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 <linux/smc.h>
17 #include <linux/pci.h>
18 #include <rdma/ib_verbs.h>
19 #include <net/genetlink.h>
20
21 #include "smc.h"
22 #include "smc_ib.h"
23
24 #define SMC_RMBS_PER_LGR_MAX 255 /* max. # of RMBs per link group */
25
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 */
30 };
31
32 enum smc_lgr_role { /* possible roles of a link group */
33 SMC_CLNT, /* client */
34 SMC_SERV /* server */
35 };
36
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 */
42 };
43
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 */
46
47 struct smc_wr_buf {
48 u8 raw[SMC_WR_BUF_SIZE];
49 };
50
51 struct smc_wr_v2_buf {
52 u8 raw[SMC_WR_BUF_V2_SIZE];
53 };
54
55 #define SMC_WR_REG_MR_WAIT_TIME (5 * HZ)/* wait time for ib_wr_reg_mr result */
56
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 */
61 };
62
63 struct smc_rdma_sge { /* sges for RDMA writes */
64 struct ib_sge wr_tx_rdma_sge[SMC_IB_MAX_SEND_SGE];
65 };
66
67 #define SMC_MAX_RDMA_WRITES 2 /* max. # of RDMA writes per
68 * message send
69 */
70
71 struct smc_rdma_sges { /* sges per message send */
72 struct smc_rdma_sge tx_rdma_sge[SMC_MAX_RDMA_WRITES];
73 };
74
75 struct smc_rdma_wr { /* work requests per message
76 * send
77 */
78 struct ib_rdma_wr wr_tx_rdma[SMC_MAX_RDMA_WRITES];
79 };
80
81 #define SMC_LGR_ID_SIZE 4
82
83 struct smc_link {
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
88 */
89 struct ib_qp *roce_qp; /* IB queue pair */
90 struct ib_qp_attr qp_attr; /* IB queue pair attributes */
91
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 */
110
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 u32 wr_rx_cnt; /* number of WR recv buffers */
119 unsigned long wr_rx_tstamp; /* jiffies when last buf rx */
120
121 struct ib_reg_wr wr_reg; /* WR register memory region */
122 wait_queue_head_t wr_reg_wait; /* wait for wr_reg result */
123 atomic_t wr_reg_refcnt; /* reg refs to link */
124 enum smc_wr_reg_state wr_reg_state; /* state of wr_reg request */
125
126 u8 gid[SMC_GID_SIZE];/* gid matching used vlan id*/
127 u8 sgid_index; /* gid index for vlan id */
128 u32 peer_qpn; /* QP number of peer */
129 enum ib_mtu path_mtu; /* used mtu */
130 enum ib_mtu peer_mtu; /* mtu size of peer */
131 u32 psn_initial; /* QP tx initial packet seqno */
132 u32 peer_psn; /* QP rx initial packet seqno */
133 u8 peer_mac[ETH_ALEN]; /* = gid[8:10||13:15] */
134 u8 peer_gid[SMC_GID_SIZE]; /* gid of peer*/
135 u8 link_id; /* unique # within link group */
136 u8 link_uid[SMC_LGR_ID_SIZE]; /* unique lnk id */
137 u8 peer_link_uid[SMC_LGR_ID_SIZE]; /* peer uid */
138 u8 link_idx; /* index in lgr link array */
139 u8 link_is_asym; /* is link asymmetric? */
140 struct smc_link_group *lgr; /* parent link group */
141 struct work_struct link_down_wrk; /* wrk to bring link down */
142 char ibname[IB_DEVICE_NAME_MAX]; /* ib device name */
143 int ndev_ifidx; /* network device ifindex */
144
145 enum smc_link_state state; /* state of link */
146 struct delayed_work llc_testlink_wrk; /* testlink worker */
147 struct completion llc_testlink_resp; /* wait for rx of testlink */
148 int llc_testlink_time; /* testlink interval */
149 atomic_t conn_cnt; /* connections on this link */
150 };
151
152 /* For now we just allow one parallel link per link group. The SMC protocol
153 * allows more (up to 8).
