xref: /kernel/linux/linux-5.10/net/smc/smc_core.h (revision 8c2ecf20)
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	atomic_t		wr_tx_refcnt;	/* tx refs to link */
98
99	struct smc_wr_buf	*wr_rx_bufs;	/* WR recv payload buffers */
100	struct ib_recv_wr	*wr_rx_ibs;	/* WR recv meta data */
101	struct ib_sge		*wr_rx_sges;	/* WR recv scatter meta data */
102	/* above three vectors have wr_rx_cnt elements and use the same index */
103	dma_addr_t		wr_rx_dma_addr;	/* DMA address of wr_rx_bufs */
104	u64			wr_rx_id;	/* seq # of last recv WR */
105	u32			wr_rx_cnt;	/* number of WR recv buffers */
106	unsigned long		wr_rx_tstamp;	/* jiffies when last buf rx */
107
108	struct ib_reg_wr	wr_reg;		/* WR register memory region */
109	wait_queue_head_t	wr_reg_wait;	/* wait for wr_reg result */
110	atomic_t		wr_reg_refcnt;	/* reg refs to link */
111	enum smc_wr_reg_state	wr_reg_state;	/* state of wr_reg request */
112
113	u8			gid[SMC_GID_SIZE];/* gid matching used vlan id*/
114	u8			sgid_index;	/* gid index for vlan id      */
115	u32			peer_qpn;	/* QP number of peer */
116	enum ib_mtu		path_mtu;	/* used mtu */
117	enum ib_mtu		peer_mtu;	/* mtu size of peer */
118	u32			psn_initial;	/* QP tx initial packet seqno */
119	u32			peer_psn;	/* QP rx initial packet seqno */
120	u8			peer_mac[ETH_ALEN];	/* = gid[8:10||13:15] */
121	u8			peer_gid[SMC_GID_SIZE];	/* gid of peer*/
122	u8			link_id;	/* unique # within link group */
123	u8			link_uid[SMC_LGR_ID_SIZE]; /* unique lnk id */
124	u8			peer_link_uid[SMC_LGR_ID_SIZE]; /* peer uid */
125	u8			link_idx;	/* index in lgr link array */
126	u8			link_is_asym;	/* is link asymmetric? */
127	struct smc_link_group	*lgr;		/* parent link group */
128	struct work_struct	link_down_wrk;	/* wrk to bring link down */
129
130	enum smc_link_state	state;		/* state of link */
131	struct delayed_work	llc_testlink_wrk; /* testlink worker */
132	struct completion	llc_testlink_resp; /* wait for rx of testlink */
133	int			llc_testlink_time; /* testlink interval */
134};
135
136/* For now we just allow one parallel link per link group. The SMC protocol
137 * allows more (up to 8).
138 */
139#define SMC_LINKS_PER_LGR_MAX	3
140#define SMC_SINGLE_LINK		0
141
142/* tx/rx buffer list element for sndbufs list and rmbs list of a lgr */
143struct smc_buf_desc {
144	struct list_head	list;
145	void			*cpu_addr;	/* virtual address of buffer */
146	struct page		*pages;
147	int			len;		/* length of buffer */
148	u32			used;		/* currently used / unused */
149	union {
150		struct { /* SMC-R */
151			struct sg_table	sgt[SMC_LINKS_PER_LGR_MAX];
152					/* virtual buffer */
153			struct ib_mr	*mr_rx[SMC_LINKS_PER_LGR_MAX];
154					/* for rmb only: memory region
155					 * incl. rkey provided to peer
156					 */
157			u32		order;	/* allocation order */
158
159			u8		is_conf_rkey;
160					/* confirm_rkey done */
161			u8		is_reg_mr[SMC_LINKS_PER_LGR_MAX];
162					/* mem region registered */
163			u8		is_map_ib[SMC_LINKS_PER_LGR_MAX];
164					/* mem region mapped to lnk */
165			u8		is_reg_err;
166					/* buffer registration err */
167		};
168		struct { /* SMC-D */
169			unsigned short	sba_idx;
170					/* SBA index number */
171			u64		token;
172					/* DMB token number */
173			dma_addr_t	dma_addr;
174					/* DMA address */
175		};
176	};
177};
178
179struct smc_rtoken {				/* address/key of remote RMB */
180	u64			dma_addr;
181	u32			rkey;
182};
183
184#define SMC_BUF_MIN_SIZE	16384	/* minimum size of an RMB */
185#define SMC_RMBE_SIZES		16	/* number of distinct RMBE sizes */
186/* theoretically, the RFC states that largest size would be 512K,
187 * i.e. compressed 5 and thus 6 sizes (0..5), despite
188 * struct smc_clc_msg_accept_confirm.rmbe_size being a 4 bit value (0..15)
189 */
190
191struct smcd_dev;
192
193enum smc_lgr_type {				/* redundancy state of lgr */
194	SMC_LGR_NONE,			/* no active links, lgr to be deleted */
195	SMC_LGR_SINGLE,			/* 1 active RNIC on each peer */
196	SMC_LGR_SYMMETRIC,		/* 2 active RNICs on each peer */
197	SMC_LGR_ASYMMETRIC_PEER,	/* local has 2, peer 1 active RNICs */
198	SMC_LGR_ASYMMETRIC_LOCAL,	/* local has 1, peer 2 active RNICs */
199};
200
201enum smc_llc_flowtype {
202	SMC_LLC_FLOW_NONE	= 0,
203	SMC_LLC_FLOW_ADD_LINK	= 2,
204	SMC_LLC_FLOW_DEL_LINK	= 4,
205	SMC_LLC_FLOW_RKEY	= 6,
206};
207
208struct smc_llc_qentry;
209
210struct smc_llc_flow {
211	enum smc_llc_flowtype type;
212	struct smc_llc_qentry *qentry;
213};
214
215struct smc_link_group {
216	struct list_head	list;
217	struct rb_root		conns_all;	/* connection tree */
218	rwlock_t		conns_lock;	/* protects conns_all */
219	unsigned int		conns_num;	/* current # of connections */
220	unsigned short		vlan_id;	/* vlan id of link group */
221
222	struct list_head	sndbufs[SMC_RMBE_SIZES];/* tx buffers */
223	struct mutex		sndbufs_lock;	/* protects tx buffers */
224	struct list_head	rmbs[SMC_RMBE_SIZES];	/* rx buffers */
225	struct mutex		rmbs_lock;	/* protects rx buffers */
226
227	u8			id[SMC_LGR_ID_SIZE];	/* unique lgr id */
228	struct delayed_work	free_work;	/* delayed freeing of an lgr */
229	struct work_struct	terminate_work;	/* abnormal lgr termination */
230	struct workqueue_struct	*tx_wq;		/* wq for conn. tx workers */
231	u8			sync_err : 1;	/* lgr no longer fits to peer */
232	u8			terminating : 1;/* lgr is terminating */
233	u8			freeing : 1;	/* lgr is being freed */
234
235	bool			is_smcd;	/* SMC-R or SMC-D */
236	u8			smc_version;
237	u8			negotiated_eid[SMC_MAX_EID_LEN];
238	u8			peer_os;	/* peer operating system */
239	u8			peer_smc_release;
240	u8			peer_hostname[SMC_MAX_HOSTNAME_LEN];
241	union {
242		struct { /* SMC-R */
243			enum smc_lgr_role	role;
244						/* client or server */
245			struct smc_link		lnk[SMC_LINKS_PER_LGR_MAX];
246						/* smc link */
247			char			peer_systemid[SMC_SYSTEMID_LEN];
248						/* unique system_id of peer */
249			struct smc_rtoken	rtokens[SMC_RMBS_PER_LGR_MAX]
250						[SMC_LINKS_PER_LGR_MAX];
251						/* remote addr/key pairs */
252			DECLARE_BITMAP(rtokens_used_mask, SMC_RMBS_PER_LGR_MAX);
253						/* used rtoken elements */
254			u8			next_link_id;
255			enum smc_lgr_type	type;
256						/* redundancy state */
257			u8			pnet_id[SMC_MAX_PNETID_LEN + 1];
258						/* pnet id of this lgr */
259			struct list_head	llc_event_q;
260						/* queue for llc events */
261			spinlock_t		llc_event_q_lock;
262						/* protects llc_event_q */
263			struct mutex		llc_conf_mutex;
264						/* protects lgr reconfig. */
265			struct work_struct	llc_add_link_work;
266			struct work_struct	llc_del_link_work;
267			struct work_struct	llc_event_work;
268						/* llc event worker */
269			wait_queue_head_t	llc_flow_waiter;
270						/* w4 next llc event */
271			wait_queue_head_t	llc_msg_waiter;
272						/* w4 next llc msg */
273			struct smc_llc_flow	llc_flow_lcl;
274						/* llc local control field */
275			struct smc_llc_flow	llc_flow_rmt;
276						/* llc remote control field */
277			struct smc_llc_qentry	*delayed_event;
278						/* arrived when flow active */
279			spinlock_t		llc_flow_lock;
280						/* protects llc flow */
281			int			llc_testlink_time;
282						/* link keep alive time */
283			u32			llc_termination_rsn;
284						/* rsn code for termination */
285		};
286		struct { /* SMC-D */
287			u64			peer_gid;
288						/* Peer GID (remote) */
289			struct smcd_dev		*smcd;
290						/* ISM device for VLAN reg. */
291			u8			peer_shutdown : 1;
292						/* peer triggered shutdownn */
293		};
294	};
295};
296
297struct smc_clc_msg_local;
298
299struct smc_init_info {
300	u8			is_smcd;
301	u8			smc_type_v1;
302	u8			smc_type_v2;
303	u8			first_contact_peer;
304	u8			first_contact_local;
305	unsigned short		vlan_id;
306	/* SMC-R */
307	struct smc_clc_msg_local *ib_lcl;
308	struct smc_ib_device	*ib_dev;
309	u8			ib_gid[SMC_GID_SIZE];
310	u8			ib_port;
311	u32			ib_clcqpn;
312	/* SMC-D */
313	u64			ism_peer_gid[SMC_MAX_ISM_DEVS + 1];
314	struct smcd_dev		*ism_dev[SMC_MAX_ISM_DEVS + 1];
315	u16			ism_chid[SMC_MAX_ISM_DEVS + 1];
316	u8			ism_offered_cnt; /* # of ISM devices offered */
317	u8			ism_selected;    /* index of selected ISM dev*/
318	u8			smcd_version;
319};
320
321/* Find the connection associated with the given alert token in the link group.
