1 /* SPDX-License-Identifier: GPL-2.0 */
3 * Shared Memory Communications over RDMA (SMC-R) and RoCE
5 * Definitions for SMC Connections, Link Groups and Links
7 * Copyright IBM Corp. 2016
9 * Author(s): Ursula Braun <ubraun@linux.vnet.ibm.com>
15 #include <linux/atomic.h>
16 #include <rdma/ib_verbs.h>
21 #define SMC_RMBS_PER_LGR_MAX 255 /* max. # of RMBs per link group */
23 struct smc_lgr_list { /* list of link group definition */
24 struct list_head list;
25 spinlock_t lock; /* protects list of link groups */
26 u32 num; /* unique link group number */
29 enum smc_lgr_role { /* possible roles of a link group */
30 SMC_CLNT, /* client */
34 enum smc_link_state { /* possible states of a link */
35 SMC_LNK_INACTIVE, /* link is inactive */
36 SMC_LNK_ACTIVATING, /* link is being activated */
37 SMC_LNK_ACTIVE /* link is active */
40 #define SMC_WR_BUF_SIZE 48 /* size of work request buffer */
43 u8 raw[SMC_WR_BUF_SIZE];
46 #define SMC_WR_REG_MR_WAIT_TIME (5 * HZ)/* wait time for ib_wr_reg_mr result */
48 enum smc_wr_reg_state {
49 POSTED, /* ib_wr_reg_mr request posted */
50 CONFIRMED, /* ib_wr_reg_mr response: successful */
51 FAILED /* ib_wr_reg_mr response: failure */
55 struct smc_ib_device *smcibdev; /* ib-device */
56 u8 ibport; /* port - values 1 | 2 */
57 struct ib_pd *roce_pd; /* IB protection domain,
58 * unique for every RoCE QP
60 struct ib_qp *roce_qp; /* IB queue pair */
61 struct ib_qp_attr qp_attr; /* IB queue pair attributes */
63 struct smc_wr_buf *wr_tx_bufs; /* WR send payload buffers */
64 struct ib_send_wr *wr_tx_ibs; /* WR send meta data */
65 struct ib_sge *wr_tx_sges; /* WR send gather meta data */
66 struct smc_wr_tx_pend *wr_tx_pends; /* WR send waiting for CQE */
67 /* above four vectors have wr_tx_cnt elements and use the same index */
68 dma_addr_t wr_tx_dma_addr; /* DMA address of wr_tx_bufs */
69 atomic_long_t wr_tx_id; /* seq # of last sent WR */
70 unsigned long *wr_tx_mask; /* bit mask of used indexes */
71 u32 wr_tx_cnt; /* number of WR send buffers */
72 wait_queue_head_t wr_tx_wait; /* wait for free WR send buf */
74 struct smc_wr_buf *wr_rx_bufs; /* WR recv payload buffers */
75 struct ib_recv_wr *wr_rx_ibs; /* WR recv meta data */
76 struct ib_sge *wr_rx_sges; /* WR recv scatter meta data */
77 /* above three vectors have wr_rx_cnt elements and use the same index */
78 dma_addr_t wr_rx_dma_addr; /* DMA address of wr_rx_bufs */
79 u64 wr_rx_id; /* seq # of last recv WR */
80 u32 wr_rx_cnt; /* number of WR recv buffers */
81 unsigned long wr_rx_tstamp; /* jiffies when last buf rx */
83 struct ib_reg_wr wr_reg; /* WR register memory region */
84 wait_queue_head_t wr_reg_wait; /* wait for wr_reg result */
85 enum smc_wr_reg_state wr_reg_state; /* state of wr_reg request */
87 u8 gid[SMC_GID_SIZE];/* gid matching used vlan id*/
88 u8 sgid_index; /* gid index for vlan id */
89 u32 peer_qpn; /* QP number of peer */
90 enum ib_mtu path_mtu; /* used mtu */
91 enum ib_mtu peer_mtu; /* mtu size of peer */
92 u32 psn_initial; /* QP tx initial packet seqno */
93 u32 peer_psn; /* QP rx initial packet seqno */
94 u8 peer_mac[ETH_ALEN]; /* = gid[8:10||13:15] */
95 u8 peer_gid[SMC_GID_SIZE]; /* gid of peer*/
96 u8 link_id; /* unique # within link group */
98 enum smc_link_state state; /* state of link */
99 struct workqueue_struct *llc_wq; /* single thread work queue */
100 struct completion llc_confirm; /* wait for rx of conf link */
101 struct completion llc_confirm_resp; /* wait 4 rx of cnf lnk rsp */
102 int llc_confirm_rc; /* rc from confirm link msg */
103 int llc_confirm_resp_rc; /* rc from conf_resp msg */
104 struct completion llc_add; /* wait for rx of add link */
105 struct completion llc_add_resp; /* wait for rx of add link rsp*/
106 struct delayed_work llc_testlink_wrk; /* testlink worker */
107 struct completion llc_testlink_resp; /* wait for rx of testlink */
108 int llc_testlink_time; /* testlink interval */
109 struct completion llc_confirm_rkey; /* wait 4 rx of cnf rkey */
110 int llc_confirm_rkey_rc; /* rc from cnf rkey msg */
113 /* For now we just allow one parallel link per link group. The SMC protocol
114 * allows more (up to 8).
116 #define SMC_LINKS_PER_LGR_MAX 1
117 #define SMC_SINGLE_LINK 0
119 #define SMC_FIRST_CONTACT 1 /* first contact to a peer */
120 #define SMC_REUSE_CONTACT 0 /* follow-on contact to a peer*/
122 /* tx/rx buffer list element for sndbufs list and rmbs list of a lgr */
123 struct smc_buf_desc {
124 struct list_head list;
125 void *cpu_addr; /* virtual address of buffer */
127 int len; /* length of buffer */
128 u32 used; /* currently used / unused */
129 u8 reused : 1; /* new created / reused */
130 u8 regerr : 1; /* err during registration */
133 struct sg_table sgt[SMC_LINKS_PER_LGR_MAX];
135 struct ib_mr *mr_rx[SMC_LINKS_PER_LGR_MAX];
136 /* for rmb only: memory region
137 * incl. rkey provided to peer
139 u32 order; /* allocation order */
142 unsigned short sba_idx;
143 /* SBA index number */
145 /* DMB token number */
152 struct smc_rtoken { /* address/key of remote RMB */
157 #define SMC_LGR_ID_SIZE 4
158 #define SMC_BUF_MIN_SIZE 16384 /* minimum size of an RMB */
159 #define SMC_RMBE_SIZES 16 /* number of distinct RMBE sizes */
160 /* theoretically, the RFC states that largest size would be 512K,
161 * i.e. compressed 5 and thus 6 sizes (0..5), despite
162 * struct smc_clc_msg_accept_confirm.rmbe_size being a 4 bit value (0..15)
167 struct smc_link_group {
168 struct list_head list;
169 struct rb_root conns_all; /* connection tree */
170 rwlock_t conns_lock; /* protects conns_all */
171 unsigned int conns_num; /* current # of connections */
172 unsigned short vlan_id; /* vlan id of link group */
174 struct list_head sndbufs[SMC_RMBE_SIZES];/* tx buffers */
175 rwlock_t sndbufs_lock; /* protects tx buffers */
176 struct list_head rmbs[SMC_RMBE_SIZES]; /* rx buffers */
177 rwlock_t rmbs_lock; /* protects rx buffers */
179 u8 id[SMC_LGR_ID_SIZE]; /* unique lgr id */
180 struct delayed_work free_work; /* delayed freeing of an lgr */
181 u8 sync_err : 1; /* lgr no longer fits to peer */
182 u8 terminating : 1;/* lgr is terminating */
184 bool is_smcd; /* SMC-R or SMC-D */
187 enum smc_lgr_role role;
188 /* client or server */
189 struct smc_link lnk[SMC_LINKS_PER_LGR_MAX];
191 char peer_systemid[SMC_SYSTEMID_LEN];
192 /* unique system_id of peer */
193 struct smc_rtoken rtokens[SMC_RMBS_PER_LGR_MAX]
194 [SMC_LINKS_PER_LGR_MAX];
195 /* remote addr/key pairs */
196 DECLARE_BITMAP(rtokens_used_mask, SMC_RMBS_PER_LGR_MAX);
197 /* used rtoken elements */
201 /* Peer GID (remote) */
202 struct smcd_dev *smcd;
203 /* ISM device for VLAN reg. */
208 /* Find the connection associated with the given alert token in the link group.
209 * To use rbtrees we have to implement our own search core.
210 * Requires @conns_lock
211 * @token alert token to search for
212 * @lgr link group to search in
213 * Returns connection associated with token if found, NULL otherwise.
215 static inline struct smc_connection *smc_lgr_find_conn(
216 u32 token, struct smc_link_group *lgr)
218 struct smc_connection *res = NULL;
219 struct rb_node *node;
221 node = lgr->conns_all.rb_node;
223 struct smc_connection *cur = rb_entry(node,
224 struct smc_connection, alert_node);
226 if (cur->alert_token_local > token) {
227 node = node->rb_left;
229 if (cur->alert_token_local < token) {
230 node = node->rb_right;
242 struct smc_clc_msg_accept_confirm;
243 struct smc_clc_msg_local;
245 void smc_lgr_free(struct smc_link_group *lgr);
246 void smc_lgr_forget(struct smc_link_group *lgr);
247 void smc_lgr_terminate(struct smc_link_group *lgr);
248 void smc_port_terminate(struct smc_ib_device *smcibdev, u8 ibport);
249 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid);
250 int smc_buf_create(struct smc_sock *smc, bool is_smcd);
251 int smc_uncompress_bufsize(u8 compressed);
252 int smc_rmb_rtoken_handling(struct smc_connection *conn,
253 struct smc_clc_msg_accept_confirm *clc);
254 int smc_rtoken_add(struct smc_link_group *lgr, __be64 nw_vaddr, __be32 nw_rkey);
255 int smc_rtoken_delete(struct smc_link_group *lgr, __be32 nw_rkey);
256 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn);
257 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn);
258 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn);
259 void smc_rmb_sync_sg_for_device(struct smc_connection *conn);
260 int smc_vlan_by_tcpsk(struct socket *clcsock, unsigned short *vlan_id);
262 void smc_conn_free(struct smc_connection *conn);
263 int smc_conn_create(struct smc_sock *smc, bool is_smcd, int srv_first_contact,
264 struct smc_ib_device *smcibdev, u8 ibport,
265 struct smc_clc_msg_local *lcl, struct smcd_dev *smcd,
267 void smcd_conn_free(struct smc_connection *conn);
268 void smc_core_exit(void);
270 static inline struct smc_link_group *smc_get_lgr(struct smc_link *link)
272 return container_of(link, struct smc_link_group, lnk[SMC_SINGLE_LINK]);