30:30 +0000430 /* Create a CMA ID and try to bind it. This catches both
431 * IB and iWARP capable NICs.
432 */
Guanglei Li2c0aa082018-02-06 10:43:21 +0800433 cm_id = rdma_create_id(&init_net, rds_rdma_cm_event_handler,
434 NULL, RDMA_PS_TCP, IB_QPT_RC);
Dan Carpenter94713ba2009-04-09 14:09:46 +0000435 if (IS_ERR(cm_id))
436 return PTR_ERR(cm_id);
Andy Groverec162272009-02-24 15:30:30 +0000437
Ka-Cheong Poon1e2b44e2018-07-23 20:51:22 -0700438 if (isv4) {
439 memset(&sin, 0, sizeof(sin));
440 sin.sin_family = AF_INET;
441 sin.sin_addr.s_addr = addr->s6_addr32[3];
442 sa = (struct sockaddr *)&sin;
443 } else {
Ka-Cheong Poone65d4d92018-07-30 22:48:42 -0700444#if IS_ENABLED(CONFIG_IPV6)
Ka-Cheong Poon1e2b44e2018-07-23 20:51:22 -0700445 memset(&sin6, 0, sizeof(sin6));
446 sin6.sin6_family = AF_INET6;
447 sin6.sin6_addr = *addr;
448 sin6.sin6_scope_id = scope_id;
449 sa = (struct sockaddr *)&sin6;
450
451 /* XXX Do a special IPv6 link local address check here. The
452 * reason is that rdma_bind_addr() always succeeds with IPv6
453 * link local address regardless it is indeed configured in a
454 * system.
455 */
456 if (ipv6_addr_type(addr) & IPV6_ADDR_LINKLOCAL) {
457 struct net_device *dev;
458
Ka-Cheong Poone65d4d92018-07-30 22:48:42 -0700459 if (scope_id == 0) {
460 ret = -EADDRNOTAVAIL;
461 goto out;
462 }
Ka-Cheong Poon1e2b44e2018-07-23 20:51:22 -0700463
464 /* Use init_net for now as RDS is not network
465 * name space aware.
466 */
467 dev = dev_get_by_index(&init_net, scope_id);
Ka-Cheong Poone65d4d92018-07-30 22:48:42 -0700468 if (!dev) {
469 ret = -EADDRNOTAVAIL;
470 goto out;
471 }
Ka-Cheong Poon1e2b44e2018-07-23 20:51:22 -0700472 if (!ipv6_chk_addr(&init_net, addr, dev, 1)) {
473 dev_put(dev);
Ka-Cheong Poone65d4d92018-07-30 22:48:42 -0700474 ret = -EADDRNOTAVAIL;
475 goto out;
Ka-Cheong Poon1e2b44e2018-07-23 20:51:22 -0700476 }
477 dev_put(dev);
478 }
Ka-Cheong Poone65d4d92018-07-30 22:48:42 -0700479#else
480 ret = -EADDRNOTAVAIL;
481 goto out;
482#endif
Ka-Cheong Poon1e2b44e2018-07-23 20:51:22 -0700483 }
Andy Groverec162272009-02-24 15:30:30 +0000484
485 /* rdma_bind_addr will only succeed for IB & iWARP devices */
Ka-Cheong Poon1e2b44e2018-07-23 20:51:22 -0700486 ret = rdma_bind_addr(cm_id, sa);
Andy Groverec162272009-02-24 15:30:30 +0000487 /* due to this, we will claim to support iWARP devices unless we
488 check node_type. */
Sasha Levinc2349752013-12-18 23:49:42 -0500489 if (ret || !cm_id->device ||
490 cm_id->device->node_type != RDMA_NODE_IB_CA)
Andy Groverec162272009-02-24 15:30:30 +0000491 ret = -EADDRNOTAVAIL;
492
Ka-Cheong Poon1e2b44e2018-07-23 20:51:22 -0700493 rdsdebug("addr %pI6c%%%u ret %d node type %d\n",
494 addr, scope_id, ret,
Ka-Cheong Pooneee2fa62018-07-23 20:51:21 -0700495 cm_id->device ? cm_id->device->node_type : -1);
Andy Groverec162272009-02-24 15:30:30 +0000496
Ka-Cheong Poone65d4d92018-07-30 22:48:42 -0700497out:
Andy Groverec162272009-02-24 15:30:30 +0000498 rdma_destroy_id(cm_id);
499
500 return ret;
501}
502
Zach Brown24fa1632010-06-25 14:59:49 -0700503static void rds_ib_unregister_client(void)
504{
505 ib_unregister_client(&rds_ib_client);
506 /* wait for rds_ib_dev_free() to complete */
507 flush_workqueue(rds_wq);
508}
509
Sowmini Varadhanebeeb1a2018-02-03 04:26:51 -0800510static void rds_ib_set_unloading(void)
511{
512 atomic_set(&rds_ib_unloading, 1);
513}
514
515static bool rds_ib_is_unloading(struct rds_connection *conn)
516{
517 struct rds_conn_path *cp = &conn->c_path[0];
518
519 return (test_bit(RDS_DESTROY_PENDING, &cp->cp_flags) ||
520 atomic_read(&rds_ib_unloading) != 0);
521}
522
Andy Groverec162272009-02-24 15:30:30 +0000523void rds_ib_exit(void)
524{
Sowmini Varadhanebeeb1a2018-02-03 04:26:51 -0800525 rds_ib_set_unloading();
526 synchronize_rcu();
Andy Groverec162272009-02-24 15:30:30 +0000527 rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
Ka-Cheong Poone65d4d92018-07-30 22:48:42 -0700528#if IS_ENABLED(CONFIG_IPV6)
Ka-Cheong Poonb7ff8b12018-07-23 20:51:23 -0700529 rds_info_deregister_func(RDS6_INFO_IB_CONNECTIONS, rds6_ib_ic_info);
Ka-Cheong Poone65d4d92018-07-30 22:48:42 -0700530#endif
Zach Brown24fa1632010-06-25 14:59:49 -0700531 rds_ib_unregister_client();
Zach Brown8aeb1ba2010-06-25 14:58:16 -0700532 rds_ib_destroy_nodev_conns();
Andy Groverec162272009-02-24 15:30:30 +0000533 rds_ib_sysctl_exit();
534 rds_ib_recv_exit();
535 rds_trans_unregister(&rds_ib_transport);
santosh.shilimkar@oracle.comf6df6832016-03-01 15:20:46 -0800536 rds_ib_mr_exit();
Andy Groverec162272009-02-24 15:30:30 +0000537}
538
Santosh Shilimkar56dc8bc2018-10-13 21:36:49 +0800539static u8 rds_ib_get_tos_map(u8 tos)
540{
541 /* 1:1 user to transport map for RDMA transport.
542 * In future, if custom map is desired, hook can export
543 * user configurable map.
544 */
545 return tos;
546}
547
Andy Groverec162272009-02-24 15:30:30 +0000548struct rds_transport rds_ib_transport = {
549 .laddr_check = rds_ib_laddr_check,
Sowmini Varadhan226f7a72016-06-30 16:11:10 -0700550 .xmit_path_complete = rds_ib_xmit_path_complete,
Andy Groverec162272009-02-24 15:30:30 +0000551 .xmit = rds_ib_xmit,
Andy Groverec162272009-02-24 15:30:30 +0000552 .xmit_rdma = rds_ib_xmit_rdma,
Andy Grover15133f62010-01-12 14:33:38 -0800553 .xmit_atomic = rds_ib_xmit_atomic,
Sowmini Varadhan2da43c42016-06-30 16:11:15 -0700554 .recv_path = rds_ib_recv_path,
Andy Groverec162272009-02-24 15:30:30 +0000555 .conn_alloc = rds_ib_conn_alloc,
556 .conn_free = rds_ib_conn_free,
Sowmini Varadhanb04e8552016-06-30 16:11:16 -0700557 .conn_path_connect = rds_ib_conn_path_connect,
Sowmini Varadhan226f7a72016-06-30 16:11:10 -0700558 .conn_path_shutdown = rds_ib_conn_path_shutdown,
Andy Groverec162272009-02-24 15:30:30 +0000559 .inc_copy_to_user = rds_ib_inc_copy_to_user,
Andy Groverec162272009-02-24 15:30:30 +0000560 .inc_free = rds_ib_inc_free,
561 .cm_initiate_connect = rds_ib_cm_initiate_connect,
562 .cm_handle_connect = rds_ib_cm_handle_connect,
563 .cm_connect_complete = rds_ib_cm_connect_complete,
564 .stats_info_copy = rds_ib_stats_info_copy,
565 .exit = rds_ib_exit,
566 .get_mr = rds_ib_get_mr,
567 .sync_mr = rds_ib_sync_mr,
568 .free_mr = rds_ib_free_mr,
569 .flush_mrs = rds_ib_flush_mrs,
Santosh Shilimkar56dc8bc2018-10-13 21:36:49 +0800570 .get_tos_map = rds_ib_get_tos_map,
Andy Groverec162272009-02-24 15:30:30 +0000571 .t_owner = THIS_MODULE,
572 .t_name = "infiniband",
Sowmini Varadhanebeeb1a2018-02-03 04:26:51 -0800573 .t_unloading = rds_ib_is_unloading,
Andy Grover335776b2009-08-21 12:28:34 +0000574 .t_type = RDS_TRANS_IB
Andy Groverec162272009-02-24 15:30:30 +0000575};
576
Zach Brownef87b7e2010-07-09 12:26:20 -0700577int rds_ib_init(void)
Andy Groverec162272009-02-24 15:30:30 +0000578{
579 int ret;
580
581 INIT_LIST_HEAD(&rds_ib_devices);
582
santosh.shilimkar@oracle.comf6df6832016-03-01 15:20:46 -0800583 ret = rds_ib_mr_init();
Zach Brown515e0792010-07-06 15:09:56 -0700584 if (ret)
Tejun Heoc534a102011-02-01 11:42:43 +0100585 goto out;
Zach Brown515e0792010-07-06 15:09:56 -0700586
santosh.shilimkar@oracle.comad1d7dc2015-08-25 12:02:01 -0700587 ret = ib_register_client(&rds_ib_client);
588 if (ret)
santosh.shilimkar@oracle.comf6df6832016-03-01 15:20:46 -0800589 goto out_mr_exit;
santosh.shilimkar@oracle.comad1d7dc2015-08-25 12:02:01 -0700590
Andy Groverec162272009-02-24 15:30:30 +0000591 ret = rds_ib_sysctl_init();
592 if (ret)
593 goto out_ibreg;
594
595 ret = rds_ib_recv_init();
596 if (ret)
597 goto out_sysctl;
598
Zhu Yanjuna8d63a52017-03-03 00:44:26 -0500599 rds_trans_register(&rds_ib_transport);
Andy Groverec162272009-02-24 15:30:30 +0000600
601 rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
Ka-Cheong Poone65d4d92018-07-30 22:48:42 -0700602#if IS_ENABLED(CONFIG_IPV6)
Ka-Cheong Poonb7ff8b12018-07-23 20:51:23 -0700603 rds_info_register_func(RDS6_INFO_IB_CONNECTIONS, rds6_ib_ic_info);
Ka-Cheong Poone65d4d92018-07-30 22:48:42 -0700604#endif
Andy Groverec162272009-02-24 15:30:30 +0000605
606 goto out;
607
Andy Groverec162272009-02-24 15:30:30 +0000608out_sysctl:
609 rds_ib_sysctl_exit();
610out_ibreg:
Zach Brown24fa1632010-06-25 14:59:49 -0700611 rds_ib_unregister_client();
santosh.shilimkar@oracle.comf6df6832016-03-01 15:20:46 -0800612out_mr_exit:
613 rds_ib_mr_exit();
Andy Groverec162272009-02-24 15:30:30 +0000614out:
615 return ret;
616}
617
618MODULE_LICENSE("GPL");