322 * To use rbtrees we have to implement our own search core.
323 * Requires @conns_lock
324 * @token	alert token to search for
325 * @lgr		 link group to search in
326 * Returns connection associated with token if found, NULL otherwise.
327 */
328static inline struct smc_connection *smc_lgr_find_conn(
329	u32 token, struct smc_link_group *lgr)
330{
331	struct smc_connection *res = NULL;
332	struct rb_node *node;
333
334	node = lgr->conns_all.rb_node;
335	while (node) {
336		struct smc_connection *cur = rb_entry(node,
337					struct smc_connection, alert_node);
338
339		if (cur->alert_token_local > token) {
340			node = node->rb_left;
341		} else {
342			if (cur->alert_token_local < token) {
343				node = node->rb_right;
344			} else {
345				res = cur;
346				break;
347			}
348		}
349	}
350
351	return res;
352}
353
354/* returns true if the specified link is usable */
355static inline bool smc_link_usable(struct smc_link *lnk)
356{
357	if (lnk->state == SMC_LNK_UNUSED || lnk->state == SMC_LNK_INACTIVE)
358		return false;
359	return true;
360}
361
362static inline bool smc_link_sendable(struct smc_link *lnk)
363{
364	return smc_link_usable(lnk) &&
365		lnk->qp_attr.cur_qp_state == IB_QPS_RTS;
366}
367
368static inline bool smc_link_active(struct smc_link *lnk)
369{
370	return lnk->state == SMC_LNK_ACTIVE;
371}
372
373struct smc_sock;
374struct smc_clc_msg_accept_confirm;
375struct smc_clc_msg_local;
376
377void smc_lgr_cleanup_early(struct smc_connection *conn);
378void smc_lgr_terminate_sched(struct smc_link_group *lgr);
379void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport);
380void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport);
381void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid,
382			unsigned short vlan);
383void smc_smcd_terminate_all(struct smcd_dev *dev);
384void smc_smcr_terminate_all(struct smc_ib_device *smcibdev);
385int smc_buf_create(struct smc_sock *smc, bool is_smcd);
386int smc_uncompress_bufsize(u8 compressed);
387int smc_rmb_rtoken_handling(struct smc_connection *conn, struct smc_link *link,
388			    struct smc_clc_msg_accept_confirm *clc);
389int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey);
390int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey);
391void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new,
392		    __be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey);
393void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id,
394		     __be64 nw_vaddr, __be32 nw_rkey);
395void smc_sndbuf_sync_sg_for_cpu(struct smc_connection *conn);
396void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn);
397void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn);
398void smc_rmb_sync_sg_for_device(struct smc_connection *conn);
399int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini);
400
401void smc_conn_free(struct smc_connection *conn);
402int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini);
403void smc_lgr_schedule_free_work_fast(struct smc_link_group *lgr);
404int smc_core_init(void);
405void smc_core_exit(void);
406
407int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
408		   u8 link_idx, struct smc_init_info *ini);
409void smcr_link_clear(struct smc_link *lnk, bool log);
410int smcr_buf_map_lgr(struct smc_link *lnk);
411int smcr_buf_reg_lgr(struct smc_link *lnk);
412void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type);
413void smcr_lgr_set_type_asym(struct smc_link_group *lgr,
414			    enum smc_lgr_type new_type, int asym_lnk_idx);
415int smcr_link_reg_rmb(struct smc_link *link, struct smc_buf_desc *rmb_desc);
416struct smc_link *smc_switch_conns(struct smc_link_group *lgr,
417				  struct smc_link *from_lnk, bool is_dev_err);
418void smcr_link_down_cond(struct smc_link *lnk);
419void smcr_link_down_cond_sched(struct smc_link *lnk);
420
421static inline struct smc_link_group *smc_get_lgr(struct smc_link *link)
422{
423	return link->lgr;
424}
425#endif
426