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 */
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)
64 * Interface Adapter -- one per transport instance
67 const struct rpcrdma_memreg_ops
*ri_ops
;
69 struct ib_device
*ri_device
;
70 struct rdma_cm_id
*ri_id
;
72 struct ib_mr
*ri_dma_mr
;
73 struct completion ri_done
;
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
;
81 * RDMA Endpoint -- one per transport instance
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
;
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)
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.
102 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
103 #define RPCRDMA_BACKWARD_WRS (8)
105 #define RPCRDMA_BACKWARD_WRS (0)
108 /* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
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.
114 struct rpcrdma_regbuf
{
116 struct rpcrdma_req
*rg_owner
;
117 struct ib_sge rg_iov
;
118 __be32 rg_base
[0] __attribute__ ((aligned(256)));
122 rdmab_addr(struct rpcrdma_regbuf
*rb
)
124 return rb
->rg_iov
.addr
;
128 rdmab_length(struct rpcrdma_regbuf
*rb
)
130 return rb
->rg_iov
.length
;
134 rdmab_lkey(struct rpcrdma_regbuf
*rb
)
136 return rb
->rg_iov
.lkey
;
139 static inline struct rpcrdma_msg
*
140 rdmab_to_msg(struct rpcrdma_regbuf
*rb
)
142 return (struct rpcrdma_msg
*)rb
->rg_base
;
145 #define RPCRDMA_DEF_GFP (GFP_NOIO | __GFP_NOWARN)
148 * struct rpcrdma_rep -- this structure encapsulates state required to recv
149 * and complete a reply, asychronously. It needs several pieces of
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)
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.
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.
164 #define RPCRDMA_MAX_DATA_SEGS ((1 * 1024 * 1024) / PAGE_SIZE)
165 #define RPCRDMA_MAX_SEGS (RPCRDMA_MAX_DATA_SEGS + 2) /* head+tail = 2 */
167 struct rpcrdma_buffer
;
170 struct ib_cqe rr_cqe
;
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
;
179 #define RPCRDMA_BAD_LEN (~0U)
182 * struct rpcrdma_mw - external memory region metadata
184 * An external memory region is any buffer or page that is registered
185 * on the fly (ie, not pre-registered).
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.
194 enum rpcrdma_frmr_state
{
195 FRMR_IS_INVALID
, /* ready to be used */
196 FRMR_IS_VALID
, /* in use */
197 FRMR_IS_STALE
, /* failed completion */
200 struct rpcrdma_frmr
{
201 struct scatterlist
*sg
;
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
;
210 struct ib_reg_wr fr_regwr
;
211 struct ib_send_wr fr_invwr
;
222 struct rpcrdma_fmr fmr
;
223 struct rpcrdma_frmr frmr
;
225 struct list_head mw_list
;
226 struct list_head mw_all
;
230 * struct rpcrdma_req -- structure central to the request/reply sequence.
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.
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.
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
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.
253 struct 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 */
266 #define RPCRDMA_MAX_IOVS (2)
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
];
280 struct ib_cqe rl_cqe
;
281 struct list_head rl_all
;
285 static inline struct rpcrdma_req
*
286 rpcr_to_rdmar(struct rpc_rqst
*rqst
)
288 void *buffer
= rqst
->rq_buffer
;
289 struct rpcrdma_regbuf
*rb
;
291 rb
= container_of(buffer
, struct rpcrdma_regbuf
, rg_base
);
296 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
297 * inline requests/replies, and client/server credits.
299 * One of these is associated with a transport instance
301 struct rpcrdma_buffer
{
302 spinlock_t rb_mwlock
; /* protect rb_mws list */
303 struct list_head rb_mws
;
304 struct list_head rb_all
;
307 spinlock_t rb_lock
; /* protect buf lists */
308 struct list_head rb_send_bufs
;
309 struct list_head rb_recv_bufs
;
311 atomic_t rb_credits
; /* most recent credit grant */
313 u32 rb_bc_srv_max_requests
;
314 spinlock_t rb_reqslock
; /* protect rb_allreqs */
315 struct list_head rb_allreqs
;
317 u32 rb_bc_max_requests
;
319 #define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia)
322 * Internal structure for transport instance creation. This
323 * exists primarily for modularity.
325 * This data should be set with mount options
327 struct 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 */
337 #define RPCRDMA_INLINE_READ_THRESHOLD(rq) \
338 (rpcx_to_rdmad(rq->rq_xprt).inline_rsize)
340 #define RPCRDMA_INLINE_WRITE_THRESHOLD(rq)\
341 (rpcx_to_rdmad(rq->rq_xprt).inline_wsize)
343 #define RPCRDMA_INLINE_PAD_VALUE(rq)\
344 rpcx_to_rdmad(rq->rq_xprt).padding
347 * Statistics for RPCRDMA
349 struct rpcrdma_stats
{
350 unsigned long read_chunk_count
;
351 unsigned long write_chunk_count
;
352 unsigned long reply_chunk_count
;
354 unsigned long long total_rdma_request
;
355 unsigned long long total_rdma_reply
;
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
;
367 * Per-registration mode operations
370 struct 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
*,
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
;
386 extern const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops
;
387 extern const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops
;
388 extern const struct rpcrdma_memreg_ops rpcrdma_physical_memreg_ops
;
391 * RPCRDMA transport -- encapsulates the structures above for
392 * integration with RPC.
394 * The contained structures are embedded, not pointers,
395 * for convenience. This structure need not be visible externally.
397 * It is allocated and initialized during mount, and released
400 struct 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
;
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)
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() */
416 extern int xprt_rdma_pad_optimize
;
419 * Interface Adapter calls - xprtrdma/verbs.c
421 int rpcrdma_ia_open(struct rpcrdma_xprt
*, struct sockaddr
*, int);
422 void rpcrdma_ia_close(struct rpcrdma_ia
*);
425 * Endpoint calls - xprtrdma/verbs.c
427 int rpcrdma_ep_create(struct rpcrdma_ep
*, struct rpcrdma_ia
*,
428 struct rpcrdma_create_data_internal
*);
429 void rpcrdma_ep_destroy(struct rpcrdma_ep
*, struct rpcrdma_ia
*);
430 int rpcrdma_ep_connect(struct rpcrdma_ep
*, struct rpcrdma_ia
*);
431 void rpcrdma_ep_disconnect(struct rpcrdma_ep
*, struct rpcrdma_ia
*);
433 int rpcrdma_ep_post(struct rpcrdma_ia
*, struct rpcrdma_ep
*,
434 struct rpcrdma_req
*);
435 int rpcrdma_ep_post_recv(struct rpcrdma_ia
*, struct rpcrdma_ep
*,
436 struct rpcrdma_rep
*);
439 * Buffer calls - xprtrdma/verbs.c
441 struct rpcrdma_req
*rpcrdma_create_req(struct rpcrdma_xprt
*);
442 struct rpcrdma_rep
*rpcrdma_create_rep(struct rpcrdma_xprt
*);
443 void rpcrdma_destroy_req(struct rpcrdma_ia
*, struct rpcrdma_req
*);
444 int rpcrdma_buffer_create(struct rpcrdma_xprt
*);
445 void rpcrdma_buffer_destroy(struct rpcrdma_buffer
*);
447 struct rpcrdma_mw
*rpcrdma_get_mw(struct rpcrdma_xprt
*);
448 void rpcrdma_put_mw(struct rpcrdma_xprt
*, struct rpcrdma_mw
*);
449 struct rpcrdma_req
*rpcrdma_buffer_get(struct rpcrdma_buffer
*);
450 void rpcrdma_buffer_put(struct rpcrdma_req
*);
451 void rpcrdma_recv_buffer_get(struct rpcrdma_req
*);
452 void rpcrdma_recv_buffer_put(struct rpcrdma_rep
*);
454 struct rpcrdma_regbuf
*rpcrdma_alloc_regbuf(struct rpcrdma_ia
*,
456 void rpcrdma_free_regbuf(struct rpcrdma_ia
*,
457 struct rpcrdma_regbuf
*);
459 unsigned int rpcrdma_max_segments(struct rpcrdma_xprt
*);
460 int rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt
*, unsigned int);
462 int frwr_alloc_recovery_wq(void);
463 void frwr_destroy_recovery_wq(void);
465 int rpcrdma_alloc_wq(void);
466 void rpcrdma_destroy_wq(void);
469 * Wrappers for chunk registration, shared by read/write chunk code.
472 void rpcrdma_mapping_error(struct rpcrdma_mr_seg
*);
474 static inline enum dma_data_direction
475 rpcrdma_data_dir(bool writing
)
477 return writing
? DMA_FROM_DEVICE
: DMA_TO_DEVICE
;
481 rpcrdma_map_one(struct ib_device
*device
, struct rpcrdma_mr_seg
*seg
,
482 enum dma_data_direction direction
)
484 seg
->mr_dir
= direction
;
485 seg
->mr_dmalen
= seg
->mr_len
;
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
);
492 seg
->mr_dma
= ib_dma_map_single(device
,
494 seg
->mr_dmalen
, seg
->mr_dir
);
496 if (ib_dma_mapping_error(device
, seg
->mr_dma
))
497 rpcrdma_mapping_error(seg
);
501 rpcrdma_unmap_one(struct ib_device
*device
, struct rpcrdma_mr_seg
*seg
)
504 ib_dma_unmap_page(device
,
505 seg
->mr_dma
, seg
->mr_dmalen
, seg
->mr_dir
);
507 ib_dma_unmap_single(device
,
508 seg
->mr_dma
, seg
->mr_dmalen
, seg
->mr_dir
);
512 * RPC/RDMA connection management calls - xprtrdma/rpc_rdma.c
514 void rpcrdma_connect_worker(struct work_struct
*);
515 void rpcrdma_conn_func(struct rpcrdma_ep
*);
516 void rpcrdma_reply_handler(struct rpcrdma_rep
*);
519 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
521 int rpcrdma_marshal_req(struct rpc_rqst
*);
523 /* RPC/RDMA module init - xprtrdma/transport.c
525 extern unsigned int xprt_rdma_max_inline_read
;
526 void xprt_rdma_format_addresses(struct rpc_xprt
*xprt
, struct sockaddr
*sap
);
527 void xprt_rdma_free_addresses(struct rpc_xprt
*xprt
);
528 void xprt_rdma_print_stats(struct rpc_xprt
*xprt
, struct seq_file
*seq
);
529 int xprt_rdma_init(void);
530 void xprt_rdma_cleanup(void);
532 /* Backchannel calls - xprtrdma/backchannel.c
534 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
535 int xprt_rdma_bc_setup(struct rpc_xprt
*, unsigned int);
536 int xprt_rdma_bc_up(struct svc_serv
*, struct net
*);
537 int rpcrdma_bc_post_recv(struct rpcrdma_xprt
*, unsigned int);
538 void rpcrdma_bc_receive_call(struct rpcrdma_xprt
*, struct rpcrdma_rep
*);
539 int rpcrdma_bc_marshal_reply(struct rpc_rqst
*);
540 void xprt_rdma_bc_free_rqst(struct rpc_rqst
*);
541 void xprt_rdma_bc_destroy(struct rpc_xprt
*, unsigned int);
542 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
544 extern struct xprt_class xprt_rdma_bc
;
546 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */