xref: /linux/net/rds/ib.h (revision 1f2367a39f17bd553a75e179a747f9b257bc9478)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _RDS_IB_H
3 #define _RDS_IB_H
4 
5 #include <rdma/ib_verbs.h>
6 #include <rdma/rdma_cm.h>
7 #include <linux/interrupt.h>
8 #include <linux/pci.h>
9 #include <linux/slab.h>
10 #include "rds.h"
11 #include "rdma_transport.h"
12 
13 #define RDS_IB_MAX_SGE			8
14 #define RDS_IB_RECV_SGE 		2
15 
16 #define RDS_IB_DEFAULT_RECV_WR		1024
17 #define RDS_IB_DEFAULT_SEND_WR		256
18 #define RDS_IB_DEFAULT_FR_WR		256
19 #define RDS_IB_DEFAULT_FR_INV_WR	256
20 
21 #define RDS_IB_DEFAULT_RETRY_COUNT	1
22 
23 #define RDS_IB_SUPPORTED_PROTOCOLS	0x00000003	/* minor versions supported */
24 
25 #define RDS_IB_RECYCLE_BATCH_COUNT	32
26 
27 #define RDS_IB_WC_MAX			32
28 
29 extern struct rw_semaphore rds_ib_devices_lock;
30 extern struct list_head rds_ib_devices;
31 
32 /*
33  * IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to
34  * try and minimize the amount of memory tied up both the device and
35  * socket receive queues.
36  */
37 struct rds_page_frag {
38 	struct list_head	f_item;
39 	struct list_head	f_cache_entry;
40 	struct scatterlist	f_sg;
41 };
42 
43 struct rds_ib_incoming {
44 	struct list_head	ii_frags;
45 	struct list_head	ii_cache_entry;
46 	struct rds_incoming	ii_inc;
47 };
48 
49 struct rds_ib_cache_head {
50 	struct list_head *first;
51 	unsigned long count;
52 };
53 
54 struct rds_ib_refill_cache {
55 	struct rds_ib_cache_head __percpu *percpu;
56 	struct list_head	 *xfer;
57 	struct list_head	 *ready;
58 };
59 
60 /* This is the common structure for the IB private data exchange in setting up
61  * an RDS connection.  The exchange is different for IPv4 and IPv6 connections.
62  * The reason is that the address size is different and the addresses
63  * exchanged are in the beginning of the structure.  Hence it is not possible
64  * for interoperability if same structure is used.
65  */
66 struct rds_ib_conn_priv_cmn {
67 	u8			ricpc_protocol_major;
68 	u8			ricpc_protocol_minor;
69 	__be16			ricpc_protocol_minor_mask;	/* bitmask */
70 	u8			ricpc_dp_toss;
71 	u8			ripc_reserved1;
72 	__be16			ripc_reserved2;
73 	__be64			ricpc_ack_seq;
74 	__be32			ricpc_credit;	/* non-zero enables flow ctl */
75 };
76 
77 struct rds_ib_connect_private {
78 	/* Add new fields at the end, and don't permute existing fields. */
79 	__be32				dp_saddr;
80 	__be32				dp_daddr;
81 	struct rds_ib_conn_priv_cmn	dp_cmn;
82 };
83 
84 struct rds6_ib_connect_private {
85 	/* Add new fields at the end, and don't permute existing fields. */
86 	struct in6_addr			dp_saddr;
87 	struct in6_addr			dp_daddr;
88 	struct rds_ib_conn_priv_cmn	dp_cmn;
89 };
90 
91 #define dp_protocol_major	dp_cmn.ricpc_protocol_major
92 #define dp_protocol_minor	dp_cmn.ricpc_protocol_minor
93 #define dp_protocol_minor_mask	dp_cmn.ricpc_protocol_minor_mask
94 #define dp_ack_seq		dp_cmn.ricpc_ack_seq
95 #define dp_credit		dp_cmn.ricpc_credit
96 
97 union rds_ib_conn_priv {
98 	struct rds_ib_connect_private	ricp_v4;
99 	struct rds6_ib_connect_private	ricp_v6;
100 };
101 
102 struct rds_ib_send_work {
103 	void			*s_op;
104 	union {
105 		struct ib_send_wr	s_wr;
106 		struct ib_rdma_wr	s_rdma_wr;
107 		struct ib_atomic_wr	s_atomic_wr;
108 	};
109 	struct ib_sge		s_sge[RDS_IB_MAX_SGE];
110 	unsigned long		s_queued;
111 };
112 
113 struct rds_ib_recv_work {
114 	struct rds_ib_incoming 	*r_ibinc;
115 	struct rds_page_frag	*r_frag;
116 	struct ib_recv_wr	r_wr;
117 	struct ib_sge		r_sge[2];
118 };
119 
120 struct rds_ib_work_ring {
121 	u32		w_nr;
122 	u32		w_alloc_ptr;
123 	u32		w_alloc_ctr;
124 	u32		w_free_ptr;
125 	atomic_t	w_free_ctr;
126 };
127 
128 /* Rings are posted with all the allocations they'll need to queue the
129  * incoming message to the receiving socket so this can't fail.
130  * All fragments start with a header, so we can make sure we're not receiving
131  * garbage, and we can tell a small 8 byte fragment from an ACK frame.
132  */
133 struct rds_ib_ack_state {
134 	u64		ack_next;
135 	u64		ack_recv;
136 	unsigned int	ack_required:1;
137 	unsigned int	ack_next_valid:1;
138 	unsigned int	ack_recv_valid:1;
139 };
140 
141 
142 struct rds_ib_device;
143 
144 struct rds_ib_connection {
145 
146 	struct list_head	ib_node;
147 	struct rds_ib_device	*rds_ibdev;
148 	struct rds_connection	*conn;
149 
150 	/* alphabet soup, IBTA style */
151 	struct rdma_cm_id	*i_cm_id;
152 	struct ib_pd		*i_pd;
153 	struct ib_cq		*i_send_cq;
154 	struct ib_cq		*i_recv_cq;
155 	struct ib_wc		i_send_wc[RDS_IB_WC_MAX];
156 	struct ib_wc		i_recv_wc[RDS_IB_WC_MAX];
157 
158 	/* To control the number of wrs from fastreg */
159 	atomic_t		i_fastreg_wrs;
160 	atomic_t		i_fastunreg_wrs;
161 
162 	/* interrupt handling */
163 	struct tasklet_struct	i_send_tasklet;
164 	struct tasklet_struct	i_recv_tasklet;
165 
166 	/* tx */
167 	struct rds_ib_work_ring	i_send_ring;
168 	struct rm_data_op	*i_data_op;
169 	struct rds_header	*i_send_hdrs;
170 	dma_addr_t		i_send_hdrs_dma;
171 	struct rds_ib_send_work *i_sends;
172 	atomic_t		i_signaled_sends;
173 
174 	/* rx */
175 	struct mutex		i_recv_mutex;
176 	struct rds_ib_work_ring	i_recv_ring;
177 	struct rds_ib_incoming	*i_ibinc;
178 	u32			i_recv_data_rem;
179 	struct rds_header	*i_recv_hdrs;
180 	dma_addr_t		i_recv_hdrs_dma;
181 	struct rds_ib_recv_work *i_recvs;
182 	u64			i_ack_recv;	/* last ACK received */
183 	struct rds_ib_refill_cache i_cache_incs;
184 	struct rds_ib_refill_cache i_cache_frags;
185 	atomic_t		i_cache_allocs;
186 
187 	/* sending acks */
188 	unsigned long		i_ack_flags;
189 #ifdef KERNEL_HAS_ATOMIC64
190 	atomic64_t		i_ack_next;	/* next ACK to send */
191 #else
192 	spinlock_t		i_ack_lock;	/* protect i_ack_next */
193 	u64			i_ack_next;	/* next ACK to send */
194 #endif
195 	struct rds_header	*i_ack;
196 	struct ib_send_wr	i_ack_wr;
197 	struct ib_sge		i_ack_sge;
198 	dma_addr_t		i_ack_dma;
199 	unsigned long		i_ack_queued;
200 
201 	/* Flow control related information
202 	 *
203 	 * Our algorithm uses a pair variables that we need to access
204 	 * atomically - one for the send credits, and one posted
205 	 * recv credits we need to transfer to remote.
206 	 * Rather than protect them using a slow spinlock, we put both into
207 	 * a single atomic_t and update it using cmpxchg
208 	 */
209 	atomic_t		i_credits;
210 
211 	/* Protocol version specific information */
212 	unsigned int		i_flowctl:1;	/* enable/disable flow ctl */
213 
214 	/* Batched completions */
215 	unsigned int		i_unsignaled_wrs;
216 
217 	/* Endpoint role in connection */
218 	bool			i_active_side;
219 	atomic_t		i_cq_quiesce;
220 
221 	/* Send/Recv vectors */
222 	int			i_scq_vector;
223 	int			i_rcq_vector;
224 };
225 
226 /* This assumes that atomic_t is at least 32 bits */
227 #define IB_GET_SEND_CREDITS(v)	((v) & 0xffff)
228 #define IB_GET_POST_CREDITS(v)	((v) >> 16)
229 #define IB_SET_SEND_CREDITS(v)	((v) & 0xffff)
230 #define IB_SET_POST_CREDITS(v)	((v) << 16)
231 
232 struct rds_ib_ipaddr {
233 	struct list_head	list;
234 	__be32			ipaddr;
235 	struct rcu_head		rcu;
236 };
237 
238 enum {
239 	RDS_IB_MR_8K_POOL,
240 	RDS_IB_MR_1M_POOL,
241 };
242 
243 struct rds_ib_device {
244 	struct list_head	list;
245 	struct list_head	ipaddr_list;
246 	struct list_head	conn_list;
247 	struct ib_device	*dev;
248 	struct ib_pd		*pd;
249 	bool                    use_fastreg;
250 
251 	unsigned int		max_mrs;
252 	struct rds_ib_mr_pool	*mr_1m_pool;
253 	struct rds_ib_mr_pool   *mr_8k_pool;
254 	unsigned int		fmr_max_remaps;
255 	unsigned int		max_8k_mrs;
256 	unsigned int		max_1m_mrs;
257 	int			max_sge;
258 	unsigned int		max_wrs;
259 	unsigned int		max_initiator_depth;
260 	unsigned int		max_responder_resources;
261 	spinlock_t		spinlock;	/* protect the above */
262 	refcount_t		refcount;
263 	struct work_struct	free_work;
264 	int			*vector_load;
265 };
266 
267 #define ibdev_to_node(ibdev) dev_to_node((ibdev)->dev.parent)
268 #define rdsibdev_to_node(rdsibdev) ibdev_to_node(rdsibdev->dev)
269 
270 /* bits for i_ack_flags */
271 #define IB_ACK_IN_FLIGHT	0
272 #define IB_ACK_REQUESTED	1
273 
274 /* Magic WR_ID for ACKs */
275 #define RDS_IB_ACK_WR_ID	(~(u64) 0)
276 
277 struct rds_ib_statistics {
278 	uint64_t	s_ib_connect_raced;
279 	uint64_t	s_ib_listen_closed_stale;
280 	uint64_t	s_ib_evt_handler_call;
281 	uint64_t	s_ib_tasklet_call;
282 	uint64_t	s_ib_tx_cq_event;
283 	uint64_t	s_ib_tx_ring_full;
284 	uint64_t	s_ib_tx_throttle;
285 	uint64_t	s_ib_tx_sg_mapping_failure;
286 	uint64_t	s_ib_tx_stalled;
287 	uint64_t	s_ib_tx_credit_updates;
288 	uint64_t	s_ib_rx_cq_event;
289 	uint64_t	s_ib_rx_ring_empty;
290 	uint64_t	s_ib_rx_refill_from_cq;
291 	uint64_t	s_ib_rx_refill_from_thread;
292 	uint64_t	s_ib_rx_alloc_limit;
293 	uint64_t	s_ib_rx_total_frags;
294 	uint64_t	s_ib_rx_total_incs;
295 	uint64_t	s_ib_rx_credit_updates;
296 	uint64_t	s_ib_ack_sent;
297 	uint64_t	s_ib_ack_send_failure;
298 	uint64_t	s_ib_ack_send_delayed;
299 	uint64_t	s_ib_ack_send_piggybacked;
300 	uint64_t	s_ib_ack_received;
301 	uint64_t	s_ib_rdma_mr_8k_alloc;
302 	uint64_t	s_ib_rdma_mr_8k_free;
303 	uint64_t	s_ib_rdma_mr_8k_used;
304 	uint64_t	s_ib_rdma_mr_8k_pool_flush;
305 	uint64_t	s_ib_rdma_mr_8k_pool_wait;
306 	uint64_t	s_ib_rdma_mr_8k_pool_depleted;
307 	uint64_t	s_ib_rdma_mr_1m_alloc;
308 	uint64_t	s_ib_rdma_mr_1m_free;
309 	uint64_t	s_ib_rdma_mr_1m_used;
310 	uint64_t	s_ib_rdma_mr_1m_pool_flush;
311 	uint64_t	s_ib_rdma_mr_1m_pool_wait;
312 	uint64_t	s_ib_rdma_mr_1m_pool_depleted;
313 	uint64_t	s_ib_rdma_mr_8k_reused;
314 	uint64_t	s_ib_rdma_mr_1m_reused;
315 	uint64_t	s_ib_atomic_cswp;
316 	uint64_t	s_ib_atomic_fadd;
317 	uint64_t	s_ib_recv_added_to_cache;
318 	uint64_t	s_ib_recv_removed_from_cache;
319 };
320 
321 extern struct workqueue_struct *rds_ib_wq;
322 
323 /*
324  * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h
325  * doesn't define it.
326  */
327 static inline void rds_ib_dma_sync_sg_for_cpu(struct ib_device *dev,
328 					      struct scatterlist *sglist,
329 					      unsigned int sg_dma_len,
330 					      int direction)
331 {
332 	struct scatterlist *sg;
333 	unsigned int i;
334 
335 	for_each_sg(sglist, sg, sg_dma_len, i) {
336 		ib_dma_sync_single_for_cpu(dev, sg_dma_address(sg),
337 					   sg_dma_len(sg), direction);
338 	}
339 }
340 #define ib_dma_sync_sg_for_cpu	rds_ib_dma_sync_sg_for_cpu
341 
342 static inline void rds_ib_dma_sync_sg_for_device(struct ib_device *dev,
343 						 struct scatterlist *sglist,
344 						 unsigned int sg_dma_len,
345 						 int direction)
346 {
347 	struct scatterlist *sg;
348 	unsigned int i;
349 
350 	for_each_sg(sglist, sg, sg_dma_len, i) {
351 		ib_dma_sync_single_for_device(dev, sg_dma_address(sg),
352 					      sg_dma_len(sg), direction);
353 	}
354 }
355 #define ib_dma_sync_sg_for_device	rds_ib_dma_sync_sg_for_device
356 
357 
358 /* ib.c */
359 extern struct rds_transport rds_ib_transport;
360 struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device);
361 void rds_ib_dev_put(struct rds_ib_device *rds_ibdev);
362 extern struct ib_client rds_ib_client;
363 
364 extern unsigned int rds_ib_retry_count;
365 
366 extern spinlock_t ib_nodev_conns_lock;
367 extern struct list_head ib_nodev_conns;
368 
369 /* ib_cm.c */
370 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp);
371 void rds_ib_conn_free(void *arg);
372 int rds_ib_conn_path_connect(struct rds_conn_path *cp);
373 void rds_ib_conn_path_shutdown(struct rds_conn_path *cp);
374 void rds_ib_state_change(struct sock *sk);
375 int rds_ib_listen_init(void);
376 void rds_ib_listen_stop(void);
377 __printf(2, 3)
378 void __rds_ib_conn_error(struct rds_connection *conn, const char *, ...);
379 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
380 			     struct rdma_cm_event *event, bool isv6);
381 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id, bool isv6);
382 void rds_ib_cm_connect_complete(struct rds_connection *conn,
383 				struct rdma_cm_event *event);
384 
385 
386 #define rds_ib_conn_error(conn, fmt...) \
387 	__rds_ib_conn_error(conn, KERN_WARNING "RDS/IB: " fmt)
388 
389 /* ib_rdma.c */
390 int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev,
391 			 struct in6_addr *ipaddr);
392 void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
393 void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
394 void rds_ib_destroy_nodev_conns(void);
395 void rds_ib_mr_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc);
396 
397 /* ib_recv.c */
398 int rds_ib_recv_init(void);
399 void rds_ib_recv_exit(void);
400 int rds_ib_recv_path(struct rds_conn_path *conn);
