xref: /illumos-gate/usr/src/uts/common/sys/ib/clients/rdsv3/ib.h (revision 4f364e7c95ee7fd9d5bbeddc1940e92405bb0e72)
1 /*
2  * This file contains definitions imported from the OFED rds header ib.h.
3  * Oracle elects to have and use the contents of ib.h under and
4  * governed by the OpenIB.org BSD license.
5  */
6 /*
7  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
8  */
9 
10 #ifndef _RDSV3_IB_H
11 #define	_RDSV3_IB_H
12 
13 #include <sys/rds.h>
14 #include <sys/ib/clients/rdsv3/rdsv3.h>
15 #include <sys/ib/clients/rdsv3/rdma_transport.h>
16 #include <sys/ib/clients/rdsv3/rdsv3_af_thr.h>
17 
18 #define	RDSV3_FMR_SIZE			256
19 #define	RDSV3_FMR_POOL_SIZE		(12 * 1024)
20 
21 #define	RDSV3_IB_MAX_SGE		8
22 #define	RDSV3_IB_RECV_SGE 		2
23 
24 #define	RDSV3_IB_DEFAULT_RECV_WR	1024
25 #define	RDSV3_IB_DEFAULT_SEND_WR	256
26 
27 #define	RDSV3_IB_DEFAULT_RETRY_COUNT	2
28 
29 /* minor versions supported */
30 #define	RDSV3_IB_SUPPORTED_PROTOCOLS	0x00000003
31 
32 #define	RDSV3_IB_MAX_RECV_ALLOC		((512 * 1024 * 1024) / RDSV3_FRAG_SIZE)
33 #define	RDSV3_IB_WC_POLL_SIZE		16
34 
35 extern struct list rdsv3_ib_devices;
36 
37 /*
38  * IB posts RDSV3_FRAG_SIZE fragments of pages to the receive queues to
39  * try and minimize the amount of memory tied up both the device and
40  * socket receive queues.
41  */
42 /* page offset of the final full frag that fits in the page */
43 #define	RDSV3_PAGE_LAST_OFF	\
44 	(((PAGE_SIZE  / RDSV3_FRAG_SIZE) - 1) * RDSV3_FRAG_SIZE)
45 struct rdsv3_page_frag {
46 	struct list_node	f_item;
47 	caddr_t			f_page;
48 	unsigned long		f_offset;
49 	ibt_wr_ds_t		f_sge;
50 	ibt_mi_hdl_t 		f_mapped;
51 };
52 
53 struct rdsv3_ib_incoming {
54 	list_node_t		ii_obj;	/* list obj of rdsv3_inc_pool list */
55 	struct list		ii_frags;
56 	struct rdsv3_incoming	ii_inc;
57 	struct rdsv3_inc_pool	*ii_pool;
58 	struct rdsv3_ib_device	*ii_ibdev;
59 };
60 
61 struct rdsv3_ib_connect_private {
62 	/* Add new fields at the end, and don't permute existing fields. */
63 	uint32_be_t		dp_saddr;
64 	uint32_be_t		dp_daddr;
65 	uint8_t			dp_protocol_major;
66 	uint8_t			dp_protocol_minor;
67 	uint16_be_t		dp_protocol_minor_mask; /* bitmask */
68 	uint32_be_t		dp_reserved1;
69 	uint32_be_t		dp_ack_seq;
70 	uint32_be_t		dp_credit;	/* non-zero enables flow ctl */
71 };
72 
73 struct rdsv3_ib_send_work {
74 	struct rdsv3_message	*s_rm;
75 	struct rdsv3_rdma_op	*s_op;
76 	ibt_wrc_opcode_t	s_opcode;
77 	unsigned long		s_queued;
78 };
79 
80 struct rdsv3_ib_recv_work {
81 	struct rdsv3_ib_incoming 	*r_ibinc;
82 	struct rdsv3_page_frag		*r_frag;
83 	ibt_wr_ds_t			r_sge[2];
84 };
85 
86 struct rdsv3_ib_work_ring {
87 	uint32_t		w_nr;
88 	uint32_t		w_alloc_ptr;
89 	uint32_t		w_alloc_ctr;
90 	uint32_t		w_free_ptr;
91 	atomic_t		w_free_ctr;
92 	rdsv3_wait_queue_t	w_empty_wait;
93 };
94 
95 /*
96  * Rings are posted with all the allocations they'll need to queue the
97  * incoming message to the receiving socket so this can't fail.
