xref: /linux/net/smc/smc_core.h (revision 132db93572821ec2fdf81e354cc40f558faf7e4f)
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 #define SMC_LGR_ID_SIZE		4
74 
75 struct 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 #define SMC_FIRST_CONTACT	1		/* first contact to a peer */
141 #define SMC_REUSE_CONTACT	0		/* follow-on contact to a peer*/
142 
143 /* tx/rx buffer list element for sndbufs list and rmbs list of a lgr */
144 struct smc_buf_desc {
145 	struct list_head	list;
146 	void			*cpu_addr;	/* virtual address of buffer */
147 	struct page		*pages;
148 	int			len;		/* length of buffer */
149 	u32			used;		/* currently used / unused */
150 	union {
151 		struct { /* SMC-R */
152 			struct sg_table	sgt[SMC_LINKS_PER_LGR_MAX];
153 					/* virtual buffer */
154 			struct ib_mr	*mr_rx[SMC_LINKS_PER_LGR_MAX];
155 					/* for rmb only: memory region
156 					 * incl. rkey provided to peer
157 					 */
158 			u32		order;	/* allocation order */
159 
160 			u8		is_conf_rkey;
161 					/* confirm_rkey done */
162 			u8		is_reg_mr[SMC_LINKS_PER_LGR_MAX];
163 					/* mem region registered */
164 			u8		is_map_ib[SMC_LINKS_PER_LGR_MAX];
165 					/* mem region mapped to lnk */
166 			u8		is_reg_err;
167 					/* buffer registration err */
168 		};
169 		struct { /* SMC-D */
170 			unsigned short	sba_idx;
171 					/* SBA index number */
172 			u64		token;
173 					/* DMB token number */
174 			dma_addr_t	dma_addr;
175 					/* DMA address */
176 		};
177 	};
178 };
179 
180 struct smc_rtoken {				/* address/key of remote RMB */
181 	u64			dma_addr;
182 	u32			rkey;
183 };
184 
185 #define SMC_BUF_MIN_SIZE	16384	/* minimum size of an RMB */
186 #define SMC_RMBE_SIZES		16	/* number of distinct RMBE sizes */
187 /* theoretically, the RFC states that largest size would be 512K,
188  * i.e. compressed 5 and thus 6 sizes (0..5), despite
189  * struct smc_clc_msg_accept_confirm.rmbe_size being a 4 bit value (0..15)
190  */
191 
192 struct smcd_dev;
193 
194 enum smc_lgr_type {				/* redundancy state of lgr */
195 	SMC_LGR_NONE,			/* no active links, lgr to be deleted */
196 	SMC_LGR_SINGLE,			/* 1 active RNIC on each peer */
197 	SMC_LGR_SYMMETRIC,		/* 2 active RNICs on each peer */
198 	SMC_LGR_ASYMMETRIC_PEER,	/* local has 2, peer 1 active RNICs */
199 	SMC_LGR_ASYMMETRIC_LOCAL,	/* local has 1, peer 2 active RNICs */
200 };
201 
202 enum smc_llc_flowtype {
203 	SMC_LLC_FLOW_NONE	= 0,
204 	SMC_LLC_FLOW_ADD_LINK	= 2,
205 	SMC_LLC_FLOW_DEL_LINK	= 4,
206 	SMC_LLC_FLOW_RKEY	= 6,
207 };
208 
209 struct smc_llc_qentry;
210 
211 struct smc_llc_flow {
212 	enum smc_llc_flowtype type;
213 	struct smc_llc_qentry *qentry;
214 };
215 
216 struct smc_link_group {
217 	struct list_head	list;
218 	struct rb_root		conns_all;	/* connection tree */
219 	rwlock_t		conns_lock;	/* protects conns_all */
220 	unsigned int		conns_num;	/* current # of connections */
221 	unsigned short		vlan_id;	/* vlan id of link group */
222 
223 	struct list_head	sndbufs[SMC_RMBE_SIZES];/* tx buffers */
224 	struct mutex		sndbufs_lock;	/* protects tx buffers */
225 	struct list_head	rmbs[SMC_RMBE_SIZES];	/* rx buffers */
226 	struct mutex		rmbs_lock;	/* protects rx buffers */
227 
228 	u8			id[SMC_LGR_ID_SIZE];	/* unique lgr id */
229 	struct delayed_work	free_work;	/* delayed freeing of an lgr */
230 	struct work_struct	terminate_work;	/* abnormal lgr termination */
231 	u8			sync_err : 1;	/* lgr no longer fits to peer */
232 	u8			terminating : 1;/* lgr is terminating */
233 	u8			freefast : 1;	/* free worker scheduled fast */
234 	u8			freeing : 1;	/* lgr is being freed */
235 
236 	bool			is_smcd;	/* SMC-R or SMC-D */
237 	union {
238 		struct { /* SMC-R */
239 			enum smc_lgr_role	role;
240 						/* client or server */
241 			struct smc_link		lnk[SMC_LINKS_PER_LGR_MAX];
242 						/* smc link */
243 			char			peer_systemid[SMC_SYSTEMID_LEN];
244 						/* unique system_id of peer */
245 			struct smc_rtoken	rtokens[SMC_RMBS_PER_LGR_MAX]
246 						[SMC_LINKS_PER_LGR_MAX];
247 						/* remote addr/key pairs */
248 			DECLARE_BITMAP(rtokens_used_mask, SMC_RMBS_PER_LGR_MAX);
249 						/* used rtoken elements */
250 			u8			next_link_id;
251 			enum smc_lgr_type	type;
252 						/* redundancy state */
253 			u8			pnet_id[SMC_MAX_PNETID_LEN + 1];
254 						/* pnet id of this lgr */
255 			struct list_head	llc_event_q;
256 						/* queue for llc events */
257 			spinlock_t		llc_event_q_lock;
258 						/* protects llc_event_q */
259 			struct mutex		llc_conf_mutex;
260 						/* protects lgr reconfig. */
261 			struct work_struct	llc_add_link_work;
262 			struct work_struct	llc_del_link_work;
263 			struct work_struct	llc_event_work;
264 						/* llc event worker */
265 			wait_queue_head_t	llc_waiter;
266 						/* w4 next llc event */
267 			struct smc_llc_flow	llc_flow_lcl;
268 						/* llc local control field */
269 			struct smc_llc_flow	llc_flow_rmt;
