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
64 struct rdma_cm_id *ri_id;
66 struct ib_mr *ri_bind_mem;
69 struct completion ri_done;
71 enum rpcrdma_memreg ri_memreg_strategy;
72 unsigned int ri_max_frmr_depth;
73 struct ib_device_attr ri_devattr;
74 struct ib_qp_attr ri_qp_attr;
75 struct ib_qp_init_attr ri_qp_init_attr;
79 * RDMA Endpoint -- one per transport instance
82 #define RPCRDMA_WC_BUDGET (128)
83 #define RPCRDMA_POLLSIZE (16)
89 struct ib_qp_init_attr rep_attr;
90 wait_queue_head_t rep_connect_wait;
91 struct rpcrdma_regbuf *rep_padbuf;
92 struct rdma_conn_param rep_remote_cma;
93 struct sockaddr_storage rep_remote_addr;
94 struct delayed_work rep_connect_worker;
95 struct ib_wc rep_send_wcs[RPCRDMA_POLLSIZE];
96 struct ib_wc rep_recv_wcs[RPCRDMA_POLLSIZE];
100 * Force a signaled SEND Work Request every so often,
101 * in case the provider needs to do some housekeeping.
103 #define RPCRDMA_MAX_UNSIGNALED_SENDS (32)
105 #define INIT_CQCOUNT(ep) atomic_set(&(ep)->rep_cqcount, (ep)->rep_cqinit)
106 #define DECR_CQCOUNT(ep) atomic_sub_return(1, &(ep)->rep_cqcount)
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;
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 enum rpcrdma_chunktype {
155 * struct rpcrdma_rep -- this structure encapsulates state required to recv
156 * and complete a reply, asychronously. It needs several pieces of
158 * o recv buffer (posted to provider)
159 * o ib_sge (also donated to provider)
160 * o status of reply (length, success or not)
161 * o bookkeeping state to get run by tasklet (list, etc)
163 * These are allocated during initialization, per-transport instance;
164 * however, the tasklet execution list itself is global, as it should
165 * always be pretty short.
167 * N of these are associated with a transport instance, and stored in
168 * struct rpcrdma_buffer. N is the max number of outstanding requests.
171 /* temporary static scatter/gather max */
172 #define RPCRDMA_MAX_DATA_SEGS (64) /* max scatter/gather */
173 #define RPCRDMA_MAX_SEGS (RPCRDMA_MAX_DATA_SEGS + 2) /* head+tail = 2 */
175 struct rpcrdma_buffer;
179 struct rpcrdma_buffer *rr_buffer;
180 struct rpc_xprt *rr_xprt;
181 void (*rr_func)(struct rpcrdma_rep *);
182 struct list_head rr_list;
183 struct rpcrdma_regbuf *rr_rdmabuf;
187 * struct rpcrdma_mw - external memory region metadata
189 * An external memory region is any buffer or page that is registered
190 * on the fly (ie, not pre-registered).
192 * Each rpcrdma_buffer has a list of free MWs anchored in rb_mws. During
193 * call_allocate, rpcrdma_buffer_get() assigns one to each segment in
194 * an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep
195 * track of registration metadata while each RPC is pending.
196 * rpcrdma_deregister_external() uses this metadata to unmap and
197 * release these resources when an RPC is complete.
199 enum rpcrdma_frmr_state {
200 FRMR_IS_INVALID, /* ready to be used */
201 FRMR_IS_VALID, /* in use */
202 FRMR_IS_STALE, /* failed completion */
205 struct rpcrdma_frmr {
206 struct ib_fast_reg_page_list *fr_pgl;
208 enum rpcrdma_frmr_state fr_state;
214 struct rpcrdma_frmr frmr;
216 struct list_head mw_list;
217 struct list_head mw_all;
221 * struct rpcrdma_req -- structure central to the request/reply sequence.
223 * N of these are associated with a transport instance, and stored in
224 * struct rpcrdma_buffer. N is the max number of outstanding requests.
226 * It includes pre-registered buffer memory for send AND recv.
227 * The recv buffer, however, is not owned by this structure, and
228 * is "donated" to the hardware when a recv is posted. When a
229 * reply is handled, the recv buffer used is given back to the
230 * struct rpcrdma_req associated with the request.
232 * In addition to the basic memory, this structure includes an array
233 * of iovs for send operations. The reason is that the iovs passed to
234 * ib_post_{send,recv} must not be modified until the work request
238 * o RPCRDMA_MAX_SEGS is the max number of addressible chunk elements we
239 * marshal. The number needed varies depending on the iov lists that
240 * are passed to us, the memory registration mode we are in, and if
241 * physical addressing is used, the layout.
244 struct rpcrdma_mr_seg { /* chunk descriptors */
245 struct rpcrdma_mw *rl_mw; /* registered MR */
246 u64 mr_base; /* registration result */
247 u32 mr_rkey; /* registration result */
248 u32 mr_len; /* length of chunk or segment */
249 int mr_nsegs; /* number of segments in chunk or 0 */
250 enum dma_data_direction mr_dir; /* segment mapping direction */
251 dma_addr_t mr_dma; /* segment mapping address */
252 size_t mr_dmalen; /* segment mapping length */
253 struct page *mr_page; /* owning page, if any */
254 char *mr_offset; /* kva if no page, else offset */
258 unsigned int rl_niovs; /* 0, 2 or 4 */
259 unsigned int rl_nchunks; /* non-zero if chunks */
260 unsigned int rl_connect_cookie; /* retry detection */
261 enum rpcrdma_chunktype rl_rtype, rl_wtype;
262 struct rpcrdma_buffer *rl_buffer; /* home base for this structure */
263 struct rpcrdma_rep *rl_reply;/* holder for reply buffer */
264 struct ib_sge rl_send_iov[4]; /* for active requests */
265 struct rpcrdma_regbuf *rl_rdmabuf;
266 struct rpcrdma_regbuf *rl_sendbuf;
267 struct rpcrdma_mr_seg rl_segments[RPCRDMA_MAX_SEGS];
270 static inline struct rpcrdma_req *
271 rpcr_to_rdmar(struct rpc_rqst *rqst)
273 void *buffer = rqst->rq_buffer;
274 struct rpcrdma_regbuf *rb;
276 rb = container_of(buffer, struct rpcrdma_regbuf, rg_base);
281 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
282 * inline requests/replies, and client/server credits.
284 * One of these is associated with a transport instance
286 struct rpcrdma_buffer {
287 spinlock_t rb_lock; /* protects indexes */
288 u32 rb_max_requests;/* client max requests */
289 struct list_head rb_mws; /* optional memory windows/fmrs/frmrs */
290 struct list_head rb_all;
292 struct rpcrdma_req **rb_send_bufs;
294 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 * RPCRDMA transport -- encapsulates the structures above for
344 * integration with RPC.
346 * The contained structures are embedded, not pointers,
347 * for convenience. This structure need not be visible externally.
349 * It is allocated and initialized during mount, and released
352 struct rpcrdma_xprt {
353 struct rpc_xprt rx_xprt;
354 struct rpcrdma_ia rx_ia;
355 struct rpcrdma_ep rx_ep;
356 struct rpcrdma_buffer rx_buf;
357 struct rpcrdma_create_data_internal rx_data;
358 struct delayed_work rx_connect_worker;
359 struct rpcrdma_stats rx_stats;
362 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
363 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data)
365 /* Setting this to 0 ensures interoperability with early servers.
366 * Setting this to 1 enhances certain unaligned read/write performance.
367 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
368 extern int xprt_rdma_pad_optimize;
371 * Interface Adapter calls - xprtrdma/verbs.c
373 int rpcrdma_ia_open(struct rpcrdma_xprt *, struct sockaddr *, int);
374 void rpcrdma_ia_close(struct rpcrdma_ia *);
377 * Endpoint calls - xprtrdma/verbs.c
379 int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *,
380 struct rpcrdma_create_data_internal *);
381 void rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *);
382 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *);
383 void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *);
385 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
386 struct rpcrdma_req *);
387 int rpcrdma_ep_post_recv(struct rpcrdma_ia *, struct rpcrdma_ep *,
388 struct rpcrdma_rep *);
391 * Buffer calls - xprtrdma/verbs.c
393 int rpcrdma_buffer_create(struct rpcrdma_xprt *);
394 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
396 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
397 void rpcrdma_buffer_put(struct rpcrdma_req *);
398 void rpcrdma_recv_buffer_get(struct rpcrdma_req *);
399 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
401 int rpcrdma_register_external(struct rpcrdma_mr_seg *,
402 int, int, struct rpcrdma_xprt *);
403 int rpcrdma_deregister_external(struct rpcrdma_mr_seg *,
404 struct rpcrdma_xprt *);
406 struct rpcrdma_regbuf *rpcrdma_alloc_regbuf(struct rpcrdma_ia *,
408 void rpcrdma_free_regbuf(struct rpcrdma_ia *,
409 struct rpcrdma_regbuf *);
412 * RPC/RDMA connection management calls - xprtrdma/rpc_rdma.c
414 void rpcrdma_connect_worker(struct work_struct *);
415 void rpcrdma_conn_func(struct rpcrdma_ep *);
416 void rpcrdma_reply_handler(struct rpcrdma_rep *);
419 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
421 ssize_t rpcrdma_marshal_chunks(struct rpc_rqst *, ssize_t);
422 int rpcrdma_marshal_req(struct rpc_rqst *);
423 size_t rpcrdma_max_payload(struct rpcrdma_xprt *);
425 /* Temporary NFS request map cache. Created in svc_rdma.c */
426 extern struct kmem_cache *svc_rdma_map_cachep;
427 /* WR context cache. Created in svc_rdma.c */
428 extern struct kmem_cache *svc_rdma_ctxt_cachep;
429 /* Workqueue created in svc_rdma.c */
430 extern struct workqueue_struct *svc_rdma_wq;
432 #if RPCSVC_MAXPAYLOAD < (RPCRDMA_MAX_DATA_SEGS << PAGE_SHIFT)
433 #define RPCSVC_MAXPAYLOAD_RDMA RPCSVC_MAXPAYLOAD
435 #define RPCSVC_MAXPAYLOAD_RDMA (RPCRDMA_MAX_DATA_SEGS << PAGE_SHIFT)
438 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */