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 /* Force completion handler to ignore the signal
100 #define RPCRDMA_IGNORE_COMPLETION (0ULL)
102 /* Pre-allocate extra Work Requests for handling backward receives
103 * and sends. This is a fixed value because the Work Queues are
104 * allocated when the forward channel is set up.
106 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
107 #define RPCRDMA_BACKWARD_WRS (8)
109 #define RPCRDMA_BACKWARD_WRS (0)
112 /* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
114 * The below structure appears at the front of a large region of kmalloc'd
115 * memory, which always starts on a good alignment boundary.
118 struct rpcrdma_regbuf {
120 struct rpcrdma_req *rg_owner;
121 struct ib_sge rg_iov;
122 __be32 rg_base[0] __attribute__ ((aligned(256)));
126 rdmab_addr(struct rpcrdma_regbuf *rb)
128 return rb->rg_iov.addr;
132 rdmab_length(struct rpcrdma_regbuf *rb)
134 return rb->rg_iov.length;
138 rdmab_lkey(struct rpcrdma_regbuf *rb)
140 return rb->rg_iov.lkey;
143 static inline struct rpcrdma_msg *
144 rdmab_to_msg(struct rpcrdma_regbuf *rb)
146 return (struct rpcrdma_msg *)rb->rg_base;
149 #define RPCRDMA_DEF_GFP (GFP_NOIO | __GFP_NOWARN)
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 #define RPCRDMA_MAX_DATA_SEGS ((1 * 1024 * 1024) / PAGE_SIZE)
169 #define RPCRDMA_MAX_SEGS (RPCRDMA_MAX_DATA_SEGS + 2) /* head+tail = 2 */
171 struct rpcrdma_buffer;
175 struct ib_device *rr_device;
176 struct rpcrdma_xprt *rr_rxprt;
177 struct work_struct rr_work;
178 struct list_head rr_list;
179 struct rpcrdma_regbuf *rr_rdmabuf;
182 #define RPCRDMA_BAD_LEN (~0U)
185 * struct rpcrdma_mw - external memory region metadata
187 * An external memory region is any buffer or page that is registered
188 * on the fly (ie, not pre-registered).
190 * Each rpcrdma_buffer has a list of free MWs anchored in rb_mws. During
191 * call_allocate, rpcrdma_buffer_get() assigns one to each segment in
192 * an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep
193 * track of registration metadata while each RPC is pending.
194 * rpcrdma_deregister_external() uses this metadata to unmap and
195 * release these resources when an RPC is complete.
197 enum rpcrdma_frmr_state {
198 FRMR_IS_INVALID, /* ready to be used */
199 FRMR_IS_VALID, /* in use */
200 FRMR_IS_STALE, /* failed completion */
203 struct rpcrdma_frmr {
204 struct scatterlist *sg;
207 enum rpcrdma_frmr_state fr_state;
208 struct work_struct fr_work;
209 struct rpcrdma_xprt *fr_xprt;
211 struct completion fr_linv_done;;
213 struct ib_reg_wr fr_regwr;
214 struct ib_send_wr fr_invwr;
225 struct rpcrdma_fmr fmr;
226 struct rpcrdma_frmr frmr;
228 void (*mw_sendcompletion)(struct ib_wc *);
229 struct list_head mw_list;
230 struct list_head mw_all;
234 * struct rpcrdma_req -- structure central to the request/reply sequence.
236 * N of these are associated with a transport instance, and stored in
237 * struct rpcrdma_buffer. N is the max number of outstanding requests.
239 * It includes pre-registered buffer memory for send AND recv.
240 * The recv buffer, however, is not owned by this structure, and
241 * is "donated" to the hardware when a recv is posted. When a
242 * reply is handled, the recv buffer used is given back to the
243 * struct rpcrdma_req associated with the request.
245 * In addition to the basic memory, this structure includes an array
246 * of iovs for send operations. The reason is that the iovs passed to
247 * ib_post_{send,recv} must not be modified until the work request
251 * o RPCRDMA_MAX_SEGS is the max number of addressible chunk elements we
252 * marshal. The number needed varies depending on the iov lists that
253 * are passed to us, the memory registration mode we are in, and if
254 * physical addressing is used, the layout.
257 struct rpcrdma_mr_seg { /* chunk descriptors */
258 struct rpcrdma_mw *rl_mw; /* registered MR */
259 u64 mr_base; /* registration result */
260 u32 mr_rkey; /* registration result */
261 u32 mr_len; /* length of chunk or segment */
262 int mr_nsegs; /* number of segments in chunk or 0 */
263 enum dma_data_direction mr_dir; /* segment mapping direction */
264 dma_addr_t mr_dma; /* segment mapping address */
265 size_t mr_dmalen; /* segment mapping length */
266 struct page *mr_page; /* owning page, if any */
267 char *mr_offset; /* kva if no page, else offset */
270 #define RPCRDMA_MAX_IOVS (2)
273 struct list_head rl_free;
274 unsigned int rl_niovs;
275 unsigned int rl_nchunks;
276 unsigned int rl_connect_cookie;
277 struct rpcrdma_buffer *rl_buffer;
278 struct rpcrdma_rep *rl_reply;/* holder for reply buffer */
279 struct ib_sge rl_send_iov[RPCRDMA_MAX_IOVS];
280 struct rpcrdma_regbuf *rl_rdmabuf;
281 struct rpcrdma_regbuf *rl_sendbuf;
282 struct rpcrdma_mr_seg rl_segments[RPCRDMA_MAX_SEGS];
284 struct list_head rl_all;
288 static inline struct rpcrdma_req *
289 rpcr_to_rdmar(struct rpc_rqst *rqst)
291 void *buffer = rqst->rq_buffer;
292 struct rpcrdma_regbuf *rb;
294 rb = container_of(buffer, struct rpcrdma_regbuf, rg_base);
299 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
300 * inline requests/replies, and client/server credits.
302 * One of these is associated with a transport instance
304 struct rpcrdma_buffer {
305 spinlock_t rb_mwlock; /* protect rb_mws list */
306 struct list_head rb_mws;
307 struct list_head rb_all;
310 spinlock_t rb_lock; /* protect buf lists */
311 struct list_head rb_send_bufs;
312 struct list_head rb_recv_bufs;
314 atomic_t rb_credits; /* most recent credit grant */
316 u32 rb_bc_srv_max_requests;
317 spinlock_t rb_reqslock; /* protect rb_allreqs */
318 struct list_head rb_allreqs;
320 u32 rb_bc_max_requests;
322 #define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia)
325 * Internal structure for transport instance creation. This
326 * exists primarily for modularity.
328 * This data should be set with mount options
330 struct rpcrdma_create_data_internal {
331 struct sockaddr_storage addr; /* RDMA server address */
332 unsigned int max_requests; /* max requests (slots) in flight */
333 unsigned int rsize; /* mount rsize - max read hdr+data */
334 unsigned int wsize; /* mount wsize - max write hdr+data */
335 unsigned int inline_rsize; /* max non-rdma read data payload */
336 unsigned int inline_wsize; /* max non-rdma write data payload */
337 unsigned int padding; /* non-rdma write header padding */
340 #define RPCRDMA_INLINE_READ_THRESHOLD(rq) \
341 (rpcx_to_rdmad(rq->rq_xprt).inline_rsize)
343 #define RPCRDMA_INLINE_WRITE_THRESHOLD(rq)\
344 (rpcx_to_rdmad(rq->rq_xprt).inline_wsize)
346 #define RPCRDMA_INLINE_PAD_VALUE(rq)\
347 rpcx_to_rdmad(rq->rq_xprt).padding
350 * Statistics for RPCRDMA
352 struct rpcrdma_stats {
353 unsigned long read_chunk_count;
354 unsigned long write_chunk_count;
355 unsigned long reply_chunk_count;
357 unsigned long long total_rdma_request;
358 unsigned long long total_rdma_reply;
360 unsigned long long pullup_copy_count;
361 unsigned long long fixup_copy_count;
362 unsigned long hardway_register_count;
363 unsigned long failed_marshal_count;
364 unsigned long bad_reply_count;
365 unsigned long nomsg_call_count;
366 unsigned long bcall_count;
370 * Per-registration mode operations
373 struct rpcrdma_memreg_ops {
374 int (*ro_map)(struct rpcrdma_xprt *,
375 struct rpcrdma_mr_seg *, int, bool);
376 void (*ro_unmap_sync)(struct rpcrdma_xprt *,
377 struct rpcrdma_req *);
378 int (*ro_unmap)(struct rpcrdma_xprt *,
379 struct rpcrdma_mr_seg *);
380 int (*ro_open)(struct rpcrdma_ia *,
382 struct rpcrdma_create_data_internal *);
383 size_t (*ro_maxpages)(struct rpcrdma_xprt *);
384 int (*ro_init)(struct rpcrdma_xprt *);
385 void (*ro_destroy)(struct rpcrdma_buffer *);
386 const char *ro_displayname;
389 extern const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops;
390 extern const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops;
391 extern const struct rpcrdma_memreg_ops rpcrdma_physical_memreg_ops;
394 * RPCRDMA transport -- encapsulates the structures above for
395 * integration with RPC.
397 * The contained structures are embedded, not pointers,
398 * for convenience. This structure need not be visible externally.
400 * It is allocated and initialized during mount, and released
403 struct rpcrdma_xprt {
404 struct rpc_xprt rx_xprt;
405 struct rpcrdma_ia rx_ia;
406 struct rpcrdma_ep rx_ep;
407 struct rpcrdma_buffer rx_buf;
408 struct rpcrdma_create_data_internal rx_data;
409 struct delayed_work rx_connect_worker;
410 struct rpcrdma_stats rx_stats;
413 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
414 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data)
416 /* Setting this to 0 ensures interoperability with early servers.
417 * Setting this to 1 enhances certain unaligned read/write performance.
418 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
419 extern int xprt_rdma_pad_optimize;
422 * Interface Adapter calls - xprtrdma/verbs.c
424 int rpcrdma_ia_open(struct rpcrdma_xprt *, struct sockaddr *, int);
425 void rpcrdma_ia_close(struct rpcrdma_ia *);
428 * Endpoint calls - xprtrdma/verbs.c
430 int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *,
431 struct rpcrdma_create_data_internal *);
432 void rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *);
433 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *);
434 void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *);
436 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
437 struct rpcrdma_req *);
438 int rpcrdma_ep_post_recv(struct rpcrdma_ia *, struct rpcrdma_ep *,
439 struct rpcrdma_rep *);
442 * Buffer calls - xprtrdma/verbs.c
444 struct rpcrdma_req *rpcrdma_create_req(struct rpcrdma_xprt *);
445 struct rpcrdma_rep *rpcrdma_create_rep(struct rpcrdma_xprt *);
446 void rpcrdma_destroy_req(struct rpcrdma_ia *, struct rpcrdma_req *);
447 int rpcrdma_buffer_create(struct rpcrdma_xprt *);
448 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
450 struct rpcrdma_mw *rpcrdma_get_mw(struct rpcrdma_xprt *);
451 void rpcrdma_put_mw(struct rpcrdma_xprt *, struct rpcrdma_mw *);
452 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
453 void rpcrdma_buffer_put(struct rpcrdma_req *);
454 void rpcrdma_recv_buffer_get(struct rpcrdma_req *);
455 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
457 struct rpcrdma_regbuf *rpcrdma_alloc_regbuf(struct rpcrdma_ia *,
459 void rpcrdma_free_regbuf(struct rpcrdma_ia *,
460 struct rpcrdma_regbuf *);
462 unsigned int rpcrdma_max_segments(struct rpcrdma_xprt *);
463 int rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt *, unsigned int);
465 int frwr_alloc_recovery_wq(void);
466 void frwr_destroy_recovery_wq(void);
468 int rpcrdma_alloc_wq(void);
469 void rpcrdma_destroy_wq(void);
472 * Wrappers for chunk registration, shared by read/write chunk code.
475 void rpcrdma_mapping_error(struct rpcrdma_mr_seg *);
477 static inline enum dma_data_direction
478 rpcrdma_data_dir(bool writing)
480 return writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
484 rpcrdma_map_one(struct ib_device *device, struct rpcrdma_mr_seg *seg,
485 enum dma_data_direction direction)
487 seg->mr_dir = direction;
488 seg->mr_dmalen = seg->mr_len;
491 seg->mr_dma = ib_dma_map_page(device,
492 seg->mr_page, offset_in_page(seg->mr_offset),
493 seg->mr_dmalen, seg->mr_dir);
495 seg->mr_dma = ib_dma_map_single(device,
497 seg->mr_dmalen, seg->mr_dir);
499 if (ib_dma_mapping_error(device, seg->mr_dma))
500 rpcrdma_mapping_error(seg);
504 rpcrdma_unmap_one(struct ib_device *device, struct rpcrdma_mr_seg *seg)
507 ib_dma_unmap_page(device,
508 seg->mr_dma, seg->mr_dmalen, seg->mr_dir);
510 ib_dma_unmap_single(device,
511 seg->mr_dma, seg->mr_dmalen, seg->mr_dir);
515 * RPC/RDMA connection management calls - xprtrdma/rpc_rdma.c
517 void rpcrdma_connect_worker(struct work_struct *);
518 void rpcrdma_conn_func(struct rpcrdma_ep *);
519 void rpcrdma_reply_handler(struct rpcrdma_rep *);
522 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
524 int rpcrdma_marshal_req(struct rpc_rqst *);
526 /* RPC/RDMA module init - xprtrdma/transport.c
528 extern unsigned int xprt_rdma_max_inline_read;
529 void xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap);
530 void xprt_rdma_free_addresses(struct rpc_xprt *xprt);
531 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq);
532 int xprt_rdma_init(void);
533 void xprt_rdma_cleanup(void);
535 /* Backchannel calls - xprtrdma/backchannel.c
537 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
538 int xprt_rdma_bc_setup(struct rpc_xprt *, unsigned int);
539 int xprt_rdma_bc_up(struct svc_serv *, struct net *);
540 int rpcrdma_bc_post_recv(struct rpcrdma_xprt *, unsigned int);
541 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *, struct rpcrdma_rep *);
542 int rpcrdma_bc_marshal_reply(struct rpc_rqst *);
543 void xprt_rdma_bc_free_rqst(struct rpc_rqst *);
544 void xprt_rdma_bc_destroy(struct rpc_xprt *, unsigned int);
545 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
547 extern struct xprt_class xprt_rdma_bc;
549 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */