xref: /openbmc/linux/net/sunrpc/xprtrdma/xprt_rdma.h (revision be709d48)
1 /* SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause */
2 /*
3  * Copyright (c) 2014-2017 Oracle.  All rights reserved.
4  * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the BSD-type
10  * license below:
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  *
16  *      Redistributions of source code must retain the above copyright
17  *      notice, this list of conditions and the following disclaimer.
18  *
19  *      Redistributions in binary form must reproduce the above
20  *      copyright notice, this list of conditions and the following
21  *      disclaimer in the documentation and/or other materials provided
22  *      with the distribution.
23  *
24  *      Neither the name of the Network Appliance, Inc. nor the names of
25  *      its contributors may be used to endorse or promote products
26  *      derived from this software without specific prior written
27  *      permission.
28  *
29  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
30  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
31  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
32  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
33  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
34  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
35  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
36  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
37  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
38  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
39  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
40  */
41 
42 #ifndef _LINUX_SUNRPC_XPRT_RDMA_H
43 #define _LINUX_SUNRPC_XPRT_RDMA_H
44 
45 #include <linux/wait.h> 		/* wait_queue_head_t, etc */
46 #include <linux/spinlock.h> 		/* spinlock_t, etc */
47 #include <linux/atomic.h>			/* atomic_t, etc */
48 #include <linux/workqueue.h>		/* struct work_struct */
49 
50 #include <rdma/rdma_cm.h>		/* RDMA connection api */
51 #include <rdma/ib_verbs.h>		/* RDMA verbs api */
52 
53 #include <linux/sunrpc/clnt.h> 		/* rpc_xprt */
54 #include <linux/sunrpc/rpc_rdma.h> 	/* RPC/RDMA protocol */
55 #include <linux/sunrpc/xprtrdma.h> 	/* xprt parameters */
56 
57 #define RDMA_RESOLVE_TIMEOUT	(5000)	/* 5 seconds */
58 #define RDMA_CONNECT_RETRY_MAX	(2)	/* retries if no listener backlog */
59 
60 #define RPCRDMA_BIND_TO		(60U * HZ)
61 #define RPCRDMA_INIT_REEST_TO	(5U * HZ)
62 #define RPCRDMA_MAX_REEST_TO	(30U * HZ)
63 #define RPCRDMA_IDLE_DISC_TO	(5U * 60 * HZ)
64 
65 /*
66  * Interface Adapter -- one per transport instance
67  */
68 struct rpcrdma_ia {
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 };
84 
85 enum {
86 	RPCRDMA_IAF_REMOVING = 0,
87 };
88 
89 /*
90  * RDMA Endpoint -- one per transport instance
91  */
92 
93 struct rpcrdma_ep {
94 	unsigned int		rep_send_count;
95 	unsigned int		rep_send_batch;
96 	int			rep_connected;
97 	struct ib_qp_init_attr	rep_attr;
98 	wait_queue_head_t 	rep_connect_wait;
99 	struct rpcrdma_connect_private	rep_cm_private;
100 	struct rdma_conn_param	rep_remote_cma;
101 	int			rep_receive_count;
102 };
103 
104 /* Pre-allocate extra Work Requests for handling backward receives
105  * and sends. This is a fixed value because the Work Queues are
106  * allocated when the forward channel is set up.
107  */
108 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
109 #define RPCRDMA_BACKWARD_WRS		(8)
110 #else
111 #define RPCRDMA_BACKWARD_WRS		(0)
112 #endif
113 
114 /* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
115  *
116  * The below structure appears at the front of a large region of kmalloc'd
117  * memory, which always starts on a good alignment boundary.
118  */
119 
120 struct rpcrdma_regbuf {
121 	struct ib_sge		rg_iov;
122 	struct ib_device	*rg_device;
123 	enum dma_data_direction	rg_direction;
124 	__be32			rg_base[0] __attribute__ ((aligned(256)));
125 };
126 
127 static inline u64
128 rdmab_addr(struct rpcrdma_regbuf *rb)
129 {
130 	return rb->rg_iov.addr;
131 }
132 
133 static inline u32
134 rdmab_length(struct rpcrdma_regbuf *rb)
135 {
136 	return rb->rg_iov.length;
137 }
138 
139 static inline u32
140 rdmab_lkey(struct rpcrdma_regbuf *rb)
141 {
142 	return rb->rg_iov.lkey;
143 }
144 
145 static inline struct ib_device *
146 rdmab_device(struct rpcrdma_regbuf *rb)
147 {
148 	return rb->rg_device;
149 }
150 
151 #define RPCRDMA_DEF_GFP		(GFP_NOIO | __GFP_NOWARN)
152 
153 /* To ensure a transport can always make forward progress,
154  * the number of RDMA segments allowed in header chunk lists
155  * is capped at 8. This prevents less-capable devices and
156  * memory registrations from overrunning the Send buffer
157  * while building chunk lists.
158  *
159  * Elements of the Read list take up more room than the
160  * Write list or Reply chunk. 8 read segments means the Read
161  * list (or Write list or Reply chunk) cannot consume more
162  * than
163  *
164  * ((8 + 2) * read segment size) + 1 XDR words, or 244 bytes.
165  *
166  * And the fixed part of the header is another 24 bytes.
167  *
168  * The smallest inline threshold is 1024 bytes, ensuring that
169  * at least 750 bytes are available for RPC messages.
170  */
171 enum {
172 	RPCRDMA_MAX_HDR_SEGS = 8,
173 	RPCRDMA_HDRBUF_SIZE = 256,
174 };
175 
176 /*
177  * struct rpcrdma_rep -- this structure encapsulates state required
178  * to receive and complete an RPC Reply, asychronously. It needs
179  * several pieces of state:
180  *
181  *   o receive buffer and ib_sge (donated to provider)
182  *   o status of receive (success or not, length, inv rkey)
183  *   o bookkeeping state to get run by reply handler (XDR stream)
184  *
185  * These structures are allocated during transport initialization.
186  * N of these are associated with a transport instance, managed by
187  * struct rpcrdma_buffer. N is the max number of outstanding RPCs.
188  */
189 
190 struct rpcrdma_rep {
191 	struct ib_cqe		rr_cqe;
192 	__be32			rr_xid;
193 	__be32			rr_vers;
194 	__be32			rr_proc;
195 	int			rr_wc_flags;
196 	u32			rr_inv_rkey;
197 	bool			rr_temp;
198 	struct rpcrdma_regbuf	*rr_rdmabuf;
199 	struct rpcrdma_xprt	*rr_rxprt;
200 	struct work_struct	rr_work;
201 	struct xdr_buf		rr_hdrbuf;
202 	struct xdr_stream	rr_stream;
203 	struct rpc_rqst		*rr_rqst;
204 	struct list_head	rr_list;
205 	struct ib_recv_wr	rr_recv_wr;
206 };
207 
208 /* To reduce the rate at which a transport invokes ib_post_recv
209  * (and thus the hardware doorbell rate), xprtrdma posts Receive
210  * WRs in batches.
211  *
212  * Setting this to zero disables Receive post batching.
213  */
214 enum {
215 	RPCRDMA_MAX_RECV_BATCH = 7,
216 };
217 
218 /* struct rpcrdma_sendctx - DMA mapped SGEs to unmap after Send completes
219  */
220 struct rpcrdma_req;
221 struct rpcrdma_xprt;
222 struct rpcrdma_sendctx {
223 	struct ib_send_wr	sc_wr;
224 	struct ib_cqe		sc_cqe;
225 	struct rpcrdma_xprt	*sc_xprt;
226 	struct rpcrdma_req	*sc_req;
227 	unsigned int		sc_unmap_count;
228 	struct ib_sge		sc_sges[];
229 };
230 
231 /* Limit the number of SGEs that can be unmapped during one
232  * Send completion. This caps the amount of work a single
233  * completion can do before returning to the provider.
234  *
235  * Setting this to zero disables Send completion batching.
236  */
237 enum {
238 	RPCRDMA_MAX_SEND_BATCH = 7,
239 };
240 
241 /*
242  * struct rpcrdma_mr - external memory region metadata
243  *
244  * An external memory region is any buffer or page that is registered
245  * on the fly (ie, not pre-registered).
246  *
247  * Each rpcrdma_buffer has a list of free MWs anchored in rb_mrs. During
248  * call_allocate, rpcrdma_buffer_get() assigns one to each segment in
249  * an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep
250  * track of registration metadata while each RPC is pending.
251  * rpcrdma_deregister_external() uses this metadata to unmap and
252  * release these resources when an RPC is complete.
253  */
254 enum rpcrdma_frwr_state {
255 	FRWR_IS_INVALID,	/* ready to be used */
256 	FRWR_IS_VALID,		/* in use */
257 	FRWR_FLUSHED_FR,	/* flushed FASTREG WR */
258 	FRWR_FLUSHED_LI,	/* flushed LOCALINV WR */
259 };
260 
261 struct rpcrdma_frwr {
262 	struct ib_mr			*fr_mr;
263 	struct ib_cqe			fr_cqe;
264 	enum rpcrdma_frwr_state		fr_state;
265 	struct completion		fr_linv_done;
266 	union {
267 		struct ib_reg_wr	fr_regwr;
268 		struct ib_send_wr	fr_invwr;
269 	};
270 };
271 
272 struct rpcrdma_mr {
273 	struct list_head	mr_list;
274 	struct scatterlist	*mr_sg;
275 	int			mr_nents;
276 	enum dma_data_direction	mr_dir;
277 	struct rpcrdma_frwr	frwr;
278 	struct rpcrdma_xprt	*mr_xprt;
279 	u32			mr_handle;
280 	u32			mr_length;
281 	u64			mr_offset;
282 	struct work_struct	mr_recycle;
283 	struct list_head	mr_all;
284 };
285 
286 /*
287  * struct rpcrdma_req -- structure central to the request/reply sequence.
288  *
289  * N of these are associated with a transport instance, and stored in
290  * struct rpcrdma_buffer. N is the max number of outstanding requests.
291  *
292  * It includes pre-registered buffer memory for send AND recv.
293  * The recv buffer, however, is not owned by this structure, and
294  * is "donated" to the hardware when a recv is posted. When a
295  * reply is handled, the recv buffer used is given back to the
296  * struct rpcrdma_req associated with the request.
297  *
298  * In addition to the basic memory, this structure includes an array
299  * of iovs for send operations. The reason is that the iovs passed to
300  * ib_post_{send,recv} must not be modified until the work request
301  * completes.
302  */
303 
304 /* Maximum number of page-sized "segments" per chunk list to be
305  * registered or invalidated. Must handle a Reply chunk:
306  */
307 enum {
308 	RPCRDMA_MAX_IOV_SEGS	= 3,
309 	RPCRDMA_MAX_DATA_SEGS	= ((1 * 1024 * 1024) / PAGE_SIZE) + 1,
310 	RPCRDMA_MAX_SEGS	= RPCRDMA_MAX_DATA_SEGS +
311 				  RPCRDMA_MAX_IOV_SEGS,
312 };
313 
314 struct rpcrdma_mr_seg {		/* chunk descriptors */
315 	u32		mr_len;		/* length of chunk or segment */
316 	struct page	*mr_page;	/* owning page, if any */
317 	char		*mr_offset;	/* kva if no page, else offset */
318 };
319 
320 /* The Send SGE array is provisioned to send a maximum size
321  * inline request:
322  * - RPC-over-RDMA header
323  * - xdr_buf head iovec
324  * - RPCRDMA_MAX_INLINE bytes, in pages
325  * - xdr_buf tail iovec
326  *
327  * The actual number of array elements consumed by each RPC
328  * depends on the device's max_sge limit.
