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 <linux/smc.h>
17 #include <linux/pci.h>
18 #include <rdma/ib_verbs.h>
19 #include <net/genetlink.h>
24 #define SMC_RMBS_PER_LGR_MAX 255 /* max. # of RMBs per link group */
26 struct smc_lgr_list { /* list of link group definition */
27 struct list_head list;
28 spinlock_t lock; /* protects list of link groups */
29 u32 num; /* unique link group number */
32 enum smc_lgr_role { /* possible roles of a link group */
33 SMC_CLNT, /* client */
37 enum smc_link_state { /* possible states of a link */
38 SMC_LNK_UNUSED, /* link is unused */
39 SMC_LNK_INACTIVE, /* link is inactive */
40 SMC_LNK_ACTIVATING, /* link is being activated */
41 SMC_LNK_ACTIVE, /* link is active */
44 #define SMC_WR_BUF_SIZE 48 /* size of work request buffer */
45 #define SMC_WR_BUF_V2_SIZE 8192 /* size of v2 work request buffer */
48 u8 raw[SMC_WR_BUF_SIZE];
51 struct smc_wr_v2_buf {
52 u8 raw[SMC_WR_BUF_V2_SIZE];
55 #define SMC_WR_REG_MR_WAIT_TIME (5 * HZ)/* wait time for ib_wr_reg_mr result */
57 enum smc_wr_reg_state {
58 POSTED, /* ib_wr_reg_mr request posted */
59 CONFIRMED, /* ib_wr_reg_mr response: successful */
60 FAILED /* ib_wr_reg_mr response: failure */
63 struct smc_rdma_sge { /* sges for RDMA writes */
64 struct ib_sge wr_tx_rdma_sge[SMC_IB_MAX_SEND_SGE];
67 #define SMC_MAX_RDMA_WRITES 2 /* max. # of RDMA writes per
71 struct smc_rdma_sges { /* sges per message send */
72 struct smc_rdma_sge tx_rdma_sge[SMC_MAX_RDMA_WRITES];
75 struct smc_rdma_wr { /* work requests per message
78 struct ib_rdma_wr wr_tx_rdma[SMC_MAX_RDMA_WRITES];
81 #define SMC_LGR_ID_SIZE 4
84 struct smc_ib_device *smcibdev; /* ib-device */
85 u8 ibport; /* port - values 1 | 2 */
86 struct ib_pd *roce_pd; /* IB protection domain,
87 * unique for every RoCE QP
89 struct ib_qp *roce_qp; /* IB queue pair */
90 struct ib_qp_attr qp_attr; /* IB queue pair attributes */
92 struct smc_wr_buf *wr_tx_bufs; /* WR send payload buffers */
93 struct ib_send_wr *wr_tx_ibs; /* WR send meta data */
94 struct ib_sge *wr_tx_sges; /* WR send gather meta data */
95 struct smc_rdma_sges *wr_tx_rdma_sges;/*RDMA WRITE gather meta data*/
96 struct smc_rdma_wr *wr_tx_rdmas; /* WR RDMA WRITE */
97 struct smc_wr_tx_pend *wr_tx_pends; /* WR send waiting for CQE */
98 struct completion *wr_tx_compl; /* WR send CQE completion */
99 /* above four vectors have wr_tx_cnt elements and use the same index */
100 struct ib_send_wr *wr_tx_v2_ib; /* WR send v2 meta data */
101 struct ib_sge *wr_tx_v2_sge; /* WR send v2 gather meta data*/
102 struct smc_wr_tx_pend *wr_tx_v2_pend; /* WR send v2 waiting for CQE */
103 dma_addr_t wr_tx_dma_addr; /* DMA address of wr_tx_bufs */
104 dma_addr_t wr_tx_v2_dma_addr; /* DMA address of v2 tx buf*/
105 atomic_long_t wr_tx_id; /* seq # of last sent WR */
106 unsigned long *wr_tx_mask; /* bit mask of used indexes */
107 u32 wr_tx_cnt; /* number of WR send buffers */
108 wait_queue_head_t wr_tx_wait; /* wait for free WR send buf */
109 atomic_t wr_tx_refcnt; /* tx refs to link */
111 struct smc_wr_buf *wr_rx_bufs; /* WR recv payload buffers */
112 struct ib_recv_wr *wr_rx_ibs; /* WR recv meta data */
113 struct ib_sge *wr_rx_sges; /* WR recv scatter meta data */
114 /* above three vectors have wr_rx_cnt elements and use the same index */
115 dma_addr_t wr_rx_dma_addr; /* DMA address of wr_rx_bufs */
116 dma_addr_t wr_rx_v2_dma_addr; /* DMA address of v2 rx buf*/
117 u64 wr_rx_id; /* seq # of last recv WR */
118 u64 wr_rx_id_compl; /* seq # of last completed WR */
119 u32 wr_rx_cnt; /* number of WR recv buffers */
120 unsigned long wr_rx_tstamp; /* jiffies when last buf rx */
121 wait_queue_head_t wr_rx_empty_wait; /* wait for RQ empty */
123 struct ib_reg_wr wr_reg; /* WR register memory region */
124 wait_queue_head_t wr_reg_wait; /* wait for wr_reg result */
125 atomic_t wr_reg_refcnt; /* reg refs to link */
126 enum smc_wr_reg_state wr_reg_state; /* state of wr_reg request */
128 u8 gid[SMC_GID_SIZE];/* gid matching used vlan id*/
129 u8 sgid_index; /* gid index for vlan id */
130 u32 peer_qpn; /* QP number of peer */
131 enum ib_mtu path_mtu; /* used mtu */
132 enum ib_mtu peer_mtu; /* mtu size of peer */
133 u32 psn_initial; /* QP tx initial packet seqno */
134 u32 peer_psn; /* QP rx initial packet seqno */
135 u8 peer_mac[ETH_ALEN]; /* = gid[8:10||13:15] */
136 u8 peer_gid[SMC_GID_SIZE]; /* gid of peer*/
137 u8 link_id; /* unique # within link group */
138 u8 link_uid[SMC_LGR_ID_SIZE]; /* unique lnk id */
139 u8 peer_link_uid[SMC_LGR_ID_SIZE]; /* peer uid */
140 u8 link_idx; /* index in lgr link array */
141 u8 link_is_asym; /* is link asymmetric? */
142 u8 clearing : 1; /* link is being cleared */
143 refcount_t refcnt; /* link reference count */
144 struct smc_link_group *lgr; /* parent link group */
145 struct work_struct link_down_wrk; /* wrk to bring link down */
146 char ibname[IB_DEVICE_NAME_MAX]; /* ib device name */
147 int ndev_ifidx; /* network device ifindex */
149 enum smc_link_state state; /* state of link */
150 struct delayed_work llc_testlink_wrk; /* testlink worker */
151 struct completion llc_testlink_resp; /* wait for rx of testlink */
152 int llc_testlink_time; /* testlink interval */
153 atomic_t conn_cnt; /* connections on this link */
156 /* For now we just allow one parallel link per link group. The SMC protocol
157 * allows more (up to 8).
159 #define SMC_LINKS_PER_LGR_MAX 3
160 #define SMC_SINGLE_LINK 0
162 /* tx/rx buffer list element for sndbufs list and rmbs list of a lgr */
163 struct smc_buf_desc {
164 struct list_head list;
165 void *cpu_addr; /* virtual address of buffer */
167 int len; /* length of buffer */
168 u32 used; /* currently used / unused */
171 struct sg_table sgt[SMC_LINKS_PER_LGR_MAX];
173 struct ib_mr *mr_rx[SMC_LINKS_PER_LGR_MAX];
174 /* for rmb only: memory region
175 * incl. rkey provided to peer
177 u32 order; /* allocation order */
180 /* confirm_rkey done */
181 u8 is_reg_mr[SMC_LINKS_PER_LGR_MAX];
182 /* mem region registered */
183 u8 is_map_ib[SMC_LINKS_PER_LGR_MAX];
184 /* mem region mapped to lnk */
186 /* buffer registration err */
189 unsigned short sba_idx;
190 /* SBA index number */
192 /* DMB token number */
199 struct smc_rtoken { /* address/key of remote RMB */
204 #define SMC_BUF_MIN_SIZE 16384 /* minimum size of an RMB */
205 #define SMC_RMBE_SIZES 16 /* number of distinct RMBE sizes */
206 /* theoretically, the RFC states that largest size would be 512K,
207 * i.e. compressed 5 and thus 6 sizes (0..5), despite
208 * struct smc_clc_msg_accept_confirm.rmbe_size being a 4 bit value (0..15)
213 enum smc_lgr_type { /* redundancy state of lgr */
214 SMC_LGR_NONE, /* no active links, lgr to be deleted */
215 SMC_LGR_SINGLE, /* 1 active RNIC on each peer */
216 SMC_LGR_SYMMETRIC, /* 2 active RNICs on each peer */
217 SMC_LGR_ASYMMETRIC_PEER, /* local has 2, peer 1 active RNICs */
218 SMC_LGR_ASYMMETRIC_LOCAL, /* local has 1, peer 2 active RNICs */
221 enum smc_llc_flowtype {
222 SMC_LLC_FLOW_NONE = 0,
223 SMC_LLC_FLOW_ADD_LINK = 2,
224 SMC_LLC_FLOW_DEL_LINK = 4,
225 SMC_LLC_FLOW_REQ_ADD_LINK = 5,
226 SMC_LLC_FLOW_RKEY = 6,
229 struct smc_llc_qentry;
231 struct smc_llc_flow {
232 enum smc_llc_flowtype type;
233 struct smc_llc_qentry *qentry;
236 struct smc_link_group {
237 struct list_head list;
238 struct rb_root conns_all; /* connection tree */
239 rwlock_t conns_lock; /* protects conns_all */
240 unsigned int conns_num; /* current # of connections */
241 unsigned short vlan_id; /* vlan id of link group */
243 struct list_head sndbufs[SMC_RMBE_SIZES];/* tx buffers */
244 struct mutex sndbufs_lock; /* protects tx buffers */
245 struct list_head rmbs[SMC_RMBE_SIZES]; /* rx buffers */
246 struct mutex rmbs_lock; /* protects rx buffers */
248 u8 id[SMC_LGR_ID_SIZE]; /* unique lgr id */
249 struct delayed_work free_work; /* delayed freeing of an lgr */
250 struct work_struct terminate_work; /* abnormal lgr termination */
251 struct workqueue_struct *tx_wq; /* wq for conn. tx workers */
252 u8 sync_err : 1; /* lgr no longer fits to peer */
253 u8 terminating : 1;/* lgr is terminating */
254 u8 freeing : 1; /* lgr is being freed */
256 refcount_t refcnt; /* lgr reference count */
257 bool is_smcd; /* SMC-R or SMC-D */
259 u8 negotiated_eid[SMC_MAX_EID_LEN];
260 u8 peer_os; /* peer operating system */
262 u8 peer_hostname[SMC_MAX_HOSTNAME_LEN];
265 enum smc_lgr_role role;
266 /* client or server */
267 struct smc_link lnk[SMC_LINKS_PER_LGR_MAX];
269 struct smc_wr_v2_buf *wr_rx_buf_v2;
270 /* WR v2 recv payload buffer */
271 struct smc_wr_v2_buf *wr_tx_buf_v2;
272 /* WR v2 send payload buffer */
273 char peer_systemid[SMC_SYSTEMID_LEN];
274 /* unique system_id of peer */
275 struct smc_rtoken rtokens[SMC_RMBS_PER_LGR_MAX]
276 [SMC_LINKS_PER_LGR_MAX];
277 /* remote addr/key pairs */
278 DECLARE_BITMAP(rtokens_used_mask, SMC_RMBS_PER_LGR_MAX);
279 /* used rtoken elements */
281 enum smc_lgr_type type;
282 /* redundancy state */
283 u8 pnet_id[SMC_MAX_PNETID_LEN + 1];
284 /* pnet id of this lgr */
285 struct list_head llc_event_q;
286 /* queue for llc events */
287 spinlock_t llc_event_q_lock;
288 /* protects llc_event_q */
289 struct mutex llc_conf_mutex;
290 /* protects lgr reconfig. */
291 struct work_struct llc_add_link_work;
292 struct work_struct llc_del_link_work;
293 struct work_struct llc_event_work;
294 /* llc event worker */
295 wait_queue_head_t llc_flow_waiter;
296 /* w4 next llc event */
297 wait_queue_head_t llc_msg_waiter;
298 /* w4 next llc msg */
299 struct smc_llc_flow llc_flow_lcl;
300 /* llc local control field */
301 struct smc_llc_flow llc_flow_rmt;
302 /* llc remote control field */
303 struct smc_llc_qentry *delayed_event;
304 /* arrived when flow active */
305 spinlock_t llc_flow_lock;
306 /* protects llc flow */
307 int llc_testlink_time;
308 /* link keep alive time */
309 u32 llc_termination_rsn;
310 /* rsn code for termination */
311 u8 nexthop_mac[ETH_ALEN];
319 /* Peer GID (remote) */
320 struct smcd_dev *smcd;
321 /* ISM device for VLAN reg. */
322 u8 peer_shutdown : 1;
323 /* peer triggered shutdownn */
328 struct smc_clc_msg_local;
330 #define GID_LIST_SIZE 2
334 u8 list[GID_LIST_SIZE][SMC_GID_SIZE];
337 struct smc_init_info_smcrv2 {
343 /* Output fields when saddr is set */
344 struct smc_ib_device *ib_dev_v2;
346 u8 ib_gid_v2[SMC_GID_SIZE];
348 /* Additional output fields when clc_sk and daddr is set as well */
350 u8 nexthop_mac[ETH_ALEN];
352 struct smc_gidlist gidlist;
355 struct smc_init_info {
359 u8 first_contact_peer;
360 u8 first_contact_local;
361 unsigned short vlan_id;
363 u8 negotiated_eid[SMC_MAX_EID_LEN];
367 u8 peer_gid[SMC_GID_SIZE];
368 u8 peer_mac[ETH_ALEN];
369 u8 peer_systemid[SMC_SYSTEMID_LEN];
370 struct smc_ib_device *ib_dev;
371 u8 ib_gid[SMC_GID_SIZE];
374 struct smc_init_info_smcrv2 smcrv2;
376 u64 ism_peer_gid[SMC_MAX_ISM_DEVS + 1];
377 struct smcd_dev *ism_dev[SMC_MAX_ISM_DEVS + 1];
378 u16 ism_chid[SMC_MAX_ISM_DEVS + 1];
379 u8 ism_offered_cnt; /* # of ISM devices offered */
380 u8 ism_selected; /* index of selected ISM dev*/
384 /* Find the connection associated with the given alert token in the link group.
385 * To use rbtrees we have to implement our own search core.
386 * Requires @conns_lock
387 * @token alert token to search for
388 * @lgr link group to search in
389 * Returns connection associated with token if found, NULL otherwise.
391 static inline struct smc_connection *smc_lgr_find_conn(
392 u32 token, struct smc_link_group *lgr)
394 struct smc_connection *res = NULL;
395 struct rb_node *node;
397 node = lgr->conns_all.rb_node;
399 struct smc_connection *cur = rb_entry(node,
400 struct smc_connection, alert_node);
402 if (cur->alert_token_local > token) {
403 node = node->rb_left;
405 if (cur->alert_token_local < token) {
406 node = node->rb_right;
417 static inline bool smc_conn_lgr_valid(struct smc_connection *conn)
419 return conn->lgr && conn->alert_token_local;
423 * Returns true if the specified link is usable.
425 * usable means the link is ready to receive RDMA messages, map memory
426 * on the link, etc. This doesn't ensure we are able to send RDMA messages
427 * on this link, if sending RDMA messages is needed, use smc_link_sendable()
429 static inline bool smc_link_usable(struct smc_link *lnk)
431 if (lnk->state == SMC_LNK_UNUSED || lnk->state == SMC_LNK_INACTIVE)
437 * Returns true if the specified link is ready to receive AND send RDMA
440 * For the client side in first contact, the underlying QP may still in
441 * RESET or RTR when the link state is ACTIVATING, checks in smc_link_usable()
442 * is not strong enough. For those places that need to send any CDC or LLC
443 * messages, use smc_link_sendable(), otherwise, use smc_link_usable() instead
445 static inline bool smc_link_sendable(struct smc_link *lnk)
447 return smc_link_usable(lnk) &&
448 lnk->qp_attr.cur_qp_state == IB_QPS_RTS;
451 static inline bool smc_link_active(struct smc_link *lnk)
453 return lnk->state == SMC_LNK_ACTIVE;
456 static inline void smc_gid_be16_convert(__u8 *buf, u8 *gid_raw)
458 sprintf(buf, "%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x",
459 be16_to_cpu(((__be16 *)gid_raw)[0]),
460 be16_to_cpu(((__be16 *)gid_raw)[1]),
461 be16_to_cpu(((__be16 *)gid_raw)[2]),
462 be16_to_cpu(((__be16 *)gid_raw)[3]),
463 be16_to_cpu(((__be16 *)gid_raw)[4]),
464 be16_to_cpu(((__be16 *)gid_raw)[5]),
465 be16_to_cpu(((__be16 *)gid_raw)[6]),
466 be16_to_cpu(((__be16 *)gid_raw)[7]));
474 __u8 pci_id[SMC_PCI_ID_STR_LEN];
477 static inline void smc_set_pci_values(struct pci_dev *pci_dev,
478 struct smc_pci_dev *smc_dev)
480 smc_dev->pci_vendor = pci_dev->vendor;
481 smc_dev->pci_device = pci_dev->device;
482 snprintf(smc_dev->pci_id, sizeof(smc_dev->pci_id), "%s",
484 #if IS_ENABLED(CONFIG_S390)
485 { /* Set s390 specific PCI information */
486 struct zpci_dev *zdev;
488 zdev = to_zpci(pci_dev);
489 smc_dev->pci_fid = zdev->fid;
490 smc_dev->pci_pchid = zdev->pchid;
496 struct smc_clc_msg_accept_confirm;
498 void smc_lgr_cleanup_early(struct smc_link_group *lgr);
499 void smc_lgr_terminate_sched(struct smc_link_group *lgr);
500 void smc_lgr_hold(struct smc_link_group *lgr);
501 void smc_lgr_put(struct smc_link_group *lgr);
502 void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport);
503 void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport);
504 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid,
505 unsigned short vlan);
506 void smc_smcd_terminate_all(struct smcd_dev *dev);
507 void smc_smcr_terminate_all(struct smc_ib_device *smcibdev);
508 int smc_buf_create(struct smc_sock *smc, bool is_smcd);
509 int smc_uncompress_bufsize(u8 compressed);
510 int smc_rmb_rtoken_handling(struct smc_connection *conn, struct smc_link *link,
511 struct smc_clc_msg_accept_confirm *clc);
512 int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey);
513 int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey);
514 void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new,
515 __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey);
516 void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id,
517 __be64 nw_vaddr, __be32 nw_rkey);
518 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn);
519 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn);
520 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn);
521 void smc_rmb_sync_sg_for_device(struct smc_connection *conn);
522 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini);
524 void smc_conn_free(struct smc_connection *conn);
525 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini);
526 void smc_lgr_schedule_free_work_fast(struct smc_link_group *lgr);
527 int smc_core_init(void);
528 void smc_core_exit(void);
530 int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
531 u8 link_idx, struct smc_init_info *ini);
532 void smcr_link_clear(struct smc_link *lnk, bool log);
533 void smcr_link_hold(struct smc_link *lnk);
534 void smcr_link_put(struct smc_link *lnk);
535 void smc_switch_link_and_count(struct smc_connection *conn,
536 struct smc_link *to_lnk);
537 int smcr_buf_map_lgr(struct smc_link *lnk);
538 int smcr_buf_reg_lgr(struct smc_link *lnk);
539 void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type);
540 void smcr_lgr_set_type_asym(struct smc_link_group *lgr,
541 enum smc_lgr_type new_type, int asym_lnk_idx);
542 int smcr_link_reg_rmb(struct smc_link *link, struct smc_buf_desc *rmb_desc);
543 struct smc_link *smc_switch_conns(struct smc_link_group *lgr,
544 struct smc_link *from_lnk, bool is_dev_err);
545 void smcr_link_down_cond(struct smc_link *lnk);
546 void smcr_link_down_cond_sched(struct smc_link *lnk);
547 int smc_nl_get_sys_info(struct sk_buff *skb, struct netlink_callback *cb);
548 int smcr_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb);
549 int smcr_nl_get_link(struct sk_buff *skb, struct netlink_callback *cb);
550 int smcd_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb);
552 static inline struct smc_link_group *smc_get_lgr(struct smc_link *link)