2 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
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
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
14 * Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
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.
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
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.
40 #ifndef _LINUX_SUNRPC_XPRT_RDMA_H
41 #define _LINUX_SUNRPC_XPRT_RDMA_H
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 */
48 #include <rdma/rdma_cm.h> /* RDMA connection api */
49 #include <rdma/ib_verbs.h> /* RDMA verbs api */
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 */
55 #define RDMA_RESOLVE_TIMEOUT (5000) /* 5 seconds */
56 #define RDMA_CONNECT_RETRY_MAX (2) /* retries if no listener backlog */
59 * Interface Adapter -- one per transport instance
62 const struct rpcrdma_memreg_ops
*ri_ops
;
64 struct ib_device
*ri_device
;
65 struct rdma_cm_id
*ri_id
;
67 struct ib_mr
*ri_dma_mr
;
68 struct completion ri_done
;
70 unsigned int ri_max_frmr_depth
;
71 struct ib_qp_attr ri_qp_attr
;
72 struct ib_qp_init_attr ri_qp_init_attr
;
76 * RDMA Endpoint -- one per transport instance
83 struct ib_qp_init_attr rep_attr
;
84 wait_queue_head_t rep_connect_wait
;
85 struct rdma_conn_param rep_remote_cma
;
86 struct sockaddr_storage rep_remote_addr
;
87 struct delayed_work rep_connect_worker
;
91 * Force a signaled SEND Work Request every so often,
92 * in case the provider needs to do some housekeeping.
94 #define RPCRDMA_MAX_UNSIGNALED_SENDS (32)
96 #define INIT_CQCOUNT(ep) atomic_set(&(ep)->rep_cqcount, (ep)->rep_cqinit)
97 #define DECR_CQCOUNT(ep) atomic_sub_return(1, &(ep)->rep_cqcount)
99 /* Force completion handler to ignore the signal
101 #define RPCRDMA_IGNORE_COMPLETION (0ULL)
103 /* Pre-allocate extra Work Requests for handling backward receives
104 * and sends. This is a fixed value because the Work Queues are
105 * allocated when the forward channel is set up.
107 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
108 #define RPCRDMA_BACKWARD_WRS (8)
110 #define RPCRDMA_BACKWARD_WRS (0)
113 /* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
115 * The below structure appears at the front of a large region of kmalloc'd
116 * memory, which always starts on a good alignment boundary.
119 struct rpcrdma_regbuf
{
121 struct rpcrdma_req
*rg_owner
;
122 struct ib_sge rg_iov
;
123 __be32 rg_base
[0] __attribute__ ((aligned(256)));
127 rdmab_addr(struct rpcrdma_regbuf
*rb
)
129 return rb
->rg_iov
.addr
;
133 rdmab_length(struct rpcrdma_regbuf
*rb
)
135 return rb
->rg_iov
.length
;
139 rdmab_lkey(struct rpcrdma_regbuf
*rb
)
141 return rb
->rg_iov
.lkey
;
144 static inline struct rpcrdma_msg
*
145 rdmab_to_msg(struct rpcrdma_regbuf
*rb
)
147 return (struct rpcrdma_msg
*)rb
->rg_base
;
151 * struct rpcrdma_rep -- this structure encapsulates state required to recv
152 * and complete a reply, asychronously. It needs several pieces of
154 * o recv buffer (posted to provider)
155 * o ib_sge (also donated to provider)
156 * o status of reply (length, success or not)
157 * o bookkeeping state to get run by tasklet (list, etc)
159 * These are allocated during initialization, per-transport instance;
160 * however, the tasklet execution list itself is global, as it should
161 * always be pretty short.
163 * N of these are associated with a transport instance, and stored in
164 * struct rpcrdma_buffer. N is the max number of outstanding requests.
167 #define RPCRDMA_MAX_DATA_SEGS ((1 * 1024 * 1024) / PAGE_SIZE)
168 #define RPCRDMA_MAX_SEGS (RPCRDMA_MAX_DATA_SEGS + 2) /* head+tail = 2 */
170 struct rpcrdma_buffer
;
174 struct ib_device
*rr_device
;
175 struct rpcrdma_xprt
*rr_rxprt
;
176 struct work_struct rr_work
;
177 struct list_head rr_list
;
178 struct rpcrdma_regbuf
*rr_rdmabuf
;
181 #define RPCRDMA_BAD_LEN (~0U)
184 * struct rpcrdma_mw - external memory region metadata
186 * An external memory region is any buffer or page that is registered
187 * on the fly (ie, not pre-registered).
189 * Each rpcrdma_buffer has a list of free MWs anchored in rb_mws. During
190 * call_allocate, rpcrdma_buffer_get() assigns one to each segment in
191 * an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep
192 * track of registration metadata while each RPC is pending.
193 * rpcrdma_deregister_external() uses this metadata to unmap and
194 * release these resources when an RPC is complete.
196 enum rpcrdma_frmr_state
{
197 FRMR_IS_INVALID
, /* ready to be used */
198 FRMR_IS_VALID
, /* in use */
199 FRMR_IS_STALE
, /* failed completion */
202 struct rpcrdma_frmr
{
203 struct scatterlist
*sg
;
206 enum rpcrdma_frmr_state fr_state
;
207 struct work_struct fr_work
;
208 struct rpcrdma_xprt
*fr_xprt
;
218 struct rpcrdma_fmr fmr
;
219 struct rpcrdma_frmr frmr
;
221 void (*mw_sendcompletion
)(struct ib_wc
*);
222 struct list_head mw_list
;
223 struct list_head mw_all
;
227 * struct rpcrdma_req -- structure central to the request/reply sequence.
229 * N of these are associated with a transport instance, and stored in
230 * struct rpcrdma_buffer. N is the max number of outstanding requests.
232 * It includes pre-registered buffer memory for send AND recv.
233 * The recv buffer, however, is not owned by this structure, and
234 * is "donated" to the hardware when a recv is posted. When a
235 * reply is handled, the recv buffer used is given back to the
236 * struct rpcrdma_req associated with the request.
238 * In addition to the basic memory, this structure includes an array
239 * of iovs for send operations. The reason is that the iovs passed to
240 * ib_post_{send,recv} must not be modified until the work request
244 * o RPCRDMA_MAX_SEGS is the max number of addressible chunk elements we
245 * marshal. The number needed varies depending on the iov lists that
246 * are passed to us, the memory registration mode we are in, and if
247 * physical addressing is used, the layout.
250 struct rpcrdma_mr_seg
{ /* chunk descriptors */
251 struct rpcrdma_mw
*rl_mw
; /* registered MR */
252 u64 mr_base
; /* registration result */
253 u32 mr_rkey
; /* registration result */
254 u32 mr_len
; /* length of chunk or segment */
255 int mr_nsegs
; /* number of segments in chunk or 0 */
256 enum dma_data_direction mr_dir
; /* segment mapping direction */
257 dma_addr_t mr_dma
; /* segment mapping address */
258 size_t mr_dmalen
; /* segment mapping length */
259 struct page
*mr_page
; /* owning page, if any */
260 char *mr_offset
; /* kva if no page, else offset */
263 #define RPCRDMA_MAX_IOVS (2)
266 struct list_head rl_free
;
267 unsigned int rl_niovs
;
268 unsigned int rl_nchunks
;
269 unsigned int rl_connect_cookie
;
270 struct rpcrdma_buffer
*rl_buffer
;
271 struct rpcrdma_rep
*rl_reply
;/* holder for reply buffer */
272 struct ib_sge rl_send_iov
[RPCRDMA_MAX_IOVS
];
273 struct rpcrdma_regbuf
*rl_rdmabuf
;
274 struct rpcrdma_regbuf
*rl_sendbuf
;
275 struct rpcrdma_mr_seg rl_segments
[RPCRDMA_MAX_SEGS
];
277 struct list_head rl_all
;
281 static inline struct rpcrdma_req
*
282 rpcr_to_rdmar(struct rpc_rqst
*rqst
)
284 void *buffer
= rqst
->rq_buffer
;
285 struct rpcrdma_regbuf
*rb
;
287 rb
= container_of(buffer
, struct rpcrdma_regbuf
, rg_base
);
292 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
293 * inline requests/replies, and client/server credits.
295 * One of these is associated with a transport instance
297 struct rpcrdma_buffer
{
298 spinlock_t rb_mwlock
; /* protect rb_mws list */
299 struct list_head rb_mws
;
300 struct list_head rb_all
;
303 spinlock_t rb_lock
; /* protect buf lists */
304 struct list_head rb_send_bufs
;
305 struct list_head rb_recv_bufs
;
308 u32 rb_bc_srv_max_requests
;
309 spinlock_t rb_reqslock
; /* protect rb_allreqs */
310 struct list_head rb_allreqs
;
312 #define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia)
315 * Internal structure for transport instance creation. This
316 * exists primarily for modularity.
318 * This data should be set with mount options
320 struct rpcrdma_create_data_internal
{
321 struct sockaddr_storage addr
; /* RDMA server address */
322 unsigned int max_requests
; /* max requests (slots) in flight */
323 unsigned int rsize
; /* mount rsize - max read hdr+data */
324 unsigned int wsize
; /* mount wsize - max write hdr+data */
325 unsigned int inline_rsize
; /* max non-rdma read data payload */
326 unsigned int inline_wsize
; /* max non-rdma write data payload */
327 unsigned int padding
; /* non-rdma write header padding */
330 #define RPCRDMA_INLINE_READ_THRESHOLD(rq) \
331 (rpcx_to_rdmad(rq->rq_xprt).inline_rsize)
333 #define RPCRDMA_INLINE_WRITE_THRESHOLD(rq)\
334 (rpcx_to_rdmad(rq->rq_xprt).inline_wsize)
336 #define RPCRDMA_INLINE_PAD_VALUE(rq)\
337 rpcx_to_rdmad(rq->rq_xprt).padding
340 * Statistics for RPCRDMA
342 struct rpcrdma_stats
{
343 unsigned long read_chunk_count
;
344 unsigned long write_chunk_count
;
345 unsigned long reply_chunk_count
;
347 unsigned long long total_rdma_request
;
348 unsigned long long total_rdma_reply
;
350 unsigned long long pullup_copy_count
;
351 unsigned long long fixup_copy_count
;
352 unsigned long hardway_register_count
;
353 unsigned long failed_marshal_count
;
354 unsigned long bad_reply_count
;
355 unsigned long nomsg_call_count
;
356 unsigned long bcall_count
;
360 * Per-registration mode operations
363 struct rpcrdma_memreg_ops
{
364 int (*ro_map
)(struct rpcrdma_xprt
*,
365 struct rpcrdma_mr_seg
*, int, bool);
366 int (*ro_unmap
)(struct rpcrdma_xprt
*,
367 struct rpcrdma_mr_seg
*);
368 int (*ro_open
)(struct rpcrdma_ia
*,
370 struct rpcrdma_create_data_internal
*);
371 size_t (*ro_maxpages
)(struct rpcrdma_xprt
*);
372 int (*ro_init
)(struct rpcrdma_xprt
*);
373 void (*ro_destroy
)(struct rpcrdma_buffer
*);
374 const char *ro_displayname
;
377 extern const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops
;
378 extern const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops
;
379 extern const struct rpcrdma_memreg_ops rpcrdma_physical_memreg_ops
;
382 * RPCRDMA transport -- encapsulates the structures above for
383 * integration with RPC.
385 * The contained structures are embedded, not pointers,
386 * for convenience. This structure need not be visible externally.
388 * It is allocated and initialized during mount, and released
391 struct rpcrdma_xprt
{
392 struct rpc_xprt rx_xprt
;
393 struct rpcrdma_ia rx_ia
;
394 struct rpcrdma_ep rx_ep
;
395 struct rpcrdma_buffer rx_buf
;
396 struct rpcrdma_create_data_internal rx_data
;
397 struct delayed_work rx_connect_worker
;
398 struct rpcrdma_stats rx_stats
;
401 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
402 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data)
404 /* Setting this to 0 ensures interoperability with early servers.
405 * Setting this to 1 enhances certain unaligned read/write performance.
406 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
407 extern int xprt_rdma_pad_optimize
;
410 * Interface Adapter calls - xprtrdma/verbs.c
412 int rpcrdma_ia_open(struct rpcrdma_xprt
*, struct sockaddr
*, int);
413 void rpcrdma_ia_close(struct rpcrdma_ia
*);
416 * Endpoint calls - xprtrdma/verbs.c
418 int rpcrdma_ep_create(struct rpcrdma_ep
*, struct rpcrdma_ia
*,
419 struct rpcrdma_create_data_internal
*);
420 void rpcrdma_ep_destroy(struct rpcrdma_ep
*, struct rpcrdma_ia
*);
421 int rpcrdma_ep_connect(struct rpcrdma_ep
*, struct rpcrdma_ia
*);
422 void rpcrdma_ep_disconnect(struct rpcrdma_ep
*, struct rpcrdma_ia
*);
424 int rpcrdma_ep_post(struct rpcrdma_ia
*, struct rpcrdma_ep
*,
425 struct rpcrdma_req
*);
426 int rpcrdma_ep_post_recv(struct rpcrdma_ia
*, struct rpcrdma_ep
*,
427 struct rpcrdma_rep
*);
430 * Buffer calls - xprtrdma/verbs.c
432 struct rpcrdma_req
*rpcrdma_create_req(struct rpcrdma_xprt
*);
433 struct rpcrdma_rep
*rpcrdma_create_rep(struct rpcrdma_xprt
*);
434 void rpcrdma_destroy_req(struct rpcrdma_ia
*, struct rpcrdma_req
*);
435 int rpcrdma_buffer_create(struct rpcrdma_xprt
*);
436 void rpcrdma_buffer_destroy(struct rpcrdma_buffer
*);
438 struct rpcrdma_mw
*rpcrdma_get_mw(struct rpcrdma_xprt
*);
439 void rpcrdma_put_mw(struct rpcrdma_xprt
*, struct rpcrdma_mw
*);
440 struct rpcrdma_req
*rpcrdma_buffer_get(struct rpcrdma_buffer
*);
441 void rpcrdma_buffer_put(struct rpcrdma_req
*);
442 void rpcrdma_recv_buffer_get(struct rpcrdma_req
*);
443 void rpcrdma_recv_buffer_put(struct rpcrdma_rep
*);
445 struct rpcrdma_regbuf
*rpcrdma_alloc_regbuf(struct rpcrdma_ia
*,
447 void rpcrdma_free_regbuf(struct rpcrdma_ia
*,
448 struct rpcrdma_regbuf
*);
450 unsigned int rpcrdma_max_segments(struct rpcrdma_xprt
*);
451 int rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt
*, unsigned int);
453 int frwr_alloc_recovery_wq(void);
454 void frwr_destroy_recovery_wq(void);
456 int rpcrdma_alloc_wq(void);
457 void rpcrdma_destroy_wq(void);
460 * Wrappers for chunk registration, shared by read/write chunk code.
463 void rpcrdma_mapping_error(struct rpcrdma_mr_seg
*);
465 static inline enum dma_data_direction
466 rpcrdma_data_dir(bool writing
)
468 return writing
? DMA_FROM_DEVICE
: DMA_TO_DEVICE
;
472 rpcrdma_map_one(struct ib_device
*device
, struct rpcrdma_mr_seg
*seg
,
473 enum dma_data_direction direction
)
475 seg
->mr_dir
= direction
;
476 seg
->mr_dmalen
= seg
->mr_len
;
479 seg
->mr_dma
= ib_dma_map_page(device
,
480 seg
->mr_page
, offset_in_page(seg
->mr_offset
),
481 seg
->mr_dmalen
, seg
->mr_dir
);
483 seg
->mr_dma
= ib_dma_map_single(device
,
485 seg
->mr_dmalen
, seg
->mr_dir
);
487 if (ib_dma_mapping_error(device
, seg
->mr_dma
))
488 rpcrdma_mapping_error(seg
);
492 rpcrdma_unmap_one(struct ib_device
*device
, struct rpcrdma_mr_seg
*seg
)
495 ib_dma_unmap_page(device
,
496 seg
->mr_dma
, seg
->mr_dmalen
, seg
->mr_dir
);
498 ib_dma_unmap_single(device
,
499 seg
->mr_dma
, seg
->mr_dmalen
, seg
->mr_dir
);
503 * RPC/RDMA connection management calls - xprtrdma/rpc_rdma.c
505 void rpcrdma_connect_worker(struct work_struct
*);
506 void rpcrdma_conn_func(struct rpcrdma_ep
*);
507 void rpcrdma_reply_handler(struct rpcrdma_rep
*);
510 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
512 int rpcrdma_marshal_req(struct rpc_rqst
*);
514 /* RPC/RDMA module init - xprtrdma/transport.c
516 int xprt_rdma_init(void);
517 void xprt_rdma_cleanup(void);
519 /* Backchannel calls - xprtrdma/backchannel.c
521 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
522 int xprt_rdma_bc_setup(struct rpc_xprt
*, unsigned int);
523 int xprt_rdma_bc_up(struct svc_serv
*, struct net
*);
524 int rpcrdma_bc_post_recv(struct rpcrdma_xprt
*, unsigned int);
525 void rpcrdma_bc_receive_call(struct rpcrdma_xprt
*, struct rpcrdma_rep
*);
526 int rpcrdma_bc_marshal_reply(struct rpc_rqst
*);
527 void xprt_rdma_bc_free_rqst(struct rpc_rqst
*);
528 void xprt_rdma_bc_destroy(struct rpc_xprt
*, unsigned int);
529 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
531 /* Temporary NFS request map cache. Created in svc_rdma.c */
532 extern struct kmem_cache
*svc_rdma_map_cachep
;
533 /* WR context cache. Created in svc_rdma.c */
534 extern struct kmem_cache
*svc_rdma_ctxt_cachep
;
535 /* Workqueue created in svc_rdma.c */
536 extern struct workqueue_struct
*svc_rdma_wq
;
538 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */