Loading...
1/*
2 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the BSD-type
8 * license below:
9 *
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
13 *
14 * Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 *
17 * Redistributions in binary form must reproduce the above
18 * copyright notice, this list of conditions and the following
19 * disclaimer in the documentation and/or other materials provided
20 * with the distribution.
21 *
22 * Neither the name of the Network Appliance, Inc. nor the names of
23 * its contributors may be used to endorse or promote products
24 * derived from this software without specific prior written
25 * permission.
26 *
27 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
28 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
29 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
30 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
31 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
32 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
33 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
34 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
35 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
36 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
37 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
38 */
39
40#ifndef _LINUX_SUNRPC_XPRT_RDMA_H
41#define _LINUX_SUNRPC_XPRT_RDMA_H
42
43#include <linux/wait.h> /* wait_queue_head_t, etc */
44#include <linux/spinlock.h> /* spinlock_t, etc */
45#include <linux/atomic.h> /* atomic_t, etc */
46#include <linux/workqueue.h> /* struct work_struct */
47
48#include <rdma/rdma_cm.h> /* RDMA connection api */
49#include <rdma/ib_verbs.h> /* RDMA verbs api */
50
51#include <linux/sunrpc/clnt.h> /* rpc_xprt */
52#include <linux/sunrpc/rpc_rdma.h> /* RPC/RDMA protocol */
53#include <linux/sunrpc/xprtrdma.h> /* xprt parameters */
54
55#define RDMA_RESOLVE_TIMEOUT (5000) /* 5 seconds */
56#define RDMA_CONNECT_RETRY_MAX (2) /* retries if no listener backlog */
57
58#define RPCRDMA_BIND_TO (60U * HZ)
59#define RPCRDMA_INIT_REEST_TO (5U * HZ)
60#define RPCRDMA_MAX_REEST_TO (30U * HZ)
61#define RPCRDMA_IDLE_DISC_TO (5U * 60 * HZ)
62
63/*
64 * Interface Adapter -- one per transport instance
65 */
66struct rpcrdma_ia {
67 const struct rpcrdma_memreg_ops *ri_ops;
68 rwlock_t ri_qplock;
69 struct ib_device *ri_device;
70 struct rdma_cm_id *ri_id;
71 struct ib_pd *ri_pd;
72 struct ib_mr *ri_dma_mr;
73 struct completion ri_done;
74 int ri_async_rc;
75 unsigned int ri_max_frmr_depth;
76 struct ib_qp_attr ri_qp_attr;
77 struct ib_qp_init_attr ri_qp_init_attr;
78};
79
80/*
81 * RDMA Endpoint -- one per transport instance
82 */
83
84struct rpcrdma_ep {
85 atomic_t rep_cqcount;
86 int rep_cqinit;
87 int rep_connected;
88 struct ib_qp_init_attr rep_attr;
89 wait_queue_head_t rep_connect_wait;
90 struct rdma_conn_param rep_remote_cma;
91 struct sockaddr_storage rep_remote_addr;
92 struct delayed_work rep_connect_worker;
93};
94
95#define INIT_CQCOUNT(ep) atomic_set(&(ep)->rep_cqcount, (ep)->rep_cqinit)
96#define DECR_CQCOUNT(ep) atomic_sub_return(1, &(ep)->rep_cqcount)
97
98/* Pre-allocate extra Work Requests for handling backward receives
99 * and sends. This is a fixed value because the Work Queues are
100 * allocated when the forward channel is set up.
101 */
102#if defined(CONFIG_SUNRPC_BACKCHANNEL)
103#define RPCRDMA_BACKWARD_WRS (8)
104#else
105#define RPCRDMA_BACKWARD_WRS (0)
106#endif
107
108/* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
109 *
110 * The below structure appears at the front of a large region of kmalloc'd
111 * memory, which always starts on a good alignment boundary.
112 */
113
114struct rpcrdma_regbuf {
115 size_t rg_size;
116 struct rpcrdma_req *rg_owner;
117 struct ib_sge rg_iov;
118 __be32 rg_base[0] __attribute__ ((aligned(256)));
119};
120
121static inline u64
122rdmab_addr(struct rpcrdma_regbuf *rb)
123{
124 return rb->rg_iov.addr;
125}
126
127static inline u32
128rdmab_length(struct rpcrdma_regbuf *rb)
129{
130 return rb->rg_iov.length;
131}
132
133static inline u32
134rdmab_lkey(struct rpcrdma_regbuf *rb)
135{
136 return rb->rg_iov.lkey;
137}
138
139static inline struct rpcrdma_msg *
140rdmab_to_msg(struct rpcrdma_regbuf *rb)
141{
142 return (struct rpcrdma_msg *)rb->rg_base;
143}
144
145#define RPCRDMA_DEF_GFP (GFP_NOIO | __GFP_NOWARN)
146
147/*
148 * struct rpcrdma_rep -- this structure encapsulates state required to recv
149 * and complete a reply, asychronously. It needs several pieces of
150 * state:
151 * o recv buffer (posted to provider)
152 * o ib_sge (also donated to provider)
153 * o status of reply (length, success or not)
154 * o bookkeeping state to get run by tasklet (list, etc)
155 *
156 * These are allocated during initialization, per-transport instance;
157 * however, the tasklet execution list itself is global, as it should
158 * always be pretty short.
159 *
160 * N of these are associated with a transport instance, and stored in
161 * struct rpcrdma_buffer. N is the max number of outstanding requests.
162 */
163
164#define RPCRDMA_MAX_DATA_SEGS ((1 * 1024 * 1024) / PAGE_SIZE)
165#define RPCRDMA_MAX_SEGS (RPCRDMA_MAX_DATA_SEGS + 2) /* head+tail = 2 */
166
167struct rpcrdma_buffer;
168
169struct rpcrdma_rep {
170 struct ib_cqe rr_cqe;
171 unsigned int rr_len;
172 struct ib_device *rr_device;
173 struct rpcrdma_xprt *rr_rxprt;
174 struct work_struct rr_work;
175 struct list_head rr_list;
176 struct rpcrdma_regbuf *rr_rdmabuf;
177};
178
179#define RPCRDMA_BAD_LEN (~0U)
180
181/*
182 * struct rpcrdma_mw - external memory region metadata
183 *
184 * An external memory region is any buffer or page that is registered
185 * on the fly (ie, not pre-registered).
186 *
187 * Each rpcrdma_buffer has a list of free MWs anchored in rb_mws. During
188 * call_allocate, rpcrdma_buffer_get() assigns one to each segment in
189 * an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep
190 * track of registration metadata while each RPC is pending.
191 * rpcrdma_deregister_external() uses this metadata to unmap and
192 * release these resources when an RPC is complete.
193 */
194enum rpcrdma_frmr_state {
195 FRMR_IS_INVALID, /* ready to be used */
196 FRMR_IS_VALID, /* in use */
197 FRMR_IS_STALE, /* failed completion */
198};
199
200struct rpcrdma_frmr {
201 struct scatterlist *sg;
202 int sg_nents;
203 struct ib_mr *fr_mr;
204 struct ib_cqe fr_cqe;
205 enum rpcrdma_frmr_state fr_state;
206 struct completion fr_linv_done;
207 struct work_struct fr_work;
208 struct rpcrdma_xprt *fr_xprt;
209 union {
210 struct ib_reg_wr fr_regwr;
211 struct ib_send_wr fr_invwr;
212 };
213};
214
215struct rpcrdma_fmr {
216 struct ib_fmr *fmr;
217 u64 *physaddrs;
218};
219
220struct rpcrdma_mw {
221 union {
222 struct rpcrdma_fmr fmr;
223 struct rpcrdma_frmr frmr;
224 };
225 struct list_head mw_list;
226 struct list_head mw_all;
227};
228
229/*
230 * struct rpcrdma_req -- structure central to the request/reply sequence.
231 *
232 * N of these are associated with a transport instance, and stored in
233 * struct rpcrdma_buffer. N is the max number of outstanding requests.
234 *
235 * It includes pre-registered buffer memory for send AND recv.
236 * The recv buffer, however, is not owned by this structure, and
237 * is "donated" to the hardware when a recv is posted. When a
238 * reply is handled, the recv buffer used is given back to the
239 * struct rpcrdma_req associated with the request.
240 *
241 * In addition to the basic memory, this structure includes an array
242 * of iovs for send operations. The reason is that the iovs passed to
243 * ib_post_{send,recv} must not be modified until the work request
244 * completes.
245 *
246 * NOTES:
247 * o RPCRDMA_MAX_SEGS is the max number of addressible chunk elements we
248 * marshal. The number needed varies depending on the iov lists that
249 * are passed to us, the memory registration mode we are in, and if
250 * physical addressing is used, the layout.
251 */
252
253struct rpcrdma_mr_seg { /* chunk descriptors */
254 struct rpcrdma_mw *rl_mw; /* registered MR */
255 u64 mr_base; /* registration result */
256 u32 mr_rkey; /* registration result */
257 u32 mr_len; /* length of chunk or segment */
258 int mr_nsegs; /* number of segments in chunk or 0 */
259 enum dma_data_direction mr_dir; /* segment mapping direction */
260 dma_addr_t mr_dma; /* segment mapping address */
261 size_t mr_dmalen; /* segment mapping length */
262 struct page *mr_page; /* owning page, if any */
263 char *mr_offset; /* kva if no page, else offset */
264};
265
266#define RPCRDMA_MAX_IOVS (2)
267
268struct rpcrdma_req {
269 struct list_head rl_free;
270 unsigned int rl_niovs;
271 unsigned int rl_nchunks;
272 unsigned int rl_connect_cookie;
273 struct rpcrdma_buffer *rl_buffer;
274 struct rpcrdma_rep *rl_reply;/* holder for reply buffer */
275 struct ib_sge rl_send_iov[RPCRDMA_MAX_IOVS];
276 struct rpcrdma_regbuf *rl_rdmabuf;
277 struct rpcrdma_regbuf *rl_sendbuf;
278 struct rpcrdma_mr_seg rl_segments[RPCRDMA_MAX_SEGS];
279
280 struct ib_cqe rl_cqe;
281 struct list_head rl_all;
282 bool rl_backchannel;
283};
284
285static inline struct rpcrdma_req *
286rpcr_to_rdmar(struct rpc_rqst *rqst)
287{
288 void *buffer = rqst->rq_buffer;
289 struct rpcrdma_regbuf *rb;
290
291 rb = container_of(buffer, struct rpcrdma_regbuf, rg_base);
292 return rb->rg_owner;
293}
294
295/*
296 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
297 * inline requests/replies, and client/server credits.
298 *
299 * One of these is associated with a transport instance
300 */
301struct rpcrdma_buffer {
302 spinlock_t rb_mwlock; /* protect rb_mws list */
303 struct list_head rb_mws;
304 struct list_head rb_all;
305 char *rb_pool;
306
307 spinlock_t rb_lock; /* protect buf lists */
308 struct list_head rb_send_bufs;
309 struct list_head rb_recv_bufs;
310 u32 rb_max_requests;
311 atomic_t rb_credits; /* most recent credit grant */
312
313 u32 rb_bc_srv_max_requests;
314 spinlock_t rb_reqslock; /* protect rb_allreqs */
315 struct list_head rb_allreqs;
316
317 u32 rb_bc_max_requests;
318};
319#define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia)
320
321/*
322 * Internal structure for transport instance creation. This
323 * exists primarily for modularity.
324 *
325 * This data should be set with mount options
326 */
327struct rpcrdma_create_data_internal {
328 struct sockaddr_storage addr; /* RDMA server address */
329 unsigned int max_requests; /* max requests (slots) in flight */
330 unsigned int rsize; /* mount rsize - max read hdr+data */
331 unsigned int wsize; /* mount wsize - max write hdr+data */
332 unsigned int inline_rsize; /* max non-rdma read data payload */
333 unsigned int inline_wsize; /* max non-rdma write data payload */
334 unsigned int padding; /* non-rdma write header padding */
335};
336
337#define RPCRDMA_INLINE_READ_THRESHOLD(rq) \
338 (rpcx_to_rdmad(rq->rq_xprt).inline_rsize)
339
340#define RPCRDMA_INLINE_WRITE_THRESHOLD(rq)\
341 (rpcx_to_rdmad(rq->rq_xprt).inline_wsize)
342
343#define RPCRDMA_INLINE_PAD_VALUE(rq)\
344 rpcx_to_rdmad(rq->rq_xprt).padding
345
346/*
347 * Statistics for RPCRDMA
348 */
349struct rpcrdma_stats {
350 unsigned long read_chunk_count;
351 unsigned long write_chunk_count;
352 unsigned long reply_chunk_count;
353
354 unsigned long long total_rdma_request;
355 unsigned long long total_rdma_reply;
356
357 unsigned long long pullup_copy_count;
358 unsigned long long fixup_copy_count;
359 unsigned long hardway_register_count;
360 unsigned long failed_marshal_count;
361 unsigned long bad_reply_count;
362 unsigned long nomsg_call_count;
363 unsigned long bcall_count;
364};
365
366/*
367 * Per-registration mode operations
368 */
369struct rpcrdma_xprt;
370struct rpcrdma_memreg_ops {
371 int (*ro_map)(struct rpcrdma_xprt *,
372 struct rpcrdma_mr_seg *, int, bool);
373 void (*ro_unmap_sync)(struct rpcrdma_xprt *,
374 struct rpcrdma_req *);
375 int (*ro_unmap)(struct rpcrdma_xprt *,
376 struct rpcrdma_mr_seg *);
377 int (*ro_open)(struct rpcrdma_ia *,
378 struct rpcrdma_ep *,
379 struct rpcrdma_create_data_internal *);
380 size_t (*ro_maxpages)(struct rpcrdma_xprt *);
381 int (*ro_init)(struct rpcrdma_xprt *);
382 void (*ro_destroy)(struct rpcrdma_buffer *);
383 const char *ro_displayname;
384};
385
386extern const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops;
387extern const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops;
388extern const struct rpcrdma_memreg_ops rpcrdma_physical_memreg_ops;
389
390/*
391 * RPCRDMA transport -- encapsulates the structures above for
392 * integration with RPC.
393 *
394 * The contained structures are embedded, not pointers,
395 * for convenience. This structure need not be visible externally.
396 *
397 * It is allocated and initialized during mount, and released
398 * during unmount.
399 */
400struct rpcrdma_xprt {
401 struct rpc_xprt rx_xprt;
402 struct rpcrdma_ia rx_ia;
403 struct rpcrdma_ep rx_ep;
404 struct rpcrdma_buffer rx_buf;
405 struct rpcrdma_create_data_internal rx_data;
406 struct delayed_work rx_connect_worker;
407 struct rpcrdma_stats rx_stats;
408};
409
410#define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
411#define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data)
412
413/* Setting this to 0 ensures interoperability with early servers.
414 * Setting this to 1 enhances certain unaligned read/write performance.
415 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
416extern int xprt_rdma_pad_optimize;
417
418/*
419 * Interface Adapter calls - xprtrdma/verbs.c
420 */
421int rpcrdma_ia_open(struct rpcrdma_xprt *, struct sockaddr *, int);
422void rpcrdma_ia_close(struct rpcrdma_ia *);
423
424/*
425 * Endpoint calls - xprtrdma/verbs.c
426 */
427int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *,
428 struct rpcrdma_create_data_internal *);
429void rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *);
430int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *);
431void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *);
432
433int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
434 struct rpcrdma_req *);
435int rpcrdma_ep_post_recv(struct rpcrdma_ia *, struct rpcrdma_ep *,
436 struct rpcrdma_rep *);
437
438/*
439 * Buffer calls - xprtrdma/verbs.c
440 */
441struct rpcrdma_req *rpcrdma_create_req(struct rpcrdma_xprt *);
442struct rpcrdma_rep *rpcrdma_create_rep(struct rpcrdma_xprt *);
443void rpcrdma_destroy_req(struct rpcrdma_ia *, struct rpcrdma_req *);
444int rpcrdma_buffer_create(struct rpcrdma_xprt *);
445void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
446
447struct rpcrdma_mw *rpcrdma_get_mw(struct rpcrdma_xprt *);
448void rpcrdma_put_mw(struct rpcrdma_xprt *, struct rpcrdma_mw *);
449struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
450void rpcrdma_buffer_put(struct rpcrdma_req *);
451void rpcrdma_recv_buffer_get(struct rpcrdma_req *);
452void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
453
454struct rpcrdma_regbuf *rpcrdma_alloc_regbuf(struct rpcrdma_ia *,
455 size_t, gfp_t);
456void rpcrdma_free_regbuf(struct rpcrdma_ia *,
457 struct rpcrdma_regbuf *);
458
459unsigned int rpcrdma_max_segments(struct rpcrdma_xprt *);
460int rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt *, unsigned int);
461
462int frwr_alloc_recovery_wq(void);
463void frwr_destroy_recovery_wq(void);
464
465int rpcrdma_alloc_wq(void);
466void rpcrdma_destroy_wq(void);
467
468/*
469 * Wrappers for chunk registration, shared by read/write chunk code.
470 */
471
472void rpcrdma_mapping_error(struct rpcrdma_mr_seg *);
473
474static inline enum dma_data_direction
475rpcrdma_data_dir(bool writing)
476{
477 return writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
478}
479
480static inline void
481rpcrdma_map_one(struct ib_device *device, struct rpcrdma_mr_seg *seg,
482 enum dma_data_direction direction)
483{
484 seg->mr_dir = direction;
485 seg->mr_dmalen = seg->mr_len;
486
487 if (seg->mr_page)
488 seg->mr_dma = ib_dma_map_page(device,
489 seg->mr_page, offset_in_page(seg->mr_offset),
490 seg->mr_dmalen, seg->mr_dir);
491 else
492 seg->mr_dma = ib_dma_map_single(device,
493 seg->mr_offset,
494 seg->mr_dmalen, seg->mr_dir);
495
496 if (ib_dma_mapping_error(device, seg->mr_dma))
497 rpcrdma_mapping_error(seg);
498}
499
500static inline void
501rpcrdma_unmap_one(struct ib_device *device, struct rpcrdma_mr_seg *seg)
502{
503 if (seg->mr_page)
504 ib_dma_unmap_page(device,
505 seg->mr_dma, seg->mr_dmalen, seg->mr_dir);
506 else
507 ib_dma_unmap_single(device,
508 seg->mr_dma, seg->mr_dmalen, seg->mr_dir);
509}
510
511/*
512 * RPC/RDMA connection management calls - xprtrdma/rpc_rdma.c
513 */
514void rpcrdma_connect_worker(struct work_struct *);
515void rpcrdma_conn_func(struct rpcrdma_ep *);
516void rpcrdma_reply_handler(struct rpcrdma_rep *);
517
518/*
519 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
520 */
521int rpcrdma_marshal_req(struct rpc_rqst *);
522
523/* RPC/RDMA module init - xprtrdma/transport.c
524 */
525extern unsigned int xprt_rdma_max_inline_read;
526void xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap);
527void xprt_rdma_free_addresses(struct rpc_xprt *xprt);
528void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq);
529int xprt_rdma_init(void);
530void xprt_rdma_cleanup(void);
531
532/* Backchannel calls - xprtrdma/backchannel.c
533 */
534#if defined(CONFIG_SUNRPC_BACKCHANNEL)
535int xprt_rdma_bc_setup(struct rpc_xprt *, unsigned int);
536int xprt_rdma_bc_up(struct svc_serv *, struct net *);
537int rpcrdma_bc_post_recv(struct rpcrdma_xprt *, unsigned int);
538void rpcrdma_bc_receive_call(struct rpcrdma_xprt *, struct rpcrdma_rep *);
539int rpcrdma_bc_marshal_reply(struct rpc_rqst *);
540void xprt_rdma_bc_free_rqst(struct rpc_rqst *);
541void xprt_rdma_bc_destroy(struct rpc_xprt *, unsigned int);
542#endif /* CONFIG_SUNRPC_BACKCHANNEL */
543
544extern struct xprt_class xprt_rdma_bc;
545
546#endif /* _LINUX_SUNRPC_XPRT_RDMA_H */
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/kref.h> /* struct kref */
49#include <linux/workqueue.h> /* struct work_struct */
50#include <linux/llist.h>
51
52#include <rdma/rdma_cm.h> /* RDMA connection api */
53#include <rdma/ib_verbs.h> /* RDMA verbs api */
54
55#include <linux/sunrpc/clnt.h> /* rpc_xprt */
56#include <linux/sunrpc/rpc_rdma.h> /* RPC/RDMA protocol */
57#include <linux/sunrpc/xprtrdma.h> /* xprt parameters */
58
59#define RDMA_RESOLVE_TIMEOUT (5000) /* 5 seconds */
60#define RDMA_CONNECT_RETRY_MAX (2) /* retries if no listener backlog */
61
62#define RPCRDMA_BIND_TO (60U * HZ)
63#define RPCRDMA_INIT_REEST_TO (5U * HZ)
64#define RPCRDMA_MAX_REEST_TO (30U * HZ)
65#define RPCRDMA_IDLE_DISC_TO (5U * 60 * HZ)
66
67/*
68 * RDMA Endpoint -- connection endpoint details
69 */
70struct rpcrdma_ep {
71 struct kref re_kref;
72 struct rdma_cm_id *re_id;
73 struct ib_pd *re_pd;
74 unsigned int re_max_rdma_segs;
75 unsigned int re_max_fr_depth;
76 bool re_implicit_roundup;
77 enum ib_mr_type re_mrtype;
78 struct completion re_done;
79 unsigned int re_send_count;
80 unsigned int re_send_batch;
81 unsigned int re_max_inline_send;
82 unsigned int re_max_inline_recv;
83 int re_async_rc;
84 int re_connect_status;
85 atomic_t re_force_disconnect;
86 struct ib_qp_init_attr re_attr;
87 wait_queue_head_t re_connect_wait;
88 struct rpc_xprt *re_xprt;
89 struct rpcrdma_connect_private
90 re_cm_private;
91 struct rdma_conn_param re_remote_cma;
92 int re_receive_count;
93 unsigned int re_max_requests; /* depends on device */
94 unsigned int re_inline_send; /* negotiated */
95 unsigned int re_inline_recv; /* negotiated */
96};
97
98/* Pre-allocate extra Work Requests for handling backward receives
99 * and sends. This is a fixed value because the Work Queues are
100 * allocated when the forward channel is set up, long before the
101 * backchannel is provisioned. This value is two times
102 * NFS4_DEF_CB_SLOT_TABLE_SIZE.
103 */
104#if defined(CONFIG_SUNRPC_BACKCHANNEL)
105#define RPCRDMA_BACKWARD_WRS (32)
106#else
107#define RPCRDMA_BACKWARD_WRS (0)
108#endif
109
110/* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
111 */
112
113struct rpcrdma_regbuf {
114 struct ib_sge rg_iov;
115 struct ib_device *rg_device;
116 enum dma_data_direction rg_direction;
117 void *rg_data;
118};
119
120static inline u64 rdmab_addr(struct rpcrdma_regbuf *rb)
121{
122 return rb->rg_iov.addr;
123}
124
125static inline u32 rdmab_length(struct rpcrdma_regbuf *rb)
126{
127 return rb->rg_iov.length;
128}
129
130static inline u32 rdmab_lkey(struct rpcrdma_regbuf *rb)
131{
132 return rb->rg_iov.lkey;
133}
134
135static inline struct ib_device *rdmab_device(struct rpcrdma_regbuf *rb)
136{
137 return rb->rg_device;
138}
139
140static inline void *rdmab_data(const struct rpcrdma_regbuf *rb)
141{
142 return rb->rg_data;
143}
144
145#define RPCRDMA_DEF_GFP (GFP_NOIO | __GFP_NOWARN)
146
147/* To ensure a transport can always make forward progress,
148 * the number of RDMA segments allowed in header chunk lists
149 * is capped at 16. This prevents less-capable devices from
150 * overrunning the Send buffer while building chunk lists.
151 *
152 * Elements of the Read list take up more room than the
153 * Write list or Reply chunk. 16 read segments means the
154 * chunk lists cannot consume more than
155 *
156 * ((16 + 2) * read segment size) + 1 XDR words,
157 *
158 * or about 400 bytes. The fixed part of the header is
159 * another 24 bytes. Thus when the inline threshold is
160 * 1024 bytes, at least 600 bytes are available for RPC
161 * message bodies.
162 */
163enum {
164 RPCRDMA_MAX_HDR_SEGS = 16,
165};
166
167/*
168 * struct rpcrdma_rep -- this structure encapsulates state required
169 * to receive and complete an RPC Reply, asychronously. It needs
170 * several pieces of state:
171 *
172 * o receive buffer and ib_sge (donated to provider)
173 * o status of receive (success or not, length, inv rkey)
174 * o bookkeeping state to get run by reply handler (XDR stream)
175 *
176 * These structures are allocated during transport initialization.
177 * N of these are associated with a transport instance, managed by
178 * struct rpcrdma_buffer. N is the max number of outstanding RPCs.
179 */
180
181struct rpcrdma_rep {
182 struct ib_cqe rr_cqe;
183 __be32 rr_xid;
184 __be32 rr_vers;
185 __be32 rr_proc;
186 int rr_wc_flags;
187 u32 rr_inv_rkey;
188 bool rr_temp;
189 struct rpcrdma_regbuf *rr_rdmabuf;
190 struct rpcrdma_xprt *rr_rxprt;
191 struct rpc_rqst *rr_rqst;
192 struct xdr_buf rr_hdrbuf;
193 struct xdr_stream rr_stream;
194 struct llist_node rr_node;
195 struct ib_recv_wr rr_recv_wr;
196 struct list_head rr_all;
197};
198
199/* To reduce the rate at which a transport invokes ib_post_recv
200 * (and thus the hardware doorbell rate), xprtrdma posts Receive
201 * WRs in batches.
202 *
203 * Setting this to zero disables Receive post batching.
204 */
205enum {
206 RPCRDMA_MAX_RECV_BATCH = 7,
207};
208
209/* struct rpcrdma_sendctx - DMA mapped SGEs to unmap after Send completes
210 */
211struct rpcrdma_req;
212struct rpcrdma_sendctx {
213 struct ib_cqe sc_cqe;
214 struct rpcrdma_req *sc_req;
215 unsigned int sc_unmap_count;
216 struct ib_sge sc_sges[];
217};
218
219/*
220 * struct rpcrdma_mr - external memory region metadata
221 *
222 * An external memory region is any buffer or page that is registered
223 * on the fly (ie, not pre-registered).
224 */
225struct rpcrdma_frwr {
226 struct ib_mr *fr_mr;
227 struct ib_cqe fr_cqe;
228 struct completion fr_linv_done;
229 union {
230 struct ib_reg_wr fr_regwr;
231 struct ib_send_wr fr_invwr;
232 };
233};
234
235struct rpcrdma_req;
236struct rpcrdma_mr {
237 struct list_head mr_list;
238 struct rpcrdma_req *mr_req;
239 struct scatterlist *mr_sg;
240 int mr_nents;
241 enum dma_data_direction mr_dir;
242 struct rpcrdma_frwr frwr;
243 struct rpcrdma_xprt *mr_xprt;
244 u32 mr_handle;
245 u32 mr_length;
246 u64 mr_offset;
247 struct list_head mr_all;
248};
249
250/*
251 * struct rpcrdma_req -- structure central to the request/reply sequence.
252 *
253 * N of these are associated with a transport instance, and stored in
254 * struct rpcrdma_buffer. N is the max number of outstanding requests.
255 *
256 * It includes pre-registered buffer memory for send AND recv.
257 * The recv buffer, however, is not owned by this structure, and
258 * is "donated" to the hardware when a recv is posted. When a
259 * reply is handled, the recv buffer used is given back to the
260 * struct rpcrdma_req associated with the request.
261 *
262 * In addition to the basic memory, this structure includes an array
263 * of iovs for send operations. The reason is that the iovs passed to
264 * ib_post_{send,recv} must not be modified until the work request
265 * completes.
266 */
267
268/* Maximum number of page-sized "segments" per chunk list to be
269 * registered or invalidated. Must handle a Reply chunk:
270 */
271enum {
272 RPCRDMA_MAX_IOV_SEGS = 3,
273 RPCRDMA_MAX_DATA_SEGS = ((1 * 1024 * 1024) / PAGE_SIZE) + 1,
274 RPCRDMA_MAX_SEGS = RPCRDMA_MAX_DATA_SEGS +
275 RPCRDMA_MAX_IOV_SEGS,
276};
277
278struct rpcrdma_mr_seg { /* chunk descriptors */
279 u32 mr_len; /* length of chunk or segment */
280 struct page *mr_page; /* owning page, if any */
281 char *mr_offset; /* kva if no page, else offset */
282};
283
284/* The Send SGE array is provisioned to send a maximum size
285 * inline request:
286 * - RPC-over-RDMA header
287 * - xdr_buf head iovec
288 * - RPCRDMA_MAX_INLINE bytes, in pages
289 * - xdr_buf tail iovec
290 *
291 * The actual number of array elements consumed by each RPC
292 * depends on the device's max_sge limit.
293 */
294enum {
295 RPCRDMA_MIN_SEND_SGES = 3,
296 RPCRDMA_MAX_PAGE_SGES = RPCRDMA_MAX_INLINE >> PAGE_SHIFT,
297 RPCRDMA_MAX_SEND_SGES = 1 + 1 + RPCRDMA_MAX_PAGE_SGES + 1,
298};
299
300struct rpcrdma_buffer;
301struct rpcrdma_req {
302 struct list_head rl_list;
303 struct rpc_rqst rl_slot;
304 struct rpcrdma_rep *rl_reply;
305 struct xdr_stream rl_stream;
306 struct xdr_buf rl_hdrbuf;
307 struct ib_send_wr rl_wr;
308 struct rpcrdma_sendctx *rl_sendctx;
309 struct rpcrdma_regbuf *rl_rdmabuf; /* xprt header */
310 struct rpcrdma_regbuf *rl_sendbuf; /* rq_snd_buf */
311 struct rpcrdma_regbuf *rl_recvbuf; /* rq_rcv_buf */
312
313 struct list_head rl_all;
314 struct kref rl_kref;
315
316 struct list_head rl_free_mrs;
317 struct list_head rl_registered;
318 struct rpcrdma_mr_seg rl_segments[RPCRDMA_MAX_SEGS];
319};
320
321static inline struct rpcrdma_req *
322rpcr_to_rdmar(const struct rpc_rqst *rqst)
323{
324 return container_of(rqst, struct rpcrdma_req, rl_slot);
325}
326
327static inline void
328rpcrdma_mr_push(struct rpcrdma_mr *mr, struct list_head *list)
329{
330 list_add(&mr->mr_list, list);
331}
332
333static inline struct rpcrdma_mr *
334rpcrdma_mr_pop(struct list_head *list)
335{
336 struct rpcrdma_mr *mr;
337
338 mr = list_first_entry_or_null(list, struct rpcrdma_mr, mr_list);
339 if (mr)
340 list_del_init(&mr->mr_list);
341 return mr;
342}
343
344/*
345 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
346 * inline requests/replies, and client/server credits.
347 *
348 * One of these is associated with a transport instance
349 */
350struct rpcrdma_buffer {
351 spinlock_t rb_lock;
352 struct list_head rb_send_bufs;
353 struct list_head rb_mrs;
354
355 unsigned long rb_sc_head;
356 unsigned long rb_sc_tail;
357 unsigned long rb_sc_last;
358 struct rpcrdma_sendctx **rb_sc_ctxs;
359
360 struct list_head rb_allreqs;
361 struct list_head rb_all_mrs;
362 struct list_head rb_all_reps;
363
364 struct llist_head rb_free_reps;
365
366 __be32 rb_max_requests;
367 u32 rb_credits; /* most recent credit grant */
368
369 u32 rb_bc_srv_max_requests;
370 u32 rb_bc_max_requests;
371
372 struct work_struct rb_refresh_worker;
373};
374
375/*
376 * Statistics for RPCRDMA
377 */
378struct rpcrdma_stats {
379 /* accessed when sending a call */
380 unsigned long read_chunk_count;
381 unsigned long write_chunk_count;
382 unsigned long reply_chunk_count;
383 unsigned long long total_rdma_request;
384
385 /* rarely accessed error counters */
386 unsigned long long pullup_copy_count;
387 unsigned long hardway_register_count;
388 unsigned long failed_marshal_count;
389 unsigned long bad_reply_count;
390 unsigned long mrs_recycled;
391 unsigned long mrs_orphaned;
392 unsigned long mrs_allocated;
393 unsigned long empty_sendctx_q;
394
395 /* accessed when receiving a reply */
396 unsigned long long total_rdma_reply;
397 unsigned long long fixup_copy_count;
398 unsigned long reply_waits_for_send;
399 unsigned long local_inv_needed;
400 unsigned long nomsg_call_count;
401 unsigned long bcall_count;
402};
403
404/*
405 * RPCRDMA transport -- encapsulates the structures above for
406 * integration with RPC.
407 *
408 * The contained structures are embedded, not pointers,
409 * for convenience. This structure need not be visible externally.
410 *
411 * It is allocated and initialized during mount, and released
412 * during unmount.
413 */
414struct rpcrdma_xprt {
415 struct rpc_xprt rx_xprt;
416 struct rpcrdma_ep *rx_ep;
417 struct rpcrdma_buffer rx_buf;
418 struct delayed_work rx_connect_worker;
419 struct rpc_timeout rx_timeout;
420 struct rpcrdma_stats rx_stats;
421};
422
423#define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
424
425static inline const char *
426rpcrdma_addrstr(const struct rpcrdma_xprt *r_xprt)
427{
428 return r_xprt->rx_xprt.address_strings[RPC_DISPLAY_ADDR];
429}
430
431static inline const char *
432rpcrdma_portstr(const struct rpcrdma_xprt *r_xprt)
433{
434 return r_xprt->rx_xprt.address_strings[RPC_DISPLAY_PORT];
435}
436
437/* Setting this to 0 ensures interoperability with early servers.
438 * Setting this to 1 enhances certain unaligned read/write performance.
439 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
440extern int xprt_rdma_pad_optimize;
441
442/* This setting controls the hunt for a supported memory
443 * registration strategy.
444 */
445extern unsigned int xprt_rdma_memreg_strategy;
446
447/*
448 * Endpoint calls - xprtrdma/verbs.c
449 */
450void rpcrdma_flush_disconnect(struct rpcrdma_xprt *r_xprt, struct ib_wc *wc);
451int rpcrdma_xprt_connect(struct rpcrdma_xprt *r_xprt);
452void rpcrdma_xprt_disconnect(struct rpcrdma_xprt *r_xprt);
453
454int rpcrdma_post_sends(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req);
455void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp);
456
457/*
458 * Buffer calls - xprtrdma/verbs.c
459 */
460struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
461 gfp_t flags);
462int rpcrdma_req_setup(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req);
463void rpcrdma_req_destroy(struct rpcrdma_req *req);
464int rpcrdma_buffer_create(struct rpcrdma_xprt *);
465void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
466struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt);
467
468struct rpcrdma_mr *rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt);
469void rpcrdma_mr_put(struct rpcrdma_mr *mr);
470void rpcrdma_mrs_refresh(struct rpcrdma_xprt *r_xprt);
471
472struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
473void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers,
474 struct rpcrdma_req *req);
475void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
476
477bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size,
478 gfp_t flags);
479bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
480 struct rpcrdma_regbuf *rb);
481
482/**
483 * rpcrdma_regbuf_is_mapped - check if buffer is DMA mapped
484 *
485 * Returns true if the buffer is now mapped to rb->rg_device.
486 */
487static inline bool rpcrdma_regbuf_is_mapped(struct rpcrdma_regbuf *rb)
488{
489 return rb->rg_device != NULL;
490}
491
492/**
493 * rpcrdma_regbuf_dma_map - DMA-map a regbuf
494 * @r_xprt: controlling transport instance
495 * @rb: regbuf to be mapped
496 *
497 * Returns true if the buffer is currently DMA mapped.
498 */
499static inline bool rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
500 struct rpcrdma_regbuf *rb)
501{
502 if (likely(rpcrdma_regbuf_is_mapped(rb)))
503 return true;
504 return __rpcrdma_regbuf_dma_map(r_xprt, rb);
505}
506
507/*
508 * Wrappers for chunk registration, shared by read/write chunk code.
509 */
510
511static inline enum dma_data_direction
512rpcrdma_data_dir(bool writing)
513{
514 return writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
515}
516
517/* Memory registration calls xprtrdma/frwr_ops.c
518 */
519void frwr_reset(struct rpcrdma_req *req);
520int frwr_query_device(struct rpcrdma_ep *ep, const struct ib_device *device);
521int frwr_mr_init(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr *mr);
522void frwr_release_mr(struct rpcrdma_mr *mr);
523struct rpcrdma_mr_seg *frwr_map(struct rpcrdma_xprt *r_xprt,
524 struct rpcrdma_mr_seg *seg,
525 int nsegs, bool writing, __be32 xid,
526 struct rpcrdma_mr *mr);
527int frwr_send(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req);
528void frwr_reminv(struct rpcrdma_rep *rep, struct list_head *mrs);
529void frwr_unmap_sync(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req);
530void frwr_unmap_async(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req);
531
532/*
533 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
534 */
535
536enum rpcrdma_chunktype {
537 rpcrdma_noch = 0,
538 rpcrdma_noch_pullup,
539 rpcrdma_noch_mapped,
540 rpcrdma_readch,
541 rpcrdma_areadch,
542 rpcrdma_writech,
543 rpcrdma_replych
544};
545
546int rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
547 struct rpcrdma_req *req, u32 hdrlen,
548 struct xdr_buf *xdr,
549 enum rpcrdma_chunktype rtype);
550void rpcrdma_sendctx_unmap(struct rpcrdma_sendctx *sc);
551int rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst);
552void rpcrdma_set_max_header_sizes(struct rpcrdma_ep *ep);
553void rpcrdma_reset_cwnd(struct rpcrdma_xprt *r_xprt);
554void rpcrdma_complete_rqst(struct rpcrdma_rep *rep);
555void rpcrdma_reply_handler(struct rpcrdma_rep *rep);
556
557static inline void rpcrdma_set_xdrlen(struct xdr_buf *xdr, size_t len)
558{
559 xdr->head[0].iov_len = len;
560 xdr->len = len;
561}
562
563/* RPC/RDMA module init - xprtrdma/transport.c
564 */
565extern unsigned int xprt_rdma_max_inline_read;
566extern unsigned int xprt_rdma_max_inline_write;
567void xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap);
568void xprt_rdma_free_addresses(struct rpc_xprt *xprt);
569void xprt_rdma_close(struct rpc_xprt *xprt);
570void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq);
571int xprt_rdma_init(void);
572void xprt_rdma_cleanup(void);
573
574/* Backchannel calls - xprtrdma/backchannel.c
575 */
576#if defined(CONFIG_SUNRPC_BACKCHANNEL)
577int xprt_rdma_bc_setup(struct rpc_xprt *, unsigned int);
578size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *);
579unsigned int xprt_rdma_bc_max_slots(struct rpc_xprt *);
580int rpcrdma_bc_post_recv(struct rpcrdma_xprt *, unsigned int);
581void rpcrdma_bc_receive_call(struct rpcrdma_xprt *, struct rpcrdma_rep *);
582int xprt_rdma_bc_send_reply(struct rpc_rqst *rqst);
583void xprt_rdma_bc_free_rqst(struct rpc_rqst *);
584void xprt_rdma_bc_destroy(struct rpc_xprt *, unsigned int);
585#endif /* CONFIG_SUNRPC_BACKCHANNEL */
586
587extern struct xprt_class xprt_rdma_bc;
588
589#endif /* _LINUX_SUNRPC_XPRT_RDMA_H */