xref: /openbmc/linux/net/sunrpc/xprtrdma/xprt_rdma.h (revision e639c869)
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_frmr_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_reminv_expected;
81 	bool			ri_implicit_roundup;
82 	enum ib_mr_type		ri_mrtype;
83 	unsigned long		ri_flags;
84 	struct ib_qp_attr	ri_qp_attr;
85 	struct ib_qp_init_attr	ri_qp_init_attr;
86 };
87 
88 enum {
89 	RPCRDMA_IAF_REMOVING = 0,
90 };
91 
92 /*
93  * RDMA Endpoint -- one per transport instance
94  */
95 
96 struct rpcrdma_ep {
97 	unsigned int		rep_send_count;
98 	unsigned int		rep_send_batch;
99 	int			rep_connected;
100 	struct ib_qp_init_attr	rep_attr;
101 	wait_queue_head_t 	rep_connect_wait;
102 	struct rpcrdma_connect_private	rep_cm_private;
103 	struct rdma_conn_param	rep_remote_cma;
104 	struct sockaddr_storage	rep_remote_addr;
105 	struct delayed_work	rep_connect_worker;
106 };
107 
108 /* Pre-allocate extra Work Requests for handling backward receives
109  * and sends. This is a fixed value because the Work Queues are
110  * allocated when the forward channel is set up.
111  */
112 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
113 #define RPCRDMA_BACKWARD_WRS		(8)
114 #else
115 #define RPCRDMA_BACKWARD_WRS		(0)
116 #endif
117 
118 /* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
119  *
120  * The below structure appears at the front of a large region of kmalloc'd
121  * memory, which always starts on a good alignment boundary.
122  */
123 
124 struct rpcrdma_regbuf {
125 	struct ib_sge		rg_iov;
126 	struct ib_device	*rg_device;
127 	enum dma_data_direction	rg_direction;
128 	__be32			rg_base[0] __attribute__ ((aligned(256)));
129 };
130 
131 static inline u64
132 rdmab_addr(struct rpcrdma_regbuf *rb)
133 {
134 	return rb->rg_iov.addr;
135 }
136 
137 static inline u32
138 rdmab_length(struct rpcrdma_regbuf *rb)
139 {
140 	return rb->rg_iov.length;
141 }
142 
143 static inline u32
144 rdmab_lkey(struct rpcrdma_regbuf *rb)
145 {
146 	return rb->rg_iov.lkey;
147 }
148 
149 static inline struct ib_device *
150 rdmab_device(struct rpcrdma_regbuf *rb)
151 {
152 	return rb->rg_device;
153 }
154 
155 #define RPCRDMA_DEF_GFP		(GFP_NOIO | __GFP_NOWARN)
156 
157 /* To ensure a transport can always make forward progress,
158  * the number of RDMA segments allowed in header chunk lists
159  * is capped at 8. This prevents less-capable devices and
160  * memory registrations from overrunning the Send buffer
161  * while building chunk lists.
162  *
163  * Elements of the Read list take up more room than the
164  * Write list or Reply chunk. 8 read segments means the Read
165  * list (or Write list or Reply chunk) cannot consume more
166  * than
167  *
168  * ((8 + 2) * read segment size) + 1 XDR words, or 244 bytes.
169  *
170  * And the fixed part of the header is another 24 bytes.
171  *
172  * The smallest inline threshold is 1024 bytes, ensuring that
173  * at least 750 bytes are available for RPC messages.
174  */
175 enum {
176 	RPCRDMA_MAX_HDR_SEGS = 8,
177 	RPCRDMA_HDRBUF_SIZE = 256,
178 };
179 
180 /*
181  * struct rpcrdma_rep -- this structure encapsulates state required
182  * to receive and complete an RPC Reply, asychronously. It needs
183  * several pieces of state:
184  *
185  *   o receive buffer and ib_sge (donated to provider)
186  *   o status of receive (success or not, length, inv rkey)
187  *   o bookkeeping state to get run by reply handler (XDR stream)
188  *
189  * These structures are allocated during transport initialization.
190  * N of these are associated with a transport instance, managed by
191  * struct rpcrdma_buffer. N is the max number of outstanding RPCs.
192  */
193 
194 struct rpcrdma_rep {
195 	struct ib_cqe		rr_cqe;
196 	__be32			rr_xid;
197 	__be32			rr_vers;
198 	__be32			rr_proc;
199 	int			rr_wc_flags;
200 	u32			rr_inv_rkey;
201 	struct rpcrdma_regbuf	*rr_rdmabuf;
202 	struct rpcrdma_xprt	*rr_rxprt;
203 	struct work_struct	rr_work;
204 	struct xdr_buf		rr_hdrbuf;
205 	struct xdr_stream	rr_stream;
206 	struct rpc_rqst		*rr_rqst;
207 	struct list_head	rr_list;
208 	struct ib_recv_wr	rr_recv_wr;
209 };
210 
211 /* struct rpcrdma_sendctx - DMA mapped SGEs to unmap after Send completes
212  */
213 struct rpcrdma_req;
214 struct rpcrdma_xprt;
215 struct rpcrdma_sendctx {
216 	struct ib_send_wr	sc_wr;
217 	struct ib_cqe		sc_cqe;
218 	struct rpcrdma_xprt	*sc_xprt;
219 	struct rpcrdma_req	*sc_req;
220 	unsigned int		sc_unmap_count;
221 	struct ib_sge		sc_sges[];
222 };
223 
224 /* Limit the number of SGEs that can be unmapped during one
225  * Send completion. This caps the amount of work a single
226  * completion can do before returning to the provider.
227  *
228  * Setting this to zero disables Send completion batching.
229  */
230 enum {
231 	RPCRDMA_MAX_SEND_BATCH = 7,
232 };
233 
234 /*
235  * struct rpcrdma_mw - external memory region metadata
236  *
237  * An external memory region is any buffer or page that is registered
238  * on the fly (ie, not pre-registered).
239  *
240  * Each rpcrdma_buffer has a list of free MWs anchored in rb_mws. During
241  * call_allocate, rpcrdma_buffer_get() assigns one to each segment in
242  * an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep
243  * track of registration metadata while each RPC is pending.
244  * rpcrdma_deregister_external() uses this metadata to unmap and
245  * release these resources when an RPC is complete.
246  */
247 enum rpcrdma_frmr_state {
248 	FRMR_IS_INVALID,	/* ready to be used */
249 	FRMR_IS_VALID,		/* in use */
250 	FRMR_FLUSHED_FR,	/* flushed FASTREG WR */
251 	FRMR_FLUSHED_LI,	/* flushed LOCALINV WR */
252 };
253 
254 struct rpcrdma_frmr {
255 	struct ib_mr			*fr_mr;
256 	struct ib_cqe			fr_cqe;
257 	enum rpcrdma_frmr_state		fr_state;
258 	struct completion		fr_linv_done;
259 	union {
260 		struct ib_reg_wr	fr_regwr;
261 		struct ib_send_wr	fr_invwr;
262 	};
263 };
264 
265 struct rpcrdma_fmr {
266 	struct ib_fmr		*fm_mr;
267 	u64			*fm_physaddrs;
268 };
269 
270 struct rpcrdma_mw {
271 	struct list_head	mw_list;
272 	struct scatterlist	*mw_sg;
273 	int			mw_nents;
274 	enum dma_data_direction	mw_dir;
275 	unsigned long		mw_flags;
276 	union {
277 		struct rpcrdma_fmr	fmr;
278 		struct rpcrdma_frmr	frmr;
279 	};
280 	struct rpcrdma_xprt	*mw_xprt;
281 	u32			mw_handle;
282 	u32			mw_length;
283 	u64			mw_offset;
284 	struct list_head	mw_all;
285 };
286 
287 /* mw_flags */
288 enum {
289 	RPCRDMA_MW_F_RI		= 1,
290 };
291 
292 /*
293  * struct rpcrdma_req -- structure central to the request/reply sequence.
294  *
295  * N of these are associated with a transport instance, and stored in
296  * struct rpcrdma_buffer. N is the max number of outstanding requests.
297  *
298  * It includes pre-registered buffer memory for send AND recv.
299  * The recv buffer, however, is not owned by this structure, and
300  * is "donated" to the hardware when a recv is posted. When a
301  * reply is handled, the recv buffer used is given back to the
302  * struct rpcrdma_req associated with the request.
303  *
304  * In addition to the basic memory, this structure includes an array
305  * of iovs for send operations. The reason is that the iovs passed to
306  * ib_post_{send,recv} must not be modified until the work request
307  * completes.
308  */
309 
310 /* Maximum number of page-sized "segments" per chunk list to be
311  * registered or invalidated. Must handle a Reply chunk:
312  */
313 enum {
314 	RPCRDMA_MAX_IOV_SEGS	= 3,
315 	RPCRDMA_MAX_DATA_SEGS	= ((1 * 1024 * 1024) / PAGE_SIZE) + 1,
316 	RPCRDMA_MAX_SEGS	= RPCRDMA_MAX_DATA_SEGS +
317 				  RPCRDMA_MAX_IOV_SEGS,
318 };
319 
320 struct rpcrdma_mr_seg {		/* chunk descriptors */
321 	u32		mr_len;		/* length of chunk or segment */
322 	struct page	*mr_page;	/* owning page, if any */
323 	char		*mr_offset;	/* kva if no page, else offset */
324 };
325 
326 /* The Send SGE array is provisioned to send a maximum size
327  * inline request:
328  * - RPC-over-RDMA header
329  * - xdr_buf head iovec
330  * - RPCRDMA_MAX_INLINE bytes, in pages
331  * - xdr_buf tail iovec
332  *
333  * The actual number of array elements consumed by each RPC
334  * depends on the device's max_sge limit.
335  */
336 enum {
337 	RPCRDMA_MIN_SEND_SGES = 3,
338 	RPCRDMA_MAX_PAGE_SGES = RPCRDMA_MAX_INLINE >> PAGE_SHIFT,
339 	RPCRDMA_MAX_SEND_SGES = 1 + 1 + RPCRDMA_MAX_PAGE_SGES + 1,
340 };
341 
342 struct rpcrdma_buffer;
343 struct rpcrdma_req {
344 	struct list_head	rl_list;
345 	unsigned int		rl_connect_cookie;
346 	struct rpcrdma_buffer	*rl_buffer;
347 	struct rpcrdma_rep	*rl_reply;
348 	struct xdr_stream	rl_stream;
349 	struct xdr_buf		rl_hdrbuf;
350 	struct rpcrdma_sendctx	*rl_sendctx;
351 	struct rpcrdma_regbuf	*rl_rdmabuf;	/* xprt header */
352 	struct rpcrdma_regbuf	*rl_sendbuf;	/* rq_snd_buf */
353 	struct rpcrdma_regbuf	*rl_recvbuf;	/* rq_rcv_buf */
354 
355 	struct list_head	rl_all;
356 	unsigned long		rl_flags;
357 
358 	struct list_head	rl_registered;	/* registered segments */
359 	struct rpcrdma_mr_seg	rl_segments[RPCRDMA_MAX_SEGS];
360 };
361 
362 /* rl_flags */
363 enum {
364 	RPCRDMA_REQ_F_BACKCHANNEL = 0,
365 	RPCRDMA_REQ_F_PENDING,
366 	RPCRDMA_REQ_F_TX_RESOURCES,
367 };
368 
369 static inline void
370 rpcrdma_set_xprtdata(struct rpc_rqst *rqst, struct rpcrdma_req *req)
371 {
372 	rqst->rq_xprtdata = req;
373 }
374 
375 static inline struct rpcrdma_req *
376 rpcr_to_rdmar(struct rpc_rqst *rqst)
377 {
378 	return rqst->rq_xprtdata;
379 }
380 
381 static inline void
382 rpcrdma_push_mw(struct rpcrdma_mw *mw, struct list_head *list)
383 {
384 	list_add_tail(&mw->mw_list, list);
385 }
386 
387 static inline struct rpcrdma_mw *
388 rpcrdma_pop_mw(struct list_head *list)
389 {
390 	struct rpcrdma_mw *mw;
391 
392 	mw = list_first_entry(list, struct rpcrdma_mw, mw_list);
393 	list_del(&mw->mw_list);
394 	return mw;
395 }
396 
397 /*
398  * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
399  * inline requests/replies, and client/server credits.
