1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
3 * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved.
4 * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved.
5 */
6
7 #include <rdma/rdma_netlink.h>
8 #include <net/addrconf.h>
9 #include "rxe.h"
10 #include "rxe_loc.h"
11
12 MODULE_AUTHOR("Bob Pearson, Frank Zago, John Groves, Kamal Heib");
13 MODULE_DESCRIPTION("Soft RDMA transport");
14 MODULE_LICENSE("Dual BSD/GPL");
15
16 /* free resources for a rxe device all objects created for this device must
17 * have been destroyed
18 */
rxe_dealloc(struct ib_device * ib_dev)19 void rxe_dealloc(struct ib_device *ib_dev)
20 {
21 struct rxe_dev *rxe = container_of(ib_dev, struct rxe_dev, ib_dev);
22
23 rxe_pool_cleanup(&rxe->uc_pool);
24 rxe_pool_cleanup(&rxe->pd_pool);
25 rxe_pool_cleanup(&rxe->ah_pool);
26 rxe_pool_cleanup(&rxe->srq_pool);
27 rxe_pool_cleanup(&rxe->qp_pool);
28 rxe_pool_cleanup(&rxe->cq_pool);
29 rxe_pool_cleanup(&rxe->mr_pool);
30 rxe_pool_cleanup(&rxe->mw_pool);
31
32 WARN_ON(!RB_EMPTY_ROOT(&rxe->mcg_tree));
33
34 mutex_destroy(&rxe->usdev_lock);
35 }
36
37 /* initialize rxe device parameters */
rxe_init_device_param(struct rxe_dev * rxe,struct net_device * ndev)38 static void rxe_init_device_param(struct rxe_dev *rxe, struct net_device *ndev)
39 {
40 rxe->max_inline_data = RXE_MAX_INLINE_DATA;
41
42 rxe->attr.vendor_id = RXE_VENDOR_ID;
43 rxe->attr.max_mr_size = RXE_MAX_MR_SIZE;
44 rxe->attr.page_size_cap = RXE_PAGE_SIZE_CAP;
45 rxe->attr.max_qp = RXE_MAX_QP;
46 rxe->attr.max_qp_wr = RXE_MAX_QP_WR;
47 rxe->attr.device_cap_flags = RXE_DEVICE_CAP_FLAGS;
48 rxe->attr.kernel_cap_flags = IBK_ALLOW_USER_UNREG;
49 rxe->attr.max_send_sge = RXE_MAX_SGE;
50 rxe->attr.max_recv_sge = RXE_MAX_SGE;
51 rxe->attr.max_sge_rd = RXE_MAX_SGE_RD;
52 rxe->attr.max_cq = RXE_MAX_CQ;
53 rxe->attr.max_cqe = (1 << RXE_MAX_LOG_CQE) - 1;
54 rxe->attr.max_mr = RXE_MAX_MR;
55 rxe->attr.max_mw = RXE_MAX_MW;
56 rxe->attr.max_pd = RXE_MAX_PD;
57 rxe->attr.max_qp_rd_atom = RXE_MAX_QP_RD_ATOM;
58 rxe->attr.max_res_rd_atom = RXE_MAX_RES_RD_ATOM;
59 rxe->attr.max_qp_init_rd_atom = RXE_MAX_QP_INIT_RD_ATOM;
60 rxe->attr.atomic_cap = IB_ATOMIC_HCA;
61 rxe->attr.max_mcast_grp = RXE_MAX_MCAST_GRP;
62 rxe->attr.max_mcast_qp_attach = RXE_MAX_MCAST_QP_ATTACH;
63 rxe->attr.max_total_mcast_qp_attach = RXE_MAX_TOT_MCAST_QP_ATTACH;
64 rxe->attr.max_ah = RXE_MAX_AH;
65 rxe->attr.max_srq = RXE_MAX_SRQ;
66 rxe->attr.max_srq_wr = RXE_MAX_SRQ_WR;
67 rxe->attr.max_srq_sge = RXE_MAX_SRQ_SGE;
68 rxe->attr.max_fast_reg_page_list_len = RXE_MAX_FMR_PAGE_LIST_LEN;
69 rxe->attr.max_pkeys = RXE_MAX_PKEYS;
70 rxe->attr.local_ca_ack_delay = RXE_LOCAL_CA_ACK_DELAY;
71
72 if (ndev->addr_len) {
73 memcpy(rxe->raw_gid, ndev->dev_addr,
74 min_t(unsigned int, ndev->addr_len, ETH_ALEN));
75 } else {
76 /*
77 * This device does not have a HW address, but
78 * connection mangagement requires a unique gid.
79 */
80 eth_random_addr(rxe->raw_gid);
81 }
82
83 addrconf_addr_eui48((unsigned char *)&rxe->attr.sys_image_guid,
84 rxe->raw_gid);
85
86 rxe->max_ucontext = RXE_MAX_UCONTEXT;
87
88 if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
89 rxe->attr.kernel_cap_flags |= IBK_ON_DEMAND_PAGING;
90
91 /* IB_ODP_SUPPORT_IMPLICIT is not supported right now. */
92 rxe->attr.odp_caps.general_caps |= IB_ODP_SUPPORT;
93
94 rxe->attr.odp_caps.per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_SEND;
95 rxe->attr.odp_caps.per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_RECV;
96 rxe->attr.odp_caps.per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;
97
98 rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SEND;
99 rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_RECV;
100 rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_WRITE;
101 rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_READ;
102 rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_ATOMIC;
103 rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;
104 rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_FLUSH;
105 rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_ATOMIC_WRITE;
106 }
107 }
108
109 /* initialize port attributes */
rxe_init_port_param(struct rxe_port * port)110 static void rxe_init_port_param(struct rxe_port *port)
111 {
112 port->attr.state = IB_PORT_DOWN;
113 port->attr.max_mtu = IB_MTU_4096;
114 port->attr.active_mtu = IB_MTU_256;
115 port->attr.gid_tbl_len = RXE_PORT_GID_TBL_LEN;
116 port->attr.port_cap_flags = RXE_PORT_PORT_CAP_FLAGS;
117 port->attr.max_msg_sz = RXE_PORT_MAX_MSG_SZ;
118 port->attr.bad_pkey_cntr = RXE_PORT_BAD_PKEY_CNTR;
119 port->attr.qkey_viol_cntr = RXE_PORT_QKEY_VIOL_CNTR;
120 port->attr.pkey_tbl_len = RXE_PORT_PKEY_TBL_LEN;
121 port->attr.lid = RXE_PORT_LID;
122 port->attr.sm_lid = RXE_PORT_SM_LID;
123 port->attr.lmc = RXE_PORT_LMC;
124 port->attr.max_vl_num = RXE_PORT_MAX_VL_NUM;
125 port->attr.sm_sl = RXE_PORT_SM_SL;
126 port->attr.subnet_timeout = RXE_PORT_SUBNET_TIMEOUT;
127 port->attr.init_type_reply = RXE_PORT_INIT_TYPE_REPLY;
128 port->attr.active_width = RXE_PORT_ACTIVE_WIDTH;
129 port->attr.active_speed = RXE_PORT_ACTIVE_SPEED;
130 port->attr.phys_state = RXE_PORT_PHYS_STATE;
131 port->mtu_cap = ib_mtu_enum_to_int(IB_MTU_256);
132 port->subnet_prefix = cpu_to_be64(RXE_PORT_SUBNET_PREFIX);
133 }
134
135 /* initialize port state, note IB convention that HCA ports are always
136 * numbered from 1
137 */
rxe_init_ports(struct rxe_dev * rxe,struct net_device * ndev)138 static void rxe_init_ports(struct rxe_dev *rxe, struct net_device *ndev)
139 {
140 struct rxe_port *port = &rxe->port;
141
142 rxe_init_port_param(port);
143 addrconf_addr_eui48((unsigned char *)&port->port_guid,
144 rxe->raw_gid);
145 spin_lock_init(&port->port_lock);
146 }
147
148 /* init pools of managed objects */
rxe_init_pools(struct rxe_dev * rxe)149 static void rxe_init_pools(struct rxe_dev *rxe)
150 {
151 rxe_pool_init(rxe, &rxe->uc_pool, RXE_TYPE_UC);
152 rxe_pool_init(rxe, &rxe->pd_pool, RXE_TYPE_PD);
153 rxe_pool_init(rxe, &rxe->ah_pool, RXE_TYPE_AH);
154 rxe_pool_init(rxe, &rxe->srq_pool, RXE_TYPE_SRQ);
155 rxe_pool_init(rxe, &rxe->qp_pool, RXE_TYPE_QP);
156 rxe_pool_init(rxe, &rxe->cq_pool, RXE_TYPE_CQ);
157 rxe_pool_init(rxe, &rxe->mr_pool, RXE_TYPE_MR);
158 rxe_pool_init(rxe, &rxe->mw_pool, RXE_TYPE_MW);
159 }
160
161 /* initialize rxe device state */
rxe_init(struct rxe_dev * rxe,struct net_device * ndev)162 static void rxe_init(struct rxe_dev *rxe, struct net_device *ndev)
163 {
164 /* init default device parameters */
165 rxe_init_device_param(rxe, ndev);
166
167 rxe_init_ports(rxe, ndev);
168 rxe_init_pools(rxe);
169
170 /* init pending mmap list */
171 spin_lock_init(&rxe->mmap_offset_lock);
172 spin_lock_init(&rxe->pending_lock);
173 INIT_LIST_HEAD(&rxe->pending_mmaps);
174
175 /* init multicast support */
176 spin_lock_init(&rxe->mcg_lock);
177 rxe->mcg_tree = RB_ROOT;
178
179 mutex_init(&rxe->usdev_lock);
180 }
181
rxe_set_mtu(struct rxe_dev * rxe,unsigned int ndev_mtu)182 void rxe_set_mtu(struct rxe_dev *rxe, unsigned int ndev_mtu)
183 {
184 struct rxe_port *port = &rxe->port;
185 enum ib_mtu mtu;
186
187 mtu = eth_mtu_int_to_enum(ndev_mtu);
188
189 /* Make sure that new MTU in range */
190 mtu = mtu ? min_t(enum ib_mtu, mtu, IB_MTU_4096) : IB_MTU_256;
191
192 port->attr.active_mtu = mtu;
193 port->mtu_cap = ib_mtu_enum_to_int(mtu);
194 }
195
196 /* called by ifc layer to create new rxe device.
197 * The caller should allocate memory for rxe by calling ib_alloc_device.
198 */
rxe_add(struct rxe_dev * rxe,unsigned int mtu,const char * ibdev_name,struct net_device * ndev)199 int rxe_add(struct rxe_dev *rxe, unsigned int mtu, const char *ibdev_name,
200 struct net_device *ndev)
201 {
202 rxe_init(rxe, ndev);
203 rxe_set_mtu(rxe, mtu);
204
205 return rxe_register_device(rxe, ibdev_name, ndev);
206 }
207
rxe_newlink(const char * ibdev_name,struct net_device * ndev)208 static int rxe_newlink(const char *ibdev_name, struct net_device *ndev)
209 {
210 struct rxe_dev *rxe;
211 int err = 0;
212
213 if (is_vlan_dev(ndev)) {
214 rxe_err("rxe creation allowed on top of a real device only\n");
215 err = -EPERM;
216 goto err;
217 }
218
219 rxe = rxe_get_dev_from_net(ndev);
220 if (rxe) {
221 ib_device_put(&rxe->ib_dev);
222 rxe_err_dev(rxe, "already configured on %s\n", ndev->name);
223 err = -EEXIST;
224 goto err;
225 }
226
227 err = rxe_net_add(ibdev_name, ndev);
228 if (err) {
229 rxe_err("failed to add %s\n", ndev->name);
230 goto err;
231 }
232 err:
233 return err;
234 }
235
236 static struct rdma_link_ops rxe_link_ops = {
237 .type = "rxe",
238 .newlink = rxe_newlink,
239 };
240
rxe_module_init(void)241 static int __init rxe_module_init(void)
242 {
243 int err;
244
245 err = rxe_alloc_wq();
246 if (err)
247 return err;
248
249 err = rxe_net_init();
250 if (err) {
251 rxe_destroy_wq();
252 return err;
253 }
254
255 rdma_link_register(&rxe_link_ops);
256 pr_info("loaded\n");
257 return 0;
258 }
259
rxe_module_exit(void)260 static void __exit rxe_module_exit(void)
261 {
262 rdma_link_unregister(&rxe_link_ops);
263 ib_unregister_driver(RDMA_DRIVER_RXE);
264 rxe_net_exit();
265 rxe_destroy_wq();
266
267 pr_info("unloaded\n");
268 }
269
270 late_initcall(rxe_module_init);
271 module_exit(rxe_module_exit);
272
273 MODULE_ALIAS_RDMA_LINK("rxe");
274