tjh.dev
Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
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
23struct 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
29enum smc_lgr_role { /* possible roles of a link group */
30 SMC_CLNT, /* client */
31 SMC_SERV /* server */
32};
33
34enum 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
43struct 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
49enum 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
55struct 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
63struct smc_rdma_sges { /* sges per message send */
64 struct smc_rdma_sge tx_rdma_sge[SMC_MAX_RDMA_WRITES];
65};
66
67struct 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#define SMC_LGR_ID_SIZE 4
74
75struct smc_link {
76 struct smc_ib_device *smcibdev; /* ib-device */
77 u8 ibport; /* port - values 1 | 2 */
78 struct ib_pd *roce_pd; /* IB protection domain,
79 * unique for every RoCE QP
80 */
81 struct ib_qp *roce_qp; /* IB queue pair */
82 struct ib_qp_attr qp_attr; /* IB queue pair attributes */
83
84 struct smc_wr_buf *wr_tx_bufs; /* WR send payload buffers */
85 struct ib_send_wr *wr_tx_ibs; /* WR send meta data */
86 struct ib_sge *wr_tx_sges; /* WR send gather meta data */
87 struct smc_rdma_sges *wr_tx_rdma_sges;/*RDMA WRITE gather meta data*/
88 struct smc_rdma_wr *wr_tx_rdmas; /* WR RDMA WRITE */
89 struct smc_wr_tx_pend *wr_tx_pends; /* WR send waiting for CQE */
90 struct completion *wr_tx_compl; /* WR send CQE completion */
91 /* above four vectors have wr_tx_cnt elements and use the same index */
92 dma_addr_t wr_tx_dma_addr; /* DMA address of wr_tx_bufs */
93 atomic_long_t wr_tx_id; /* seq # of last sent WR */
94 unsigned long *wr_tx_mask; /* bit mask of used indexes */
95 u32 wr_tx_cnt; /* number of WR send buffers */
96 wait_queue_head_t wr_tx_wait; /* wait for free WR send buf */
97
98 struct smc_wr_buf *wr_rx_bufs; /* WR recv payload buffers */
99 struct ib_recv_wr *wr_rx_ibs; /* WR recv meta data */
100 struct ib_sge *wr_rx_sges; /* WR recv scatter meta data */
101 /* above three vectors have wr_rx_cnt elements and use the same index */
102 dma_addr_t wr_rx_dma_addr; /* DMA address of wr_rx_bufs */
103 u64 wr_rx_id; /* seq # of last recv WR */
104 u32 wr_rx_cnt; /* number of WR recv buffers */
105 unsigned long wr_rx_tstamp; /* jiffies when last buf rx */
106
107 struct ib_reg_wr wr_reg; /* WR register memory region */
108 wait_queue_head_t wr_reg_wait; /* wait for wr_reg result */
109 enum smc_wr_reg_state wr_reg_state; /* state of wr_reg request */
110
111 u8 gid[SMC_GID_SIZE];/* gid matching used vlan id*/
112 u8 sgid_index; /* gid index for vlan id */
113 u32 peer_qpn; /* QP number of peer */
114 enum ib_mtu path_mtu; /* used mtu */
115 enum ib_mtu peer_mtu; /* mtu size of peer */
116 u32 psn_initial; /* QP tx initial packet seqno */
117 u32 peer_psn; /* QP rx initial packet seqno */
118 u8 peer_mac[ETH_ALEN]; /* = gid[8:10||13:15] */
119 u8 peer_gid[SMC_GID_SIZE]; /* gid of peer*/
120 u8 link_id; /* unique # within link group */
121 u8 link_uid[SMC_LGR_ID_SIZE]; /* unique lnk id */
122 u8 peer_link_uid[SMC_LGR_ID_SIZE]; /* peer uid */
123 u8 link_idx; /* index in lgr link array */
124 u8 link_is_asym; /* is link asymmetric? */
125 struct smc_link_group *lgr; /* parent link group */
126 struct work_struct link_down_wrk; /* wrk to bring link down */
127
128 enum smc_link_state state; /* state of link */
129 struct delayed_work llc_testlink_wrk; /* testlink worker */
130 struct completion llc_testlink_resp; /* wait for rx of testlink */
131 int llc_testlink_time; /* testlink interval */
132};
133
134/* For now we just allow one parallel link per link group. The SMC protocol
135 * allows more (up to 8).