400  *
401  * One of these is associated with a transport instance
402  */
403 struct rpcrdma_buffer {
404 	spinlock_t		rb_mwlock;	/* protect rb_mws list */
405 	struct list_head	rb_mws;
406 	struct list_head	rb_all;
407 
408 	unsigned long		rb_sc_head;
409 	unsigned long		rb_sc_tail;
410 	unsigned long		rb_sc_last;
411 	struct rpcrdma_sendctx	**rb_sc_ctxs;
412 
413 	spinlock_t		rb_lock;	/* protect buf lists */
414 	int			rb_send_count, rb_recv_count;
415 	struct list_head	rb_send_bufs;
416 	struct list_head	rb_recv_bufs;
417 	u32			rb_max_requests;
418 	u32			rb_credits;	/* most recent credit grant */
419 
420 	u32			rb_bc_srv_max_requests;
421 	spinlock_t		rb_reqslock;	/* protect rb_allreqs */
422 	struct list_head	rb_allreqs;
423 
424 	u32			rb_bc_max_requests;
425 
426 	spinlock_t		rb_recovery_lock; /* protect rb_stale_mrs */
427 	struct list_head	rb_stale_mrs;
428 	struct delayed_work	rb_recovery_worker;
429 	struct delayed_work	rb_refresh_worker;
430 };
431 #define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia)
432 
433 /*
434  * Internal structure for transport instance creation. This
435  * exists primarily for modularity.
436  *
437  * This data should be set with mount options
438  */
439 struct rpcrdma_create_data_internal {
440 	struct sockaddr_storage	addr;	/* RDMA server address */
441 	unsigned int	max_requests;	/* max requests (slots) in flight */
442 	unsigned int	rsize;		/* mount rsize - max read hdr+data */
443 	unsigned int	wsize;		/* mount wsize - max write hdr+data */
444 	unsigned int	inline_rsize;	/* max non-rdma read data payload */
445 	unsigned int	inline_wsize;	/* max non-rdma write data payload */
446 	unsigned int	padding;	/* non-rdma write header padding */
447 };
448 
449 /*
450  * Statistics for RPCRDMA
451  */
452 struct rpcrdma_stats {
453 	/* accessed when sending a call */
454 	unsigned long		read_chunk_count;
455 	unsigned long		write_chunk_count;
456 	unsigned long		reply_chunk_count;
457 	unsigned long long	total_rdma_request;
458 
459 	/* rarely accessed error counters */
460 	unsigned long long	pullup_copy_count;
461 	unsigned long		hardway_register_count;
462 	unsigned long		failed_marshal_count;
463 	unsigned long		bad_reply_count;
464 	unsigned long		mrs_recovered;
465 	unsigned long		mrs_orphaned;
466 	unsigned long		mrs_allocated;
467 	unsigned long		empty_sendctx_q;
468 
469 	/* accessed when receiving a reply */
470 	unsigned long long	total_rdma_reply;
471 	unsigned long long	fixup_copy_count;
472 	unsigned long		reply_waits_for_send;
473 	unsigned long		local_inv_needed;
474 	unsigned long		nomsg_call_count;
475 	unsigned long		bcall_count;
476 };
477 
478 /*
479  * Per-registration mode operations
480  */
481 struct rpcrdma_xprt;
482 struct rpcrdma_memreg_ops {
483 	struct rpcrdma_mr_seg *
484 			(*ro_map)(struct rpcrdma_xprt *,
485 				  struct rpcrdma_mr_seg *, int, bool,
486 				  struct rpcrdma_mw **);
487 	void		(*ro_unmap_sync)(struct rpcrdma_xprt *,
488 					 struct list_head *);
489 	void		(*ro_recover_mr)(struct rpcrdma_mw *);
490 	int		(*ro_open)(struct rpcrdma_ia *,
491 				   struct rpcrdma_ep *,
492 				   struct rpcrdma_create_data_internal *);
493 	size_t		(*ro_maxpages)(struct rpcrdma_xprt *);
494 	int		(*ro_init_mr)(struct rpcrdma_ia *,
495 				      struct rpcrdma_mw *);
496 	void		(*ro_release_mr)(struct rpcrdma_mw *);
497 	const char	*ro_displayname;
498 	const int	ro_send_w_inv_ok;
499 };
500 
501 extern const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops;
502 extern const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops;
503 
504 /*
505  * RPCRDMA transport -- encapsulates the structures above for
506  * integration with RPC.
507  *
508  * The contained structures are embedded, not pointers,
509  * for convenience. This structure need not be visible externally.
510  *
511  * It is allocated and initialized during mount, and released
512  * during unmount.
513  */
514 struct rpcrdma_xprt {
515 	struct rpc_xprt		rx_xprt;
516 	struct rpcrdma_ia	rx_ia;
517 	struct rpcrdma_ep	rx_ep;
518 	struct rpcrdma_buffer	rx_buf;
519 	struct rpcrdma_create_data_internal rx_data;
520 	struct delayed_work	rx_connect_worker;
521 	struct rpcrdma_stats	rx_stats;
522 };
523 
524 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
525 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data)
526 
527 /* Setting this to 0 ensures interoperability with early servers.
528  * Setting this to 1 enhances certain unaligned read/write performance.
529  * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
530 extern int xprt_rdma_pad_optimize;
531 
532 /* This setting controls the hunt for a supported memory
533  * registration strategy.
534  */
535 extern unsigned int xprt_rdma_memreg_strategy;
536 
537 /*
538  * Interface Adapter calls - xprtrdma/verbs.c
539  */
540 int rpcrdma_ia_open(struct rpcrdma_xprt *xprt, struct sockaddr *addr);
541 void rpcrdma_ia_remove(struct rpcrdma_ia *ia);
542 void rpcrdma_ia_close(struct rpcrdma_ia *);
543 bool frwr_is_supported(struct rpcrdma_ia *);
544 bool fmr_is_supported(struct rpcrdma_ia *);
545 
546 extern struct workqueue_struct *rpcrdma_receive_wq;
547 
548 /*
549  * Endpoint calls - xprtrdma/verbs.c
550  */
551 int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *,
552 				struct rpcrdma_create_data_internal *);
553 void rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *);
554 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *);
555 void rpcrdma_conn_func(struct rpcrdma_ep *ep);
556 void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *);
557 
558 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
559 				struct rpcrdma_req *);
560 int rpcrdma_ep_post_recv(struct rpcrdma_ia *, struct rpcrdma_rep *);
561 
562 /*
563  * Buffer calls - xprtrdma/verbs.c
564  */
565 struct rpcrdma_req *rpcrdma_create_req(struct rpcrdma_xprt *);
566 struct rpcrdma_rep *rpcrdma_create_rep(struct rpcrdma_xprt *);
567 void rpcrdma_destroy_req(struct rpcrdma_req *);
568 int rpcrdma_buffer_create(struct rpcrdma_xprt *);
569 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
570 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_buffer *buf);
571 void rpcrdma_sendctx_put_locked(struct rpcrdma_sendctx *sc);
572 
573 struct rpcrdma_mw *rpcrdma_get_mw(struct rpcrdma_xprt *);
574 void rpcrdma_put_mw(struct rpcrdma_xprt *, struct rpcrdma_mw *);
575 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
576 void rpcrdma_buffer_put(struct rpcrdma_req *);
577 void rpcrdma_recv_buffer_get(struct rpcrdma_req *);
578 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
579 
580 void rpcrdma_defer_mr_recovery(struct rpcrdma_mw *);
581 
582 struct rpcrdma_regbuf *rpcrdma_alloc_regbuf(size_t, enum dma_data_direction,
583 					    gfp_t);
584 bool __rpcrdma_dma_map_regbuf(struct rpcrdma_ia *, struct rpcrdma_regbuf *);
585 void rpcrdma_free_regbuf(struct rpcrdma_regbuf *);
586 
587 static inline bool
588 rpcrdma_regbuf_is_mapped(struct rpcrdma_regbuf *rb)
589 {
590 	return rb->rg_device != NULL;
591 }
592 
593 static inline bool
594 rpcrdma_dma_map_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
595 {
596 	if (likely(rpcrdma_regbuf_is_mapped(rb)))
597 		return true;
598 	return __rpcrdma_dma_map_regbuf(ia, rb);
599 }
600 
601 int rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt *, unsigned int);
602 
603 int rpcrdma_alloc_wq(void);
604 void rpcrdma_destroy_wq(void);
605 
606 /*
607  * Wrappers for chunk registration, shared by read/write chunk code.
608  */
609 
610 static inline enum dma_data_direction
611 rpcrdma_data_dir(bool writing)
612 {
613 	return writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
614 }
615 
616 /*
617  * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
618  */
619 
620 enum rpcrdma_chunktype {
621 	rpcrdma_noch = 0,
622 	rpcrdma_readch,
623 	rpcrdma_areadch,
624 	rpcrdma_writech,
625 	rpcrdma_replych
626 };
627 
628 int rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
629 			      struct rpcrdma_req *req, u32 hdrlen,
630 			      struct xdr_buf *xdr,
631 			      enum rpcrdma_chunktype rtype);
632 void rpcrdma_unmap_sendctx(struct rpcrdma_sendctx *sc);
633 int rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst);
634 void rpcrdma_set_max_header_sizes(struct rpcrdma_xprt *);
635 void rpcrdma_complete_rqst(struct rpcrdma_rep *rep);
636 void rpcrdma_reply_handler(struct rpcrdma_rep *rep);
637 void rpcrdma_release_rqst(struct rpcrdma_xprt *r_xprt,
638 			  struct rpcrdma_req *req);
639 void rpcrdma_deferred_completion(struct work_struct *work);
640 
641 static inline void rpcrdma_set_xdrlen(struct xdr_buf *xdr, size_t len)
642 {
643 	xdr->head[0].iov_len = len;
644 	xdr->len = len;
645 }
646 
647 /* RPC/RDMA module init - xprtrdma/transport.c
648  */
649 extern unsigned int xprt_rdma_max_inline_read;
650 void xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap);
651 void xprt_rdma_free_addresses(struct rpc_xprt *xprt);
652 void rpcrdma_connect_worker(struct work_struct *work);
653 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq);
654 int xprt_rdma_init(void);
655 void xprt_rdma_cleanup(void);
656 
657 /* Backchannel calls - xprtrdma/backchannel.c
658  */
659 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
660 int xprt_rdma_bc_setup(struct rpc_xprt *, unsigned int);
661 int xprt_rdma_bc_up(struct svc_serv *, struct net *);
662 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *);
663 int rpcrdma_bc_post_recv(struct rpcrdma_xprt *, unsigned int);
664 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *, struct rpcrdma_rep *);
665 int rpcrdma_bc_marshal_reply(struct rpc_rqst *);
666 void xprt_rdma_bc_free_rqst(struct rpc_rqst *);
667 void xprt_rdma_bc_destroy(struct rpc_xprt *, unsigned int);
668 #endif	/* CONFIG_SUNRPC_BACKCHANNEL */
669 
670 extern struct xprt_class xprt_rdma_bc;
671 
672 #endif				/* _LINUX_SUNRPC_XPRT_RDMA_H */
673