xref: /linux/net/sunrpc/xprtrdma/xprt_rdma.h (revision ca55b2fef3a9373fcfc30f82fd26bc7fccbda732)
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 <linux/atomic.h>			/* atomic_t, etc */
46 #include <linux/workqueue.h>		/* struct work_struct */
47 
48 #include <rdma/rdma_cm.h>		/* RDMA connection api */
49 #include <rdma/ib_verbs.h>		/* RDMA verbs api */
50 
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 
55 #define RDMA_RESOLVE_TIMEOUT	(5000)	/* 5 seconds */
56 #define RDMA_CONNECT_RETRY_MAX	(2)	/* retries if no listener backlog */
57 
58 /*
59  * Interface Adapter -- one per transport instance
60  */
61 struct rpcrdma_ia {
62 	const struct rpcrdma_memreg_ops	*ri_ops;
63 	rwlock_t		ri_qplock;
64 	struct ib_device	*ri_device;
65 	struct rdma_cm_id 	*ri_id;
66 	struct ib_pd		*ri_pd;
67 	struct ib_mr		*ri_dma_mr;
68 	struct completion	ri_done;
69 	int			ri_async_rc;
70 	unsigned int		ri_max_frmr_depth;
71 	struct ib_device_attr	ri_devattr;
72 	struct ib_qp_attr	ri_qp_attr;
73 	struct ib_qp_init_attr	ri_qp_init_attr;
74 };
75 
76 /*
77  * RDMA Endpoint -- one per transport instance
78  */
79 
80 #define RPCRDMA_WC_BUDGET	(128)
81 #define RPCRDMA_POLLSIZE	(16)
82 
83 struct rpcrdma_ep {
84 	atomic_t		rep_cqcount;
85 	int			rep_cqinit;
86 	int			rep_connected;
87 	struct ib_qp_init_attr	rep_attr;
88 	wait_queue_head_t 	rep_connect_wait;
89 	struct rdma_conn_param	rep_remote_cma;
90 	struct sockaddr_storage	rep_remote_addr;
91 	struct delayed_work	rep_connect_worker;
92 	struct ib_wc		rep_send_wcs[RPCRDMA_POLLSIZE];
93 	struct ib_wc		rep_recv_wcs[RPCRDMA_POLLSIZE];
94 };
95 
96 /*
97  * Force a signaled SEND Work Request every so often,
98  * in case the provider needs to do some housekeeping.
99  */
100 #define RPCRDMA_MAX_UNSIGNALED_SENDS	(32)
101 
102 #define INIT_CQCOUNT(ep) atomic_set(&(ep)->rep_cqcount, (ep)->rep_cqinit)
103 #define DECR_CQCOUNT(ep) atomic_sub_return(1, &(ep)->rep_cqcount)
104 
105 /* Force completion handler to ignore the signal
106  */
107 #define RPCRDMA_IGNORE_COMPLETION	(0ULL)
108 
109 /* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
110  *
111  * The below structure appears at the front of a large region of kmalloc'd
112  * memory, which always starts on a good alignment boundary.
113  */
114 
115 struct rpcrdma_regbuf {
116 	size_t			rg_size;
117 	struct rpcrdma_req	*rg_owner;
118 	struct ib_sge		rg_iov;
119 	__be32			rg_base[0] __attribute__ ((aligned(256)));
120 };
121 
122 static inline u64
123 rdmab_addr(struct rpcrdma_regbuf *rb)
124 {
125 	return rb->rg_iov.addr;
126 }
127 
128 static inline u32
129 rdmab_length(struct rpcrdma_regbuf *rb)
130 {
131 	return rb->rg_iov.length;
132 }
133 
134 static inline u32
135 rdmab_lkey(struct rpcrdma_regbuf *rb)
136 {
137 	return rb->rg_iov.lkey;
138 }
139 
140 static inline struct rpcrdma_msg *
141 rdmab_to_msg(struct rpcrdma_regbuf *rb)
142 {
143 	return (struct rpcrdma_msg *)rb->rg_base;
144 }
145 
146 /*
147  * struct rpcrdma_rep -- this structure encapsulates state required to recv
148  * and complete a reply, asychronously. It needs several pieces of
149  * state:
150  *   o recv buffer (posted to provider)
151  *   o ib_sge (also donated to provider)
152  *   o status of reply (length, success or not)
153  *   o bookkeeping state to get run by tasklet (list, etc)
154  *
155  * These are allocated during initialization, per-transport instance;
156  * however, the tasklet execution list itself is global, as it should
157  * always be pretty short.
158  *
159  * N of these are associated with a transport instance, and stored in
160  * struct rpcrdma_buffer. N is the max number of outstanding requests.
161  */
162 
163 #define RPCRDMA_MAX_DATA_SEGS	((1 * 1024 * 1024) / PAGE_SIZE)
164 #define RPCRDMA_MAX_SEGS 	(RPCRDMA_MAX_DATA_SEGS + 2) /* head+tail = 2 */
165 
166 struct rpcrdma_buffer;
167 
168 struct rpcrdma_rep {
169 	unsigned int		rr_len;
170 	struct ib_device	*rr_device;
171 	struct rpcrdma_xprt	*rr_rxprt;
172 	struct list_head	rr_list;
173 	struct rpcrdma_regbuf	*rr_rdmabuf;
174 };
175 
176 /*
177  * struct rpcrdma_mw - external memory region metadata
178  *
179  * An external memory region is any buffer or page that is registered
180  * on the fly (ie, not pre-registered).
181  *
182  * Each rpcrdma_buffer has a list of free MWs anchored in rb_mws. During
183  * call_allocate, rpcrdma_buffer_get() assigns one to each segment in
184  * an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep
185  * track of registration metadata while each RPC is pending.
186  * rpcrdma_deregister_external() uses this metadata to unmap and
187  * release these resources when an RPC is complete.
188  */
189 enum rpcrdma_frmr_state {
190 	FRMR_IS_INVALID,	/* ready to be used */
191 	FRMR_IS_VALID,		/* in use */
192 	FRMR_IS_STALE,		/* failed completion */
193 };
194 
195 struct rpcrdma_frmr {
196 	struct ib_fast_reg_page_list	*fr_pgl;
197 	struct ib_mr			*fr_mr;
198 	enum rpcrdma_frmr_state		fr_state;
199 	struct work_struct		fr_work;
200 	struct rpcrdma_xprt		*fr_xprt;
201 };
202 
203 struct rpcrdma_fmr {
204 	struct ib_fmr		*fmr;
205 	u64			*physaddrs;
206 };
207 
208 struct rpcrdma_mw {
209 	union {
210 		struct rpcrdma_fmr	fmr;
211 		struct rpcrdma_frmr	frmr;
212 	} r;
213 	void			(*mw_sendcompletion)(struct ib_wc *);
214 	struct list_head	mw_list;
215 	struct list_head	mw_all;
216 };
217 
218 /*
219  * struct rpcrdma_req -- structure central to the request/reply sequence.
220  *
221  * N of these are associated with a transport instance, and stored in
222  * struct rpcrdma_buffer. N is the max number of outstanding requests.
223  *
224  * It includes pre-registered buffer memory for send AND recv.
225  * The recv buffer, however, is not owned by this structure, and
226  * is "donated" to the hardware when a recv is posted. When a
227  * reply is handled, the recv buffer used is given back to the
228  * struct rpcrdma_req associated with the request.
229  *
230  * In addition to the basic memory, this structure includes an array
231  * of iovs for send operations. The reason is that the iovs passed to
232  * ib_post_{send,recv} must not be modified until the work request
233  * completes.
234  *
235  * NOTES:
236  *   o RPCRDMA_MAX_SEGS is the max number of addressible chunk elements we
237  *     marshal. The number needed varies depending on the iov lists that
238  *     are passed to us, the memory registration mode we are in, and if
239  *     physical addressing is used, the layout.
240  */
241 
242 struct rpcrdma_mr_seg {		/* chunk descriptors */
243 	struct rpcrdma_mw *rl_mw;	/* registered MR */
244 	u64		mr_base;	/* registration result */
245 	u32		mr_rkey;	/* registration result */
246 	u32		mr_len;		/* length of chunk or segment */
247 	int		mr_nsegs;	/* number of segments in chunk or 0 */
248 	enum dma_data_direction	mr_dir;	/* segment mapping direction */
249 	dma_addr_t	mr_dma;		/* segment mapping address */
250 	size_t		mr_dmalen;	/* segment mapping length */
251 	struct page	*mr_page;	/* owning page, if any */
252 	char		*mr_offset;	/* kva if no page, else offset */
253 };
254 
255 #define RPCRDMA_MAX_IOVS	(2)
256 
257 struct rpcrdma_req {
258 	unsigned int		rl_niovs;
259 	unsigned int		rl_nchunks;
260 	unsigned int		rl_connect_cookie;
261 	struct rpcrdma_buffer	*rl_buffer;
262 	struct rpcrdma_rep	*rl_reply;/* holder for reply buffer */
263 	struct ib_sge		rl_send_iov[RPCRDMA_MAX_IOVS];
264 	struct rpcrdma_regbuf	*rl_rdmabuf;
265 	struct rpcrdma_regbuf	*rl_sendbuf;
266 	struct rpcrdma_mr_seg	rl_segments[RPCRDMA_MAX_SEGS];
267 };
268 
269 static inline struct rpcrdma_req *
270 rpcr_to_rdmar(struct rpc_rqst *rqst)
271 {
272 	void *buffer = rqst->rq_buffer;
273 	struct rpcrdma_regbuf *rb;
274 
275 	rb = container_of(buffer, struct rpcrdma_regbuf, rg_base);
276 	return rb->rg_owner;
277 }
278 
279 /*
280  * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
281  * inline requests/replies, and client/server credits.
282  *
283  * One of these is associated with a transport instance
284  */
285 struct rpcrdma_buffer {
286 	spinlock_t		rb_mwlock;	/* protect rb_mws list */
287 	struct list_head	rb_mws;
288 	struct list_head	rb_all;
289 	char			*rb_pool;
290 
291 	spinlock_t		rb_lock;	/* protect buf arrays */
292 	u32			rb_max_requests;
293 	int			rb_send_index;
294 	int			rb_recv_index;
295 	struct rpcrdma_req	**rb_send_bufs;
296 	struct rpcrdma_rep	**rb_recv_bufs;
297 };
298 #define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia)
299 
300 /*
301  * Internal structure for transport instance creation. This
302  * exists primarily for modularity.
303  *
304  * This data should be set with mount options
305  */
306 struct rpcrdma_create_data_internal {
307 	struct sockaddr_storage	addr;	/* RDMA server address */
308 	unsigned int	max_requests;	/* max requests (slots) in flight */
309 	unsigned int	rsize;		/* mount rsize - max read hdr+data */
310 	unsigned int	wsize;		/* mount wsize - max write hdr+data */
311 	unsigned int	inline_rsize;	/* max non-rdma read data payload */
312 	unsigned int	inline_wsize;	/* max non-rdma write data payload */
313 	unsigned int	padding;	/* non-rdma write header padding */
314 };
315 
316 #define RPCRDMA_INLINE_READ_THRESHOLD(rq) \
317 	(rpcx_to_rdmad(rq->rq_xprt).inline_rsize)
318 
319 #define RPCRDMA_INLINE_WRITE_THRESHOLD(rq)\
320 	(rpcx_to_rdmad(rq->rq_xprt).inline_wsize)
321 
322 #define RPCRDMA_INLINE_PAD_VALUE(rq)\
323 	rpcx_to_rdmad(rq->rq_xprt).padding
324 
325 /*
326  * Statistics for RPCRDMA
327  */
328 struct rpcrdma_stats {
329 	unsigned long		read_chunk_count;
330 	unsigned long		write_chunk_count;
331 	unsigned long		reply_chunk_count;
332 
333 	unsigned long long	total_rdma_request;
334 	unsigned long long	total_rdma_reply;
335 
336 	unsigned long long	pullup_copy_count;
337 	unsigned long long	fixup_copy_count;
338 	unsigned long		hardway_register_count;
339 	unsigned long		failed_marshal_count;
340 	unsigned long		bad_reply_count;
341 	unsigned long		nomsg_call_count;
342 };
343 
344 /*
345  * Per-registration mode operations
346  */
347 struct rpcrdma_xprt;
348 struct rpcrdma_memreg_ops {
349 	int		(*ro_map)(struct rpcrdma_xprt *,
350 				  struct rpcrdma_mr_seg *, int, bool);
351 	int		(*ro_unmap)(struct rpcrdma_xprt *,
352 				    struct rpcrdma_mr_seg *);
353 	int		(*ro_open)(struct rpcrdma_ia *,
354 				   struct rpcrdma_ep *,
355 				   struct rpcrdma_create_data_internal *);
356 	size_t		(*ro_maxpages)(struct rpcrdma_xprt *);
357 	int		(*ro_init)(struct rpcrdma_xprt *);
358 	void		(*ro_destroy)(struct rpcrdma_buffer *);
359 	const char	*ro_displayname;
360 };
361 
362 extern const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops;
363 extern const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops;
364 extern const struct rpcrdma_memreg_ops rpcrdma_physical_memreg_ops;
365 
366 /*
367  * RPCRDMA transport -- encapsulates the structures above for
368  * integration with RPC.
