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
;
68 struct ib_device
*ri_device
;
69 struct rdma_cm_id
*ri_id
;
71 struct ib_mr
*ri_dma_mr
;
72 struct completion ri_done
;
74 unsigned int ri_max_frmr_depth
;
75 unsigned int ri_max_inline_write
;
76 unsigned int ri_max_inline_read
;
77 struct ib_qp_attr ri_qp_attr
;
78 struct ib_qp_init_attr ri_qp_init_attr
;
82 * RDMA Endpoint -- one per transport instance
89 struct ib_qp_init_attr rep_attr
;
90 wait_queue_head_t rep_connect_wait
;
91 struct rdma_conn_param rep_remote_cma
;
92 struct sockaddr_storage rep_remote_addr
;
93 struct delayed_work rep_connect_worker
;
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 /* Pre-allocate extra Work Requests for handling backward receives
100 * and sends. This is a fixed value because the Work Queues are
101 * allocated when the forward channel is set up.
103 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
104 #define RPCRDMA_BACKWARD_WRS (8)
106 #define RPCRDMA_BACKWARD_WRS (0)
109 /* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
111 * The below structure appears at the front of a large region of kmalloc'd
112 * memory, which always starts on a good alignment boundary.
115 struct rpcrdma_regbuf
{
117 struct rpcrdma_req
*rg_owner
;
118 struct ib_sge rg_iov
;
119 __be32 rg_base
[0] __attribute__ ((aligned(256)));
123 rdmab_addr(struct rpcrdma_regbuf
*rb
)
125 return rb
->rg_iov
.addr
;
129 rdmab_length(struct rpcrdma_regbuf
*rb
)
131 return rb
->rg_iov
.length
;
135 rdmab_lkey(struct rpcrdma_regbuf
*rb
)
137 return rb
->rg_iov
.lkey
;
140 static inline struct rpcrdma_msg
*
141 rdmab_to_msg(struct rpcrdma_regbuf
*rb
)
143 return (struct rpcrdma_msg
*)rb
->rg_base
;
146 #define RPCRDMA_DEF_GFP (GFP_NOIO | __GFP_NOWARN)
148 /* To ensure a transport can always make forward progress,
149 * the number of RDMA segments allowed in header chunk lists
150 * is capped at 8. This prevents less-capable devices and
151 * memory registrations from overrunning the Send buffer
152 * while building chunk lists.
154 * Elements of the Read list take up more room than the
155 * Write list or Reply chunk. 8 read segments means the Read
156 * list (or Write list or Reply chunk) cannot consume more
159 * ((8 + 2) * read segment size) + 1 XDR words, or 244 bytes.
161 * And the fixed part of the header is another 24 bytes.
163 * The smallest inline threshold is 1024 bytes, ensuring that
164 * at least 750 bytes are available for RPC messages.
166 #define RPCRDMA_MAX_HDR_SEGS (8)
169 * struct rpcrdma_rep -- this structure encapsulates state required to recv
170 * and complete a reply, asychronously. It needs several pieces of
172 * o recv buffer (posted to provider)
173 * o ib_sge (also donated to provider)
174 * o status of reply (length, success or not)
175 * o bookkeeping state to get run by tasklet (list, etc)
177 * These are allocated during initialization, per-transport instance;
178 * however, the tasklet execution list itself is global, as it should
179 * always be pretty short.
181 * N of these are associated with a transport instance, and stored in
182 * struct rpcrdma_buffer. N is the max number of outstanding requests.
185 #define RPCRDMA_MAX_DATA_SEGS ((1 * 1024 * 1024) / PAGE_SIZE)
187 /* data segments + head/tail for Call + head/tail for Reply */
188 #define RPCRDMA_MAX_SEGS (RPCRDMA_MAX_DATA_SEGS + 4)
190 struct rpcrdma_buffer
;
193 struct ib_cqe rr_cqe
;
195 struct ib_device
*rr_device
;
196 struct rpcrdma_xprt
*rr_rxprt
;
197 struct work_struct rr_work
;
198 struct list_head rr_list
;
199 struct rpcrdma_regbuf
*rr_rdmabuf
;
202 #define RPCRDMA_BAD_LEN (~0U)
205 * struct rpcrdma_mw - external memory region metadata
207 * An external memory region is any buffer or page that is registered
208 * on the fly (ie, not pre-registered).
210 * Each rpcrdma_buffer has a list of free MWs anchored in rb_mws. During
211 * call_allocate, rpcrdma_buffer_get() assigns one to each segment in
212 * an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep
213 * track of registration metadata while each RPC is pending.
214 * rpcrdma_deregister_external() uses this metadata to unmap and
215 * release these resources when an RPC is complete.
217 enum rpcrdma_frmr_state
{
218 FRMR_IS_INVALID
, /* ready to be used */
219 FRMR_IS_VALID
, /* in use */
220 FRMR_IS_STALE
, /* failed completion */
223 struct rpcrdma_frmr
{
224 struct scatterlist
*fr_sg
;
226 enum dma_data_direction fr_dir
;
228 struct ib_cqe fr_cqe
;
229 enum rpcrdma_frmr_state fr_state
;
230 struct completion fr_linv_done
;
232 struct ib_reg_wr fr_regwr
;
233 struct ib_send_wr fr_invwr
;
244 struct rpcrdma_fmr fmr
;
245 struct rpcrdma_frmr frmr
;
247 struct work_struct mw_work
;
248 struct rpcrdma_xprt
*mw_xprt
;
249 struct list_head mw_list
;
250 struct list_head mw_all
;
254 * struct rpcrdma_req -- structure central to the request/reply sequence.
256 * N of these are associated with a transport instance, and stored in
257 * struct rpcrdma_buffer. N is the max number of outstanding requests.
259 * It includes pre-registered buffer memory for send AND recv.
260 * The recv buffer, however, is not owned by this structure, and
261 * is "donated" to the hardware when a recv is posted. When a
262 * reply is handled, the recv buffer used is given back to the
263 * struct rpcrdma_req associated with the request.
265 * In addition to the basic memory, this structure includes an array
266 * of iovs for send operations. The reason is that the iovs passed to
267 * ib_post_{send,recv} must not be modified until the work request
271 * o RPCRDMA_MAX_SEGS is the max number of addressible chunk elements we
272 * marshal. The number needed varies depending on the iov lists that
273 * are passed to us, the memory registration mode we are in, and if
274 * physical addressing is used, the layout.
277 struct rpcrdma_mr_seg
{ /* chunk descriptors */
278 struct rpcrdma_mw
*rl_mw
; /* registered MR */
279 u64 mr_base
; /* registration result */
280 u32 mr_rkey
; /* registration result */
281 u32 mr_len
; /* length of chunk or segment */
282 int mr_nsegs
; /* number of segments in chunk or 0 */
283 enum dma_data_direction mr_dir
; /* segment mapping direction */
284 dma_addr_t mr_dma
; /* segment mapping address */
285 size_t mr_dmalen
; /* segment mapping length */
286 struct page
*mr_page
; /* owning page, if any */
287 char *mr_offset
; /* kva if no page, else offset */
290 #define RPCRDMA_MAX_IOVS (2)
293 struct list_head rl_free
;
294 unsigned int rl_niovs
;
295 unsigned int rl_nchunks
;
296 unsigned int rl_connect_cookie
;
297 struct rpc_task
*rl_task
;
298 struct rpcrdma_buffer
*rl_buffer
;
299 struct rpcrdma_rep
*rl_reply
;/* holder for reply buffer */
300 struct ib_sge rl_send_iov
[RPCRDMA_MAX_IOVS
];
301 struct rpcrdma_regbuf
*rl_rdmabuf
;
302 struct rpcrdma_regbuf
*rl_sendbuf
;
303 struct rpcrdma_mr_seg rl_segments
[RPCRDMA_MAX_SEGS
];
304 struct rpcrdma_mr_seg
*rl_nextseg
;
306 struct ib_cqe rl_cqe
;
307 struct list_head rl_all
;
311 static inline struct rpcrdma_req
*
312 rpcr_to_rdmar(struct rpc_rqst
*rqst
)
314 void *buffer
= rqst
->rq_buffer
;
315 struct rpcrdma_regbuf
*rb
;
317 rb
= container_of(buffer
, struct rpcrdma_regbuf
, rg_base
);
322 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
323 * inline requests/replies, and client/server credits.
325 * One of these is associated with a transport instance
327 struct rpcrdma_buffer
{
328 spinlock_t rb_mwlock
; /* protect rb_mws list */
329 struct list_head rb_mws
;
330 struct list_head rb_all
;
333 spinlock_t rb_lock
; /* protect buf lists */
334 struct list_head rb_send_bufs
;
335 struct list_head rb_recv_bufs
;
337 atomic_t rb_credits
; /* most recent credit grant */
339 u32 rb_bc_srv_max_requests
;
340 spinlock_t rb_reqslock
; /* protect rb_allreqs */
341 struct list_head rb_allreqs
;
343 u32 rb_bc_max_requests
;
345 #define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia)
348 * Internal structure for transport instance creation. This
349 * exists primarily for modularity.
351 * This data should be set with mount options
353 struct rpcrdma_create_data_internal
{
354 struct sockaddr_storage addr
; /* RDMA server address */
355 unsigned int max_requests
; /* max requests (slots) in flight */
356 unsigned int rsize
; /* mount rsize - max read hdr+data */
357 unsigned int wsize
; /* mount wsize - max write hdr+data */
358 unsigned int inline_rsize
; /* max non-rdma read data payload */
359 unsigned int inline_wsize
; /* max non-rdma write data payload */
360 unsigned int padding
; /* non-rdma write header padding */
363 #define RPCRDMA_INLINE_READ_THRESHOLD(rq) \
364 (rpcx_to_rdmad(rq->rq_xprt).inline_rsize)
366 #define RPCRDMA_INLINE_WRITE_THRESHOLD(rq)\
367 (rpcx_to_rdmad(rq->rq_xprt).inline_wsize)
369 #define RPCRDMA_INLINE_PAD_VALUE(rq)\
370 rpcx_to_rdmad(rq->rq_xprt).padding
373 * Statistics for RPCRDMA
375 struct rpcrdma_stats
{
376 unsigned long read_chunk_count
;
377 unsigned long write_chunk_count
;
378 unsigned long reply_chunk_count
;
380 unsigned long long total_rdma_request
;
381 unsigned long long total_rdma_reply
;
383 unsigned long long pullup_copy_count
;
384 unsigned long long fixup_copy_count
;
385 unsigned long hardway_register_count
;
386 unsigned long failed_marshal_count
;
387 unsigned long bad_reply_count
;
388 unsigned long nomsg_call_count
;
389 unsigned long bcall_count
;
393 * Per-registration mode operations
396 struct rpcrdma_memreg_ops
{
397 int (*ro_map
)(struct rpcrdma_xprt
*,
398 struct rpcrdma_mr_seg
*, int, bool);
399 void (*ro_unmap_sync
)(struct rpcrdma_xprt
*,
400 struct rpcrdma_req
*);
401 void (*ro_unmap_safe
)(struct rpcrdma_xprt
*,
402 struct rpcrdma_req
*, bool);
403 int (*ro_open
)(struct rpcrdma_ia
*,
405 struct rpcrdma_create_data_internal
*);
406 size_t (*ro_maxpages
)(struct rpcrdma_xprt
*);
407 int (*ro_init
)(struct rpcrdma_xprt
*);
408 void (*ro_destroy
)(struct rpcrdma_buffer
*);
409 const char *ro_displayname
;
412 extern const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops
;
413 extern const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops
;
414 extern const struct rpcrdma_memreg_ops rpcrdma_physical_memreg_ops
;
417 * RPCRDMA transport -- encapsulates the structures above for
418 * integration with RPC.
420 * The contained structures are embedded, not pointers,
421 * for convenience. This structure need not be visible externally.
423 * It is allocated and initialized during mount, and released
426 struct rpcrdma_xprt
{
427 struct rpc_xprt rx_xprt
;
428 struct rpcrdma_ia rx_ia
;
429 struct rpcrdma_ep rx_ep
;
430 struct rpcrdma_buffer rx_buf
;
431 struct rpcrdma_create_data_internal rx_data
;
432 struct delayed_work rx_connect_worker
;
433 struct rpcrdma_stats rx_stats
;
436 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
437 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data)
439 /* Setting this to 0 ensures interoperability with early servers.
440 * Setting this to 1 enhances certain unaligned read/write performance.
441 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
442 extern int xprt_rdma_pad_optimize
;
445 * Interface Adapter calls - xprtrdma/verbs.c
447 int rpcrdma_ia_open(struct rpcrdma_xprt
*, struct sockaddr
*, int);
448 void rpcrdma_ia_close(struct rpcrdma_ia
*);
451 * Endpoint calls - xprtrdma/verbs.c
453 int rpcrdma_ep_create(struct rpcrdma_ep
*, struct rpcrdma_ia
*,
454 struct rpcrdma_create_data_internal
*);
455 void rpcrdma_ep_destroy(struct rpcrdma_ep
*, struct rpcrdma_ia
*);
456 int rpcrdma_ep_connect(struct rpcrdma_ep
*, struct rpcrdma_ia
*);
457 void rpcrdma_ep_disconnect(struct rpcrdma_ep
*, struct rpcrdma_ia
*);
459 int rpcrdma_ep_post(struct rpcrdma_ia
*, struct rpcrdma_ep
*,
460 struct rpcrdma_req
*);
461 int rpcrdma_ep_post_recv(struct rpcrdma_ia
*, struct rpcrdma_ep
*,
462 struct rpcrdma_rep
*);
465 * Buffer calls - xprtrdma/verbs.c
467 struct rpcrdma_req
*rpcrdma_create_req(struct rpcrdma_xprt
*);
468 struct rpcrdma_rep
*rpcrdma_create_rep(struct rpcrdma_xprt
*);
469 void rpcrdma_destroy_req(struct rpcrdma_ia
*, struct rpcrdma_req
*);
470 int rpcrdma_buffer_create(struct rpcrdma_xprt
*);
471 void rpcrdma_buffer_destroy(struct rpcrdma_buffer
*);
473 struct rpcrdma_mw
*rpcrdma_get_mw(struct rpcrdma_xprt
*);
474 void rpcrdma_put_mw(struct rpcrdma_xprt
*, struct rpcrdma_mw
*);
475 struct rpcrdma_req
*rpcrdma_buffer_get(struct rpcrdma_buffer
*);
476 void rpcrdma_buffer_put(struct rpcrdma_req
*);
477 void rpcrdma_recv_buffer_get(struct rpcrdma_req
*);
478 void rpcrdma_recv_buffer_put(struct rpcrdma_rep
*);
480 struct rpcrdma_regbuf
*rpcrdma_alloc_regbuf(struct rpcrdma_ia
*,
482 void rpcrdma_free_regbuf(struct rpcrdma_ia
*,
483 struct rpcrdma_regbuf
*);
485 int rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt
*, unsigned int);
487 int frwr_alloc_recovery_wq(void);
488 void frwr_destroy_recovery_wq(void);
490 int rpcrdma_alloc_wq(void);
491 void rpcrdma_destroy_wq(void);
494 * Wrappers for chunk registration, shared by read/write chunk code.
497 void rpcrdma_mapping_error(struct rpcrdma_mr_seg
*);
499 static inline enum dma_data_direction
500 rpcrdma_data_dir(bool writing
)
502 return writing
? DMA_FROM_DEVICE
: DMA_TO_DEVICE
;
506 rpcrdma_map_one(struct ib_device
*device
, struct rpcrdma_mr_seg
*seg
,
507 enum dma_data_direction direction
)
509 seg
->mr_dir
= direction
;
510 seg
->mr_dmalen
= seg
->mr_len
;
513 seg
->mr_dma
= ib_dma_map_page(device
,
514 seg
->mr_page
, offset_in_page(seg
->mr_offset
),
515 seg
->mr_dmalen
, seg
->mr_dir
);
517 seg
->mr_dma
= ib_dma_map_single(device
,
519 seg
->mr_dmalen
, seg
->mr_dir
);
521 if (ib_dma_mapping_error(device
, seg
->mr_dma
))
522 rpcrdma_mapping_error(seg
);
526 rpcrdma_unmap_one(struct ib_device
*device
, struct rpcrdma_mr_seg
*seg
)
529 ib_dma_unmap_page(device
,
530 seg
->mr_dma
, seg
->mr_dmalen
, seg
->mr_dir
);
532 ib_dma_unmap_single(device
,
533 seg
->mr_dma
, seg
->mr_dmalen
, seg
->mr_dir
);
537 * RPC/RDMA connection management calls - xprtrdma/rpc_rdma.c
539 void rpcrdma_connect_worker(struct work_struct
*);
540 void rpcrdma_conn_func(struct rpcrdma_ep
*);
541 void rpcrdma_reply_handler(struct rpcrdma_rep
*);
544 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
546 int rpcrdma_marshal_req(struct rpc_rqst
*);
547 void rpcrdma_set_max_header_sizes(struct rpcrdma_ia
*,
548 struct rpcrdma_create_data_internal
*,
551 /* RPC/RDMA module init - xprtrdma/transport.c
553 extern unsigned int xprt_rdma_max_inline_read
;
554 void xprt_rdma_format_addresses(struct rpc_xprt
*xprt
, struct sockaddr
*sap
);
555 void xprt_rdma_free_addresses(struct rpc_xprt
*xprt
);
556 void xprt_rdma_print_stats(struct rpc_xprt
*xprt
, struct seq_file
*seq
);
557 int xprt_rdma_init(void);
558 void xprt_rdma_cleanup(void);
560 /* Backchannel calls - xprtrdma/backchannel.c
562 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
563 int xprt_rdma_bc_setup(struct rpc_xprt
*, unsigned int);
564 int xprt_rdma_bc_up(struct svc_serv
*, struct net
*);
565 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt
*);
566 int rpcrdma_bc_post_recv(struct rpcrdma_xprt
*, unsigned int);
567 void rpcrdma_bc_receive_call(struct rpcrdma_xprt
*, struct rpcrdma_rep
*);
568 int rpcrdma_bc_marshal_reply(struct rpc_rqst
*);
569 void xprt_rdma_bc_free_rqst(struct rpc_rqst
*);
570 void xprt_rdma_bc_destroy(struct rpc_xprt
*, unsigned int);
571 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
573 extern struct xprt_class xprt_rdma_bc
;
575 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */