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 */
110 struct percpu_ref wr_tx_refs;
111 } ____cacheline_aligned_in_smp;
112 struct completion tx_ref_comp;
114 struct smc_wr_buf *wr_rx_bufs; /* WR recv payload buffers */
115 struct ib_recv_wr *wr_rx_ibs; /* WR recv meta data */
116 struct ib_sge *wr_rx_sges; /* WR recv scatter meta data */
117 /* above three vectors have wr_rx_cnt elements and use the same index */
118 dma_addr_t wr_rx_dma_addr; /* DMA address of wr_rx_bufs */
119 dma_addr_t wr_rx_v2_dma_addr; /* DMA address of v2 rx buf*/
120 u64 wr_rx_id; /* seq # of last recv WR */
121 u64 wr_rx_id_compl; /* seq # of last completed WR */
122 u32 wr_rx_cnt; /* number of WR recv buffers */
123 unsigned long wr_rx_tstamp; /* jiffies when last buf rx */
124 wait_queue_head_t wr_rx_empty_wait; /* wait for RQ empty */
126 struct ib_reg_wr wr_reg; /* WR register memory region */
127 wait_queue_head_t wr_reg_wait; /* wait for wr_reg result */
129 struct percpu_ref wr_reg_refs;
130 } ____cacheline_aligned_in_smp;
131 struct completion reg_ref_comp;
132 enum smc_wr_reg_state wr_reg_state; /* state of wr_reg request */
134 u8 gid[SMC_GID_SIZE];/* gid matching used vlan id*/
135 u8 sgid_index; /* gid index for vlan id */
136 u32 peer_qpn; /* QP number of peer */
137 enum ib_mtu path_mtu; /* used mtu */
138 enum ib_mtu peer_mtu; /* mtu size of peer */
139 u32 psn_initial; /* QP tx initial packet seqno */
140 u32 peer_psn; /* QP rx initial packet seqno */
141 u8 peer_mac[ETH_ALEN]; /* = gid[8:10||13:15] */
142 u8 peer_gid[SMC_GID_SIZE]; /* gid of peer*/
143 u8 link_id; /* unique # within link group */
144 u8 link_uid[SMC_LGR_ID_SIZE]; /* unique lnk id */
145 u8 peer_link_uid[SMC_LGR_ID_SIZE]; /* peer uid */
146 u8 link_idx; /* index in lgr link array */
147 u8 link_is_asym; /* is link asymmetric? */
148 u8 clearing : 1; /* link is being cleared */
149 refcount_t refcnt; /* link reference count */
150 struct smc_link_group *lgr; /* parent link group */
151 struct work_struct link_down_wrk; /* wrk to bring link down */
152 char ibname[IB_DEVICE_NAME_MAX]; /* ib device name */
153 int ndev_ifidx; /* network device ifindex */
155 enum smc_link_state state; /* state of link */
156 struct delayed_work llc_testlink_wrk; /* testlink worker */
157 struct completion llc_testlink_resp; /* wait for rx of testlink */
158 int llc_testlink_time; /* testlink interval */
159 atomic_t conn_cnt; /* connections on this link */
162 /* For now we just allow one parallel link per link group. The SMC protocol
163 * allows more (up to 8).
165 #define SMC_LINKS_PER_LGR_MAX 3
166 #define SMC_SINGLE_LINK 0
168 /* tx/rx buffer list element for sndbufs list and rmbs list of a lgr */
169 struct smc_buf_desc {
170 struct list_head list;
171 void *cpu_addr; /* virtual address of buffer */
173 int len; /* length of buffer */
174 u32 used; /* currently used / unused */
177 struct sg_table sgt[SMC_LINKS_PER_LGR_MAX];
179 struct ib_mr *mr[SMC_LINKS_PER_LGR_MAX];
180 /* memory region: for rmb and
182 * incl. rkey provided to peer
183 * and lkey provided to local
185 u32 order; /* allocation order */
188 /* confirm_rkey done */
189 u8 is_reg_mr[SMC_LINKS_PER_LGR_MAX];
190 /* mem region registered */
191 u8 is_map_ib[SMC_LINKS_PER_LGR_MAX];
192 /* mem region mapped to lnk */
195 /* buffer registration err */
197 /* virtually contiguous */
200 unsigned short sba_idx;
201 /* SBA index number */
203 /* DMB token number */
210 struct smc_rtoken { /* address/key of remote RMB */
215 #define SMC_BUF_MIN_SIZE 16384 /* minimum size of an RMB */
216 #define SMC_RMBE_SIZES 16 /* number of distinct RMBE sizes */
217 /* theoretically, the RFC states that largest size would be 512K,
218 * i.e. compressed 5 and thus 6 sizes (0..5), despite
219 * struct smc_clc_msg_accept_confirm.rmbe_size being a 4 bit value (0..15)
224 enum smc_lgr_type { /* redundancy state of lgr */
225 SMC_LGR_NONE, /* no active links, lgr to be deleted */
226 SMC_LGR_SINGLE, /* 1 active RNIC on each peer */
227 SMC_LGR_SYMMETRIC, /* 2 active RNICs on each peer */
228 SMC_LGR_ASYMMETRIC_PEER, /* local has 2, peer 1 active RNICs */
229 SMC_LGR_ASYMMETRIC_LOCAL, /* local has 1, peer 2 active RNICs */
232 enum smcr_buf_type { /* types of SMC-R sndbufs and RMBs */
233 SMCR_PHYS_CONT_BUFS = 0,
234 SMCR_VIRT_CONT_BUFS = 1,
238 enum smc_llc_flowtype {
239 SMC_LLC_FLOW_NONE = 0,
240 SMC_LLC_FLOW_ADD_LINK = 2,
241 SMC_LLC_FLOW_DEL_LINK = 4,
242 SMC_LLC_FLOW_REQ_ADD_LINK = 5,
243 SMC_LLC_FLOW_RKEY = 6,
246 struct smc_llc_qentry;
248 struct smc_llc_flow {
249 enum smc_llc_flowtype type;
250 struct smc_llc_qentry *qentry;
253 struct smc_link_group {
254 struct list_head list;
255 struct rb_root conns_all; /* connection tree */
256 rwlock_t conns_lock; /* protects conns_all */
257 unsigned int conns_num; /* current # of connections */
258 unsigned short vlan_id; /* vlan id of link group */
260 struct list_head sndbufs[SMC_RMBE_SIZES];/* tx buffers */
261 struct rw_semaphore sndbufs_lock; /* protects tx buffers */
262 struct list_head rmbs[SMC_RMBE_SIZES]; /* rx buffers */
263 struct rw_semaphore rmbs_lock; /* protects rx buffers */
265 u8 id[SMC_LGR_ID_SIZE]; /* unique lgr id */
266 struct delayed_work free_work; /* delayed freeing of an lgr */
267 struct work_struct terminate_work; /* abnormal lgr termination */
268 struct workqueue_struct *tx_wq; /* wq for conn. tx workers */
269 u8 sync_err : 1; /* lgr no longer fits to peer */
270 u8 terminating : 1;/* lgr is terminating */
271 u8 freeing : 1; /* lgr is being freed */
273 refcount_t refcnt; /* lgr reference count */
274 bool is_smcd; /* SMC-R or SMC-D */
276 u8 negotiated_eid[SMC_MAX_EID_LEN];
277 u8 peer_os; /* peer operating system */
279 u8 peer_hostname[SMC_MAX_HOSTNAME_LEN];
282 enum smc_lgr_role role;
283 /* client or server */
284 struct smc_link lnk[SMC_LINKS_PER_LGR_MAX];
286 struct smc_wr_v2_buf *wr_rx_buf_v2;
287 /* WR v2 recv payload buffer */
288 struct smc_wr_v2_buf *wr_tx_buf_v2;
289 /* WR v2 send payload buffer */
290 char peer_systemid[SMC_SYSTEMID_LEN];
291 /* unique system_id of peer */
292 struct smc_rtoken rtokens[SMC_RMBS_PER_LGR_MAX]
293 [SMC_LINKS_PER_LGR_MAX];
294 /* remote addr/key pairs */
295 DECLARE_BITMAP(rtokens_used_mask, SMC_RMBS_PER_LGR_MAX);
296 /* used rtoken elements */
298 enum smc_lgr_type type;
299 enum smcr_buf_type buf_type;
300 /* redundancy state */
301 u8 pnet_id[SMC_MAX_PNETID_LEN + 1];
302 /* pnet id of this lgr */
303 struct list_head llc_event_q;
304 /* queue for llc events */
305 spinlock_t llc_event_q_lock;
306 /* protects llc_event_q */
307 struct rw_semaphore llc_conf_mutex;
308 /* protects lgr reconfig. */
309 struct work_struct llc_add_link_work;
310 struct work_struct llc_del_link_work;
311 struct work_struct llc_event_work;
312 /* llc event worker */
313 wait_queue_head_t llc_flow_waiter;
314 /* w4 next llc event */
315 wait_queue_head_t llc_msg_waiter;
316 /* w4 next llc msg */
317 struct smc_llc_flow llc_flow_lcl;
318 /* llc local control field */
319 struct smc_llc_flow llc_flow_rmt;
320 /* llc remote control field */
321 struct smc_llc_qentry *delayed_event;
322 /* arrived when flow active */
323 spinlock_t llc_flow_lock;
324 /* protects llc flow */
325 int llc_testlink_time;
326 /* link keep alive time */
327 u32 llc_termination_rsn;
328 /* rsn code for termination */
329 u8 nexthop_mac[ETH_ALEN];
337 /* Peer GID (remote) */
338 struct smcd_dev *smcd;
339 /* ISM device for VLAN reg. */
340 u8 peer_shutdown : 1;
341 /* peer triggered shutdownn */
346 struct smc_clc_msg_local;
348 #define GID_LIST_SIZE 2
352 u8 list[GID_LIST_SIZE][SMC_GID_SIZE];
355 struct smc_init_info_smcrv2 {
361 /* Output fields when saddr is set */
362 struct smc_ib_device *ib_dev_v2;
364 u8 ib_gid_v2[SMC_GID_SIZE];
366 /* Additional output fields when clc_sk and daddr is set as well */
368 u8 nexthop_mac[ETH_ALEN];
370 struct smc_gidlist gidlist;
373 struct smc_init_info {
377 u8 first_contact_peer;
378 u8 first_contact_local;
379 unsigned short vlan_id;
381 u8 negotiated_eid[SMC_MAX_EID_LEN];
385 u8 peer_gid[SMC_GID_SIZE];
386 u8 peer_mac[ETH_ALEN];
387 u8 peer_systemid[SMC_SYSTEMID_LEN];
388 struct smc_ib_device *ib_dev;
389 u8 ib_gid[SMC_GID_SIZE];
392 struct smc_init_info_smcrv2 smcrv2;
394 u64 ism_peer_gid[SMC_MAX_ISM_DEVS + 1];
395 struct smcd_dev *ism_dev[SMC_MAX_ISM_DEVS + 1];
396 u16 ism_chid[SMC_MAX_ISM_DEVS + 1];
397 u8 ism_offered_cnt; /* # of ISM devices offered */
398 u8 ism_selected; /* index of selected ISM dev*/
402 /* Find the connection associated with the given alert token in the link group.
403 * To use rbtrees we have to implement our own search core.
404 * Requires @conns_lock
405 * @token alert token to search for
406 * @lgr link group to search in
407 * Returns connection associated with token if found, NULL otherwise.
409 static inline struct smc_connection *smc_lgr_find_conn(
410 u32 token, struct smc_link_group *lgr)
412 struct smc_connection *res = NULL;
413 struct rb_node *node;
415 node = lgr->conns_all.rb_node;
417 struct smc_connection *cur = rb_entry(node,
418 struct smc_connection, alert_node);
420 if (cur->alert_token_local > token) {
421 node = node->rb_left;
423 if (cur->alert_token_local < token) {
424 node = node->rb_right;
435 static inline bool smc_conn_lgr_valid(struct smc_connection *conn)
437 return conn->lgr && conn->alert_token_local;
441 * Returns true if the specified link is usable.
443 * usable means the link is ready to receive RDMA messages, map memory
444 * on the link, etc. This doesn't ensure we are able to send RDMA messages
445 * on this link, if sending RDMA messages is needed, use smc_link_sendable()
447 static inline bool smc_link_usable(struct smc_link *lnk)
449 if (lnk->state == SMC_LNK_UNUSED || lnk->state == SMC_LNK_INACTIVE)
455 * Returns true if the specified link is ready to receive AND send RDMA
458 * For the client side in first contact, the underlying QP may still in
459 * RESET or RTR when the link state is ACTIVATING, checks in smc_link_usable()
460 * is not strong enough. For those places that need to send any CDC or LLC
461 * messages, use smc_link_sendable(), otherwise, use smc_link_usable() instead
463 static inline bool smc_link_sendable(struct smc_link *lnk)
465 return smc_link_usable(lnk) &&
466 lnk->qp_attr.cur_qp_state == IB_QPS_RTS;
469 static inline bool smc_link_active(struct smc_link *lnk)
471 return lnk->state == SMC_LNK_ACTIVE;
474 static inline void smc_gid_be16_convert(__u8 *buf, u8 *gid_raw)
476 sprintf(buf, "%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x",
477 be16_to_cpu(((__be16 *)gid_raw)[0]),
478 be16_to_cpu(((__be16 *)gid_raw)[1]),
479 be16_to_cpu(((__be16 *)gid_raw)[2]),
480 be16_to_cpu(((__be16 *)gid_raw)[3]),
481 be16_to_cpu(((__be16 *)gid_raw)[4]),
482 be16_to_cpu(((__be16 *)gid_raw)[5]),
483 be16_to_cpu(((__be16 *)gid_raw)[6]),
484 be16_to_cpu(((__be16 *)gid_raw)[7]));
492 __u8 pci_id[SMC_PCI_ID_STR_LEN];
495 static inline void smc_set_pci_values(struct pci_dev *pci_dev,
496 struct smc_pci_dev *smc_dev)
498 smc_dev->pci_vendor = pci_dev->vendor;
499 smc_dev->pci_device = pci_dev->device;
500 snprintf(smc_dev->pci_id, sizeof(smc_dev->pci_id), "%s",
502 #if IS_ENABLED(CONFIG_S390)
503 { /* Set s390 specific PCI information */
504 struct zpci_dev *zdev;
506 zdev = to_zpci(pci_dev);
507 smc_dev->pci_fid = zdev->fid;
508 smc_dev->pci_pchid = zdev->pchid;
514 struct smc_clc_msg_accept_confirm;
516 void smc_lgr_cleanup_early(struct smc_link_group *lgr);
517 void smc_lgr_terminate_sched(struct smc_link_group *lgr);
518 void smc_lgr_hold(struct smc_link_group *lgr);
519 void smc_lgr_put(struct smc_link_group *lgr);
520 void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport);
521 void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport);
522 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid,
523 unsigned short vlan);
524 void smc_smcd_terminate_all(struct smcd_dev *dev);
525 void smc_smcr_terminate_all(struct smc_ib_device *smcibdev);
526 int smc_buf_create(struct smc_sock *smc, bool is_smcd);
527 int smc_uncompress_bufsize(u8 compressed);
528 int smc_rmb_rtoken_handling(struct smc_connection *conn, struct smc_link *link,
529 struct smc_clc_msg_accept_confirm *clc);
530 int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey);
531 int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey);
532 void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new,
533 __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey);
534 void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id,
535 __be64 nw_vaddr, __be32 nw_rkey);
536 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn);
537 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn);
538 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini);
540 void smc_conn_free(struct smc_connection *conn);
541 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini);
542 void smc_lgr_schedule_free_work_fast(struct smc_link_group *lgr);
543 int smc_core_init(void);
544 void smc_core_exit(void);
546 int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
547 u8 link_idx, struct smc_init_info *ini);
548 void smcr_link_clear(struct smc_link *lnk, bool log);
549 void smcr_link_hold(struct smc_link *lnk);
550 void smcr_link_put(struct smc_link *lnk);
551 void smc_switch_link_and_count(struct smc_connection *conn,
552 struct smc_link *to_lnk);
553 int smcr_buf_map_lgr(struct smc_link *lnk);
554 int smcr_buf_reg_lgr(struct smc_link *lnk);
555 void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type);
556 void smcr_lgr_set_type_asym(struct smc_link_group *lgr,
557 enum smc_lgr_type new_type, int asym_lnk_idx);
558 int smcr_link_reg_buf(struct smc_link *link, struct smc_buf_desc *rmb_desc);
559 struct smc_link *smc_switch_conns(struct smc_link_group *lgr,
560 struct smc_link *from_lnk, bool is_dev_err);
561 void smcr_link_down_cond(struct smc_link *lnk);
562 void smcr_link_down_cond_sched(struct smc_link *lnk);
563 int smc_nl_get_sys_info(struct sk_buff *skb, struct netlink_callback *cb);
564 int smcr_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb);
565 int smcr_nl_get_link(struct sk_buff *skb, struct netlink_callback *cb);
566 int smcd_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb);
568 static inline struct smc_link_group *smc_get_lgr(struct smc_link *link)