154 */
155 #define SMC_LINKS_PER_LGR_MAX 3
156 #define SMC_SINGLE_LINK 0
157
158 /* tx/rx buffer list element for sndbufs list and rmbs list of a lgr */
159 struct smc_buf_desc {
160 struct list_head list;
161 void *cpu_addr; /* virtual address of buffer */
162 struct page *pages;
163 int len; /* length of buffer */
164 u32 used; /* currently used / unused */
165 union {
166 struct { /* SMC-R */
167 struct sg_table sgt[SMC_LINKS_PER_LGR_MAX];
168 /* virtual buffer */
169 struct ib_mr *mr_rx[SMC_LINKS_PER_LGR_MAX];
170 /* for rmb only: memory region
171 * incl. rkey provided to peer
172 */
173 u32 order; /* allocation order */
174
175 u8 is_conf_rkey;
176 /* confirm_rkey done */
177 u8 is_reg_mr[SMC_LINKS_PER_LGR_MAX];
178 /* mem region registered */
179 u8 is_map_ib[SMC_LINKS_PER_LGR_MAX];
180 /* mem region mapped to lnk */
181 u8 is_reg_err;
182 /* buffer registration err */
183 };
184 struct { /* SMC-D */
185 unsigned short sba_idx;
186 /* SBA index number */
187 u64 token;
188 /* DMB token number */
189 dma_addr_t dma_addr;
190 /* DMA address */
191 };
192 };
193 };
194
195 struct smc_rtoken { /* address/key of remote RMB */
196 u64 dma_addr;
197 u32 rkey;
198 };
199
200 #define SMC_BUF_MIN_SIZE 16384 /* minimum size of an RMB */
201 #define SMC_RMBE_SIZES 16 /* number of distinct RMBE sizes */
202 /* theoretically, the RFC states that largest size would be 512K,
203 * i.e. compressed 5 and thus 6 sizes (0..5), despite
204 * struct smc_clc_msg_accept_confirm.rmbe_size being a 4 bit value (0..15)
205 */
206
207 struct smcd_dev;
208
209 enum smc_lgr_type { /* redundancy state of lgr */
210 SMC_LGR_NONE, /* no active links, lgr to be deleted */
211 SMC_LGR_SINGLE, /* 1 active RNIC on each peer */
212 SMC_LGR_SYMMETRIC, /* 2 active RNICs on each peer */
213 SMC_LGR_ASYMMETRIC_PEER, /* local has 2, peer 1 active RNICs */
214 SMC_LGR_ASYMMETRIC_LOCAL, /* local has 1, peer 2 active RNICs */
215 };
216
217 enum smc_llc_flowtype {
218 SMC_LLC_FLOW_NONE = 0,
219 SMC_LLC_FLOW_ADD_LINK = 2,
220 SMC_LLC_FLOW_DEL_LINK = 4,
221 SMC_LLC_FLOW_REQ_ADD_LINK = 5,
222 SMC_LLC_FLOW_RKEY = 6,
223 };
224
225 struct smc_llc_qentry;
226
227 struct smc_llc_flow {
228 enum smc_llc_flowtype type;
229 struct smc_llc_qentry *qentry;
230 };
231
232 struct smc_link_group {
233 struct list_head list;
234 struct rb_root conns_all; /* connection tree */
235 rwlock_t conns_lock; /* protects conns_all */
236 unsigned int conns_num; /* current # of connections */
237 unsigned short vlan_id; /* vlan id of link group */
238
239 struct list_head sndbufs[SMC_RMBE_SIZES];/* tx buffers */
240 struct mutex sndbufs_lock; /* protects tx buffers */
241 struct list_head rmbs[SMC_RMBE_SIZES]; /* rx buffers */
242 struct mutex rmbs_lock; /* protects rx buffers */
243
244 u8 id[SMC_LGR_ID_SIZE]; /* unique lgr id */
245 struct delayed_work free_work; /* delayed freeing of an lgr */
246 struct work_struct terminate_work; /* abnormal lgr termination */
247 struct workqueue_struct *tx_wq; /* wq for conn. tx workers */
248 u8 sync_err : 1; /* lgr no longer fits to peer */
249 u8 terminating : 1;/* lgr is terminating */
250 u8 freeing : 1; /* lgr is being freed */
251
252 bool is_smcd; /* SMC-R or SMC-D */
253 u8 smc_version;
254 u8 negotiated_eid[SMC_MAX_EID_LEN];
255 u8 peer_os; /* peer operating system */
256 u8 peer_smc_release;
257 u8 peer_hostname[SMC_MAX_HOSTNAME_LEN];
258 union {
259 struct { /* SMC-R */
260 enum smc_lgr_role role;
261 /* client or server */
262 struct smc_link lnk[SMC_LINKS_PER_LGR_MAX];
263 /* smc link */
264 struct smc_wr_v2_buf *wr_rx_buf_v2;
265 /* WR v2 recv payload buffer */
266 struct smc_wr_v2_buf *wr_tx_buf_v2;
267 /* WR v2 send payload buffer */
268 char peer_systemid[SMC_SYSTEMID_LEN];
269 /* unique system_id of peer */
270 struct smc_rtoken rtokens[SMC_RMBS_PER_LGR_MAX]
271 [SMC_LINKS_PER_LGR_MAX];
272 /* remote addr/key pairs */
273 DECLARE_BITMAP(rtokens_used_mask, SMC_RMBS_PER_LGR_MAX);
274 /* used rtoken elements */
275 u8 next_link_id;
276 enum smc_lgr_type type;
277 /* redundancy state */
278 u8 pnet_id[SMC_MAX_PNETID_LEN + 1];
279 /* pnet id of this lgr */
280 struct list_head llc_event_q;
281 /* queue for llc events */
282 spinlock_t llc_event_q_lock;
283 /* protects llc_event_q */
284 struct mutex llc_conf_mutex;
285 /* protects lgr reconfig. */
286 struct work_struct llc_add_link_work;
287 struct work_struct llc_del_link_work;
288 struct work_struct llc_event_work;
289 /* llc event worker */
290 wait_queue_head_t llc_flow_waiter;
291 /* w4 next llc event */
292 wait_queue_head_t llc_msg_waiter;
293 /* w4 next llc msg */
294 struct smc_llc_flow llc_flow_lcl;
295 /* llc local control field */
296 struct smc_llc_flow llc_flow_rmt;
297 /* llc remote control field */
298 struct smc_llc_qentry *delayed_event;
299 /* arrived when flow active */
300 spinlock_t llc_flow_lock;
301 /* protects llc flow */
302 int llc_testlink_time;
303 /* link keep alive time */
304 u32 llc_termination_rsn;
305 /* rsn code for termination */
306 u8 nexthop_mac[ETH_ALEN];
307 u8 uses_gateway;
308 __be32 saddr;
309 };
310 struct { /* SMC-D */
311 u64 peer_gid;
312 /* Peer GID (remote) */
313 struct smcd_dev *smcd;
314 /* ISM device for VLAN reg. */
315 u8 peer_shutdown : 1;
316 /* peer triggered shutdownn */
317 };
318 };
319 };
320
321 struct smc_clc_msg_local;
322
323 #define GID_LIST_SIZE 2
324
325 struct smc_gidlist {
326 u8 len;
327 u8 list[GID_LIST_SIZE][SMC_GID_SIZE];
328 };
329
330 struct smc_init_info_smcrv2 {
331 /* Input fields */
332 __be32 saddr;
333 struct sock *clc_sk;
334 __be32 daddr;
335
336 /* Output fields when saddr is set */
337 struct smc_ib_device *ib_dev_v2;
338 u8 ib_port_v2;
339 u8 ib_gid_v2[SMC_GID_SIZE];
340
341 /* Additional output fields when clc_sk and daddr is set as well */
342 u8 uses_gateway;
343 u8 nexthop_mac[ETH_ALEN];
344
345 struct smc_gidlist gidlist;
346 };
347
348 struct smc_init_info {
349 u8 is_smcd;
350 u8 smc_type_v1;
351 u8 smc_type_v2;
352 u8 first_contact_peer;
353 u8 first_contact_local;
354 unsigned short vlan_id;
355 u32 rc;
356 u8 negotiated_eid[SMC_MAX_EID_LEN];
357 /* SMC-R */
358 u8 smcr_version;
359 u8 check_smcrv2;
360 u8 peer_gid[SMC_GID_SIZE];
361 u8 peer_mac[ETH_ALEN];
362 u8 peer_systemid[SMC_SYSTEMID_LEN];
363 struct smc_ib_device *ib_dev;
364 u8 ib_gid[SMC_GID_SIZE];
365 u8 ib_port;
366 u32 ib_clcqpn;
367 struct smc_init_info_smcrv2 smcrv2;
368 /* SMC-D */
369 u64 ism_peer_gid[SMC_MAX_ISM_DEVS + 1];
370 struct smcd_dev *ism_dev[SMC_MAX_ISM_DEVS + 1];
371 u16 ism_chid[SMC_MAX_ISM_DEVS + 1];
372 u8 ism_offered_cnt; /* # of ISM devices offered */
373 u8 ism_selected; /* index of selected ISM dev*/
374 u8 smcd_version;
375 };
376
377 /* Find the connection associated with the given alert token in the link group.
378 * To use rbtrees we have to implement our own search core.
379 * Requires @conns_lock
380 * @token alert token to search for
381 * @lgr link group to search in
382 * Returns connection associated with token if found, NULL otherwise.
383 */
smc_lgr_find_conn(u32 token,struct smc_link_group * lgr)384 static inline struct smc_connection *smc_lgr_find_conn(
385 u32 token, struct smc_link_group *lgr)
386 {
387 struct smc_connection *res = NULL;
388 struct rb_node *node;
389
390 node = lgr->conns_all.rb_node;
391 while (node) {
392 struct smc_connection *cur = rb_entry(node,
393 struct smc_connection, alert_node);
394
395 if (cur->alert_token_local > token) {
396 node = node->rb_left;
397 } else {
398 if (cur->alert_token_local < token) {
399 node = node->rb_right;
400 } else {
401 res = cur;
402 break;
403 }
404 }
405 }
406
407 return res;
408 }
409
410 /* returns true if the specified link is usable */
smc_link_usable(struct smc_link * lnk)411 static inline bool smc_link_usable(struct smc_link *lnk)
412 {
413 if (lnk->state == SMC_LNK_UNUSED || lnk->state == SMC_LNK_INACTIVE)
414 return false;
415 return true;
416 }
417
smc_link_sendable(struct smc_link * lnk)418 static inline bool smc_link_sendable(struct smc_link *lnk)
419 {
420 return smc_link_usable(lnk) &&
421 lnk->qp_attr.cur_qp_state == IB_QPS_RTS;
422 }
423
smc_link_active(struct smc_link * lnk)424 static inline bool smc_link_active(struct smc_link *lnk)
425 {
426 return lnk->state == SMC_LNK_ACTIVE;
427 }
428
smc_gid_be16_convert(__u8 * buf,u8 * gid_raw)429 static inline void smc_gid_be16_convert(__u8 *buf, u8 *gid_raw)
430 {
431 sprintf(buf, "%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x",
432 be16_to_cpu(((__be16 *)gid_raw)[0]),
433 be16_to_cpu(((__be16 *)gid_raw)[1]),
434 be16_to_cpu(((__be16 *)gid_raw)[2]),
435 be16_to_cpu(((__be16 *)gid_raw)[3]),
436 be16_to_cpu(((__be16 *)gid_raw)[4]),
437 be16_to_cpu(((__be16 *)gid_raw)[5]),
438 be16_to_cpu(((__be16 *)gid_raw)[6]),
439 be16_to_cpu(((__be16 *)gid_raw)[7]));
440 }
441
442 struct smc_pci_dev {
443 __u32 pci_fid;
444 __u16 pci_pchid;
445 __u16 pci_vendor;
446 __u16 pci_device;
447 __u8 pci_id[SMC_PCI_ID_STR_LEN];
448 };
449
smc_set_pci_values(struct pci_dev * pci_dev,struct smc_pci_dev * smc_dev)450 static inline void smc_set_pci_values(struct pci_dev *pci_dev,
451 struct smc_pci_dev *smc_dev)
452 {
453 smc_dev->pci_vendor = pci_dev->vendor;
454 smc_dev->pci_device = pci_dev->device;
455 snprintf(smc_dev->pci_id, sizeof(smc_dev->pci_id), "%s",
456 pci_name(pci_dev));
457 #if IS_ENABLED(CONFIG_S390)
458 { /* Set s390 specific PCI information */
459 struct zpci_dev *zdev;
460
461 zdev = to_zpci(pci_dev);
462 smc_dev->pci_fid = zdev->fid;
463 smc_dev->pci_pchid = zdev->pchid;
464 }
465 #endif
466 }
467
468 struct smc_sock;
469 struct smc_clc_msg_accept_confirm;
470
471 void smc_lgr_cleanup_early(struct smc_connection *conn);
472 void smc_lgr_terminate_sched(struct smc_link_group *lgr);
473 void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport);
474 void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport);
475 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid,
476 unsigned short vlan);
477 void smc_smcd_terminate_all(struct smcd_dev *dev);
478 void smc_smcr_terminate_all(struct smc_ib_device *smcibdev);
479 int smc_buf_create(struct smc_sock *smc, bool is_smcd);
480 int smc_uncompress_bufsize(u8 compressed);
481 int smc_rmb_rtoken_handling(struct smc_connection *conn, struct smc_link *link,
482 struct smc_clc_msg_accept_confirm *clc);
483 int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey);
484 int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey);
485 void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new,
486 __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey);
487 void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id,
488 __be64 nw_vaddr, __be32 nw_rkey);
489 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn);
490 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn);
491 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn);
492 void smc_rmb_sync_sg_for_device(struct smc_connection *conn);
493 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini);
494
495 void smc_conn_free(struct smc_connection *conn);
496 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini);
497 void smc_lgr_schedule_free_work_fast(struct smc_link_group *lgr);
498 int smc_core_init(void);
499 void smc_core_exit(void);
500
501 int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
502 u8 link_idx, struct smc_init_info *ini);
503 void smcr_link_clear(struct smc_link *lnk, bool log);
504 void smc_switch_link_and_count(struct smc_connection *conn,
505 struct smc_link *to_lnk);
506 int smcr_buf_map_lgr(struct smc_link *lnk);
507 int smcr_buf_reg_lgr(struct smc_link *lnk);
508 void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type);
509 void smcr_lgr_set_type_asym(struct smc_link_group *lgr,
510 enum smc_lgr_type new_type, int asym_lnk_idx);
511 int smcr_link_reg_rmb(struct smc_link *link, struct smc_buf_desc *rmb_desc);
512 struct smc_link *smc_switch_conns(struct smc_link_group *lgr,
513 struct smc_link *from_lnk, bool is_dev_err);
514 void smcr_link_down_cond(struct smc_link *lnk);
515 void smcr_link_down_cond_sched(struct smc_link *lnk);
516 int smc_nl_get_sys_info(struct sk_buff *skb, struct netlink_callback *cb);
517 int smcr_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb);
518 int smcr_nl_get_link(struct sk_buff *skb, struct netlink_callback *cb);
519 int smcd_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb);
520
smc_get_lgr(struct smc_link * link)521 static inline struct smc_link_group *smc_get_lgr(struct smc_link *link)
522 {
523 return link->lgr;
524 }
525 #endif
526