xref: /openbmc/linux/net/sunrpc/xprtrdma/xprt_rdma.h (revision b627b4ed)
1 /*
2  * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
3  *
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
8  * license below:
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  *
14  *      Redistributions of source code must retain the above copyright
15  *      notice, this list of conditions and the following disclaimer.
16  *
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.
21  *
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
25  *      permission.
26  *
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.
38  */
39 
40 #ifndef _LINUX_SUNRPC_XPRT_RDMA_H
41 #define _LINUX_SUNRPC_XPRT_RDMA_H
42 
43 #include <linux/wait.h> 		/* wait_queue_head_t, etc */
44 #include <linux/spinlock.h> 		/* spinlock_t, etc */
45 #include <asm/atomic.h>			/* atomic_t, etc */
46 
47 #include <rdma/rdma_cm.h>		/* RDMA connection api */
48 #include <rdma/ib_verbs.h>		/* RDMA verbs api */
49 
50 #include <linux/sunrpc/clnt.h> 		/* rpc_xprt */
51 #include <linux/sunrpc/rpc_rdma.h> 	/* RPC/RDMA protocol */
52 #include <linux/sunrpc/xprtrdma.h> 	/* xprt parameters */
53 
54 #define RDMA_RESOLVE_TIMEOUT	(5000)	/* 5 seconds */
55 #define RDMA_CONNECT_RETRY_MAX	(2)	/* retries if no listener backlog */
56 
57 /*
58  * Interface Adapter -- one per transport instance
59  */
60 struct rpcrdma_ia {
61 	struct rdma_cm_id 	*ri_id;
62 	struct ib_pd		*ri_pd;
63 	struct ib_mr		*ri_bind_mem;
64 	u32			ri_dma_lkey;
65 	int			ri_have_dma_lkey;
66 	struct completion	ri_done;
67 	int			ri_async_rc;
68 	enum rpcrdma_memreg	ri_memreg_strategy;
69 };
70 
71 /*
72  * RDMA Endpoint -- one per transport instance
73  */
74 
75 struct rpcrdma_ep {
76 	atomic_t		rep_cqcount;
77 	int			rep_cqinit;
78 	int			rep_connected;
79 	struct rpcrdma_ia	*rep_ia;
80 	struct ib_cq		*rep_cq;
81 	struct ib_qp_init_attr	rep_attr;
82 	wait_queue_head_t 	rep_connect_wait;
83 	struct ib_sge		rep_pad;	/* holds zeroed pad */
84 	struct ib_mr		*rep_pad_mr;	/* holds zeroed pad */
85 	void			(*rep_func)(struct rpcrdma_ep *);
86 	struct rpc_xprt		*rep_xprt;	/* for rep_func */
87 	struct rdma_conn_param	rep_remote_cma;
88 	struct sockaddr_storage	rep_remote_addr;
89 };
90 
91 #define INIT_CQCOUNT(ep) atomic_set(&(ep)->rep_cqcount, (ep)->rep_cqinit)
92 #define DECR_CQCOUNT(ep) atomic_sub_return(1, &(ep)->rep_cqcount)
93 
94 /*
95  * struct rpcrdma_rep -- this structure encapsulates state required to recv
96  * and complete a reply, asychronously. It needs several pieces of
97  * state:
98  *   o recv buffer (posted to provider)
99  *   o ib_sge (also donated to provider)
100  *   o status of reply (length, success or not)
101  *   o bookkeeping state to get run by tasklet (list, etc)
102  *
103  * These are allocated during initialization, per-transport instance;
104  * however, the tasklet execution list itself is global, as it should
105  * always be pretty short.
106  *
107  * N of these are associated with a transport instance, and stored in
108  * struct rpcrdma_buffer. N is the max number of outstanding requests.
109  */
110 
111 /* temporary static scatter/gather max */
112 #define RPCRDMA_MAX_DATA_SEGS	(8)	/* max scatter/gather */
113 #define RPCRDMA_MAX_SEGS 	(RPCRDMA_MAX_DATA_SEGS + 2) /* head+tail = 2 */
114 #define MAX_RPCRDMAHDR	(\
115 	/* max supported RPC/RDMA header */ \
116 	sizeof(struct rpcrdma_msg) + (2 * sizeof(u32)) + \
117 	(sizeof(struct rpcrdma_read_chunk) * RPCRDMA_MAX_SEGS) + sizeof(u32))
118 
119 struct rpcrdma_buffer;
120 
121 struct rpcrdma_rep {
122 	unsigned int	rr_len;		/* actual received reply length */
123 	struct rpcrdma_buffer *rr_buffer; /* home base for this structure */
124 	struct rpc_xprt	*rr_xprt;	/* needed for request/reply matching */
125 	void (*rr_func)(struct rpcrdma_rep *);/* called by tasklet in softint */
126 	struct list_head rr_list;	/* tasklet list */
127 	wait_queue_head_t rr_unbind;	/* optional unbind wait */
128 	struct ib_sge	rr_iov;		/* for posting */
129 	struct ib_mr	*rr_handle;	/* handle for mem in rr_iov */
130 	char	rr_base[MAX_RPCRDMAHDR]; /* minimal inline receive buffer */
131 };
132 
133 /*
134  * struct rpcrdma_req -- structure central to the request/reply sequence.
