xref: /openbmc/linux/net/sunrpc/xprtrdma/xprt_rdma.h (revision 9c6d26df1fae6ad4718d51c48e6517913304ed27)
1 /*
2  * Copyright (c) 2014-2017 Oracle.  All rights reserved.
3  * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the BSD-type
9  * license below:
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  *
15  *      Redistributions of source code must retain the above copyright
16  *      notice, this list of conditions and the following disclaimer.
17  *
18  *      Redistributions in binary form must reproduce the above
19  *      copyright notice, this list of conditions and the following
20  *      disclaimer in the documentation and/or other materials provided
21  *      with the distribution.
22  *
23  *      Neither the name of the Network Appliance, Inc. nor the names of
24  *      its contributors may be used to endorse or promote products
25  *      derived from this software without specific prior written
26  *      permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
29  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
30  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
31  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
32  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
33  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
34  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
35  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
36  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
37  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
38  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
39  */
40 
41 #ifndef _LINUX_SUNRPC_XPRT_RDMA_H
42 #define _LINUX_SUNRPC_XPRT_RDMA_H
43 
44 #include <linux/wait.h> 		/* wait_queue_head_t, etc */
45 #include <linux/spinlock.h> 		/* spinlock_t, etc */
46 #include <linux/atomic.h>			/* atomic_t, etc */
47 #include <linux/workqueue.h>		/* struct work_struct */
48 
49 #include <rdma/rdma_cm.h>		/* RDMA connection api */
50 #include <rdma/ib_verbs.h>		/* RDMA verbs api */
51 
52 #include <linux/sunrpc/clnt.h> 		/* rpc_xprt */
53 #include <linux/sunrpc/rpc_rdma.h> 	/* RPC/RDMA protocol */
54 #include <linux/sunrpc/xprtrdma.h> 	/* xprt parameters */
55 
56 #define RDMA_RESOLVE_TIMEOUT	(5000)	/* 5 seconds */
57 #define RDMA_CONNECT_RETRY_MAX	(2)	/* retries if no listener backlog */
58 
59 #define RPCRDMA_BIND_TO		(60U * HZ)
60 #define RPCRDMA_INIT_REEST_TO	(5U * HZ)
61 #define RPCRDMA_MAX_REEST_TO	(30U * HZ)
62 #define RPCRDMA_IDLE_DISC_TO	(5U * 60 * HZ)
63 
64 /*
65  * Interface Adapter -- one per transport instance
66  */
67 struct rpcrdma_ia {
68 	const struct rpcrdma_memreg_ops	*ri_ops;
69 	struct ib_device	*ri_device;
70 	struct rdma_cm_id 	*ri_id;
71 	struct ib_pd		*ri_pd;
72 	struct completion	ri_done;
73 	struct completion	ri_remove_done;
74 	int			ri_async_rc;
75 	unsigned int		ri_max_segs;
76 	unsigned int		ri_max_frwr_depth;
77 	unsigned int		ri_max_inline_write;
78 	unsigned int		ri_max_inline_read;
79 	unsigned int		ri_max_send_sges;
80 	bool			ri_implicit_roundup;
81 	enum ib_mr_type		ri_mrtype;
82 	unsigned long		ri_flags;
83 	struct ib_qp_attr	ri_qp_attr;
84 	struct ib_qp_init_attr	ri_qp_init_attr;
85 };
86 
87 enum {
88 	RPCRDMA_IAF_REMOVING = 0,
89 };
90 
91 /*
92  * RDMA Endpoint -- one per transport instance
93  */
94 
95 struct rpcrdma_ep {
96 	unsigned int		rep_send_count;
97 	unsigned int		rep_send_batch;
98 	int			rep_connected;
99 	struct ib_qp_init_attr	rep_attr;
100 	wait_queue_head_t 	rep_connect_wait;
101 	struct rpcrdma_connect_private	rep_cm_private;
102 	struct rdma_conn_param	rep_remote_cma;
103 	struct delayed_work	rep_connect_worker;
104 };
105 
106 /* Pre-allocate extra Work Requests for handling backward receives
107  * and sends. This is a fixed value because the Work Queues are
108  * allocated when the forward channel is set up.
109  */
110 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
111 #define RPCRDMA_BACKWARD_WRS		(8)
112 #else
113 #define RPCRDMA_BACKWARD_WRS		(0)
114 #endif
115 
116 /* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
117  *
118  * The below structure appears at the front of a large region of kmalloc'd
119  * memory, which always starts on a good alignment boundary.
120  */
121 
122 struct rpcrdma_regbuf {
123 	struct ib_sge		rg_iov;
124 	struct ib_device	*rg_device;
125 	enum dma_data_direction	rg_direction;
126 	__be32			rg_base[0] __attribute__ ((aligned(256)));
127 };
128 
129 static inline u64
130 rdmab_addr(struct rpcrdma_regbuf *rb)
131 {
132 	return rb->rg_iov.addr;
133 }
134 
135 static inline u32
136 rdmab_length(struct rpcrdma_regbuf *rb)
137 {
138 	return rb->rg_iov.length;
139 }
140 
141 static inline u32
142 rdmab_lkey(struct rpcrdma_regbuf *rb)
143 {
144 	return rb->rg_iov.lkey;
145 }
146 
147 static inline struct ib_device *
148 rdmab_device(struct rpcrdma_regbuf *rb)
149 {
150 	return rb->rg_device;
151 }
152 
153 #define RPCRDMA_DEF_GFP		(GFP_NOIO | __GFP_NOWARN)
154 
155 /* To ensure a transport can always make forward progress,
156  * the number of RDMA segments allowed in header chunk lists
157  * is capped at 8. This prevents less-capable devices and
158  * memory registrations from overrunning the Send buffer
159  * while building chunk lists.
160  *
161  * Elements of the Read list take up more room than the
162  * Write list or Reply chunk. 8 read segments means the Read
163  * list (or Write list or Reply chunk) cannot consume more
164  * than
165  *
166  * ((8 + 2) * read segment size) + 1 XDR words, or 244 bytes.
167  *
168  * And the fixed part of the header is another 24 bytes.
169  *
170  * The smallest inline threshold is 1024 bytes, ensuring that
171  * at least 750 bytes are available for RPC messages.
172  */
173 enum {
174 	RPCRDMA_MAX_HDR_SEGS = 8,
175 	RPCRDMA_HDRBUF_SIZE = 256,
176 };
177 
178 /*
179  * struct rpcrdma_rep -- this structure encapsulates state required
180  * to receive and complete an RPC Reply, asychronously. It needs
181  * several pieces of state:
182  *
183  *   o receive buffer and ib_sge (donated to provider)
184  *   o status of receive (success or not, length, inv rkey)
185  *   o bookkeeping state to get run by reply handler (XDR stream)
186  *
187  * These structures are allocated during transport initialization.
188  * N of these are associated with a transport instance, managed by
189  * struct rpcrdma_buffer. N is the max number of outstanding RPCs.
190  */
191 
192 struct rpcrdma_rep {
193 	struct ib_cqe		rr_cqe;
194 	__be32			rr_xid;
195 	__be32			rr_vers;
196 	__be32			rr_proc;
197 	int			rr_wc_flags;
198 	u32			rr_inv_rkey;
199 	struct rpcrdma_regbuf	*rr_rdmabuf;
200 	struct rpcrdma_xprt	*rr_rxprt;
201 	struct work_struct	rr_work;
202 	struct xdr_buf		rr_hdrbuf;
203 	struct xdr_stream	rr_stream;
204 	struct rpc_rqst		*rr_rqst;
205 	struct list_head	rr_list;
206 	struct ib_recv_wr	rr_recv_wr;
207 };
208 
209 /* struct rpcrdma_sendctx - DMA mapped SGEs to unmap after Send completes
210  */
211 struct rpcrdma_req;
212 struct rpcrdma_xprt;
213 struct rpcrdma_sendctx {
214 	struct ib_send_wr	sc_wr;
215 	struct ib_cqe		sc_cqe;
216 	struct rpcrdma_xprt	*sc_xprt;
217 	struct rpcrdma_req	*sc_req;
218 	unsigned int		sc_unmap_count;
219 	struct ib_sge		sc_sges[];
220 };
221 
222 /* Limit the number of SGEs that can be unmapped during one
223  * Send completion. This caps the amount of work a single
224  * completion can do before returning to the provider.
225  *
226  * Setting this to zero disables Send completion batching.
227  */
228 enum {
229 	RPCRDMA_MAX_SEND_BATCH = 7,
230 };
231 
232 /*
233  * struct rpcrdma_mr - external memory region metadata
234  *
235  * An external memory region is any buffer or page that is registered
236  * on the fly (ie, not pre-registered).
237  *
238  * Each rpcrdma_buffer has a list of free MWs anchored in rb_mrs. During
239  * call_allocate, rpcrdma_buffer_get() assigns one to each segment in
240  * an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep
241  * track of registration metadata while each RPC is pending.
