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 enum rpcrdma_memreg ri_memreg_strategy;
74 unsigned int ri_max_frmr_depth;
75 struct ib_device_attr ri_devattr;
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
84 #define RPCRDMA_WC_BUDGET (128)
85 #define RPCRDMA_POLLSIZE (16)
91 struct ib_qp_init_attr rep_attr;
92 wait_queue_head_t rep_connect_wait;
93 struct rpcrdma_regbuf *rep_padbuf;
94 struct rdma_conn_param rep_remote_cma;
95 struct sockaddr_storage rep_remote_addr;
96 struct delayed_work rep_connect_worker;
97 struct ib_wc rep_send_wcs[RPCRDMA_POLLSIZE];
98 struct ib_wc rep_recv_wcs[RPCRDMA_POLLSIZE];
102 * Force a signaled SEND Work Request every so often,
103 * in case the provider needs to do some housekeeping.
105 #define RPCRDMA_MAX_UNSIGNALED_SENDS (32)
107 #define INIT_CQCOUNT(ep) atomic_set(&(ep)->rep_cqcount, (ep)->rep_cqinit)
108 #define DECR_CQCOUNT(ep) atomic_sub_return(1, &(ep)->rep_cqcount)
110 /* Force completion handler to ignore the signal
112 #define RPCRDMA_IGNORE_COMPLETION (0ULL)
114 /* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
116 * The below structure appears at the front of a large region of kmalloc'd
117 * memory, which always starts on a good alignment boundary.
120 struct rpcrdma_regbuf {
122 struct rpcrdma_req *rg_owner;
124 struct ib_sge rg_iov;
125 __be32 rg_base[0] __attribute__ ((aligned(256)));
129 rdmab_addr(struct rpcrdma_regbuf *rb)
131 return rb->rg_iov.addr;
135 rdmab_length(struct rpcrdma_regbuf *rb)
137 return rb->rg_iov.length;
141 rdmab_lkey(struct rpcrdma_regbuf *rb)
143 return rb->rg_iov.lkey;
146 static inline struct rpcrdma_msg *
147 rdmab_to_msg(struct rpcrdma_regbuf *rb)
149 return (struct rpcrdma_msg *)rb->rg_base;
153 * struct rpcrdma_rep -- this structure encapsulates state required to recv
154 * and complete a reply, asychronously. It needs several pieces of
156 * o recv buffer (posted to provider)
157 * o ib_sge (also donated to provider)
158 * o status of reply (length, success or not)
159 * o bookkeeping state to get run by tasklet (list, etc)
161 * These are allocated during initialization, per-transport instance;
162 * however, the tasklet execution list itself is global, as it should
163 * always be pretty short.
165 * N of these are associated with a transport instance, and stored in
166 * struct rpcrdma_buffer. N is the max number of outstanding requests.
169 /* temporary static scatter/gather max */
170 #define RPCRDMA_MAX_DATA_SEGS (64) /* max scatter/gather */
171 #define RPCRDMA_MAX_SEGS (RPCRDMA_MAX_DATA_SEGS + 2) /* head+tail = 2 */
173 struct rpcrdma_buffer;
177 struct ib_device *rr_device;
178 struct rpcrdma_xprt *rr_rxprt;
179 struct list_head rr_list;
180 struct rpcrdma_regbuf *rr_rdmabuf;
184 * struct rpcrdma_mw - external memory region metadata
186 * An external memory region is any buffer or page that is registered
187 * on the fly (ie, not pre-registered).
189 * Each rpcrdma_buffer has a list of free MWs anchored in rb_mws. During
190 * call_allocate, rpcrdma_buffer_get() assigns one to each segment in
191 * an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep
192 * track of registration metadata while each RPC is pending.
193 * rpcrdma_deregister_external() uses this metadata to unmap and
194 * release these resources when an RPC is complete.
196 enum rpcrdma_frmr_state {
197 FRMR_IS_INVALID, /* ready to be used */
198 FRMR_IS_VALID, /* in use */
199 FRMR_IS_STALE, /* failed completion */
202 struct rpcrdma_frmr {
203 struct ib_fast_reg_page_list *fr_pgl;
205 enum rpcrdma_frmr_state fr_state;
206 struct work_struct fr_work;
207 struct rpcrdma_xprt *fr_xprt;
213 struct rpcrdma_frmr frmr;
215 void (*mw_sendcompletion)(struct ib_wc *);
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 struct rpcrdma_buffer *rl_buffer; /* home base for this structure */
262 struct rpcrdma_rep *rl_reply;/* holder for reply buffer */
263 struct ib_sge rl_send_iov[4]; /* for active requests */
264 struct rpcrdma_regbuf *rl_rdmabuf;
265 struct rpcrdma_regbuf *rl_sendbuf;
266 struct rpcrdma_mr_seg rl_segments[RPCRDMA_MAX_SEGS];
269 static inline struct rpcrdma_req *
270 rpcr_to_rdmar(struct rpc_rqst *rqst)
272 void *buffer = rqst->rq_buffer;
273 struct rpcrdma_regbuf *rb;
275 rb = container_of(buffer, struct rpcrdma_regbuf, rg_base);
280 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
281 * inline requests/replies, and client/server credits.
283 * One of these is associated with a transport instance
285 struct rpcrdma_buffer {
286 spinlock_t rb_lock; /* protects indexes */
287 u32 rb_max_requests;/* client max requests */
288 struct list_head rb_mws; /* optional memory windows/fmrs/frmrs */
289 struct list_head rb_all;
291 struct rpcrdma_req **rb_send_bufs;
293 struct rpcrdma_rep **rb_recv_bufs;
296 #define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia)
299 * Internal structure for transport instance creation. This
300 * exists primarily for modularity.
302 * This data should be set with mount options
304 struct rpcrdma_create_data_internal {
305 struct sockaddr_storage addr; /* RDMA server address */
306 unsigned int max_requests; /* max requests (slots) in flight */
307 unsigned int rsize; /* mount rsize - max read hdr+data */
308 unsigned int wsize; /* mount wsize - max write hdr+data */
309 unsigned int inline_rsize; /* max non-rdma read data payload */
310 unsigned int inline_wsize; /* max non-rdma write data payload */
311 unsigned int padding; /* non-rdma write header padding */
314 #define RPCRDMA_INLINE_READ_THRESHOLD(rq) \
315 (rpcx_to_rdmad(rq->rq_xprt).inline_rsize)
317 #define RPCRDMA_INLINE_WRITE_THRESHOLD(rq)\
318 (rpcx_to_rdmad(rq->rq_xprt).inline_wsize)
320 #define RPCRDMA_INLINE_PAD_VALUE(rq)\
321 rpcx_to_rdmad(rq->rq_xprt).padding
324 * Statistics for RPCRDMA
326 struct rpcrdma_stats {
327 unsigned long read_chunk_count;
328 unsigned long write_chunk_count;
329 unsigned long reply_chunk_count;
331 unsigned long long total_rdma_request;
332 unsigned long long total_rdma_reply;
334 unsigned long long pullup_copy_count;
335 unsigned long long fixup_copy_count;
336 unsigned long hardway_register_count;
337 unsigned long failed_marshal_count;
338 unsigned long bad_reply_count;
342 * Per-registration mode operations
345 struct rpcrdma_memreg_ops {
346 int (*ro_map)(struct rpcrdma_xprt *,
347 struct rpcrdma_mr_seg *, int, bool);
348 int (*ro_unmap)(struct rpcrdma_xprt *,
349 struct rpcrdma_mr_seg *);
350 int (*ro_open)(struct rpcrdma_ia *,
352 struct rpcrdma_create_data_internal *);
353 size_t (*ro_maxpages)(struct rpcrdma_xprt *);
354 int (*ro_init)(struct rpcrdma_xprt *);
355 void (*ro_destroy)(struct rpcrdma_buffer *);
356 const char *ro_displayname;
359 extern const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops;
360 extern const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops;
361 extern const struct rpcrdma_memreg_ops rpcrdma_physical_memreg_ops;
364 * RPCRDMA transport -- encapsulates the structures above for
365 * integration with RPC.
367 * The contained structures are embedded, not pointers,
368 * for convenience. This structure need not be visible externally.
370 * It is allocated and initialized during mount, and released
373 struct rpcrdma_xprt {
374 struct rpc_xprt rx_xprt;
375 struct rpcrdma_ia rx_ia;
376 struct rpcrdma_ep rx_ep;
377 struct rpcrdma_buffer rx_buf;
378 struct rpcrdma_create_data_internal rx_data;
379 struct delayed_work rx_connect_worker;
380 struct rpcrdma_stats rx_stats;
383 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
384 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data)
386 /* Setting this to 0 ensures interoperability with early servers.
387 * Setting this to 1 enhances certain unaligned read/write performance.
388 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
389 extern int xprt_rdma_pad_optimize;
392 * Interface Adapter calls - xprtrdma/verbs.c
394 int rpcrdma_ia_open(struct rpcrdma_xprt *, struct sockaddr *, int);
395 void rpcrdma_ia_close(struct rpcrdma_ia *);
398 * Endpoint calls - xprtrdma/verbs.c
400 int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *,
401 struct rpcrdma_create_data_internal *);
402 void rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *);
403 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *);
404 void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *);
406 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
407 struct rpcrdma_req *);
408 int rpcrdma_ep_post_recv(struct rpcrdma_ia *, struct rpcrdma_ep *,
409 struct rpcrdma_rep *);
412 * Buffer calls - xprtrdma/verbs.c
414 int rpcrdma_buffer_create(struct rpcrdma_xprt *);
415 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
417 struct rpcrdma_mw *rpcrdma_get_mw(struct rpcrdma_xprt *);
418 void rpcrdma_put_mw(struct rpcrdma_xprt *, struct rpcrdma_mw *);
419 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
420 void rpcrdma_buffer_put(struct rpcrdma_req *);
421 void rpcrdma_recv_buffer_get(struct rpcrdma_req *);
422 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
424 struct rpcrdma_regbuf *rpcrdma_alloc_regbuf(struct rpcrdma_ia *,
426 void rpcrdma_free_regbuf(struct rpcrdma_ia *,
427 struct rpcrdma_regbuf *);
429 unsigned int rpcrdma_max_segments(struct rpcrdma_xprt *);
431 int frwr_alloc_recovery_wq(void);
432 void frwr_destroy_recovery_wq(void);
435 * Wrappers for chunk registration, shared by read/write chunk code.
438 void rpcrdma_mapping_error(struct rpcrdma_mr_seg *);
440 static inline enum dma_data_direction
441 rpcrdma_data_dir(bool writing)
443 return writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
447 rpcrdma_map_one(struct ib_device *device, struct rpcrdma_mr_seg *seg,
448 enum dma_data_direction direction)
450 seg->mr_dir = direction;
451 seg->mr_dmalen = seg->mr_len;
454 seg->mr_dma = ib_dma_map_page(device,
455 seg->mr_page, offset_in_page(seg->mr_offset),
456 seg->mr_dmalen, seg->mr_dir);
458 seg->mr_dma = ib_dma_map_single(device,
460 seg->mr_dmalen, seg->mr_dir);
462 if (ib_dma_mapping_error(device, seg->mr_dma))
463 rpcrdma_mapping_error(seg);
467 rpcrdma_unmap_one(struct ib_device *device, struct rpcrdma_mr_seg *seg)
470 ib_dma_unmap_page(device,
471 seg->mr_dma, seg->mr_dmalen, seg->mr_dir);
473 ib_dma_unmap_single(device,
474 seg->mr_dma, seg->mr_dmalen, seg->mr_dir);
478 * RPC/RDMA connection management calls - xprtrdma/rpc_rdma.c
480 void rpcrdma_connect_worker(struct work_struct *);
481 void rpcrdma_conn_func(struct rpcrdma_ep *);
482 void rpcrdma_reply_handler(struct rpcrdma_rep *);
485 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
487 int rpcrdma_marshal_req(struct rpc_rqst *);
489 /* Temporary NFS request map cache. Created in svc_rdma.c */
490 extern struct kmem_cache *svc_rdma_map_cachep;
491 /* WR context cache. Created in svc_rdma.c */
492 extern struct kmem_cache *svc_rdma_ctxt_cachep;
493 /* Workqueue created in svc_rdma.c */
494 extern struct workqueue_struct *svc_rdma_wq;
496 #if RPCSVC_MAXPAYLOAD < (RPCRDMA_MAX_DATA_SEGS << PAGE_SHIFT)
497 #define RPCSVC_MAXPAYLOAD_RDMA RPCSVC_MAXPAYLOAD
499 #define RPCSVC_MAXPAYLOAD_RDMA (RPCRDMA_MAX_DATA_SEGS << PAGE_SHIFT)
502 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */