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/*
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
47#include <rdma/rdma_cm.h> /* RDMA connection api */
48#include <rdma/ib_verbs.h> /* RDMA verbs api */
49
50#include <linux/sunrpc/clnt.h> /* rpc_xprt */
51#include <linux/sunrpc/rpc_rdma.h> /* RPC/RDMA protocol */
52#include <linux/sunrpc/xprtrdma.h> /* xprt parameters */
53
54#define RDMA_RESOLVE_TIMEOUT (5000) /* 5 seconds */
55#define RDMA_CONNECT_RETRY_MAX (2) /* retries if no listener backlog */
56
57/*
58 * Interface Adapter -- one per transport instance
59 */
60struct rpcrdma_ia {
61 struct rdma_cm_id *ri_id;
62 struct ib_pd *ri_pd;
63 struct ib_mr *ri_bind_mem;
64 u32 ri_dma_lkey;
65 int ri_have_dma_lkey;
66 struct completion ri_done;
67 int ri_async_rc;
68 enum rpcrdma_memreg ri_memreg_strategy;
69};
70
71/*
72 * RDMA Endpoint -- one per transport instance
73 */
74
75struct rpcrdma_ep {
76 atomic_t rep_cqcount;
77 int rep_cqinit;
78 int rep_connected;
79 struct rpcrdma_ia *rep_ia;
80 struct ib_cq *rep_cq;
81 struct ib_qp_init_attr rep_attr;
82 wait_queue_head_t rep_connect_wait;
83 struct ib_sge rep_pad; /* holds zeroed pad */
84 struct ib_mr *rep_pad_mr; /* holds zeroed pad */
85 void (*rep_func)(struct rpcrdma_ep *);
86 struct rpc_xprt *rep_xprt; /* for rep_func */
87 struct rdma_conn_param rep_remote_cma;
88 struct sockaddr_storage rep_remote_addr;
89};
90
91#define INIT_CQCOUNT(ep) atomic_set(&(ep)->rep_cqcount, (ep)->rep_cqinit)
92#define DECR_CQCOUNT(ep) atomic_sub_return(1, &(ep)->rep_cqcount)
93
94/*
95 * struct rpcrdma_rep -- this structure encapsulates state required to recv
96 * and complete a reply, asychronously. It needs several pieces of
97 * state:
98 * o recv buffer (posted to provider)
99 * o ib_sge (also donated to provider)
100 * o status of reply (length, success or not)
101 * o bookkeeping state to get run by tasklet (list, etc)
102 *
103 * These are allocated during initialization, per-transport instance;
104 * however, the tasklet execution list itself is global, as it should
105 * always be pretty short.
106 *
107 * N of these are associated with a transport instance, and stored in
108 * struct rpcrdma_buffer. N is the max number of outstanding requests.
109 */
110
111/* temporary static scatter/gather max */
112#define RPCRDMA_MAX_DATA_SEGS (64) /* max scatter/gather */
113#define RPCRDMA_MAX_SEGS (RPCRDMA_MAX_DATA_SEGS + 2) /* head+tail = 2 */
114#define MAX_RPCRDMAHDR (\
115 /* max supported RPC/RDMA header */ \
116 sizeof(struct rpcrdma_msg) + (2 * sizeof(u32)) + \
117 (sizeof(struct rpcrdma_read_chunk) * RPCRDMA_MAX_SEGS) + sizeof(u32))
118
119struct rpcrdma_buffer;
120
121struct rpcrdma_rep {
122 unsigned int rr_len; /* actual received reply length */
123 struct rpcrdma_buffer *rr_buffer; /* home base for this structure */
124 struct rpc_xprt *rr_xprt; /* needed for request/reply matching */
125 void (*rr_func)(struct rpcrdma_rep *);/* called by tasklet in softint */
126 struct list_head rr_list; /* tasklet list */
127 wait_queue_head_t rr_unbind; /* optional unbind wait */
128 struct ib_sge rr_iov; /* for posting */
129 struct ib_mr *rr_handle; /* handle for mem in rr_iov */
130 char rr_base[MAX_RPCRDMAHDR]; /* minimal inline receive buffer */
131};
132
133/*
134 * struct rpcrdma_req -- structure central to the request/reply sequence.
135 *
136 * N of these are associated with a transport instance, and stored in
137 * struct rpcrdma_buffer. N is the max number of outstanding requests.
138 *
139 * It includes pre-registered buffer memory for send AND recv.
140 * The recv buffer, however, is not owned by this structure, and
141 * is "donated" to the hardware when a recv is posted. When a
142 * reply is handled, the recv buffer used is given back to the
143 * struct rpcrdma_req associated with the request.
144 *
145 * In addition to the basic memory, this structure includes an array
146 * of iovs for send operations. The reason is that the iovs passed to
147 * ib_post_{send,recv} must not be modified until the work request
148 * completes.
149 *
150 * NOTES:
151 * o RPCRDMA_MAX_SEGS is the max number of addressible chunk elements we
152 * marshal. The number needed varies depending on the iov lists that
153 * are passed to us, the memory registration mode we are in, and if
154 * physical addressing is used, the layout.
155 */
156
157struct rpcrdma_mr_seg { /* chunk descriptors */
158 union { /* chunk memory handles */
159 struct ib_mr *rl_mr; /* if registered directly */
160 struct rpcrdma_mw { /* if registered from region */
161 union {
162 struct ib_mw *mw;
163 struct ib_fmr *fmr;
164 struct {
165 struct ib_fast_reg_page_list *fr_pgl;
166 struct ib_mr *fr_mr;
167 enum { FRMR_IS_INVALID, FRMR_IS_VALID } state;
168 } frmr;
169 } r;
170 struct list_head mw_list;
171 } *rl_mw;
172 } mr_chunk;
173 u64 mr_base; /* registration result */
174 u32 mr_rkey; /* registration result */
175 u32 mr_len; /* length of chunk or segment */
176 int mr_nsegs; /* number of segments in chunk or 0 */
177 enum dma_data_direction mr_dir; /* segment mapping direction */
178 dma_addr_t mr_dma; /* segment mapping address */
179 size_t mr_dmalen; /* segment mapping length */
180 struct page *mr_page; /* owning page, if any */
181 char *mr_offset; /* kva if no page, else offset */
182};
183
184struct rpcrdma_req {
185 size_t rl_size; /* actual length of buffer */
186 unsigned int rl_niovs; /* 0, 2 or 4 */
187 unsigned int rl_nchunks; /* non-zero if chunks */
188 unsigned int rl_connect_cookie; /* retry detection */
189 struct rpcrdma_buffer *rl_buffer; /* home base for this structure */
190 struct rpcrdma_rep *rl_reply;/* holder for reply buffer */
191 struct rpcrdma_mr_seg rl_segments[RPCRDMA_MAX_SEGS];/* chunk segments */
192 struct ib_sge rl_send_iov[4]; /* for active requests */
193 struct ib_sge rl_iov; /* for posting */
194 struct ib_mr *rl_handle; /* handle for mem in rl_iov */
195 char rl_base[MAX_RPCRDMAHDR]; /* start of actual buffer */
196 __u32 rl_xdr_buf[0]; /* start of returned rpc rq_buffer */
197};
198#define rpcr_to_rdmar(r) \
199 container_of((r)->rq_buffer, struct rpcrdma_req, rl_xdr_buf[0])
200
201/*
202 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
203 * inline requests/replies, and client/server credits.
204 *
205 * One of these is associated with a transport instance
206 */
207struct rpcrdma_buffer {
208 spinlock_t rb_lock; /* protects indexes */
209 atomic_t rb_credits; /* most recent server credits */
210 unsigned long rb_cwndscale; /* cached framework rpc_cwndscale */
211 int rb_max_requests;/* client max requests */
212 struct list_head rb_mws; /* optional memory windows/fmrs/frmrs */
213 int rb_send_index;
214 struct rpcrdma_req **rb_send_bufs;
215 int rb_recv_index;
216 struct rpcrdma_rep **rb_recv_bufs;
217 char *rb_pool;
218};
219#define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia)
220
221/*
222 * Internal structure for transport instance creation. This
223 * exists primarily for modularity.
224 *
225 * This data should be set with mount options
226 */
227struct rpcrdma_create_data_internal {
228 struct sockaddr_storage addr; /* RDMA server address */
229 unsigned int max_requests; /* max requests (slots) in flight */
230 unsigned int rsize; /* mount rsize - max read hdr+data */
231 unsigned int wsize; /* mount wsize - max write hdr+data */
232 unsigned int inline_rsize; /* max non-rdma read data payload */
233 unsigned int inline_wsize; /* max non-rdma write data payload */
234 unsigned int padding; /* non-rdma write header padding */
235};
236
237#define RPCRDMA_INLINE_READ_THRESHOLD(rq) \
238 (rpcx_to_rdmad(rq->rq_task->tk_xprt).inline_rsize)
239
240#define RPCRDMA_INLINE_WRITE_THRESHOLD(rq)\
241 (rpcx_to_rdmad(rq->rq_task->tk_xprt).inline_wsize)
242
243#define RPCRDMA_INLINE_PAD_VALUE(rq)\
244 rpcx_to_rdmad(rq->rq_task->tk_xprt).padding
245
246/*
247 * Statistics for RPCRDMA
248 */
249struct rpcrdma_stats {
250 unsigned long read_chunk_count;
251 unsigned long write_chunk_count;
252 unsigned long reply_chunk_count;
253
254 unsigned long long total_rdma_request;
255 unsigned long long total_rdma_reply;
256
257 unsigned long long pullup_copy_count;
258 unsigned long long fixup_copy_count;
259 unsigned long hardway_register_count;
260 unsigned long failed_marshal_count;
261 unsigned long bad_reply_count;
262};
263
264/*
265 * RPCRDMA transport -- encapsulates the structures above for
266 * integration with RPC.
267 *
268 * The contained structures are embedded, not pointers,
269 * for convenience. This structure need not be visible externally.
270 *
271 * It is allocated and initialized during mount, and released
272 * during unmount.
273 */
274struct rpcrdma_xprt {
275 struct rpc_xprt xprt;
276 struct rpcrdma_ia rx_ia;
277 struct rpcrdma_ep rx_ep;
278 struct rpcrdma_buffer rx_buf;
279 struct rpcrdma_create_data_internal rx_data;
280 struct delayed_work rdma_connect;
281 struct rpcrdma_stats rx_stats;
282};
283
284#define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, xprt)
285#define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data)
286
287/* Setting this to 0 ensures interoperability with early servers.
288 * Setting this to 1 enhances certain unaligned read/write performance.
289 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
290extern int xprt_rdma_pad_optimize;
291
292/*
293 * Interface Adapter calls - xprtrdma/verbs.c
294 */
295int rpcrdma_ia_open(struct rpcrdma_xprt *, struct sockaddr *, int);
296void rpcrdma_ia_close(struct rpcrdma_ia *);
297
298/*
299 * Endpoint calls - xprtrdma/verbs.c
300 */
301int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *,
302 struct rpcrdma_create_data_internal *);
303int rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *);
304int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *);
305int rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *);
306
307int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
308 struct rpcrdma_req *);
309int rpcrdma_ep_post_recv(struct rpcrdma_ia *, struct rpcrdma_ep *,
310 struct rpcrdma_rep *);
311
312/*
313 * Buffer calls - xprtrdma/verbs.c
314 */
315int rpcrdma_buffer_create(struct rpcrdma_buffer *, struct rpcrdma_ep *,
316 struct rpcrdma_ia *,
317 struct rpcrdma_create_data_internal *);
318void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
319
320struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
321void rpcrdma_buffer_put(struct rpcrdma_req *);
322void rpcrdma_recv_buffer_get(struct rpcrdma_req *);
323void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
324
325int rpcrdma_register_internal(struct rpcrdma_ia *, void *, int,
326 struct ib_mr **, struct ib_sge *);
327int rpcrdma_deregister_internal(struct rpcrdma_ia *,
328 struct ib_mr *, struct ib_sge *);
329
330int rpcrdma_register_external(struct rpcrdma_mr_seg *,
331 int, int, struct rpcrdma_xprt *);
332int rpcrdma_deregister_external(struct rpcrdma_mr_seg *,
333 struct rpcrdma_xprt *, void *);
334
335/*
336 * RPC/RDMA connection management calls - xprtrdma/rpc_rdma.c
337 */
338void rpcrdma_conn_func(struct rpcrdma_ep *);
339void rpcrdma_reply_handler(struct rpcrdma_rep *);
340
341/*
342 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
343 */
344int rpcrdma_marshal_req(struct rpc_rqst *);
345
346/* Temporary NFS request map cache. Created in svc_rdma.c */
347extern struct kmem_cache *svc_rdma_map_cachep;
348/* WR context cache. Created in svc_rdma.c */
349extern struct kmem_cache *svc_rdma_ctxt_cachep;
350/* Workqueue created in svc_rdma.c */
351extern struct workqueue_struct *svc_rdma_wq;
352
353#endif /* _LINUX_SUNRPC_XPRT_RDMA_H */