270 						/* llc remote control field */
271 			struct smc_llc_qentry	*delayed_event;
272 						/* arrived when flow active */
273 			spinlock_t		llc_flow_lock;
274 						/* protects llc flow */
275 			int			llc_testlink_time;
276 						/* link keep alive time */
277 			u32			llc_termination_rsn;
278 						/* rsn code for termination */
279 		};
280 		struct { /* SMC-D */
281 			u64			peer_gid;
282 						/* Peer GID (remote) */
283 			struct smcd_dev		*smcd;
284 						/* ISM device for VLAN reg. */
285 			u8			peer_shutdown : 1;
286 						/* peer triggered shutdownn */
287 		};
288 	};
289 };
290 
291 struct smc_clc_msg_local;
292 
293 struct smc_init_info {
294 	u8			is_smcd;
295 	unsigned short		vlan_id;
296 	int			srv_first_contact;
297 	int			cln_first_contact;
298 	/* SMC-R */
299 	struct smc_clc_msg_local *ib_lcl;
300 	struct smc_ib_device	*ib_dev;
301 	u8			ib_gid[SMC_GID_SIZE];
302 	u8			ib_port;
303 	u32			ib_clcqpn;
304 	/* SMC-D */
305 	u64			ism_gid;
306 	struct smcd_dev		*ism_dev;
307 };
308 
309 /* Find the connection associated with the given alert token in the link group.
310  * To use rbtrees we have to implement our own search core.
311  * Requires @conns_lock
312  * @token	alert token to search for
313  * @lgr		 link group to search in
314  * Returns connection associated with token if found, NULL otherwise.
315  */
316 static inline struct smc_connection *smc_lgr_find_conn(
317 	u32 token, struct smc_link_group *lgr)
318 {
319 	struct smc_connection *res = NULL;
320 	struct rb_node *node;
321 
322 	node = lgr->conns_all.rb_node;
323 	while (node) {
324 		struct smc_connection *cur = rb_entry(node,
325 					struct smc_connection, alert_node);
326 
327 		if (cur->alert_token_local > token) {
328 			node = node->rb_left;
329 		} else {
330 			if (cur->alert_token_local < token) {
331 				node = node->rb_right;
332 			} else {
333 				res = cur;
334 				break;
335 			}
336 		}
337 	}
338 
339 	return res;
340 }
341 
342 /* returns true if the specified link is usable */
343 static inline bool smc_link_usable(struct smc_link *lnk)
344 {
345 	if (lnk->state == SMC_LNK_UNUSED || lnk->state == SMC_LNK_INACTIVE)
346 		return false;
347 	return true;
348 }
349 
350 struct smc_sock;
351 struct smc_clc_msg_accept_confirm;
352 struct smc_clc_msg_local;
353 
354 void smc_lgr_cleanup_early(struct smc_connection *conn);
355 void smc_lgr_terminate_sched(struct smc_link_group *lgr);
356 void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport);
357 void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport);
358 void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid,
359 			unsigned short vlan);
360 void smc_smcd_terminate_all(struct smcd_dev *dev);
361 void smc_smcr_terminate_all(struct smc_ib_device *smcibdev);
362 int smc_buf_create(struct smc_sock *smc, bool is_smcd);
363 int smc_uncompress_bufsize(u8 compressed);
364 int smc_rmb_rtoken_handling(struct smc_connection *conn, struct smc_link *link,
365 			    struct smc_clc_msg_accept_confirm *clc);
366 int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey);
367 int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey);
368 void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new,
369 		    __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey);
370 void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id,
371 		     __be64 nw_vaddr, __be32 nw_rkey);
372 void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn);
373 void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn);
374 void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn);
375 void smc_rmb_sync_sg_for_device(struct smc_connection *conn);
376 int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini);
377 
378 void smc_conn_free(struct smc_connection *conn);
379 int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini);
380 void smc_lgr_schedule_free_work_fast(struct smc_link_group *lgr);
381 int smc_core_init(void);
382 void smc_core_exit(void);
383 
384 int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
385 		   u8 link_idx, struct smc_init_info *ini);
386 void smcr_link_clear(struct smc_link *lnk, bool log);
387 int smcr_buf_map_lgr(struct smc_link *lnk);
388 int smcr_buf_reg_lgr(struct smc_link *lnk);
389 void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type);
390 void smcr_lgr_set_type_asym(struct smc_link_group *lgr,
391 			    enum smc_lgr_type new_type, int asym_lnk_idx);
392 int smcr_link_reg_rmb(struct smc_link *link, struct smc_buf_desc *rmb_desc);
393 struct smc_link *smc_switch_conns(struct smc_link_group *lgr,
394 				  struct smc_link *from_lnk, bool is_dev_err);
395 void smcr_link_down_cond(struct smc_link *lnk);
396 void smcr_link_down_cond_sched(struct smc_link *lnk);
397 
398 static inline struct smc_link_group *smc_get_lgr(struct smc_link *link)
399 {
400 	return link->lgr;
401 }
402 #endif
403