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 rdma_cm_id *ri_id;
67 struct ib_mr *ri_bind_mem;
70 struct completion ri_done;
72 enum rpcrdma_memreg ri_memreg_strategy;
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 rpcrdma_xprt *rr_rxprt;
177 struct list_head rr_list;
178 struct rpcrdma_regbuf *rr_rdmabuf;
182 * struct rpcrdma_mw - external memory region metadata
184 * An external memory region is any buffer or page that is registered
185 * on the fly (ie, not pre-registered).
187 * Each rpcrdma_buffer has a list of free MWs anchored in rb_mws. During
188 * call_allocate, rpcrdma_buffer_get() assigns one to each segment in
189 * an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep
190 * track of registration metadata while each RPC is pending.
191 * rpcrdma_deregister_external() uses this metadata to unmap and
192 * release these resources when an RPC is complete.
194 enum rpcrdma_frmr_state {
195 FRMR_IS_INVALID, /* ready to be used */
196 FRMR_IS_VALID, /* in use */
197 FRMR_IS_STALE, /* failed completion */
200 struct rpcrdma_frmr {
201 struct ib_fast_reg_page_list *fr_pgl;
203 enum rpcrdma_frmr_state fr_state;
209 struct rpcrdma_frmr frmr;
211 void (*mw_sendcompletion)(struct ib_wc *);
212 struct list_head mw_list;
213 struct list_head mw_all;
217 * struct rpcrdma_req -- structure central to the request/reply sequence.
219 * N of these are associated with a transport instance, and stored in
220 * struct rpcrdma_buffer. N is the max number of outstanding requests.
222 * It includes pre-registered buffer memory for send AND recv.
223 * The recv buffer, however, is not owned by this structure, and
224 * is "donated" to the hardware when a recv is posted. When a
225 * reply is handled, the recv buffer used is given back to the
226 * struct rpcrdma_req associated with the request.
228 * In addition to the basic memory, this structure includes an array
229 * of iovs for send operations. The reason is that the iovs passed to
230 * ib_post_{send,recv} must not be modified until the work request
234 * o RPCRDMA_MAX_SEGS is the max number of addressible chunk elements we
235 * marshal. The number needed varies depending on the iov lists that
236 * are passed to us, the memory registration mode we are in, and if
237 * physical addressing is used, the layout.
240 struct rpcrdma_mr_seg { /* chunk descriptors */
241 struct rpcrdma_mw *rl_mw; /* registered MR */
242 u64 mr_base; /* registration result */
243 u32 mr_rkey; /* registration result */
244 u32 mr_len; /* length of chunk or segment */
245 int mr_nsegs; /* number of segments in chunk or 0 */
246 enum dma_data_direction mr_dir; /* segment mapping direction */
247 dma_addr_t mr_dma; /* segment mapping address */
248 size_t mr_dmalen; /* segment mapping length */
249 struct page *mr_page; /* owning page, if any */
250 char *mr_offset; /* kva if no page, else offset */
254 unsigned int rl_niovs; /* 0, 2 or 4 */
255 unsigned int rl_nchunks; /* non-zero if chunks */
256 unsigned int rl_connect_cookie; /* retry detection */
257 struct rpcrdma_buffer *rl_buffer; /* home base for this structure */
258 struct rpcrdma_rep *rl_reply;/* holder for reply buffer */
259 struct ib_sge rl_send_iov[4]; /* for active requests */
260 struct rpcrdma_regbuf *rl_rdmabuf;
261 struct rpcrdma_regbuf *rl_sendbuf;
262 struct rpcrdma_mr_seg rl_segments[RPCRDMA_MAX_SEGS];
265 static inline struct rpcrdma_req *
266 rpcr_to_rdmar(struct rpc_rqst *rqst)
268 void *buffer = rqst->rq_buffer;
269 struct rpcrdma_regbuf *rb;
271 rb = container_of(buffer, struct rpcrdma_regbuf, rg_base);
276 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
277 * inline requests/replies, and client/server credits.
279 * One of these is associated with a transport instance
281 struct rpcrdma_buffer {
282 spinlock_t rb_lock; /* protects indexes */
283 u32 rb_max_requests;/* client max requests */
284 struct list_head rb_mws; /* optional memory windows/fmrs/frmrs */
285 struct list_head rb_all;
287 struct rpcrdma_req **rb_send_bufs;
289 struct rpcrdma_rep **rb_recv_bufs;
292 #define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia)
295 * Internal structure for transport instance creation. This
296 * exists primarily for modularity.
