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 static const struct ib_device_ops rxe_ib_dev_odp_ops = {
38 .advise_mr = rxe_ib_advise_mr,
39 };
40
41 /* initialize rxe device parameters */
rxe_init_device_param(struct rxe_dev * rxe,struct net_device * ndev)42 static void rxe_init_device_param(struct rxe_dev *rxe, struct net_device *ndev)
43 {
44 rxe->max_inline_data = RXE_MAX_INLINE_DATA;
45
46 rxe->attr.vendor_id = RXE_VENDOR_ID;
47 rxe->attr.max_mr_size = RXE_MAX_MR_SIZE;
48 rxe->attr.page_size_cap = RXE_PAGE_SIZE_CAP;
49 rxe->attr.max_qp = RXE_MAX_QP;
50 rxe->attr.max_qp_wr = RXE_MAX_QP_WR;
51 rxe->attr.device_cap_flags = RXE_DEVICE_CAP_FLAGS;
52 rxe->attr.kernel_cap_flags = IBK_ALLOW_USER_UNREG;
53 rxe->attr.max_send_sge = RXE_MAX_SGE;
54 rxe->attr.max_recv_sge = RXE_MAX_SGE;
55 rxe->attr.max_sge_rd = RXE_MAX_SGE_RD;
56 rxe->attr.max_cq = RXE_MAX_CQ;
57 rxe->attr.max_cqe = (1 << RXE_MAX_LOG_CQE) - 1;
58 rxe->attr.max_mr = RXE_MAX_MR;
59 rxe->attr.max_mw = RXE_MAX_MW;
60 rxe->attr.max_pd = RXE_MAX_PD;
61 rxe->attr.max_qp_rd_atom = RXE_MAX_QP_RD_ATOM;
62 rxe->attr.max_res_rd_atom = RXE_MAX_RES_RD_ATOM;
63 rxe->attr.max_qp_init_rd_atom = RXE_MAX_QP_INIT_RD_ATOM;
64 rxe->attr.atomic_cap = IB_ATOMIC_HCA;
65 rxe->attr.max_mcast_grp = RXE_MAX_MCAST_GRP;
66 rxe->attr.max_mcast_qp_attach = RXE_MAX_MCAST_QP_ATTACH;
67 rxe->attr.max_total_mcast_qp_attach = RXE_MAX_TOT_MCAST_QP_ATTACH;
68 rxe->attr.max_ah = RXE_MAX_AH;
69 rxe->attr.max_srq = RXE_MAX_SRQ;
70 rxe->attr.max_srq_wr = RXE_MAX_SRQ_WR;
71 rxe->attr.max_srq_sge = RXE_MAX_SRQ_SGE;
72 rxe->attr.max_fast_reg_page_list_len = RXE_MAX_FMR_PAGE_LIST_LEN;
73 rxe->attr.max_pkeys = RXE_MAX_PKEYS;
74 rxe->attr.local_ca_ack_delay = RXE_LOCAL_CA_ACK_DELAY;
75
76 if (ndev->addr_len) {
77 memcpy(rxe->raw_gid, ndev->dev_addr,
78 min_t(unsigned int, ndev->addr_len, ETH_ALEN));
79 } else {
80 /*
81 * This device does not have a HW address, but
82 * connection mangagement requires a unique gid.
83 */
84 eth_random_addr(rxe->raw_gid);
85 }
86
87 addrconf_addr_eui48((unsigned char *)&rxe->attr.sys_image_guid,
88 rxe->raw_gid);
89
90 rxe->max_ucontext = RXE_MAX_UCONTEXT;
91
92 if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) {
93 rxe->attr.kernel_cap_flags |= IBK_ON_DEMAND_PAGING;
94
95 /* IB_ODP_SUPPORT_IMPLICIT is not supported right now. */
96 rxe->attr.odp_caps.general_caps |= IB_ODP_SUPPORT;
97
98 rxe->attr.odp_caps.per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_SEND;
99 rxe->attr.odp_caps.per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_RECV;
100 rxe->attr.odp_caps.per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;
101
102 rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SEND;
103 rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_RECV;
104 rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_WRITE;
105 rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_READ;
106 rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_ATOMIC;
107 rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV;
108 rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_FLUSH;
109 rxe->attr.odp_caps.per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_ATOMIC_WRITE;
110
111 /* set handler for ODP prefetching API - ibv_advise_mr(3) */
112 ib_set_device_ops(&rxe->ib_dev, &rxe_ib_dev_odp_ops);
113 }
114 }
115
116 /* initialize port attributes */
rxe_init_port_param(struct rxe_port * port)117 static void rxe_init_port_param(struct rxe_port *port)
118 {
119 port->attr.state = IB_PORT_DOWN;
120 port->attr.max_mtu = IB_MTU_4096;
121 port->attr.active_mtu = IB_MTU_256;
122 port->attr.gid_tbl_len = RXE_PORT_GID_TBL_LEN;
123 port->attr.port_cap_flags = RXE_PORT_PORT_CAP_FLAGS;
124 port->attr.max_msg_sz = RXE_PORT_MAX_MSG_SZ;
125 port->attr.bad_pkey_cntr = RXE_PORT_BAD_PKEY_CNTR;
126 port->attr.qkey_viol_cntr = RXE_PORT_QKEY_VIOL_CNTR;
127 port->attr.pkey_tbl_len = RXE_PORT_PKEY_TBL_LEN;
128 port->attr.lid = RXE_PORT_LID;
129 port->attr.sm_lid = RXE_PORT_SM_LID;
130 port->attr.lmc = RXE_PORT_LMC;
131 port->attr.max_vl_num = RXE_PORT_MAX_VL_NUM;
132 port->attr.sm_sl = RXE_PORT_SM_SL;
133 port->attr.subnet_timeout = RXE_PORT_SUBNET_TIMEOUT;
134 port->attr.init_type_reply = RXE_PORT_INIT_TYPE_REPLY;
135 port->attr.active_width = RXE_PORT_ACTIVE_WIDTH;
136 port->attr.active_speed = RXE_PORT_ACTIVE_SPEED;
137 port->attr.phys_state = RXE_PORT_PHYS_STATE;
138 port->mtu_cap = ib_mtu_enum_to_int(IB_MTU_256);
139 port->subnet_prefix = cpu_to_be64(RXE_PORT_SUBNET_PREFIX);
140 }
141
142 /* initialize port state, note IB convention that HCA ports are always
143 * numbered from 1
144 */
rxe_init_ports(struct rxe_dev * rxe,struct net_device * ndev)145 static void rxe_init_ports(struct rxe_dev *rxe, struct net_device *ndev)
146 {
147 struct rxe_port *port = &rxe->port;
148
149 rxe_init_port_param(port);
150 addrconf_addr_eui48((unsigned char *)&port->port_guid,
151 rxe->raw_gid);
152 spin_lock_init(&port->port_lock);
153 }
154
155 /* init pools of managed objects */
rxe_init_pools(struct rxe_dev * rxe)156 static void rxe_init_pools(struct rxe_dev *rxe)
157 {
158 rxe_pool_init(rxe, &rxe->uc_pool, RXE_TYPE_UC);
159 rxe_pool_init(rxe, &rxe->pd_pool, RXE_TYPE_PD);
160 rxe_pool_init(rxe, &rxe->ah_pool, RXE_TYPE_AH);
161 rxe_pool_init(rxe, &rxe->srq_pool, RXE_TYPE_SRQ);
162 rxe_pool_init(rxe, &rxe->qp_pool, RXE_TYPE_QP);
163 rxe_pool_init(rxe, &rxe->cq_pool, RXE_TYPE_CQ);
164 rxe_pool_init(rxe, &rxe->mr_pool, RXE_TYPE_MR);
165 rxe_pool_init(rxe, &rxe->mw_pool, RXE_TYPE_MW);
166 }
167
168 /* initialize rxe device state */
rxe_init(struct rxe_dev * rxe,struct net_device * ndev)169 static void rxe_init(struct rxe_dev *rxe, struct net_device *ndev)
170 {
171 /* init default device parameters */
172 rxe_init_device_param(rxe, ndev);
173
174 rxe_init_ports(rxe, ndev);
175 rxe_init_pools(rxe);
176
177 /* init pending mmap list */
178 spin_lock_init(&rxe->mmap_offset_lock);
179 spin_lock_init(&rxe->pending_lock);
180 INIT_LIST_HEAD(&rxe->pending_mmaps);
181
182 /* init multicast support */
183 spin_lock_init(&rxe->mcg_lock);
184 rxe->mcg_tree = RB_ROOT;
185
186 mutex_init(&rxe->usdev_lock);
187 }
188
rxe_set_mtu(struct rxe_dev * rxe,unsigned int ndev_mtu)189 void rxe_set_mtu(struct rxe_dev *rxe, unsigned int ndev_mtu)
190 {
191 struct rxe_port *port = &rxe->port;
192 enum ib_mtu mtu;
193
194 mtu = eth_mtu_int_to_enum(ndev_mtu);
195
196 /* Make sure that new MTU in range */
197 mtu = mtu ? min_t(enum ib_mtu, mtu, IB_MTU_4096) : IB_MTU_256;
198
199 port->attr.active_mtu = mtu;
200 port->mtu_cap = ib_mtu_enum_to_int(mtu);
201 }
202
203 /* called by ifc layer to create new rxe device.
204 * The caller should allocate memory for rxe by calling ib_alloc_device.
205 */
rxe_add(struct rxe_dev * rxe,unsigned int mtu,const char * ibdev_name,struct net_device * ndev)206 int rxe_add(struct rxe_dev *rxe, unsigned int mtu, const char *ibdev_name,
207 struct net_device *ndev)
208 {
209 rxe_init(rxe, ndev);
210 rxe_set_mtu(rxe, mtu);
211
212 return rxe_register_device(rxe, ibdev_name, ndev);
213 }
214
rxe_newlink(const char * ibdev_name,struct net_device * ndev)215 static int rxe_newlink(const char *ibdev_name, struct net_device *ndev)
216 {
217 struct rxe_dev *rxe;
218 int err = 0;
219
220 if (is_vlan_dev(ndev)) {
221 rxe_err("rxe creation allowed on top of a real device only\n");
222 err = -EPERM;
223 goto err;
224 }
225
226 rxe = rxe_get_dev_from_net(ndev);
227 if (rxe) {
228 ib_device_put(&rxe->ib_dev);
229 rxe_err_dev(rxe, "already configured on %s\n", ndev->name);
230 err = -EEXIST;
231 goto err;
232 }
233
234 err = rxe_net_add(ibdev_name, ndev);
235 if (err) {
236 rxe_err("failed to add %s\n", ndev->name);
237 goto err;
238 }
239 err:
240 return err;
241 }
242
243 static struct rdma_link_ops rxe_link_ops = {
244 .type = "rxe",
245 .newlink = rxe_newlink,
246 };
247
rxe_module_init(void)248 static int __init rxe_module_init(void)
249 {
250 int err;
251
252 err = rxe_alloc_wq();
253 if (err)
254 return err;
255
256 err = rxe_net_init();
257 if (err) {
258 rxe_destroy_wq();
259 return err;
260 }
261
262 rdma_link_register(&rxe_link_ops);
263 pr_info("loaded\n");
264 return 0;
265 }
266
rxe_module_exit(void)267 static void __exit rxe_module_exit(void)
268 {
269 rdma_link_unregister(&rxe_link_ops);
270 ib_unregister_driver(RDMA_DRIVER_RXE);
271 rxe_net_exit();
272 rxe_destroy_wq();
273
274 pr_info("unloaded\n");
275 }
276
277 late_initcall(rxe_module_init);
278 module_exit(rxe_module_exit);
279
280 MODULE_ALIAS_RDMA_LINK("rxe");
281