98  * All fragments start with a header, so we can make sure we're not receiving
99  * garbage, and we can tell a small 8 byte fragment from an ACK frame.
100  */
101 struct rdsv3_ib_ack_state {
102 	uint64_t	ack_next;
103 	uint64_t	ack_recv;
104 	unsigned int	ack_required:1;
105 	unsigned int	ack_next_valid:1;
106 	unsigned int	ack_recv_valid:1;
107 };
108 
109 struct rdsv3_ib_device;
110 
111 struct rdsv3_ib_connection {
112 
113 	struct list_node	ib_node;
114 	boolean_t		i_on_dev_list;
115 	struct rdsv3_ib_device	*rds_ibdev;
116 	struct rdsv3_connection	*conn;
117 
118 	/* alphabet soup, IBTA style */
119 	struct rdma_cm_id	*i_cm_id;
120 	struct ib_pd		*i_pd;
121 	struct rdsv3_hdrs_mr	*i_mr;
122 	struct ib_cq		*i_cq;
123 	struct ib_cq		*i_snd_cq;
124 
125 	/* tx */
126 	struct rdsv3_ib_work_ring	i_send_ring;
127 	struct rdsv3_message	*i_rm;
128 	struct rdsv3_header	*i_send_hdrs;
129 	uint64_t		i_send_hdrs_dma;
130 	struct rdsv3_ib_send_work *i_sends;
131 	ibt_send_wr_t		*i_send_wrs;
132 
133 	/* soft CQ */
134 	rdsv3_af_thr_t		*i_soft_cq;
135 	rdsv3_af_thr_t		*i_snd_soft_cq;
136 	rdsv3_af_thr_t		*i_refill_rq;
137 
138 	/* rx */
139 	struct mutex		i_recv_mutex;
140 	struct rdsv3_ib_work_ring	i_recv_ring;
141 	struct rdsv3_ib_incoming	*i_ibinc;
142 	uint32_t		i_recv_data_rem;
143 	struct rdsv3_header	*i_recv_hdrs;
144 	uint64_t		i_recv_hdrs_dma;
145 	struct rdsv3_ib_recv_work *i_recvs;
146 	ibt_recv_wr_t		*i_recv_wrs;
147 	struct rdsv3_page_frag	i_frag;
148 	uint64_t		i_ack_recv;	/* last ACK received */
149 
150 	/* sending acks */
151 	unsigned long		i_ack_flags;
152 #ifdef KERNEL_HAS_ATOMIC64
153 	atomic64_t		i_ack_next;	/* next ACK to send */
154 #else
155 	kmutex_t		i_ack_lock;	/* protect i_ack_next */
156 	uint64_t		i_ack_next;	/* next ACK to send */
157 #endif
158 	struct rdsv3_header	*i_ack;
159 	ibt_send_wr_t		i_ack_wr;
160 	ibt_wr_ds_t		i_ack_sge;
161 	uint64_t		i_ack_dma;
162 	unsigned long		i_ack_queued;
163 
164 	/*
165 	 * Flow control related information
166 	 *
167 	 * Our algorithm uses a pair variables that we need to access
168 	 * atomically - one for the send credits, and one posted
169 	 * recv credits we need to transfer to remote.