135  *
136  * N of these are associated with a transport instance, and stored in
137  * struct rpcrdma_buffer. N is the max number of outstanding requests.
138  *
139  * It includes pre-registered buffer memory for send AND recv.
140  * The recv buffer, however, is not owned by this structure, and
141  * is "donated" to the hardware when a recv is posted. When a
142  * reply is handled, the recv buffer used is given back to the
143  * struct rpcrdma_req associated with the request.
144  *
145  * In addition to the basic memory, this structure includes an array
146  * of iovs for send operations. The reason is that the iovs passed to
147  * ib_post_{send,recv} must not be modified until the work request
148  * completes.
149  *
150  * NOTES:
151  *   o RPCRDMA_MAX_SEGS is the max number of addressible chunk elements we
152  *     marshal. The number needed varies depending on the iov lists that
153  *     are passed to us, the memory registration mode we are in, and if
154  *     physical addressing is used, the layout.
155  */
156 
157 struct rpcrdma_mr_seg {		/* chunk descriptors */
158 	union {				/* chunk memory handles */
159 		struct ib_mr	*rl_mr;		/* if registered directly */
160 		struct rpcrdma_mw {		/* if registered from region */
161 			union {
162 				struct ib_mw	*mw;
163 				struct ib_fmr	*fmr;
164 				struct {
165 					struct ib_fast_reg_page_list *fr_pgl;
166 					struct ib_mr *fr_mr;
167 				} frmr;
168 			} r;
169 			struct list_head mw_list;
170 		} *rl_mw;
171 	} mr_chunk;
172 	u64		mr_base;	/* registration result */
173 	u32		mr_rkey;	/* registration result */
174 	u32		mr_len;		/* length of chunk or segment */
175 	int		mr_nsegs;	/* number of segments in chunk or 0 */
176 	enum dma_data_direction	mr_dir;	/* segment mapping direction */
177 	dma_addr_t	mr_dma;		/* segment mapping address */
178 	size_t		mr_dmalen;	/* segment mapping length */
179 	struct page	*mr_page;	/* owning page, if any */
180 	char		*mr_offset;	/* kva if no page, else offset */
181 };
182 
183 struct rpcrdma_req {
184 	size_t 		rl_size;	/* actual length of buffer */
185 	unsigned int	rl_niovs;	/* 0, 2 or 4 */
186 	unsigned int	rl_nchunks;	/* non-zero if chunks */
187 	unsigned int	rl_connect_cookie;	/* retry detection */
188 	struct rpcrdma_buffer *rl_buffer; /* home base for this structure */
189 	struct rpcrdma_rep	*rl_reply;/* holder for reply buffer */
190 	struct rpcrdma_mr_seg rl_segments[RPCRDMA_MAX_SEGS];/* chunk segments */
191 	struct ib_sge	rl_send_iov[4];	/* for active requests */
192 	struct ib_sge	rl_iov;		/* for posting */
193 	struct ib_mr	*rl_handle;	/* handle for mem in rl_iov */
194 	char		rl_base[MAX_RPCRDMAHDR]; /* start of actual buffer */
195 	__u32 		rl_xdr_buf[0];	/* start of returned rpc rq_buffer */
196 };
197 #define rpcr_to_rdmar(r) \
198 	container_of((r)->rq_buffer, struct rpcrdma_req, rl_xdr_buf[0])
199 
200 /*
201  * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
202  * inline requests/replies, and client/server credits.
203  *
204  * One of these is associated with a transport instance
205  */
206 struct rpcrdma_buffer {
207 	spinlock_t	rb_lock;	/* protects indexes */
208 	atomic_t	rb_credits;	/* most recent server credits */
209 	unsigned long	rb_cwndscale;	/* cached framework rpc_cwndscale */
210 	int		rb_max_requests;/* client max requests */
211 	struct list_head rb_mws;	/* optional memory windows/fmrs/frmrs */
212 	int		rb_send_index;
213 	struct rpcrdma_req	**rb_send_bufs;
214 	int		rb_recv_index;
215 	struct rpcrdma_rep	**rb_recv_bufs;
216 	char		*rb_pool;
217 };
218 #define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia)
219 
220 /*
221  * Internal structure for transport instance creation. This
222  * exists primarily for modularity.
223  *
224  * This data should be set with mount options
225  */
226 struct rpcrdma_create_data_internal {
227 	struct sockaddr_storage	addr;	/* RDMA server address */
228 	unsigned int	max_requests;	/* max requests (slots) in flight */
229 	unsigned int	rsize;		/* mount rsize - max read hdr+data */
230 	unsigned int	wsize;		/* mount wsize - max write hdr+data */
231 	unsigned int	inline_rsize;	/* max non-rdma read data payload */
232 	unsigned int	inline_wsize;	/* max non-rdma write data payload */
233 	unsigned int	padding;	/* non-rdma write header padding */
234 };
235 
236 #define RPCRDMA_INLINE_READ_THRESHOLD(rq) \
237 	(rpcx_to_rdmad(rq->rq_task->tk_xprt).inline_rsize)
238 
239 #define RPCRDMA_INLINE_WRITE_THRESHOLD(rq)\
240 	(rpcx_to_rdmad(rq->rq_task->tk_xprt).inline_wsize)
241 
242 #define RPCRDMA_INLINE_PAD_VALUE(rq)\
243 	rpcx_to_rdmad(rq->rq_task->tk_xprt).padding
244 
245 /*
246  * Statistics for RPCRDMA
247  */
248 struct rpcrdma_stats {
249 	unsigned long		read_chunk_count;
250 	unsigned long		write_chunk_count;
251 	unsigned long		reply_chunk_count;
252 
253 	unsigned long long	total_rdma_request;
254 	unsigned long long	total_rdma_reply;
255 
256 	unsigned long long	pullup_copy_count;
257 	unsigned long long	fixup_copy_count;
258 	unsigned long		hardway_register_count;
259 	unsigned long		failed_marshal_count;
260 	unsigned long		bad_reply_count;
261 };
262 
263 /*
264  * RPCRDMA transport -- encapsulates the structures above for
265  * integration with RPC.
266  *
267  * The contained structures are embedded, not pointers,
268  * for convenience. This structure need not be visible externally.
269  *
270  * It is allocated and initialized during mount, and released
271  * during unmount.
272  */
273 struct rpcrdma_xprt {
274 	struct rpc_xprt		xprt;
275 	struct rpcrdma_ia	rx_ia;
276 	struct rpcrdma_ep	rx_ep;
277 	struct rpcrdma_buffer	rx_buf;
278 	struct rpcrdma_create_data_internal rx_data;
279 	struct delayed_work	rdma_connect;
280 	struct rpcrdma_stats	rx_stats;
281 };
282 
283 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, xprt)
284 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data)
285 
286 /* Setting this to 0 ensures interoperability with early servers.
287  * Setting this to 1 enhances certain unaligned read/write performance.
288  * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
289 extern int xprt_rdma_pad_optimize;
290 
291 /*
292  * Interface Adapter calls - xprtrdma/verbs.c
293  */
294 int rpcrdma_ia_open(struct rpcrdma_xprt *, struct sockaddr *, int);
295 void rpcrdma_ia_close(struct rpcrdma_ia *);
296 
297 /*
298  * Endpoint calls - xprtrdma/verbs.c
299  */
300 int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *,
301 				struct rpcrdma_create_data_internal *);
302 int rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *);
303 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *);
304 int rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *);
305 
306 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
307 				struct rpcrdma_req *);
308 int rpcrdma_ep_post_recv(struct rpcrdma_ia *, struct rpcrdma_ep *,
309 				struct rpcrdma_rep *);
310 
311 /*
312  * Buffer calls - xprtrdma/verbs.c
313  */
314 int rpcrdma_buffer_create(struct rpcrdma_buffer *, struct rpcrdma_ep *,
315 				struct rpcrdma_ia *,
316 				struct rpcrdma_create_data_internal *);
317 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
318 
319 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
320 void rpcrdma_buffer_put(struct rpcrdma_req *);
321 void rpcrdma_recv_buffer_get(struct rpcrdma_req *);
322 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
323 
324 int rpcrdma_register_internal(struct rpcrdma_ia *, void *, int,
325 				struct ib_mr **, struct ib_sge *);
326 int rpcrdma_deregister_internal(struct rpcrdma_ia *,
327 				struct ib_mr *, struct ib_sge *);
328 
329 int rpcrdma_register_external(struct rpcrdma_mr_seg *,
330 				int, int, struct rpcrdma_xprt *);
331 int rpcrdma_deregister_external(struct rpcrdma_mr_seg *,
332 				struct rpcrdma_xprt *, void *);
333 
334 /*
335  * RPC/RDMA connection management calls - xprtrdma/rpc_rdma.c
336  */
337 void rpcrdma_conn_func(struct rpcrdma_ep *);
338 void rpcrdma_reply_handler(struct rpcrdma_rep *);
339 
340 /*
341  * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
342  */
343 int rpcrdma_marshal_req(struct rpc_rqst *);
344 
345 #endif				/* _LINUX_SUNRPC_XPRT_RDMA_H */
346