329  */
330 enum {
331 	RPCRDMA_MIN_SEND_SGES = 3,
332 	RPCRDMA_MAX_PAGE_SGES = RPCRDMA_MAX_INLINE >> PAGE_SHIFT,
333 	RPCRDMA_MAX_SEND_SGES = 1 + 1 + RPCRDMA_MAX_PAGE_SGES + 1,
334 };
335 
336 struct rpcrdma_buffer;
337 struct rpcrdma_req {
338 	struct list_head	rl_list;
339 	struct rpc_rqst		rl_slot;
340 	struct rpcrdma_buffer	*rl_buffer;
341 	struct rpcrdma_rep	*rl_reply;
342 	struct xdr_stream	rl_stream;
343 	struct xdr_buf		rl_hdrbuf;
344 	struct rpcrdma_sendctx	*rl_sendctx;
345 	struct rpcrdma_regbuf	*rl_rdmabuf;	/* xprt header */
346 	struct rpcrdma_regbuf	*rl_sendbuf;	/* rq_snd_buf */
347 	struct rpcrdma_regbuf	*rl_recvbuf;	/* rq_rcv_buf */
348 
349 	struct list_head	rl_all;
350 	unsigned long		rl_flags;
351 
352 	struct list_head	rl_registered;	/* registered segments */
353 	struct rpcrdma_mr_seg	rl_segments[RPCRDMA_MAX_SEGS];
354 };
355 
356 /* rl_flags */
357 enum {
358 	RPCRDMA_REQ_F_PENDING = 0,
359 	RPCRDMA_REQ_F_TX_RESOURCES,
360 };
361 
362 static inline struct rpcrdma_req *
363 rpcr_to_rdmar(const struct rpc_rqst *rqst)
364 {
365 	return container_of(rqst, struct rpcrdma_req, rl_slot);
366 }
367 
368 static inline void
369 rpcrdma_mr_push(struct rpcrdma_mr *mr, struct list_head *list)
370 {
371 	list_add_tail(&mr->mr_list, list);
372 }
373 
374 static inline struct rpcrdma_mr *
375 rpcrdma_mr_pop(struct list_head *list)
376 {
377 	struct rpcrdma_mr *mr;
378 
379 	mr = list_first_entry(list, struct rpcrdma_mr, mr_list);
380 	list_del_init(&mr->mr_list);
381 	return mr;
382 }
383 
384 /*
385  * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
386  * inline requests/replies, and client/server credits.
387  *
388  * One of these is associated with a transport instance
389  */
390 struct rpcrdma_buffer {
391 	spinlock_t		rb_mrlock;	/* protect rb_mrs list */
392 	struct list_head	rb_mrs;
393 	struct list_head	rb_all;
394 
395 	unsigned long		rb_sc_head;
396 	unsigned long		rb_sc_tail;
397 	unsigned long		rb_sc_last;
398 	struct rpcrdma_sendctx	**rb_sc_ctxs;
399 
400 	spinlock_t		rb_lock;	/* protect buf lists */
401 	struct list_head	rb_send_bufs;
402 	struct list_head	rb_recv_bufs;
403 	struct list_head	rb_allreqs;
404 
405 	unsigned long		rb_flags;
406 	u32			rb_max_requests;
407 	u32			rb_credits;	/* most recent credit grant */
408 
409 	u32			rb_bc_srv_max_requests;
410 	u32			rb_bc_max_requests;
411 
412 	struct workqueue_struct *rb_completion_wq;
413 	struct delayed_work	rb_refresh_worker;
414 };
415 
416 /* rb_flags */
417 enum {
418 	RPCRDMA_BUF_F_EMPTY_SCQ = 0,
419 };
420 
421 /*
422  * Internal structure for transport instance creation. This
423  * exists primarily for modularity.
424  *
425  * This data should be set with mount options
426  */
427 struct rpcrdma_create_data_internal {
428 	unsigned int	max_requests;	/* max requests (slots) in flight */
429 	unsigned int	rsize;		/* mount rsize - max read hdr+data */
430 	unsigned int	wsize;		/* mount wsize - max write hdr+data */
431 	unsigned int	inline_rsize;	/* max non-rdma read data payload */
432 	unsigned int	inline_wsize;	/* max non-rdma write data payload */
433 };
434 
435 /*
436  * Statistics for RPCRDMA
437  */
438 struct rpcrdma_stats {
439 	/* accessed when sending a call */
440 	unsigned long		read_chunk_count;
441 	unsigned long		write_chunk_count;
442 	unsigned long		reply_chunk_count;
443 	unsigned long long	total_rdma_request;
444 
445 	/* rarely accessed error counters */
446 	unsigned long long	pullup_copy_count;
447 	unsigned long		hardway_register_count;
448 	unsigned long		failed_marshal_count;
449 	unsigned long		bad_reply_count;
450 	unsigned long		mrs_recycled;
451 	unsigned long		mrs_orphaned;
452 	unsigned long		mrs_allocated;
453 	unsigned long		empty_sendctx_q;
454 
455 	/* accessed when receiving a reply */
456 	unsigned long long	total_rdma_reply;
457 	unsigned long long	fixup_copy_count;
458 	unsigned long		reply_waits_for_send;
459 	unsigned long		local_inv_needed;
460 	unsigned long		nomsg_call_count;
461 	unsigned long		bcall_count;
462 };
463 
464 /*
465  * RPCRDMA transport -- encapsulates the structures above for
466  * integration with RPC.
467  *
468  * The contained structures are embedded, not pointers,
469  * for convenience. This structure need not be visible externally.
470  *
471  * It is allocated and initialized during mount, and released
472  * during unmount.
473  */
474 struct rpcrdma_xprt {
475 	struct rpc_xprt		rx_xprt;
476 	struct rpcrdma_ia	rx_ia;
477 	struct rpcrdma_ep	rx_ep;
478 	struct rpcrdma_buffer	rx_buf;
479 	struct rpcrdma_create_data_internal rx_data;
480 	struct delayed_work	rx_connect_worker;
481 	struct rpcrdma_stats	rx_stats;
482 };
483 
484 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
485 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data)
486 
487 static inline const char *
488 rpcrdma_addrstr(const struct rpcrdma_xprt *r_xprt)
489 {
490 	return r_xprt->rx_xprt.address_strings[RPC_DISPLAY_ADDR];
491 }
492 
493 static inline const char *
494 rpcrdma_portstr(const struct rpcrdma_xprt *r_xprt)
495 {
496 	return r_xprt->rx_xprt.address_strings[RPC_DISPLAY_PORT];
497 }
498 
499 /* Setting this to 0 ensures interoperability with early servers.
500  * Setting this to 1 enhances certain unaligned read/write performance.
501  * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
502 extern int xprt_rdma_pad_optimize;
503 
504 /* This setting controls the hunt for a supported memory
505  * registration strategy.
506  */
507 extern unsigned int xprt_rdma_memreg_strategy;
508 
509 /*
510  * Interface Adapter calls - xprtrdma/verbs.c
511  */
512 int rpcrdma_ia_open(struct rpcrdma_xprt *xprt);
513 void rpcrdma_ia_remove(struct rpcrdma_ia *ia);
514 void rpcrdma_ia_close(struct rpcrdma_ia *);
515 
516 /*
517  * Endpoint calls - xprtrdma/verbs.c
518  */
519 int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *,
520 				struct rpcrdma_create_data_internal *);
521 void rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *);
522 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *);
523 void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *);
524 
525 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
526 				struct rpcrdma_req *);
527 
528 /*
529  * Buffer calls - xprtrdma/verbs.c
530  */
531 struct rpcrdma_req *rpcrdma_create_req(struct rpcrdma_xprt *);
532 void rpcrdma_req_destroy(struct rpcrdma_req *req);
533 int rpcrdma_buffer_create(struct rpcrdma_xprt *);
534 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
535 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_buffer *buf);
536 
537 struct rpcrdma_mr *rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt);
538 void rpcrdma_mr_put(struct rpcrdma_mr *mr);
539 void rpcrdma_mr_unmap_and_put(struct rpcrdma_mr *mr);
540 
541 static inline void
542 rpcrdma_mr_recycle(struct rpcrdma_mr *mr)
543 {
544 	schedule_work(&mr->mr_recycle);
545 }
546 
547 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
548 void rpcrdma_buffer_put(struct rpcrdma_req *);
549 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
550 
551 struct rpcrdma_regbuf *rpcrdma_alloc_regbuf(size_t, enum dma_data_direction,
552 					    gfp_t);
553 bool __rpcrdma_dma_map_regbuf(struct rpcrdma_ia *, struct rpcrdma_regbuf *);
554 void rpcrdma_free_regbuf(struct rpcrdma_regbuf *);
555 
556 static inline bool
557 rpcrdma_regbuf_is_mapped(struct rpcrdma_regbuf *rb)
558 {
559 	return rb->rg_device != NULL;
560 }
561 
562 static inline bool
563 rpcrdma_dma_map_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
564 {
565 	if (likely(rpcrdma_regbuf_is_mapped(rb)))
566 		return true;
567 	return __rpcrdma_dma_map_regbuf(ia, rb);
568 }
569 
570 /*
571  * Wrappers for chunk registration, shared by read/write chunk code.
