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 */
54 #include <linux/sunrpc/svc.h> /* RPCSVC_MAXPAYLOAD */
56 #define RDMA_RESOLVE_TIMEOUT (5000) /* 5 seconds */
57 #define RDMA_CONNECT_RETRY_MAX (2) /* retries if no listener backlog */
60 * Interface Adapter -- one per transport instance
63 const struct rpcrdma_memreg_ops
*ri_ops
;
65 struct ib_device
*ri_device
;
66 struct rdma_cm_id
*ri_id
;
68 struct ib_mr
*ri_bind_mem
;
71 struct completion ri_done
;
73 unsigned int ri_max_frmr_depth
;
74 struct ib_device_attr ri_devattr
;
75 struct ib_qp_attr ri_qp_attr
;
76 struct ib_qp_init_attr ri_qp_init_attr
;
80 * RDMA Endpoint -- one per transport instance
83 #define RPCRDMA_WC_BUDGET (128)
84 #define RPCRDMA_POLLSIZE (16)
90 struct ib_qp_init_attr rep_attr
;
91 wait_queue_head_t rep_connect_wait
;
92 struct rpcrdma_regbuf
*rep_padbuf
;
93 struct rdma_conn_param rep_remote_cma
;
94 struct sockaddr_storage rep_remote_addr
;
95 struct delayed_work rep_connect_worker
;
96 struct ib_wc rep_send_wcs
[RPCRDMA_POLLSIZE
];
97 struct ib_wc rep_recv_wcs
[RPCRDMA_POLLSIZE
];
101 * Force a signaled SEND Work Request every so often,
102 * in case the provider needs to do some housekeeping.
104 #define RPCRDMA_MAX_UNSIGNALED_SENDS (32)
106 #define INIT_CQCOUNT(ep) atomic_set(&(ep)->rep_cqcount, (ep)->rep_cqinit)
107 #define DECR_CQCOUNT(ep) atomic_sub_return(1, &(ep)->rep_cqcount)
109 /* Force completion handler to ignore the signal
111 #define RPCRDMA_IGNORE_COMPLETION (0ULL)
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
;
123 struct ib_sge rg_iov
;
124 __be32 rg_base
[0] __attribute__ ((aligned(256)));
128 rdmab_addr(struct rpcrdma_regbuf
*rb
)
130 return rb
->rg_iov
.addr
;
134 rdmab_length(struct rpcrdma_regbuf
*rb
)
136 return rb
->rg_iov
.length
;
140 rdmab_lkey(struct rpcrdma_regbuf
*rb
)
142 return rb
->rg_iov
.lkey
;
145 static inline struct rpcrdma_msg
*
146 rdmab_to_msg(struct rpcrdma_regbuf
*rb
)
148 return (struct rpcrdma_msg
*)rb
->rg_base
;
152 * struct rpcrdma_rep -- this structure encapsulates state required to recv
153 * and complete a reply, asychronously. It needs several pieces of
155 * o recv buffer (posted to provider)
156 * o ib_sge (also donated to provider)
157 * o status of reply (length, success or not)
158 * o bookkeeping state to get run by tasklet (list, etc)
160 * These are allocated during initialization, per-transport instance;
161 * however, the tasklet execution list itself is global, as it should
162 * always be pretty short.
164 * N of these are associated with a transport instance, and stored in
165 * struct rpcrdma_buffer. N is the max number of outstanding requests.
168 /* temporary static scatter/gather max */
169 #define RPCRDMA_MAX_DATA_SEGS (64) /* max scatter/gather */
170 #define RPCRDMA_MAX_SEGS (RPCRDMA_MAX_DATA_SEGS + 2) /* head+tail = 2 */
172 struct rpcrdma_buffer
;
176 struct ib_device
*rr_device
;
177 struct rpcrdma_xprt
*rr_rxprt
;
178 struct list_head rr_list
;
179 struct rpcrdma_regbuf
*rr_rdmabuf
;
183 * struct rpcrdma_mw - external memory region metadata
185 * An external memory region is any buffer or page that is registered
186 * on the fly (ie, not pre-registered).
188 * Each rpcrdma_buffer has a list of free MWs anchored in rb_mws. During
189 * call_allocate, rpcrdma_buffer_get() assigns one to each segment in
190 * an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep
191 * track of registration metadata while each RPC is pending.
192 * rpcrdma_deregister_external() uses this metadata to unmap and
193 * release these resources when an RPC is complete.
195 enum rpcrdma_frmr_state
{
196 FRMR_IS_INVALID
, /* ready to be used */
197 FRMR_IS_VALID
, /* in use */
198 FRMR_IS_STALE
, /* failed completion */
201 struct rpcrdma_frmr
{
202 struct ib_fast_reg_page_list
*fr_pgl
;
204 enum rpcrdma_frmr_state fr_state
;
205 struct work_struct fr_work
;
206 struct rpcrdma_xprt
*fr_xprt
;
212 struct rpcrdma_frmr frmr
;
214 void (*mw_sendcompletion
)(struct ib_wc
*);
215 struct list_head mw_list
;
216 struct list_head mw_all
;
220 * struct rpcrdma_req -- structure central to the request/reply sequence.
222 * N of these are associated with a transport instance, and stored in
223 * struct rpcrdma_buffer. N is the max number of outstanding requests.
225 * It includes pre-registered buffer memory for send AND recv.
226 * The recv buffer, however, is not owned by this structure, and
227 * is "donated" to the hardware when a recv is posted. When a
228 * reply is handled, the recv buffer used is given back to the
229 * struct rpcrdma_req associated with the request.
231 * In addition to the basic memory, this structure includes an array
232 * of iovs for send operations. The reason is that the iovs passed to
233 * ib_post_{send,recv} must not be modified until the work request
237 * o RPCRDMA_MAX_SEGS is the max number of addressible chunk elements we
238 * marshal. The number needed varies depending on the iov lists that
239 * are passed to us, the memory registration mode we are in, and if
240 * physical addressing is used, the layout.
243 struct rpcrdma_mr_seg
{ /* chunk descriptors */
244 struct rpcrdma_mw
*rl_mw
; /* registered MR */
245 u64 mr_base
; /* registration result */
246 u32 mr_rkey
; /* registration result */
247 u32 mr_len
; /* length of chunk or segment */
248 int mr_nsegs
; /* number of segments in chunk or 0 */
249 enum dma_data_direction mr_dir
; /* segment mapping direction */
250 dma_addr_t mr_dma
; /* segment mapping address */
251 size_t mr_dmalen
; /* segment mapping length */
252 struct page
*mr_page
; /* owning page, if any */
253 char *mr_offset
; /* kva if no page, else offset */
257 unsigned int rl_niovs
; /* 0, 2 or 4 */
258 unsigned int rl_nchunks
; /* non-zero if chunks */
259 unsigned int rl_connect_cookie
; /* retry detection */
260 struct rpcrdma_buffer
*rl_buffer
; /* home base for this structure */
261 struct rpcrdma_rep
*rl_reply
;/* holder for reply buffer */
262 struct ib_sge rl_send_iov
[4]; /* for active requests */
263 struct rpcrdma_regbuf
*rl_rdmabuf
;
264 struct rpcrdma_regbuf
*rl_sendbuf
;
265 struct rpcrdma_mr_seg rl_segments
[RPCRDMA_MAX_SEGS
];
268 static inline struct rpcrdma_req
*
269 rpcr_to_rdmar(struct rpc_rqst
*rqst
)
271 void *buffer
= rqst
->rq_buffer
;
272 struct rpcrdma_regbuf
*rb
;
274 rb
= container_of(buffer
, struct rpcrdma_regbuf
, rg_base
);
279 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
280 * inline requests/replies, and client/server credits.
282 * One of these is associated with a transport instance
284 struct rpcrdma_buffer
{
285 spinlock_t rb_mwlock
; /* protect rb_mws list */
286 struct list_head rb_mws
;
287 struct list_head rb_all
;
290 spinlock_t rb_lock
; /* protect buf arrays */
294 struct rpcrdma_req
**rb_send_bufs
;
295 struct rpcrdma_rep
**rb_recv_bufs
;
297 #define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia)
300 * Internal structure for transport instance creation. This
301 * exists primarily for modularity.
303 * This data should be set with mount options
305 struct rpcrdma_create_data_internal
{
306 struct sockaddr_storage addr
; /* RDMA server address */
307 unsigned int max_requests
; /* max requests (slots) in flight */
308 unsigned int rsize
; /* mount rsize - max read hdr+data */
309 unsigned int wsize
; /* mount wsize - max write hdr+data */
310 unsigned int inline_rsize
; /* max non-rdma read data payload */
311 unsigned int inline_wsize
; /* max non-rdma write data payload */
312 unsigned int padding
; /* non-rdma write header padding */
315 #define RPCRDMA_INLINE_READ_THRESHOLD(rq) \
316 (rpcx_to_rdmad(rq->rq_xprt).inline_rsize)
318 #define RPCRDMA_INLINE_WRITE_THRESHOLD(rq)\
319 (rpcx_to_rdmad(rq->rq_xprt).inline_wsize)
321 #define RPCRDMA_INLINE_PAD_VALUE(rq)\
322 rpcx_to_rdmad(rq->rq_xprt).padding
325 * Statistics for RPCRDMA
327 struct rpcrdma_stats
{
328 unsigned long read_chunk_count
;
329 unsigned long write_chunk_count
;
330 unsigned long reply_chunk_count
;
332 unsigned long long total_rdma_request
;
333 unsigned long long total_rdma_reply
;
335 unsigned long long pullup_copy_count
;
336 unsigned long long fixup_copy_count
;
337 unsigned long hardway_register_count
;
338 unsigned long failed_marshal_count
;
339 unsigned long bad_reply_count
;
343 * Per-registration mode operations
346 struct rpcrdma_memreg_ops
{
347 int (*ro_map
)(struct rpcrdma_xprt
*,
348 struct rpcrdma_mr_seg
*, int, bool);
349 int (*ro_unmap
)(struct rpcrdma_xprt
*,
350 struct rpcrdma_mr_seg
*);
351 int (*ro_open
)(struct rpcrdma_ia
*,
353 struct rpcrdma_create_data_internal
*);
354 size_t (*ro_maxpages
)(struct rpcrdma_xprt
*);
355 int (*ro_init
)(struct rpcrdma_xprt
*);
356 void (*ro_destroy
)(struct rpcrdma_buffer
*);
357 const char *ro_displayname
;
360 extern const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops
;
361 extern const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops
;
362 extern const struct rpcrdma_memreg_ops rpcrdma_physical_memreg_ops
;
365 * RPCRDMA transport -- encapsulates the structures above for
366 * integration with RPC.
368 * The contained structures are embedded, not pointers,
369 * for convenience. This structure need not be visible externally.
371 * It is allocated and initialized during mount, and released
374 struct rpcrdma_xprt
{
375 struct rpc_xprt rx_xprt
;
376 struct rpcrdma_ia rx_ia
;
377 struct rpcrdma_ep rx_ep
;
378 struct rpcrdma_buffer rx_buf
;
379 struct rpcrdma_create_data_internal rx_data
;
380 struct delayed_work rx_connect_worker
;
381 struct rpcrdma_stats rx_stats
;
384 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
385 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data)
387 /* Setting this to 0 ensures interoperability with early servers.
388 * Setting this to 1 enhances certain unaligned read/write performance.
389 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
390 extern int xprt_rdma_pad_optimize
;
393 * Interface Adapter calls - xprtrdma/verbs.c
395 int rpcrdma_ia_open(struct rpcrdma_xprt
*, struct sockaddr
*, int);
396 void rpcrdma_ia_close(struct rpcrdma_ia
*);
399 * Endpoint calls - xprtrdma/verbs.c
401 int rpcrdma_ep_create(struct rpcrdma_ep
*, struct rpcrdma_ia
*,
402 struct rpcrdma_create_data_internal
*);
403 void rpcrdma_ep_destroy(struct rpcrdma_ep
*, struct rpcrdma_ia
*);
404 int rpcrdma_ep_connect(struct rpcrdma_ep
*, struct rpcrdma_ia
*);
405 void rpcrdma_ep_disconnect(struct rpcrdma_ep
*, struct rpcrdma_ia
*);
407 int rpcrdma_ep_post(struct rpcrdma_ia
*, struct rpcrdma_ep
*,
408 struct rpcrdma_req
*);
409 int rpcrdma_ep_post_recv(struct rpcrdma_ia
*, struct rpcrdma_ep
*,
410 struct rpcrdma_rep
*);
413 * Buffer calls - xprtrdma/verbs.c
415 int rpcrdma_buffer_create(struct rpcrdma_xprt
*);
416 void rpcrdma_buffer_destroy(struct rpcrdma_buffer
*);
418 struct rpcrdma_mw
*rpcrdma_get_mw(struct rpcrdma_xprt
*);
419 void rpcrdma_put_mw(struct rpcrdma_xprt
*, struct rpcrdma_mw
*);
420 struct rpcrdma_req
*rpcrdma_buffer_get(struct rpcrdma_buffer
*);
421 void rpcrdma_buffer_put(struct rpcrdma_req
*);
422 void rpcrdma_recv_buffer_get(struct rpcrdma_req
*);
423 void rpcrdma_recv_buffer_put(struct rpcrdma_rep
*);
425 struct rpcrdma_regbuf
*rpcrdma_alloc_regbuf(struct rpcrdma_ia
*,
427 void rpcrdma_free_regbuf(struct rpcrdma_ia
*,
428 struct rpcrdma_regbuf
*);
430 unsigned int rpcrdma_max_segments(struct rpcrdma_xprt
*);
432 int frwr_alloc_recovery_wq(void);
433 void frwr_destroy_recovery_wq(void);
436 * Wrappers for chunk registration, shared by read/write chunk code.
439 void rpcrdma_mapping_error(struct rpcrdma_mr_seg
*);
441 static inline enum dma_data_direction
442 rpcrdma_data_dir(bool writing
)
444 return writing
? DMA_FROM_DEVICE
: DMA_TO_DEVICE
;
448 rpcrdma_map_one(struct ib_device
*device
, struct rpcrdma_mr_seg
*seg
,
449 enum dma_data_direction direction
)
451 seg
->mr_dir
= direction
;
452 seg
->mr_dmalen
= seg
->mr_len
;
455 seg
->mr_dma
= ib_dma_map_page(device
,
456 seg
->mr_page
, offset_in_page(seg
->mr_offset
),
457 seg
->mr_dmalen
, seg
->mr_dir
);
459 seg
->mr_dma
= ib_dma_map_single(device
,
461 seg
->mr_dmalen
, seg
->mr_dir
);
463 if (ib_dma_mapping_error(device
, seg
->mr_dma
))
464 rpcrdma_mapping_error(seg
);
468 rpcrdma_unmap_one(struct ib_device
*device
, struct rpcrdma_mr_seg
*seg
)
471 ib_dma_unmap_page(device
,
472 seg
->mr_dma
, seg
->mr_dmalen
, seg
->mr_dir
);
474 ib_dma_unmap_single(device
,
475 seg
->mr_dma
, seg
->mr_dmalen
, seg
->mr_dir
);
479 * RPC/RDMA connection management calls - xprtrdma/rpc_rdma.c
481 void rpcrdma_connect_worker(struct work_struct
*);
482 void rpcrdma_conn_func(struct rpcrdma_ep
*);
483 void rpcrdma_reply_handler(struct rpcrdma_rep
*);
486 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
488 int rpcrdma_marshal_req(struct rpc_rqst
*);
490 /* Temporary NFS request map cache. Created in svc_rdma.c */
491 extern struct kmem_cache
*svc_rdma_map_cachep
;
492 /* WR context cache. Created in svc_rdma.c */
493 extern struct kmem_cache
*svc_rdma_ctxt_cachep
;
494 /* Workqueue created in svc_rdma.c */
495 extern struct workqueue_struct
*svc_rdma_wq
;
497 #if RPCSVC_MAXPAYLOAD < (RPCRDMA_MAX_DATA_SEGS << PAGE_SHIFT)
498 #define RPCSVC_MAXPAYLOAD_RDMA RPCSVC_MAXPAYLOAD
500 #define RPCSVC_MAXPAYLOAD_RDMA (RPCRDMA_MAX_DATA_SEGS << PAGE_SHIFT)
503 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */