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 const struct rpcrdma_memreg_ops *ri_ops;
70 struct ib_device *ri_device;
71 struct rdma_cm_id *ri_id;
72 struct ib_pd *ri_pd;
73 struct completion ri_done;
74 struct completion ri_remove_done;
75 int ri_async_rc;
76 unsigned int ri_max_segs;
77 unsigned int ri_max_frwr_depth;
78 unsigned int ri_max_inline_write;
79 unsigned int ri_max_inline_read;
80 unsigned int ri_max_send_sges;
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 delayed_work rep_connect_worker;
105 };
106
107 /* Pre-allocate extra Work Requests for handling backward receives
108 * and sends. This is a fixed value because the Work Queues are
109 * allocated when the forward channel is set up.
110 */
111 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
112 #define RPCRDMA_BACKWARD_WRS (8)
113 #else
114 #define RPCRDMA_BACKWARD_WRS (0)
115 #endif
116
117 /* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
118 *
119 * The below structure appears at the front of a large region of kmalloc'd
120 * memory, which always starts on a good alignment boundary.
121 */
122
123 struct rpcrdma_regbuf {
124 struct ib_sge rg_iov;
125 struct ib_device *rg_device;
126 enum dma_data_direction rg_direction;
127 __be32 rg_base[0] __attribute__ ((aligned(256)));
128 };
129
130 static inline u64
rdmab_addr(struct rpcrdma_regbuf * rb)131 rdmab_addr(struct rpcrdma_regbuf *rb)
132 {
133 return rb->rg_iov.addr;
134 }
135
136 static inline u32
rdmab_length(struct rpcrdma_regbuf * rb)137 rdmab_length(struct rpcrdma_regbuf *rb)
138 {
139 return rb->rg_iov.length;
140 }
141
142 static inline u32
rdmab_lkey(struct rpcrdma_regbuf * rb)143 rdmab_lkey(struct rpcrdma_regbuf *rb)
144 {
145 return rb->rg_iov.lkey;
146 }
147
148 static inline struct ib_device *
rdmab_device(struct rpcrdma_regbuf * rb)149 rdmab_device(struct rpcrdma_regbuf *rb)
150 {
151 return rb->rg_device;
152 }
153
154 #define RPCRDMA_DEF_GFP (GFP_NOIO | __GFP_NOWARN)
155
156 /* To ensure a transport can always make forward progress,
157 * the number of RDMA segments allowed in header chunk lists
158 * is capped at 8. This prevents less-capable devices and
159 * memory registrations from overrunning the Send buffer
160 * while building chunk lists.
161 *
162 * Elements of the Read list take up more room than the
163 * Write list or Reply chunk. 8 read segments means the Read
164 * list (or Write list or Reply chunk) cannot consume more
165 * than
166 *
167 * ((8 + 2) * read segment size) + 1 XDR words, or 244 bytes.
168 *
169 * And the fixed part of the header is another 24 bytes.
170 *
171 * The smallest inline threshold is 1024 bytes, ensuring that
172 * at least 750 bytes are available for RPC messages.
173 */
174 enum {
175 RPCRDMA_MAX_HDR_SEGS = 8,
176 RPCRDMA_HDRBUF_SIZE = 256,
177 };
178
179 /*
180 * struct rpcrdma_rep -- this structure encapsulates state required
181 * to receive and complete an RPC Reply, asychronously. It needs
182 * several pieces of state:
183 *
184 * o receive buffer and ib_sge (donated to provider)
185 * o status of receive (success or not, length, inv rkey)
186 * o bookkeeping state to get run by reply handler (XDR stream)
187 *
188 * These structures are allocated during transport initialization.
189 * N of these are associated with a transport instance, managed by
190 * struct rpcrdma_buffer. N is the max number of outstanding RPCs.
191 */
192
193 struct rpcrdma_rep {
194 struct ib_cqe rr_cqe;
195 __be32 rr_xid;
196 __be32 rr_vers;
197 __be32 rr_proc;
198 int rr_wc_flags;
199 u32 rr_inv_rkey;
200 bool rr_temp;
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_mr - 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_mrs. 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_frwr_state {
248 FRWR_IS_INVALID, /* ready to be used */
249 FRWR_IS_VALID, /* in use */
250 FRWR_FLUSHED_FR, /* flushed FASTREG WR */
251 FRWR_FLUSHED_LI, /* flushed LOCALINV WR */
252 };
253
254 struct rpcrdma_frwr {
255 struct ib_mr *fr_mr;
256 struct ib_cqe fr_cqe;
257 enum rpcrdma_frwr_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_mr {
271 struct list_head mr_list;
272 struct scatterlist *mr_sg;
273 int mr_nents;
274 enum dma_data_direction mr_dir;
275 union {
276 struct rpcrdma_fmr fmr;
277 struct rpcrdma_frwr frwr;
278 };
279 struct rpcrdma_xprt *mr_xprt;
280 u32 mr_handle;
281 u32 mr_length;
282 u64 mr_offset;
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 *
rpcr_to_rdmar(const struct rpc_rqst * rqst)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
rpcrdma_mr_push(struct rpcrdma_mr * mr,struct list_head * list)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 *
rpcrdma_mr_pop(struct list_head * list)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 unsigned long rb_flags;
404 u32 rb_max_requests;
405 u32 rb_credits; /* most recent credit grant */
406 int rb_posted_receives;
407
408 u32 rb_bc_srv_max_requests;
409 spinlock_t rb_reqslock; /* protect rb_allreqs */
410 struct list_head rb_allreqs;
411
412 u32 rb_bc_max_requests;
413
414 spinlock_t rb_recovery_lock; /* protect rb_stale_mrs */
415 struct list_head rb_stale_mrs;
416 struct delayed_work rb_recovery_worker;
417 struct delayed_work rb_refresh_worker;
418 };
419 #define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia)
420
421 /* rb_flags */
422 enum {
423 RPCRDMA_BUF_F_EMPTY_SCQ = 0,
424 };
425
426 /*
427 * Internal structure for transport instance creation. This
428 * exists primarily for modularity.
429 *
430 * This data should be set with mount options
431 */
432 struct rpcrdma_create_data_internal {
433 unsigned int max_requests; /* max requests (slots) in flight */
434 unsigned int rsize; /* mount rsize - max read hdr+data */
435 unsigned int wsize; /* mount wsize - max write hdr+data */
436 unsigned int inline_rsize; /* max non-rdma read data payload */
437 unsigned int inline_wsize; /* max non-rdma write data payload */
438 };
439
440 /*
441 * Statistics for RPCRDMA
442 */
443 struct rpcrdma_stats {
444 /* accessed when sending a call */
445 unsigned long read_chunk_count;
446 unsigned long write_chunk_count;
447 unsigned long reply_chunk_count;
448 unsigned long long total_rdma_request;
449
450 /* rarely accessed error counters */
451 unsigned long long pullup_copy_count;
452 unsigned long hardway_register_count;
453 unsigned long failed_marshal_count;
454 unsigned long bad_reply_count;
455 unsigned long mrs_recovered;
456 unsigned long mrs_orphaned;
457 unsigned long mrs_allocated;
458 unsigned long empty_sendctx_q;
459
460 /* accessed when receiving a reply */
461 unsigned long long total_rdma_reply;
462 unsigned long long fixup_copy_count;
463 unsigned long reply_waits_for_send;
464 unsigned long local_inv_needed;
465 unsigned long nomsg_call_count;
466 unsigned long bcall_count;
467 };
468
469 /*
470 * Per-registration mode operations
471 */
472 struct rpcrdma_xprt;
473 struct rpcrdma_memreg_ops {
474 struct rpcrdma_mr_seg *
475 (*ro_map)(struct rpcrdma_xprt *,
476 struct rpcrdma_mr_seg *, int, bool,
477 struct rpcrdma_mr **);
478 int (*ro_send)(struct rpcrdma_ia *ia,
479 struct rpcrdma_req *req);
480 void (*ro_reminv)(struct rpcrdma_rep *rep,
481 struct list_head *mrs);
482 void (*ro_unmap_sync)(struct rpcrdma_xprt *,
483 struct list_head *);
484 void (*ro_recover_mr)(struct rpcrdma_mr *mr);
485 int (*ro_open)(struct rpcrdma_ia *,
486 struct rpcrdma_ep *,
487 struct rpcrdma_create_data_internal *);
488 size_t (*ro_maxpages)(struct rpcrdma_xprt *);
489 int (*ro_init_mr)(struct rpcrdma_ia *,
490 struct rpcrdma_mr *);
491 void (*ro_release_mr)(struct rpcrdma_mr *mr);
492 const char *ro_displayname;
493 const int ro_send_w_inv_ok;
494 };
495
496 extern const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops;
497 extern const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops;
498
499 /*
500 * RPCRDMA transport -- encapsulates the structures above for
501 * integration with RPC.
502 *
503 * The contained structures are embedded, not pointers,
504 * for convenience. This structure need not be visible externally.
505 *
506 * It is allocated and initialized during mount, and released
507 * during unmount.
508 */
509 struct rpcrdma_xprt {
510 struct rpc_xprt rx_xprt;
511 struct rpcrdma_ia rx_ia;
512 struct rpcrdma_ep rx_ep;
513 struct rpcrdma_buffer rx_buf;
514 struct rpcrdma_create_data_internal rx_data;
515 struct delayed_work rx_connect_worker;
516 struct rpcrdma_stats rx_stats;
517 };
518
519 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
520 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data)
521
522 static inline const char *
rpcrdma_addrstr(const struct rpcrdma_xprt * r_xprt)523 rpcrdma_addrstr(const struct rpcrdma_xprt *r_xprt)
524 {
525 return r_xprt->rx_xprt.address_strings[RPC_DISPLAY_ADDR];
526 }
527
528 static inline const char *
rpcrdma_portstr(const struct rpcrdma_xprt * r_xprt)529 rpcrdma_portstr(const struct rpcrdma_xprt *r_xprt)
530 {
531 return r_xprt->rx_xprt.address_strings[RPC_DISPLAY_PORT];
532 }
533
534 /* Setting this to 0 ensures interoperability with early servers.
535 * Setting this to 1 enhances certain unaligned read/write performance.
536 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
537 extern int xprt_rdma_pad_optimize;
538
539 /* This setting controls the hunt for a supported memory
540 * registration strategy.
541 */
542 extern unsigned int xprt_rdma_memreg_strategy;
543
544 /*
545 * Interface Adapter calls - xprtrdma/verbs.c
546 */
547 int rpcrdma_ia_open(struct rpcrdma_xprt *xprt);
548 void rpcrdma_ia_remove(struct rpcrdma_ia *ia);
549 void rpcrdma_ia_close(struct rpcrdma_ia *);
550 bool frwr_is_supported(struct rpcrdma_ia *);
551 bool fmr_is_supported(struct rpcrdma_ia *);
552
553 extern struct workqueue_struct *rpcrdma_receive_wq;
554
555 /*
556 * Endpoint calls - xprtrdma/verbs.c
557 */
558 int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *,
559 struct rpcrdma_create_data_internal *);
560 void rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *);
561 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *);
562 void rpcrdma_conn_func(struct rpcrdma_ep *ep);
563 void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *);
564
565 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
566 struct rpcrdma_req *);
567 void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp);
568
569 /*
570 * Buffer calls - xprtrdma/verbs.c
571 */
572 struct rpcrdma_req *rpcrdma_create_req(struct rpcrdma_xprt *);
573 void rpcrdma_destroy_req(struct rpcrdma_req *);
574 int rpcrdma_buffer_create(struct rpcrdma_xprt *);
575 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
576 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_buffer *buf);
577
578 struct rpcrdma_mr *rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt);
579 void rpcrdma_mr_put(struct rpcrdma_mr *mr);
580 void rpcrdma_mr_unmap_and_put(struct rpcrdma_mr *mr);
581 void rpcrdma_mr_defer_recovery(struct rpcrdma_mr *mr);
582
583 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
584 void rpcrdma_buffer_put(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
rpcrdma_regbuf_is_mapped(struct rpcrdma_regbuf * rb)593 rpcrdma_regbuf_is_mapped(struct rpcrdma_regbuf *rb)
594 {
595 return rb->rg_device != NULL;
596 }
597
598 static inline bool
rpcrdma_dma_map_regbuf(struct rpcrdma_ia * ia,struct rpcrdma_regbuf * rb)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_alloc_wq(void);
607 void rpcrdma_destroy_wq(void);
608
609 /*
610 * Wrappers for chunk registration, shared by read/write chunk code.
611 */
612
613 static inline enum dma_data_direction
rpcrdma_data_dir(bool writing)614 rpcrdma_data_dir(bool writing)
615 {
616 return writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
617 }
618
619 /*
620 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
621 */
622
623 enum rpcrdma_chunktype {
624 rpcrdma_noch = 0,
625 rpcrdma_readch,
626 rpcrdma_areadch,
627 rpcrdma_writech,
628 rpcrdma_replych
629 };
630
631 int rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
632 struct rpcrdma_req *req, u32 hdrlen,
633 struct xdr_buf *xdr,
634 enum rpcrdma_chunktype rtype);
635 void rpcrdma_unmap_sendctx(struct rpcrdma_sendctx *sc);
636 int rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst);
637 void rpcrdma_set_max_header_sizes(struct rpcrdma_xprt *);
638 void rpcrdma_complete_rqst(struct rpcrdma_rep *rep);
639 void rpcrdma_reply_handler(struct rpcrdma_rep *rep);
640 void rpcrdma_release_rqst(struct rpcrdma_xprt *r_xprt,
641 struct rpcrdma_req *req);
642 void rpcrdma_deferred_completion(struct work_struct *work);
643
rpcrdma_set_xdrlen(struct xdr_buf * xdr,size_t len)644 static inline void rpcrdma_set_xdrlen(struct xdr_buf *xdr, size_t len)
645 {
646 xdr->head[0].iov_len = len;
647 xdr->len = len;
648 }
649
650 /* RPC/RDMA module init - xprtrdma/transport.c
651 */
652 extern unsigned int xprt_rdma_max_inline_read;
653 void xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap);
654 void xprt_rdma_free_addresses(struct rpc_xprt *xprt);
655 void rpcrdma_connect_worker(struct work_struct *work);
656 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq);
657 int xprt_rdma_init(void);
658 void xprt_rdma_cleanup(void);
659
660 /* Backchannel calls - xprtrdma/backchannel.c
661 */
662 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
663 int xprt_rdma_bc_setup(struct rpc_xprt *, unsigned int);
664 int xprt_rdma_bc_up(struct svc_serv *, struct net *);
665 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *);
666 int rpcrdma_bc_post_recv(struct rpcrdma_xprt *, unsigned int);
667 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *, struct rpcrdma_rep *);
668 int xprt_rdma_bc_send_reply(struct rpc_rqst *rqst);
669 void xprt_rdma_bc_free_rqst(struct rpc_rqst *);
670 void xprt_rdma_bc_destroy(struct rpc_xprt *, unsigned int);
671 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
672
673 extern struct xprt_class xprt_rdma_bc;
674
675 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */
676