572  */
573 
574 static inline enum dma_data_direction
575 rpcrdma_data_dir(bool writing)
576 {
577 	return writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
578 }
579 
580 /* Memory registration calls xprtrdma/frwr_ops.c
581  */
582 bool frwr_is_supported(struct rpcrdma_ia *);
583 int frwr_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep,
584 	      struct rpcrdma_create_data_internal *cdata);
585 int frwr_init_mr(struct rpcrdma_ia *ia, struct rpcrdma_mr *mr);
586 void frwr_release_mr(struct rpcrdma_mr *mr);
587 size_t frwr_maxpages(struct rpcrdma_xprt *r_xprt);
588 struct rpcrdma_mr_seg *frwr_map(struct rpcrdma_xprt *r_xprt,
589 				struct rpcrdma_mr_seg *seg,
590 				int nsegs, bool writing, __be32 xid,
591 				struct rpcrdma_mr **mr);
592 int frwr_send(struct rpcrdma_ia *ia, struct rpcrdma_req *req);
593 void frwr_reminv(struct rpcrdma_rep *rep, struct list_head *mrs);
594 void frwr_unmap_sync(struct rpcrdma_xprt *r_xprt,
595 		     struct list_head *mrs);
596 
597 /*
598  * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
599  */
600 
601 enum rpcrdma_chunktype {
602 	rpcrdma_noch = 0,
603 	rpcrdma_readch,
604 	rpcrdma_areadch,
605 	rpcrdma_writech,
606 	rpcrdma_replych
607 };
608 
609 int rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
610 			      struct rpcrdma_req *req, u32 hdrlen,
611 			      struct xdr_buf *xdr,
612 			      enum rpcrdma_chunktype rtype);
613 void rpcrdma_unmap_sendctx(struct rpcrdma_sendctx *sc);
614 int rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst);
615 void rpcrdma_set_max_header_sizes(struct rpcrdma_xprt *);
616 void rpcrdma_complete_rqst(struct rpcrdma_rep *rep);
617 void rpcrdma_reply_handler(struct rpcrdma_rep *rep);
618 void rpcrdma_release_rqst(struct rpcrdma_xprt *r_xprt,
619 			  struct rpcrdma_req *req);
620 void rpcrdma_deferred_completion(struct work_struct *work);
621 
622 static inline void rpcrdma_set_xdrlen(struct xdr_buf *xdr, size_t len)
623 {
624 	xdr->head[0].iov_len = len;
625 	xdr->len = len;
626 }
627 
628 /* RPC/RDMA module init - xprtrdma/transport.c
629  */
630 extern unsigned int xprt_rdma_max_inline_read;
631 void xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap);
632 void xprt_rdma_free_addresses(struct rpc_xprt *xprt);
633 void xprt_rdma_close(struct rpc_xprt *xprt);
634 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq);
635 int xprt_rdma_init(void);
636 void xprt_rdma_cleanup(void);
637 
638 /* Backchannel calls - xprtrdma/backchannel.c
639  */
640 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
641 int xprt_rdma_bc_setup(struct rpc_xprt *, unsigned int);
642 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *);
643 int rpcrdma_bc_post_recv(struct rpcrdma_xprt *, unsigned int);
644 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *, struct rpcrdma_rep *);
645 int xprt_rdma_bc_send_reply(struct rpc_rqst *rqst);
646 void xprt_rdma_bc_free_rqst(struct rpc_rqst *);
647 void xprt_rdma_bc_destroy(struct rpc_xprt *, unsigned int);
648 #endif	/* CONFIG_SUNRPC_BACKCHANNEL */
649 
650 extern struct xprt_class xprt_rdma_bc;
651 
652 #endif				/* _LINUX_SUNRPC_XPRT_RDMA_H */
653