242  * rpcrdma_deregister_external() uses this metadata to unmap and
243  * release these resources when an RPC is complete.
244  */
245 enum rpcrdma_frwr_state {
246 	FRWR_IS_INVALID,	/* ready to be used */
247 	FRWR_IS_VALID,		/* in use */
248 	FRWR_FLUSHED_FR,	/* flushed FASTREG WR */
249 	FRWR_FLUSHED_LI,	/* flushed LOCALINV WR */
250 };
251 
252 struct rpcrdma_frwr {
253 	struct ib_mr			*fr_mr;
254 	struct ib_cqe			fr_cqe;
255 	enum rpcrdma_frwr_state		fr_state;
256 	struct completion		fr_linv_done;
257 	union {
258 		struct ib_reg_wr	fr_regwr;
259 		struct ib_send_wr	fr_invwr;
260 	};
261 };
262 
263 struct rpcrdma_fmr {
264 	struct ib_fmr		*fm_mr;
265 	u64			*fm_physaddrs;
266 };
267 
268 struct rpcrdma_mr {
269 	struct list_head	mr_list;
270 	struct scatterlist	*mr_sg;
271 	int			mr_nents;
272 	enum dma_data_direction	mr_dir;
273 	union {
274 		struct rpcrdma_fmr	fmr;
275 		struct rpcrdma_frwr	frwr;
276 	};
277 	struct rpcrdma_xprt	*mr_xprt;
278 	u32			mr_handle;
279 	u32			mr_length;
280 	u64			mr_offset;
281 	struct list_head	mr_all;
282 };
283 
284 /*
285  * struct rpcrdma_req -- structure central to the request/reply sequence.
286  *
287  * N of these are associated with a transport instance, and stored in
288  * struct rpcrdma_buffer. N is the max number of outstanding requests.
289  *
290  * It includes pre-registered buffer memory for send AND recv.
291  * The recv buffer, however, is not owned by this structure, and
292  * is "donated" to the hardware when a recv is posted. When a
293  * reply is handled, the recv buffer used is given back to the
294  * struct rpcrdma_req associated with the request.
295  *
296  * In addition to the basic memory, this structure includes an array
297  * of iovs for send operations. The reason is that the iovs passed to
298  * ib_post_{send,recv} must not be modified until the work request
299  * completes.
300  */
301 
302 /* Maximum number of page-sized "segments" per chunk list to be
303  * registered or invalidated. Must handle a Reply chunk:
304  */
305 enum {
306 	RPCRDMA_MAX_IOV_SEGS	= 3,
307 	RPCRDMA_MAX_DATA_SEGS	= ((1 * 1024 * 1024) / PAGE_SIZE) + 1,
308 	RPCRDMA_MAX_SEGS	= RPCRDMA_MAX_DATA_SEGS +
309 				  RPCRDMA_MAX_IOV_SEGS,
310 };
311 
312 struct rpcrdma_mr_seg {		/* chunk descriptors */
313 	u32		mr_len;		/* length of chunk or segment */
314 	struct page	*mr_page;	/* owning page, if any */
315 	char		*mr_offset;	/* kva if no page, else offset */
316 };
317 
318 /* The Send SGE array is provisioned to send a maximum size
319  * inline request:
320  * - RPC-over-RDMA header
321  * - xdr_buf head iovec
322  * - RPCRDMA_MAX_INLINE bytes, in pages
323  * - xdr_buf tail iovec
324  *
325  * The actual number of array elements consumed by each RPC
326  * depends on the device's max_sge limit.
327  */
328 enum {
329 	RPCRDMA_MIN_SEND_SGES = 3,
330 	RPCRDMA_MAX_PAGE_SGES = RPCRDMA_MAX_INLINE >> PAGE_SHIFT,
331 	RPCRDMA_MAX_SEND_SGES = 1 + 1 + RPCRDMA_MAX_PAGE_SGES + 1,
332 };
333 
334 struct rpcrdma_buffer;
335 struct rpcrdma_req {
336 	struct list_head	rl_list;
337 	struct rpcrdma_buffer	*rl_buffer;
338 	struct rpcrdma_rep	*rl_reply;
339 	struct xdr_stream	rl_stream;
340 	struct xdr_buf		rl_hdrbuf;
341 	struct rpcrdma_sendctx	*rl_sendctx;
342 	struct rpcrdma_regbuf	*rl_rdmabuf;	/* xprt header */
343 	struct rpcrdma_regbuf	*rl_sendbuf;	/* rq_snd_buf */
344 	struct rpcrdma_regbuf	*rl_recvbuf;	/* rq_rcv_buf */
345 
346 	struct list_head	rl_all;
347 	unsigned long		rl_flags;
348 
349 	struct list_head	rl_registered;	/* registered segments */
350 	struct rpcrdma_mr_seg	rl_segments[RPCRDMA_MAX_SEGS];
351 };
352 
353 /* rl_flags */
354 enum {
355 	RPCRDMA_REQ_F_PENDING = 0,
356 	RPCRDMA_REQ_F_TX_RESOURCES,
357 };
358 
359 static inline void
360 rpcrdma_set_xprtdata(struct rpc_rqst *rqst, struct rpcrdma_req *req)
361 {
362 	rqst->rq_xprtdata = req;
363 }
364 
365 static inline struct rpcrdma_req *
366 rpcr_to_rdmar(const struct rpc_rqst *rqst)
367 {
368 	return rqst->rq_xprtdata;
369 }
370 
371 static inline void
372 rpcrdma_mr_push(struct rpcrdma_mr *mr, struct list_head *list)
373 {
374 	list_add_tail(&mr->mr_list, list);
375 }
376 
377 static inline struct rpcrdma_mr *
378 rpcrdma_mr_pop(struct list_head *list)
379 {
380 	struct rpcrdma_mr *mr;
381 
382 	mr = list_first_entry(list, struct rpcrdma_mr, mr_list);
383 	list_del(&mr->mr_list);
384 	return mr;
385 }
386 
387 /*
388  * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
389  * inline requests/replies, and client/server credits.
390  *
391  * One of these is associated with a transport instance
392  */
393 struct rpcrdma_buffer {
394 	spinlock_t		rb_mrlock;	/* protect rb_mrs list */
395 	struct list_head	rb_mrs;
396 	struct list_head	rb_all;
397 
398 	unsigned long		rb_sc_head;
399 	unsigned long		rb_sc_tail;
400 	unsigned long		rb_sc_last;
401 	struct rpcrdma_sendctx	**rb_sc_ctxs;
402 
403 	spinlock_t		rb_lock;	/* protect buf lists */
404 	int			rb_send_count, rb_recv_count;
405 	struct list_head	rb_send_bufs;
406 	struct list_head	rb_recv_bufs;
407 	u32			rb_max_requests;
408 	u32			rb_credits;	/* most recent credit grant */
409 
410 	u32			rb_bc_srv_max_requests;
411 	spinlock_t		rb_reqslock;	/* protect rb_allreqs */
412 	struct list_head	rb_allreqs;
413 
414 	u32			rb_bc_max_requests;
415 
416 	spinlock_t		rb_recovery_lock; /* protect rb_stale_mrs */
417 	struct list_head	rb_stale_mrs;
418 	struct delayed_work	rb_recovery_worker;
419 	struct delayed_work	rb_refresh_worker;
420 };
421 #define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia)
422 
423 /*
424  * Internal structure for transport instance creation. This
425  * exists primarily for modularity.
426  *
427  * This data should be set with mount options
428  */
429 struct rpcrdma_create_data_internal {
430 	unsigned int	max_requests;	/* max requests (slots) in flight */
431 	unsigned int	rsize;		/* mount rsize - max read hdr+data */
432 	unsigned int	wsize;		/* mount wsize - max write hdr+data */
433 	unsigned int	inline_rsize;	/* max non-rdma read data payload */
434 	unsigned int	inline_wsize;	/* max non-rdma write data payload */
435 };
436 
437 /*
438  * Statistics for RPCRDMA
439  */
440 struct rpcrdma_stats {
441 	/* accessed when sending a call */
442 	unsigned long		read_chunk_count;
443 	unsigned long		write_chunk_count;
444 	unsigned long		reply_chunk_count;
445 	unsigned long long	total_rdma_request;
446 
447 	/* rarely accessed error counters */
448 	unsigned long long	pullup_copy_count;
449 	unsigned long		hardway_register_count;
450 	unsigned long		failed_marshal_count;
451 	unsigned long		bad_reply_count;
452 	unsigned long		mrs_recovered;
453 	unsigned long		mrs_orphaned;
454 	unsigned long		mrs_allocated;
455 	unsigned long		empty_sendctx_q;
456 
457 	/* accessed when receiving a reply */
458 	unsigned long long	total_rdma_reply;
459 	unsigned long long	fixup_copy_count;
460 	unsigned long		reply_waits_for_send;
461 	unsigned long		local_inv_needed;
462 	unsigned long		nomsg_call_count;
463 	unsigned long		bcall_count;
464 };
465 
466 /*
467  * Per-registration mode operations
468  */
469 struct rpcrdma_xprt;
470 struct rpcrdma_memreg_ops {
471 	struct rpcrdma_mr_seg *
472 			(*ro_map)(struct rpcrdma_xprt *,
473 				  struct rpcrdma_mr_seg *, int, bool,
474 				  struct rpcrdma_mr **);
475 	int		(*ro_send)(struct rpcrdma_ia *ia,
476 				   struct rpcrdma_req *req);
477 	void		(*ro_reminv)(struct rpcrdma_rep *rep,
478 				     struct list_head *mrs);
479 	void		(*ro_unmap_sync)(struct rpcrdma_xprt *,
480 					 struct list_head *);
481 	void		(*ro_recover_mr)(struct rpcrdma_mr *mr);
482 	int		(*ro_open)(struct rpcrdma_ia *,
483 				   struct rpcrdma_ep *,
484 				   struct rpcrdma_create_data_internal *);
485 	size_t		(*ro_maxpages)(struct rpcrdma_xprt *);
486 	int		(*ro_init_mr)(struct rpcrdma_ia *,
487 				      struct rpcrdma_mr *);
488 	void		(*ro_release_mr)(struct rpcrdma_mr *mr);
489 	const char	*ro_displayname;
490 	const int	ro_send_w_inv_ok;
491 };
492 
493 extern const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops;
494 extern const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops;
495 
496 /*
497  * RPCRDMA transport -- encapsulates the structures above for
498  * integration with RPC.
499  *
500  * The contained structures are embedded, not pointers,
501  * for convenience. This structure need not be visible externally.
502  *
503  * It is allocated and initialized during mount, and released
504  * during unmount.
505  */
506 struct rpcrdma_xprt {
507 	struct rpc_xprt		rx_xprt;
508 	struct rpcrdma_ia	rx_ia;
509 	struct rpcrdma_ep	rx_ep;
510 	struct rpcrdma_buffer	rx_buf;
511 	struct rpcrdma_create_data_internal rx_data;
512 	struct delayed_work	rx_connect_worker;
513 	struct rpcrdma_stats	rx_stats;
514 };
515 
516 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
517 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data)
518 
519 static inline const char *
520 rpcrdma_addrstr(const struct rpcrdma_xprt *r_xprt)
521 {
522 	return r_xprt->rx_xprt.address_strings[RPC_DISPLAY_ADDR];
523 }
524 
525 static inline const char *
526 rpcrdma_portstr(const struct rpcrdma_xprt *r_xprt)
527 {
528 	return r_xprt->rx_xprt.address_strings[RPC_DISPLAY_PORT];
529 }
530 
531 /* Setting this to 0 ensures interoperability with early servers.
532  * Setting this to 1 enhances certain unaligned read/write performance.
533  * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
534 extern int xprt_rdma_pad_optimize;
535 
536 /* This setting controls the hunt for a supported memory
537  * registration strategy.
538  */
539 extern unsigned int xprt_rdma_memreg_strategy;
540 
541 /*
542  * Interface Adapter calls - xprtrdma/verbs.c
543  */
544 int rpcrdma_ia_open(struct rpcrdma_xprt *xprt);
545 void rpcrdma_ia_remove(struct rpcrdma_ia *ia);
546 void rpcrdma_ia_close(struct rpcrdma_ia *);
547 bool frwr_is_supported(struct rpcrdma_ia *);
548 bool fmr_is_supported(struct rpcrdma_ia *);
549 
550 extern struct workqueue_struct *rpcrdma_receive_wq;
551 
552 /*
553  * Endpoint calls - xprtrdma/verbs.c
554  */
555 int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *,
556 				struct rpcrdma_create_data_internal *);
557 void rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *);
558 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *);
559 void rpcrdma_conn_func(struct rpcrdma_ep *ep);
560 void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *);
561 
562 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
563 				struct rpcrdma_req *);
564 int rpcrdma_ep_post_recv(struct rpcrdma_ia *, struct rpcrdma_rep *);
565 
566 /*
567  * Buffer calls - xprtrdma/verbs.c
568  */
569 struct rpcrdma_req *rpcrdma_create_req(struct rpcrdma_xprt *);
570 void rpcrdma_destroy_req(struct rpcrdma_req *);
571 int rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt);
572 int rpcrdma_buffer_create(struct rpcrdma_xprt *);
573 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
574 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_buffer *buf);
575 void rpcrdma_sendctx_put_locked(struct rpcrdma_sendctx *sc);
576 
577 struct rpcrdma_mr *rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt);
578 void rpcrdma_mr_put(struct rpcrdma_mr *mr);
579 void rpcrdma_mr_unmap_and_put(struct rpcrdma_mr *mr);
580 void rpcrdma_mr_defer_recovery(struct rpcrdma_mr *mr);
581 
582 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
583 void rpcrdma_buffer_put(struct rpcrdma_req *);
584 void rpcrdma_recv_buffer_get(struct rpcrdma_req *);
585 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
586 
587 struct rpcrdma_regbuf *rpcrdma_alloc_regbuf(size_t, enum dma_data_direction,
588 					    gfp_t);
589 bool __rpcrdma_dma_map_regbuf(struct rpcrdma_ia *, struct rpcrdma_regbuf *);
590 void rpcrdma_free_regbuf(struct rpcrdma_regbuf *);
591 
592 static inline bool
593 rpcrdma_regbuf_is_mapped(struct rpcrdma_regbuf *rb)
594 {
595 	return rb->rg_device != NULL;
596 }
597 
598 static inline bool
599 rpcrdma_dma_map_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
600 {
601 	if (likely(rpcrdma_regbuf_is_mapped(rb)))
602 		return true;
603 	return __rpcrdma_dma_map_regbuf(ia, rb);
604 }
605 
606 int rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt *, unsigned int);
607 
608 int rpcrdma_alloc_wq(void);
609 void rpcrdma_destroy_wq(void);
610 
611 /*
612  * Wrappers for chunk registration, shared by read/write chunk code.
613  */
614 
615 static inline enum dma_data_direction
616 rpcrdma_data_dir(bool writing)
617 {
618 	return writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
619 }
620 
621 /*
622  * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
623  */
624 
625 enum rpcrdma_chunktype {
626 	rpcrdma_noch = 0,
627 	rpcrdma_readch,
628 	rpcrdma_areadch,
629 	rpcrdma_writech,
630 	rpcrdma_replych
631 };
632 
633 int rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
634 			      struct rpcrdma_req *req, u32 hdrlen,
635 			      struct xdr_buf *xdr,
636 			      enum rpcrdma_chunktype rtype);
637 void rpcrdma_unmap_sendctx(struct rpcrdma_sendctx *sc);
638 int rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst);
639 void rpcrdma_set_max_header_sizes(struct rpcrdma_xprt *);
640 void rpcrdma_complete_rqst(struct rpcrdma_rep *rep);
641 void rpcrdma_reply_handler(struct rpcrdma_rep *rep);
642 void rpcrdma_release_rqst(struct rpcrdma_xprt *r_xprt,
643 			  struct rpcrdma_req *req);
644 void rpcrdma_deferred_completion(struct work_struct *work);
645 
646 static inline void rpcrdma_set_xdrlen(struct xdr_buf *xdr, size_t len)
647 {
648 	xdr->head[0].iov_len = len;
649 	xdr->len = len;
650 }
651 
652 /* RPC/RDMA module init - xprtrdma/transport.c
653  */
654 extern unsigned int xprt_rdma_max_inline_read;
655 void xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap);
656 void xprt_rdma_free_addresses(struct rpc_xprt *xprt);
657 void rpcrdma_connect_worker(struct work_struct *work);
658 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq);
659 int xprt_rdma_init(void);
660 void xprt_rdma_cleanup(void);
661 
662 /* Backchannel calls - xprtrdma/backchannel.c
663  */
664 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
665 int xprt_rdma_bc_setup(struct rpc_xprt *, unsigned int);
666 int xprt_rdma_bc_up(struct svc_serv *, struct net *);
667 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *);
668 int rpcrdma_bc_post_recv(struct rpcrdma_xprt *, unsigned int);
669 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *, struct rpcrdma_rep *);
670 int xprt_rdma_bc_send_reply(struct rpc_rqst *rqst);
671 void xprt_rdma_bc_free_rqst(struct rpc_rqst *);
672 void xprt_rdma_bc_destroy(struct rpc_xprt *, unsigned int);
673 #endif	/* CONFIG_SUNRPC_BACKCHANNEL */
674 
675 extern struct xprt_class xprt_rdma_bc;
676 
677 #endif				/* _LINUX_SUNRPC_XPRT_RDMA_H */
678 
679 #include <trace/events/rpcrdma.h>
680