136 */
137#define SMC_LINKS_PER_LGR_MAX 3
138#define SMC_SINGLE_LINK 0
139
140/* tx/rx buffer list element for sndbufs list and rmbs list of a lgr */
141struct smc_buf_desc {
142 struct list_head list;
143 void *cpu_addr; /* virtual address of buffer */
144 struct page *pages;
145 int len; /* length of buffer */
146 u32 used; /* currently used / unused */
147 union {
148 struct { /* SMC-R */
149 struct sg_table sgt[SMC_LINKS_PER_LGR_MAX];
150 /* virtual buffer */
151 struct ib_mr *mr_rx[SMC_LINKS_PER_LGR_MAX];
152 /* for rmb only: memory region
153 * incl. rkey provided to peer
154 */
155 u32 order; /* allocation order */
156
157 u8 is_conf_rkey;
158 /* confirm_rkey done */
159 u8 is_reg_mr[SMC_LINKS_PER_LGR_MAX];
160 /* mem region registered */
161 u8 is_map_ib[SMC_LINKS_PER_LGR_MAX];
162 /* mem region mapped to lnk */
163 u8 is_reg_err;
164 /* buffer registration err */
165 };
166 struct { /* SMC-D */
167 unsigned short sba_idx;
168 /* SBA index number */
169 u64 token;
170 /* DMB token number */
171 dma_addr_t dma_addr;
172 /* DMA address */
173 };
174 };
175};
176
177struct smc_rtoken { /* address/key of remote RMB */
178 u64 dma_addr;
179 u32 rkey;
180};
181
182#define SMC_BUF_MIN_SIZE 16384 /* minimum size of an RMB */
183#define SMC_RMBE_SIZES 16 /* number of distinct RMBE sizes */
184/* theoretically, the RFC states that largest size would be 512K,
185 * i.e. compressed 5 and thus 6 sizes (0..5), despite
186 * struct smc_clc_msg_accept_confirm.rmbe_size being a 4 bit value (0..15)
187 */
188
189struct smcd_dev;
190
191enum smc_lgr_type { /* redundancy state of lgr */
192 SMC_LGR_NONE, /* no active links, lgr to be deleted */
193 SMC_LGR_SINGLE, /* 1 active RNIC on each peer */
194 SMC_LGR_SYMMETRIC, /* 2 active RNICs on each peer */
195 SMC_LGR_ASYMMETRIC_PEER, /* local has 2, peer 1 active RNICs */
196 SMC_LGR_ASYMMETRIC_LOCAL, /* local has 1, peer 2 active RNICs */
197};
198
199enum smc_llc_flowtype {
200 SMC_LLC_FLOW_NONE = 0,
201 SMC_LLC_FLOW_ADD_LINK = 2,
202 SMC_LLC_FLOW_DEL_LINK = 4,
203 SMC_LLC_FLOW_RKEY = 6,
204};
205
206struct smc_llc_qentry;
207
208struct smc_llc_flow {
209 enum smc_llc_flowtype type;
210 struct smc_llc_qentry *qentry;
211};
212
213struct smc_link_group {
214 struct list_head list;
215 struct rb_root conns_all; /* connection tree */
216 rwlock_t conns_lock; /* protects conns_all */
217 unsigned int conns_num; /* current # of connections */
218 unsigned short vlan_id; /* vlan id of link group */
219
220 struct list_head sndbufs[SMC_RMBE_SIZES];/* tx buffers */
221 struct mutex sndbufs_lock; /* protects tx buffers */
222 struct list_head rmbs[SMC_RMBE_SIZES]; /* rx buffers */
223 struct mutex rmbs_lock; /* protects rx buffers */
224
225 u8 id[SMC_LGR_ID_SIZE]; /* unique lgr id */
226 struct delayed_work free_work; /* delayed freeing of an lgr */
227 struct work_struct terminate_work; /* abnormal lgr termination */
228 struct workqueue_struct *tx_wq; /* wq for conn. tx workers */
229 u8 sync_err : 1; /* lgr no longer fits to peer */
230 u8 terminating : 1;/* lgr is terminating */
231 u8 freeing : 1; /* lgr is being freed */
232
233 bool is_smcd; /* SMC-R or SMC-D */
234 u8 smc_version;
235 u8 negotiated_eid[SMC_MAX_EID_LEN];
236 u8 peer_os; /* peer operating system */
237 u8 peer_smc_release;
238 u8 peer_hostname[SMC_MAX_HOSTNAME_LEN];
239 union {
240 struct { /* SMC-R */
241 enum smc_lgr_role role;
242 /* client or server */
243 struct smc_link lnk[SMC_LINKS_PER_LGR_MAX];
244 /* smc link */
245 char peer_systemid[SMC_SYSTEMID_LEN];
246 /* unique system_id of peer */
247 struct smc_rtoken rtokens[SMC_RMBS_PER_LGR_MAX]
248 [SMC_LINKS_PER_LGR_MAX];
249 /* remote addr/key pairs */
250 DECLARE_BITMAP(rtokens_used_mask, SMC_RMBS_PER_LGR_MAX);
251 /* used rtoken elements */
252 u8 next_link_id;
253 enum smc_lgr_type type;
254 /* redundancy state */
255 u8 pnet_id[SMC_MAX_PNETID_LEN + 1];
256 /* pnet id of this lgr */
257 struct list_head llc_event_q;
258 /* queue for llc events */
259 spinlock_t llc_event_q_lock;
260 /* protects llc_event_q */
261 struct mutex llc_conf_mutex;
262 /* protects lgr reconfig. */
263 struct work_struct llc_add_link_work;
264 struct work_struct llc_del_link_work;
265 struct work_struct llc_event_work;
266 /* llc event worker */
267 wait_queue_head_t llc_flow_waiter;
268 /* w4 next llc event */
269 wait_queue_head_t llc_msg_waiter;
270 /* w4 next llc msg */
271 struct smc_llc_flow llc_flow_lcl;
272 /* llc local control field */
273 struct smc_llc_flow llc_flow_rmt;
274 /* llc remote control field */
275 struct smc_llc_qentry *delayed_event;
276 /* arrived when flow active */
277 spinlock_t llc_flow_lock;
278 /* protects llc flow */
279 int llc_testlink_time;
280 /* link keep alive time */
281 u32 llc_termination_rsn;
282 /* rsn code for termination */
283 };
284 struct { /* SMC-D */
285 u64 peer_gid;
286 /* Peer GID (remote) */
287 struct smcd_dev *smcd;
288 /* ISM device for VLAN reg. */
289 u8 peer_shutdown : 1;
290 /* peer triggered shutdownn */
291 };
292 };
293};
294
295struct smc_clc_msg_local;
296
297struct smc_init_info {
298 u8 is_smcd;
299 u8 smc_type_v1;
300 u8 smc_type_v2;
301 u8 first_contact_peer;
302 u8 first_contact_local;
303 unsigned short vlan_id;
304 /* SMC-R */
305 struct smc_clc_msg_local *ib_lcl;
306 struct smc_ib_device *ib_dev;
307 u8 ib_gid[SMC_GID_SIZE];
308 u8 ib_port;
309 u32 ib_clcqpn;
310 /* SMC-D */
311 u64 ism_peer_gid[SMC_MAX_ISM_DEVS + 1];
312 struct smcd_dev *ism_dev[SMC_MAX_ISM_DEVS + 1];
313 u16 ism_chid[SMC_MAX_ISM_DEVS + 1];
314 u8 ism_offered_cnt; /* # of ISM devices offered */
315 u8 ism_selected; /* index of selected ISM dev*/
316 u8 smcd_version;
317};
318
319/* Find the connection associated with the given alert token in the link group.
320 * To use rbtrees we have to implement our own search core.
321 * Requires @conns_lock
322 * @token alert token to search for
323 * @lgr link group to search in
324 * Returns connection associated with token if found, NULL otherwise.
325 */
326static inline struct smc_connection *smc_lgr_find_conn(
327 u32 token, struct smc_link_group *lgr)
328{
329 struct smc_connection *res = NULL;
330 struct rb_node *node;
331
332 node = lgr->conns_all.rb_node;
333 while (node) {
334 struct smc_connection *cur = rb_entry(node,
335 struct smc_connection, alert_node);
336
337 if (cur->alert_token_local > token) {
338 node = node->rb_left;
339 } else {
340 if (cur->alert_token_local < token) {
341 node = node->rb_right;
342 } else {
343 res = cur;
344 break;
345 }
346 }
347 }
348
349 return res;
350}
351
352/* returns true if the specified link is usable */
353static inline bool smc_link_usable(struct smc_link *lnk)
354{
355 if (lnk->state == SMC_LNK_UNUSED || lnk->state == SMC_LNK_INACTIVE)
356 return false;
357 return true;
358}
359
360static inline bool smc_link_active(struct smc_link *lnk)
361{
362 return lnk->state == SMC_LNK_ACTIVE;
363}
364
365struct smc_sock;
366struct smc_clc_msg_accept_confirm;
367struct smc_clc_msg_local;
368
369void smc_lgr_cleanup_early(struct smc_connection *conn);
370void smc_lgr_terminate_sched(struct smc_link_group *lgr);
371void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport);
372void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport);
373void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid,
374 unsigned short vlan);
375void smc_smcd_terminate_all(struct smcd_dev *dev);
376void smc_smcr_terminate_all(struct smc_ib_device *smcibdev);
377int smc_buf_create(struct smc_sock *smc, bool is_smcd);
378int smc_uncompress_bufsize(u8 compressed);
379int smc_rmb_rtoken_handling(struct smc_connection *conn, struct smc_link *link,
380 struct smc_clc_msg_accept_confirm *clc);
381int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey);
382int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey);
383void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new,
384 __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey);
385void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id,
386 __be64 nw_vaddr, __be32 nw_rkey);
387void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn);
388void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn);
389void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn);
390void smc_rmb_sync_sg_for_device(struct smc_connection *conn);
391int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini);
392
393void smc_conn_free(struct smc_connection *conn);
394int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini);
395void smc_lgr_schedule_free_work_fast(struct smc_link_group *lgr);
396int smc_core_init(void);
397void smc_core_exit(void);
398
399int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
400 u8 link_idx, struct smc_init_info *ini);
401void smcr_link_clear(struct smc_link *lnk, bool log);
402int smcr_buf_map_lgr(struct smc_link *lnk);
403int smcr_buf_reg_lgr(struct smc_link *lnk);
404void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type);
405void smcr_lgr_set_type_asym(struct smc_link_group *lgr,
406 enum smc_lgr_type new_type, int asym_lnk_idx);
407int smcr_link_reg_rmb(struct smc_link *link, struct smc_buf_desc *rmb_desc);
408struct smc_link *smc_switch_conns(struct smc_link_group *lgr,
409 struct smc_link *from_lnk, bool is_dev_err);
410void smcr_link_down_cond(struct smc_link *lnk);
411void smcr_link_down_cond_sched(struct smc_link *lnk);
412
413static inline struct smc_link_group *smc_get_lgr(struct smc_link *link)
414{
415 return link->lgr;
416}
417#endif