401 int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic, gfp_t gfp);
402 void rds_ib_recv_free_caches(struct rds_ib_connection *ic);
403 void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp);
404 void rds_ib_inc_free(struct rds_incoming *inc);
405 int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to);
406 void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc,
407 			     struct rds_ib_ack_state *state);
408 void rds_ib_recv_tasklet_fn(unsigned long data);
409 void rds_ib_recv_init_ring(struct rds_ib_connection *ic);
410 void rds_ib_recv_clear_ring(struct rds_ib_connection *ic);
411 void rds_ib_recv_init_ack(struct rds_ib_connection *ic);
412 void rds_ib_attempt_ack(struct rds_ib_connection *ic);
413 void rds_ib_ack_send_complete(struct rds_ib_connection *ic);
414 u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic);
415 void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required);
416 
417 /* ib_ring.c */
418 void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr);
419 void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr);
420 u32 rds_ib_ring_alloc(struct rds_ib_work_ring *ring, u32 val, u32 *pos);
421 void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val);
422 void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val);
423 int rds_ib_ring_empty(struct rds_ib_work_ring *ring);
424 int rds_ib_ring_low(struct rds_ib_work_ring *ring);
425 u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring);
426 u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest);
427 extern wait_queue_head_t rds_ib_ring_empty_wait;
428 
429 /* ib_send.c */
430 void rds_ib_xmit_path_complete(struct rds_conn_path *cp);
431 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
432 		unsigned int hdr_off, unsigned int sg, unsigned int off);
433 void rds_ib_send_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc);
434 void rds_ib_send_init_ring(struct rds_ib_connection *ic);
435 void rds_ib_send_clear_ring(struct rds_ib_connection *ic);
436 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op);
437 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits);
438 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted);
439 int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted,
440 			     u32 *adv_credits, int need_posted, int max_posted);
441 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op);
442 
443 /* ib_stats.c */
444 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_ib_statistics, rds_ib_stats);
445 #define rds_ib_stats_inc(member) rds_stats_inc_which(rds_ib_stats, member)
446 #define rds_ib_stats_add(member, count) \
447 		rds_stats_add_which(rds_ib_stats, member, count)
448 unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter,
449 				    unsigned int avail);
450 
451 /* ib_sysctl.c */
452 int rds_ib_sysctl_init(void);
453 void rds_ib_sysctl_exit(void);
454 extern unsigned long rds_ib_sysctl_max_send_wr;
455 extern unsigned long rds_ib_sysctl_max_recv_wr;
456 extern unsigned long rds_ib_sysctl_max_unsig_wrs;
457 extern unsigned long rds_ib_sysctl_max_unsig_bytes;
458 extern unsigned long rds_ib_sysctl_max_recv_allocation;
459 extern unsigned int rds_ib_sysctl_flow_control;
460 
461 #endif
462