170 	 * Rather than protect them using a slow spinlock, we put both into
171 	 * a single atomic_t and update it using cmpxchg
172 	 */
173 	atomic_t		i_credits;
174 
175 	/* Protocol version specific information */
176 	unsigned int		i_flowctl:1;	/* enable/disable flow ctl */
177 
178 	/* Batched completions */
179 	unsigned int		i_unsignaled_wrs;
180 	long			i_unsignaled_bytes;
181 
182 	unsigned long		i_max_recv_alloc;
183 };
184 
185 /* This assumes that atomic_t is at least 32 bits */
186 #define	IB_GET_SEND_CREDITS(v)	((v) & 0xffff)
187 #define	IB_GET_POST_CREDITS(v)	((v) >> 16)
188 #define	IB_SET_SEND_CREDITS(v)	((v) & 0xffff)
189 #define	IB_SET_POST_CREDITS(v)	((v) << 16)
190 
191 struct rdsv3_ib_ipaddr {
192 	struct list_node	list;
193 	uint32_be_t		ipaddr;
194 };
195 
196 struct rdsv3_ib_device {
197 	struct list_node	list;
198 	struct list		ipaddr_list;
199 	struct list		conn_list;
200 	ib_device_t		*dev;
201 	struct ib_pd		*pd;
202 	struct kmem_cache	*ib_frag_slab;
203 	kmutex_t		spinlock;	/* protect the above */
204 	krwlock_t		rwlock;		/* protect paddr_list */
205 	unsigned int		fmr_max_remaps;
206 	unsigned int		max_fmrs;
207 	unsigned int		fmr_message_size;
208 	int			max_sge;
209 	unsigned int		max_wrs;
210 	unsigned int		max_initiator_depth;
211 	unsigned int		max_responder_resources;
212 	struct rdsv3_fmr_pool	*fmr_pool;
213 	struct rdsv3_inc_pool	*inc_pool;
214 	ibt_fmr_pool_hdl_t	fmr_pool_hdl;
215 	ibt_hca_attr_t		hca_attr;
216 	rdsv3_af_thr_t		*fmr_soft_cq;
217 	rdsv3_af_thr_t		*inc_soft_cq;
218 	ibt_hca_hdl_t		ibt_hca_hdl;
219 	rdsv3_af_grp_t		*aft_hcagp;
220 };
221 
222 /* bits for i_ack_flags */
223 #define	IB_ACK_IN_FLIGHT	0
224 #define	IB_ACK_REQUESTED	1
225 
226 #define	RDSV3_IB_SEND_OP	(1ULL << 63)
227 
228 /* Magic WR_ID for ACKs */
229 #define	RDSV3_IB_ACK_WR_ID	(~(uint64_t)0)
230 
231 struct rdsv3_ib_statistics {
232 	uint64_t	s_ib_connect_raced;
233 	uint64_t	s_ib_listen_closed_stale;
234 	uint64_t	s_ib_evt_handler_call;
235 	uint64_t	s_ib_tasklet_call;
236 	uint64_t	s_ib_tx_cq_event;
237 	uint64_t	s_ib_tx_ring_full;
238 	uint64_t	s_ib_tx_throttle;
239 	uint64_t	s_ib_tx_sg_mapping_failure;
240 	uint64_t	s_ib_tx_stalled;
241 	uint64_t	s_ib_tx_credit_updates;
242 	uint64_t	s_ib_rx_cq_event;
243 	uint64_t	s_ib_rx_ring_empty;
244 	uint64_t	s_ib_rx_refill_from_cq;
245 	uint64_t	s_ib_rx_refill_from_thread;
246 	uint64_t	s_ib_rx_alloc_limit;
247 	uint64_t	s_ib_rx_credit_updates;
248 	uint64_t	s_ib_ack_sent;
249 	uint64_t	s_ib_ack_send_failure;
250 	uint64_t	s_ib_ack_send_delayed;
251 	uint64_t	s_ib_ack_send_piggybacked;
252 	uint64_t	s_ib_ack_received;
253 	uint64_t	s_ib_rdma_mr_alloc;
254 	uint64_t	s_ib_rdma_mr_free;
255 	uint64_t	s_ib_rdma_mr_used;
256 	uint64_t	s_ib_rdma_mr_pool_flush;
257 	uint64_t	s_ib_rdma_mr_pool_wait;
258 	uint64_t	s_ib_rdma_mr_pool_depleted;
259 };
260 
261 extern struct rdsv3_workqueue_struct_s *rds_ib_wq;
262 
263 /* ib.c */
264 extern struct rdsv3_transport rdsv3_ib_transport;
265 extern void rdsv3_ib_add_one(ib_device_t *device);
266 extern void rdsv3_ib_remove_one(ib_device_t *device);
267 extern struct ib_client rdsv3_ib_client;
268 
269 extern unsigned int fmr_pool_size;
270 extern unsigned int fmr_message_size;
271 extern unsigned int rdsv3_ib_retry_count;
272 
273 extern kmutex_t ib_nodev_conns_lock;
274 extern struct list ib_nodev_conns;
275 
276 /* ib_cm.c */
277 int rdsv3_ib_conn_alloc(struct rdsv3_connection *conn, int gfp);
278 void rdsv3_ib_conn_free(void *arg);
279 int rdsv3_ib_conn_connect(struct rdsv3_connection *conn);
280 void rdsv3_ib_conn_shutdown(struct rdsv3_connection *conn);
281 void rdsv3_conn_drop(struct rdsv3_connection *conn);
282 int rdsv3_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
283     struct rdma_cm_event *event);
284 int rdsv3_ib_cm_initiate_connect(struct rdma_cm_id *cm_id);
285 void rdsv3_ib_cm_connect_complete(struct rdsv3_connection *conn,
286     struct rdma_cm_event *event);
287 void rdsv3_ib_tasklet_fn(void *data);
288 void rdsv3_ib_snd_tasklet_fn(void *data);
289 void rdsv3_ib_refill_fn(void *data);
290 
291 /* ib_rdma.c */
292 int rdsv3_ib_update_ipaddr(struct rdsv3_ib_device *rds_ibdev,
293     uint32_be_t ipaddr);
294 void rdsv3_ib_add_conn(struct rdsv3_ib_device *rds_ibdev,
295     struct rdsv3_connection *conn);
296 void rdsv3_ib_remove_conn(struct rdsv3_ib_device *rds_ibdev,
297     struct rdsv3_connection *conn);
298 void __rdsv3_ib_destroy_conns(struct list *list, kmutex_t *list_lock);
299 static inline void rdsv3_ib_destroy_nodev_conns(void)
300 {
301 	__rdsv3_ib_destroy_conns(&ib_nodev_conns, &ib_nodev_conns_lock);
302 }
303 static inline void rdsv3_ib_destroy_conns(struct rdsv3_ib_device *rds_ibdev)
304 {
305 	__rdsv3_ib_destroy_conns(&rds_ibdev->conn_list, &rds_ibdev->spinlock);
306 }
307 
308 int rdsv3_ib_create_mr_pool(struct rdsv3_ib_device *);
309 void rdsv3_ib_destroy_mr_pool(struct rdsv3_ib_device *);
310 void rdsv3_ib_get_mr_info(struct rdsv3_ib_device *rds_ibdev,
311 	struct rds_info_rdma_connection *iinfo);
312 void *rdsv3_ib_get_mr(struct rds_iovec *args, unsigned long nents,
313 	struct rdsv3_sock *rs, uint32_t *key_ret);
314 void rdsv3_ib_sync_mr(void *trans_private, int dir);
315 void rdsv3_ib_free_mr(void *trans_private, int invalidate);
316 void rdsv3_ib_flush_mrs(void);
317 void rdsv3_ib_drain_mrlist_fn(void *data);
318 
319 /* ib_recv.c */
320 int rdsv3_ib_recv_init(void);
321 void rdsv3_ib_recv_exit(void);
322 int rdsv3_ib_recv(struct rdsv3_connection *conn);
323 int rdsv3_ib_recv_refill(struct rdsv3_connection *conn, int prefill);
324 void rdsv3_ib_inc_free(struct rdsv3_incoming *inc);
325 int rdsv3_ib_inc_copy_to_user(struct rdsv3_incoming *inc, uio_t *uiop,
326     size_t size);
327 void rdsv3_ib_recv_cqe_handler(struct rdsv3_ib_connection *ic, ibt_wc_t *wc,
328     struct rdsv3_ib_ack_state *state);
329 void rdsv3_ib_recv_init_ring(struct rdsv3_ib_connection *ic);
330 void rdsv3_ib_recv_clear_ring(struct rdsv3_ib_connection *ic);
331 void rdsv3_ib_recv_init_ack(struct rdsv3_ib_connection *ic);
332 void rdsv3_ib_attempt_ack(struct rdsv3_ib_connection *ic);
333 void rdsv3_ib_ack_send_complete(struct rdsv3_ib_connection *ic);
334 uint64_t rdsv3_ib_piggyb_ack(struct rdsv3_ib_connection *ic);
335 void rdsv3_ib_set_ack(struct rdsv3_ib_connection *ic, uint64_t seq,
336     int ack_required);
337 int rdsv3_ib_create_inc_pool(struct rdsv3_ib_device *);
338 void rdsv3_ib_destroy_inc_pool(struct rdsv3_ib_device *);
339 void rdsv3_ib_drain_inclist(void *);
340 
341 /* ib_ring.c */
342 void rdsv3_ib_ring_init(struct rdsv3_ib_work_ring *ring, uint32_t nr);
343 void rdsv3_ib_ring_resize(struct rdsv3_ib_work_ring *ring, uint32_t nr);
344 uint32_t rdsv3_ib_ring_alloc(struct rdsv3_ib_work_ring *ring, uint32_t val,
345     uint32_t *pos);
346 void rdsv3_ib_ring_free(struct rdsv3_ib_work_ring *ring, uint32_t val);
347 void rdsv3_ib_ring_unalloc(struct rdsv3_ib_work_ring *ring, uint32_t val);
348 int rdsv3_ib_ring_empty(struct rdsv3_ib_work_ring *ring);
349 int rdsv3_ib_ring_low(struct rdsv3_ib_work_ring *ring);
350 uint32_t rdsv3_ib_ring_oldest(struct rdsv3_ib_work_ring *ring);
351 uint32_t rdsv3_ib_ring_completed(struct rdsv3_ib_work_ring *ring,
352     uint32_t wr_id, uint32_t oldest);
353 
354 /* ib_send.c */
355 void rdsv3_ib_xmit_complete(struct rdsv3_connection *conn);
356 int rdsv3_ib_xmit(struct rdsv3_connection *conn, struct rdsv3_message *rm,
357     unsigned int hdr_off, unsigned int sg, unsigned int off);
358 void rdsv3_ib_send_cqe_handler(struct rdsv3_ib_connection *ic, ibt_wc_t *wc);
359 void rdsv3_ib_send_init_ring(struct rdsv3_ib_connection *ic);
360 void rdsv3_ib_send_clear_ring(struct rdsv3_ib_connection *ic);
361 int rdsv3_ib_xmit_rdma(struct rdsv3_connection *conn, struct rdsv3_rdma_op *op);
362 void rdsv3_ib_send_add_credits(struct rdsv3_connection *conn,
363     unsigned int credits);
364 void rdsv3_ib_advertise_credits(struct rdsv3_connection *conn,
365     unsigned int posted);
366 int rdsv3_ib_send_grab_credits(struct rdsv3_ib_connection *ic, uint32_t wanted,
367     uint32_t *adv_credits, int need_posted);
368 
369 /* ib_stats.c */
370 extern struct rdsv3_ib_statistics	*rdsv3_ib_stats;
371 #define	rdsv3_ib_stats_inc(member) \
372 	rdsv3_stats_add_which(rdsv3_ib_stats, member, 1)
373 unsigned int rdsv3_ib_stats_info_copy(struct rdsv3_info_iterator *iter,
374     unsigned int avail);
375 
376 /* ib_sysctl.c */
377 int rdsv3_ib_sysctl_init(void);
378 void rdsv3_ib_sysctl_exit(void);
379 extern unsigned long rdsv3_ib_sysctl_max_send_wr;
380 extern unsigned long rdsv3_ib_sysctl_max_recv_wr;
381 extern unsigned long rdsv3_ib_sysctl_max_unsig_wrs;
382 extern unsigned long rdsv3_ib_sysctl_max_unsig_bytes;
383 extern unsigned long rdsv3_ib_sysctl_max_recv_allocation;
384 extern unsigned int rdsv3_ib_sysctl_flow_control;
385 
386 #endif /* _RDSV3_IB_H */
387