298 * This data should be set with mount options
300 struct rpcrdma_create_data_internal {
301 struct sockaddr_storage addr; /* RDMA server address */
302 unsigned int max_requests; /* max requests (slots) in flight */
303 unsigned int rsize; /* mount rsize - max read hdr+data */
304 unsigned int wsize; /* mount wsize - max write hdr+data */
305 unsigned int inline_rsize; /* max non-rdma read data payload */
306 unsigned int inline_wsize; /* max non-rdma write data payload */
307 unsigned int padding; /* non-rdma write header padding */
310 #define RPCRDMA_INLINE_READ_THRESHOLD(rq) \
311 (rpcx_to_rdmad(rq->rq_xprt).inline_rsize)
313 #define RPCRDMA_INLINE_WRITE_THRESHOLD(rq)\
314 (rpcx_to_rdmad(rq->rq_xprt).inline_wsize)
316 #define RPCRDMA_INLINE_PAD_VALUE(rq)\
317 rpcx_to_rdmad(rq->rq_xprt).padding
320 * Statistics for RPCRDMA
322 struct rpcrdma_stats {
323 unsigned long read_chunk_count;
324 unsigned long write_chunk_count;
325 unsigned long reply_chunk_count;
327 unsigned long long total_rdma_request;
328 unsigned long long total_rdma_reply;
330 unsigned long long pullup_copy_count;
331 unsigned long long fixup_copy_count;
332 unsigned long hardway_register_count;
333 unsigned long failed_marshal_count;
334 unsigned long bad_reply_count;
338 * Per-registration mode operations
341 struct rpcrdma_memreg_ops {
342 int (*ro_map)(struct rpcrdma_xprt *,
343 struct rpcrdma_mr_seg *, int, bool);
344 int (*ro_unmap)(struct rpcrdma_xprt *,
345 struct rpcrdma_mr_seg *);
346 int (*ro_open)(struct rpcrdma_ia *,
348 struct rpcrdma_create_data_internal *);
349 size_t (*ro_maxpages)(struct rpcrdma_xprt *);
350 int (*ro_init)(struct rpcrdma_xprt *);
351 void (*ro_reset)(struct rpcrdma_xprt *);
352 void (*ro_destroy)(struct rpcrdma_buffer *);
353 const char *ro_displayname;
356 extern const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops;
357 extern const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops;
358 extern const struct rpcrdma_memreg_ops rpcrdma_physical_memreg_ops;
361 * RPCRDMA transport -- encapsulates the structures above for
362 * integration with RPC.
364 * The contained structures are embedded, not pointers,
365 * for convenience. This structure need not be visible externally.
367 * It is allocated and initialized during mount, and released
370 struct rpcrdma_xprt {
371 struct rpc_xprt rx_xprt;
372 struct rpcrdma_ia rx_ia;
373 struct rpcrdma_ep rx_ep;
374 struct rpcrdma_buffer rx_buf;
375 struct rpcrdma_create_data_internal rx_data;
376 struct delayed_work rx_connect_worker;
377 struct rpcrdma_stats rx_stats;
380 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
381 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data)
383 /* Setting this to 0 ensures interoperability with early servers.
384 * Setting this to 1 enhances certain unaligned read/write performance.
385 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
386 extern int xprt_rdma_pad_optimize;
389 * Interface Adapter calls - xprtrdma/verbs.c
391 int rpcrdma_ia_open(struct rpcrdma_xprt *, struct sockaddr *, int);
392 void rpcrdma_ia_close(struct rpcrdma_ia *);
395 * Endpoint calls - xprtrdma/verbs.c
397 int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *,
398 struct rpcrdma_create_data_internal *);
399 void rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *);
400 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *);
401 void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *);
403 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
404 struct rpcrdma_req *);
405 int rpcrdma_ep_post_recv(struct rpcrdma_ia *, struct rpcrdma_ep *,
406 struct rpcrdma_rep *);
409 * Buffer calls - xprtrdma/verbs.c
411 int rpcrdma_buffer_create(struct rpcrdma_xprt *);
412 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
414 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
415 void rpcrdma_buffer_put(struct rpcrdma_req *);
416 void rpcrdma_recv_buffer_get(struct rpcrdma_req *);
417 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
419 struct rpcrdma_regbuf *rpcrdma_alloc_regbuf(struct rpcrdma_ia *,
421 void rpcrdma_free_regbuf(struct rpcrdma_ia *,
422 struct rpcrdma_regbuf *);
424 unsigned int rpcrdma_max_segments(struct rpcrdma_xprt *);
427 * Wrappers for chunk registration, shared by read/write chunk code.
430 void rpcrdma_mapping_error(struct rpcrdma_mr_seg *);
432 static inline enum dma_data_direction
433 rpcrdma_data_dir(bool writing)
435 return writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
439 rpcrdma_map_one(struct ib_device *device, struct rpcrdma_mr_seg *seg,
440 enum dma_data_direction direction)
442 seg->mr_dir = direction;
443 seg->mr_dmalen = seg->mr_len;
446 seg->mr_dma = ib_dma_map_page(device,
447 seg->mr_page, offset_in_page(seg->mr_offset),
448 seg->mr_dmalen, seg->mr_dir);
450 seg->mr_dma = ib_dma_map_single(device,
452 seg->mr_dmalen, seg->mr_dir);
454 if (ib_dma_mapping_error(device, seg->mr_dma))
455 rpcrdma_mapping_error(seg);
459 rpcrdma_unmap_one(struct ib_device *device, struct rpcrdma_mr_seg *seg)
462 ib_dma_unmap_page(device,
463 seg->mr_dma, seg->mr_dmalen, seg->mr_dir);
465 ib_dma_unmap_single(device,
466 seg->mr_dma, seg->mr_dmalen, seg->mr_dir);
470 * RPC/RDMA connection management calls - xprtrdma/rpc_rdma.c
472 void rpcrdma_connect_worker(struct work_struct *);
473 void rpcrdma_conn_func(struct rpcrdma_ep *);
474 void rpcrdma_reply_handler(struct rpcrdma_rep *);
477 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
479 int rpcrdma_marshal_req(struct rpc_rqst *);
481 /* Temporary NFS request map cache. Created in svc_rdma.c */
482 extern struct kmem_cache *svc_rdma_map_cachep;
483 /* WR context cache. Created in svc_rdma.c */
484 extern struct kmem_cache *svc_rdma_ctxt_cachep;
485 /* Workqueue created in svc_rdma.c */
486 extern struct workqueue_struct *svc_rdma_wq;
488 #if RPCSVC_MAXPAYLOAD < (RPCRDMA_MAX_DATA_SEGS << PAGE_SHIFT)
489 #define RPCSVC_MAXPAYLOAD_RDMA RPCSVC_MAXPAYLOAD
491 #define RPCSVC_MAXPAYLOAD_RDMA (RPCRDMA_MAX_DATA_SEGS << PAGE_SHIFT)
494 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */