1 /* SPDX-License-Identifier: GPL-2.0 */ 2 /* 3 * Shared Memory Communications over RDMA (SMC-R) and RoCE 4 * 5 * Definitions for SMC Connections, Link Groups and Links 6 * 7 * Copyright IBM Corp. 2016 8 * 9 * Author(s): Ursula Braun <ubraun@linux.vnet.ibm.com> 10 */ 11 12 #ifndef _SMC_CORE_H 13 #define _SMC_CORE_H 14 15 #include <linux/atomic.h> 16 #include <rdma/ib_verbs.h> 17 18 #include "smc.h" 19 #include "smc_ib.h" 20 21 #define SMC_RMBS_PER_LGR_MAX 255 /* max. # of RMBs per link group */ 22 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 */ 27 }; 28 29 enum smc_lgr_role { /* possible roles of a link group */ 30 SMC_CLNT, /* client */ 31 SMC_SERV /* server */ 32 }; 33 34 enum smc_link_state { /* possible states of a link */ 35 SMC_LNK_UNUSED, /* link is unused */ 36 SMC_LNK_INACTIVE, /* link is inactive */ 37 SMC_LNK_ACTIVATING, /* link is being activated */ 38 SMC_LNK_ACTIVE, /* link is active */ 39 }; 40 41 #define SMC_WR_BUF_SIZE 48 /* size of work request buffer */ 42 43 struct smc_wr_buf { 44 u8 raw[SMC_WR_BUF_SIZE]; 45 }; 46 47 #define SMC_WR_REG_MR_WAIT_TIME (5 * HZ)/* wait time for ib_wr_reg_mr result */ 48 49 enum smc_wr_reg_state { 50 POSTED, /* ib_wr_reg_mr request posted */ 51 CONFIRMED, /* ib_wr_reg_mr response: successful */ 52 FAILED /* ib_wr_reg_mr response: failure */ 53 }; 54 55 struct smc_rdma_sge { /* sges for RDMA writes */ 56 struct ib_sge wr_tx_rdma_sge[SMC_IB_MAX_SEND_SGE]; 57 }; 58 59 #define SMC_MAX_RDMA_WRITES 2 /* max. # of RDMA writes per 60 * message send 61 */ 62 63 struct smc_rdma_sges { /* sges per message send */ 64 struct smc_rdma_sge tx_rdma_sge[SMC_MAX_RDMA_WRITES]; 65 }; 66 67 struct smc_rdma_wr { /* work requests per message 68 * send 69 */ 70 struct ib_rdma_wr wr_tx_rdma[SMC_MAX_RDMA_WRITES]; 71 }; 72 73 struct smc_link { 74 struct smc_ib_device *smcibdev; /* ib-device */ 75 u8 ibport; /* port - values 1 | 2 */ 76 struct ib_pd *roce_pd; /* IB protection domain, 77 * unique for every RoCE QP 78 */ 79 struct ib_qp *roce_qp; /* IB queue pair */ 80 struct ib_qp_attr qp_attr; /* IB queue pair attributes */ 81 82 struct smc_wr_buf *wr_tx_bufs; /* WR send payload buffers */ 83 struct ib_send_wr *wr_tx_ibs; /* WR send meta data */ 84 struct ib_sge *wr_tx_sges; /* WR send gather meta data */ 85 struct smc_rdma_sges *wr_tx_rdma_sges;/*RDMA WRITE gather meta data*/ 86 struct smc_rdma_wr *wr_tx_rdmas; /* WR RDMA WRITE */ 87 struct smc_wr_tx_pend *wr_tx_pends; /* WR send waiting for CQE */ 88 /* above four vectors have wr_tx_cnt elements and use the same index */ 89 dma_addr_t wr_tx_dma_addr; /* DMA address of wr_tx_bufs */ 90 atomic_long_t wr_tx_id; /* seq # of last sent WR */ 91 unsigned long *wr_tx_mask; /* bit mask of used indexes */ 92 u32 wr_tx_cnt; /* number of WR send buffers */ 93 wait_queue_head_t wr_tx_wait; /* wait for free WR send buf */ 94 95 struct smc_wr_buf *wr_rx_bufs; /* WR recv payload buffers */ 96 struct ib_recv_wr *wr_rx_ibs; /* WR recv meta data */ 97 struct ib_sge *wr_rx_sges; /* WR recv scatter meta data */ 98 /* above three vectors have wr_rx_cnt elements and use the same index */ 99 dma_addr_t wr_rx_dma_addr; /* DMA address of wr_rx_bufs */ 100 u64 wr_rx_id; /* seq # of last recv WR */ 101 u32 wr_rx_cnt; /* number of WR recv buffers */ 102 unsigned long wr_rx_tstamp; /* jiffies when last buf rx */ 103 104 struct ib_reg_wr wr_reg; /* WR register memory region */ 105 wait_queue_head_t wr_reg_wait; /* wait for wr_reg result */ 106 enum smc_wr_reg_state wr_reg_state; /* state of wr_reg request */ 107 108 u8 gid[SMC_GID_SIZE];/* gid matching used vlan id*/ 109 u8 sgid_index; /* gid index for vlan id */ 110 u32 peer_qpn; /* QP number of peer */ 111 enum ib_mtu path_mtu; /* used mtu */ 112 enum ib_mtu peer_mtu; /* mtu size of peer */ 113 u32 psn_initial; /* QP tx initial packet seqno */ 114 u32 peer_psn; /* QP rx initial packet seqno */ 115 u8 peer_mac[ETH_ALEN]; /* = gid[8:10||13:15] */ 116 u8 peer_gid[SMC_GID_SIZE]; /* gid of peer*/ 117 u8 link_id; /* unique # within link group */ 118 u8 link_idx; /* index in lgr link array */ 119 struct smc_link_group *lgr; /* parent link group */ 120 struct work_struct link_down_wrk; /* wrk to bring link down */ 121 122 enum smc_link_state state; /* state of link */ 123 struct delayed_work llc_testlink_wrk; /* testlink worker */ 124 struct completion llc_testlink_resp; /* wait for rx of testlink */ 125 int llc_testlink_time; /* testlink interval */ 126 }; 127 128 /* For now we just allow one parallel link per link group. The SMC protocol 129 * allows more (up to 8). 130 */ 131 #define SMC_LINKS_PER_LGR_MAX 3 132 #define SMC_SINGLE_LINK 0 133 134 #define SMC_FIRST_CONTACT 1 /* first contact to a peer */ 135 #define SMC_REUSE_CONTACT 0 /* follow-on contact to a peer*/ 136 137 /* tx/rx buffer list element for sndbufs list and rmbs list of a lgr */ 138 struct smc_buf_desc { 139 struct list_head list; 140 void *cpu_addr; /* virtual address of buffer */ 141 struct page *pages; 142 int len; /* length of buffer */ 143 u32 used; /* currently used / unused */ 144 union { 145 struct { /* SMC-R */ 146 struct sg_table sgt[SMC_LINKS_PER_LGR_MAX]; 147 /* virtual buffer */ 148 struct ib_mr *mr_rx[SMC_LINKS_PER_LGR_MAX]; 149 /* for rmb only: memory region 150 * incl. rkey provided to peer 151 */ 152 u32 order; /* allocation order */ 153 154 u8 is_conf_rkey; 155 /* confirm_rkey done */ 156 u8 is_reg_mr[SMC_LINKS_PER_LGR_MAX]; 157 /* mem region registered */ 158 u8 is_map_ib[SMC_LINKS_PER_LGR_MAX]; 159 /* mem region mapped to lnk */ 160 u8 is_reg_err; 161 /* buffer registration err */ 162 }; 163 struct { /* SMC-D */ 164 unsigned short sba_idx; 165 /* SBA index number */ 166 u64 token; 167 /* DMB token number */ 168 dma_addr_t dma_addr; 169 /* DMA address */ 170 }; 171 }; 172 }; 173 174 struct smc_rtoken { /* address/key of remote RMB */ 175 u64 dma_addr; 176 u32 rkey; 177 }; 178 179 #define SMC_LGR_ID_SIZE 4 180 #define SMC_BUF_MIN_SIZE 16384 /* minimum size of an RMB */ 181 #define SMC_RMBE_SIZES 16 /* number of distinct RMBE sizes */ 182 /* theoretically, the RFC states that largest size would be 512K, 183 * i.e. compressed 5 and thus 6 sizes (0..5), despite 184 * struct smc_clc_msg_accept_confirm.rmbe_size being a 4 bit value (0..15) 185 */ 186 187 struct smcd_dev; 188 189 enum smc_lgr_type { /* redundancy state of lgr */ 190 SMC_LGR_NONE, /* no active links, lgr to be deleted */ 191 SMC_LGR_SINGLE, /* 1 active RNIC on each peer */ 192 SMC_LGR_SYMMETRIC, /* 2 active RNICs on each peer */ 193 SMC_LGR_ASYMMETRIC_PEER, /* local has 2, peer 1 active RNICs */ 194 SMC_LGR_ASYMMETRIC_LOCAL, /* local has 1, peer 2 active RNICs */ 195 }; 196 197 enum smc_llc_flowtype { 198 SMC_LLC_FLOW_NONE = 0, 199 SMC_LLC_FLOW_ADD_LINK = 2, 200 SMC_LLC_FLOW_DEL_LINK = 4, 201 SMC_LLC_FLOW_RKEY = 6, 202 }; 203 204 struct smc_llc_qentry; 205 206 struct smc_llc_flow { 207 enum smc_llc_flowtype type; 208 struct smc_llc_qentry *qentry; 209 }; 210 211 struct smc_link_group { 212 struct list_head list; 213 struct rb_root conns_all; /* connection tree */ 214 rwlock_t conns_lock; /* protects conns_all */ 215 unsigned int conns_num; /* current # of connections */ 216 unsigned short vlan_id; /* vlan id of link group */ 217 218 struct list_head sndbufs[SMC_RMBE_SIZES];/* tx buffers */ 219 struct mutex sndbufs_lock; /* protects tx buffers */ 220 struct list_head rmbs[SMC_RMBE_SIZES]; /* rx buffers */ 221 struct mutex rmbs_lock; /* protects rx buffers */ 222 223 u8 id[SMC_LGR_ID_SIZE]; /* unique lgr id */ 224 struct delayed_work free_work; /* delayed freeing of an lgr */ 225 struct work_struct terminate_work; /* abnormal lgr termination */ 226 u8 sync_err : 1; /* lgr no longer fits to peer */ 227 u8 terminating : 1;/* lgr is terminating */ 228 u8 freefast : 1; /* free worker scheduled fast */ 229 u8 freeing : 1; /* lgr is being freed */ 230 231 bool is_smcd; /* SMC-R or SMC-D */ 232 union { 233 struct { /* SMC-R */ 234 enum smc_lgr_role role; 235 /* client or server */ 236 struct smc_link lnk[SMC_LINKS_PER_LGR_MAX]; 237 /* smc link */ 238 char peer_systemid[SMC_SYSTEMID_LEN]; 239 /* unique system_id of peer */ 240 struct smc_rtoken rtokens[SMC_RMBS_PER_LGR_MAX] 241 [SMC_LINKS_PER_LGR_MAX]; 242 /* remote addr/key pairs */ 243 DECLARE_BITMAP(rtokens_used_mask, SMC_RMBS_PER_LGR_MAX); 244 /* used rtoken elements */ 245 u8 next_link_id; 246 enum smc_lgr_type type; 247 /* redundancy state */ 248 u8 pnet_id[SMC_MAX_PNETID_LEN + 1]; 249 /* pnet id of this lgr */ 250 struct list_head llc_event_q; 251 /* queue for llc events */ 252 spinlock_t llc_event_q_lock; 253 /* protects llc_event_q */ 254 struct mutex llc_conf_mutex; 255 /* protects lgr reconfig. */ 256 struct work_struct llc_add_link_work; 257 struct work_struct llc_del_link_work; 258 struct work_struct llc_event_work; 259 /* llc event worker */ 260 wait_queue_head_t llc_waiter; 261 /* w4 next llc event */ 262 struct smc_llc_flow llc_flow_lcl; 263 /* llc local control field */ 264 struct smc_llc_flow llc_flow_rmt; 265 /* llc remote control field */ 266 struct smc_llc_qentry *delayed_event; 267 /* arrived when flow active */ 268 spinlock_t llc_flow_lock; 269 /* protects llc flow */ 270 int llc_testlink_time; 271 /* link keep alive time */ 272 }; 273 struct { /* SMC-D */ 274 u64 peer_gid; 275 /* Peer GID (remote) */ 276 struct smcd_dev *smcd; 277 /* ISM device for VLAN reg. */ 278 u8 peer_shutdown : 1; 279 /* peer triggered shutdownn */ 280 }; 281 }; 282 }; 283 284 struct smc_clc_msg_local; 285 286 struct smc_init_info { 287 u8 is_smcd; 288 unsigned short vlan_id; 289 int srv_first_contact; 290 int cln_first_contact; 291 /* SMC-R */ 292 struct smc_clc_msg_local *ib_lcl; 293 struct smc_ib_device *ib_dev; 294 u8 ib_gid[SMC_GID_SIZE]; 295 u8 ib_port; 296 u32 ib_clcqpn; 297 /* SMC-D */ 298 u64 ism_gid; 299 struct smcd_dev *ism_dev; 300 }; 301 302 /* Find the connection associated with the given alert token in the link group. 303 * To use rbtrees we have to implement our own search core. 304 * Requires @conns_lock 305 * @token alert token to search for 306 * @lgr link group to search in 307 * Returns connection associated with token if found, NULL otherwise. 308 */ 309 static inline struct smc_connection *smc_lgr_find_conn( 310 u32 token, struct smc_link_group *lgr) 311 { 312 struct smc_connection *res = NULL; 313 struct rb_node *node; 314 315 node = lgr->conns_all.rb_node; 316 while (node) { 317 struct smc_connection *cur = rb_entry(node, 318 struct smc_connection, alert_node); 319 320 if (cur->alert_token_local > token) { 321 node = node->rb_left; 322 } else { 323 if (cur->alert_token_local < token) { 324 node = node->rb_right; 325 } else { 326 res = cur; 327 break; 328 } 329 } 330 } 331 332 return res; 333 } 334 335 /* returns true if the specified link is usable */ 336 static inline bool smc_link_usable(struct smc_link *lnk) 337 { 338 if (lnk->state == SMC_LNK_UNUSED || lnk->state == SMC_LNK_INACTIVE) 339 return false; 340 return true; 341 } 342 343 struct smc_sock; 344 struct smc_clc_msg_accept_confirm; 345 struct smc_clc_msg_local; 346 347 void smc_lgr_cleanup_early(struct smc_connection *conn); 348 void smc_lgr_terminate_sched(struct smc_link_group *lgr); 349 void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport); 350 void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport); 351 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid, 352 unsigned short vlan); 353 void smc_smcd_terminate_all(struct smcd_dev *dev); 354 void smc_smcr_terminate_all(struct smc_ib_device *smcibdev); 355 int smc_buf_create(struct smc_sock *smc, bool is_smcd); 356 int smc_uncompress_bufsize(u8 compressed); 357 int smc_rmb_rtoken_handling(struct smc_connection *conn, struct smc_link *link, 358 struct smc_clc_msg_accept_confirm *clc); 359 int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey); 360 int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey); 361 void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new, 362 __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey); 363 void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id, 364 __be64 nw_vaddr, __be32 nw_rkey); 365 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn); 366 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn); 367 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn); 368 void smc_rmb_sync_sg_for_device(struct smc_connection *conn); 369 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini); 370 371 void smc_conn_free(struct smc_connection *conn); 372 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini); 373 void smc_lgr_schedule_free_work_fast(struct smc_link_group *lgr); 374 int smc_core_init(void); 375 void smc_core_exit(void); 376 377 int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk, 378 u8 link_idx, struct smc_init_info *ini); 379 void smcr_link_clear(struct smc_link *lnk); 380 int smcr_buf_map_lgr(struct smc_link *lnk); 381 int smcr_buf_reg_lgr(struct smc_link *lnk); 382 int smcr_link_reg_rmb(struct smc_link *link, struct smc_buf_desc *rmb_desc); 383 void smcr_link_down_cond(struct smc_link *lnk); 384 void smcr_link_down_cond_sched(struct smc_link *lnk); 385 386 static inline struct smc_link_group *smc_get_lgr(struct smc_link *link) 387 { 388 return link->lgr; 389 } 390 #endif 391