369  *
370  * The contained structures are embedded, not pointers,
371  * for convenience. This structure need not be visible externally.
372  *
373  * It is allocated and initialized during mount, and released
374  * during unmount.
375  */
376 struct rpcrdma_xprt {
377 	struct rpc_xprt		rx_xprt;
378 	struct rpcrdma_ia	rx_ia;
379 	struct rpcrdma_ep	rx_ep;
380 	struct rpcrdma_buffer	rx_buf;
381 	struct rpcrdma_create_data_internal rx_data;
382 	struct delayed_work	rx_connect_worker;
383 	struct rpcrdma_stats	rx_stats;
384 };
385 
386 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
387 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data)
388 
389 /* Setting this to 0 ensures interoperability with early servers.
390  * Setting this to 1 enhances certain unaligned read/write performance.
391  * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
392 extern int xprt_rdma_pad_optimize;
393 
394 /*
395  * Interface Adapter calls - xprtrdma/verbs.c
396  */
397 int rpcrdma_ia_open(struct rpcrdma_xprt *, struct sockaddr *, int);
398 void rpcrdma_ia_close(struct rpcrdma_ia *);
399 
400 /*
401  * Endpoint calls - xprtrdma/verbs.c
402  */
403 int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *,
404 				struct rpcrdma_create_data_internal *);
405 void rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *);
406 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *);
407 void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *);
408 
409 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
410 				struct rpcrdma_req *);
411 int rpcrdma_ep_post_recv(struct rpcrdma_ia *, struct rpcrdma_ep *,
412 				struct rpcrdma_rep *);
413 
414 /*
415  * Buffer calls - xprtrdma/verbs.c
416  */
417 int rpcrdma_buffer_create(struct rpcrdma_xprt *);
418 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
419 
420 struct rpcrdma_mw *rpcrdma_get_mw(struct rpcrdma_xprt *);
421 void rpcrdma_put_mw(struct rpcrdma_xprt *, struct rpcrdma_mw *);
422 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
423 void rpcrdma_buffer_put(struct rpcrdma_req *);
424 void rpcrdma_recv_buffer_get(struct rpcrdma_req *);
425 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
426 
427 struct rpcrdma_regbuf *rpcrdma_alloc_regbuf(struct rpcrdma_ia *,
428 					    size_t, gfp_t);
429 void rpcrdma_free_regbuf(struct rpcrdma_ia *,
430 			 struct rpcrdma_regbuf *);
431 
432 unsigned int rpcrdma_max_segments(struct rpcrdma_xprt *);
433 
434 int frwr_alloc_recovery_wq(void);
435 void frwr_destroy_recovery_wq(void);
436 
437 /*
438  * Wrappers for chunk registration, shared by read/write chunk code.
439  */
440 
441 void rpcrdma_mapping_error(struct rpcrdma_mr_seg *);
442 
443 static inline enum dma_data_direction
444 rpcrdma_data_dir(bool writing)
445 {
446 	return writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
447 }
448 
449 static inline void
450 rpcrdma_map_one(struct ib_device *device, struct rpcrdma_mr_seg *seg,
451 		enum dma_data_direction direction)
452 {
453 	seg->mr_dir = direction;
454 	seg->mr_dmalen = seg->mr_len;
455 
456 	if (seg->mr_page)
457 		seg->mr_dma = ib_dma_map_page(device,
458 				seg->mr_page, offset_in_page(seg->mr_offset),
459 				seg->mr_dmalen, seg->mr_dir);
460 	else
461 		seg->mr_dma = ib_dma_map_single(device,
462 				seg->mr_offset,
463 				seg->mr_dmalen, seg->mr_dir);
464 
465 	if (ib_dma_mapping_error(device, seg->mr_dma))
466 		rpcrdma_mapping_error(seg);
467 }
468 
469 static inline void
470 rpcrdma_unmap_one(struct ib_device *device, struct rpcrdma_mr_seg *seg)
471 {
472 	if (seg->mr_page)
473 		ib_dma_unmap_page(device,
474 				  seg->mr_dma, seg->mr_dmalen, seg->mr_dir);
475 	else
476 		ib_dma_unmap_single(device,
477 				    seg->mr_dma, seg->mr_dmalen, seg->mr_dir);
478 }
479 
480 /*
481  * RPC/RDMA connection management calls - xprtrdma/rpc_rdma.c
482  */
483 void rpcrdma_connect_worker(struct work_struct *);
484 void rpcrdma_conn_func(struct rpcrdma_ep *);
485 void rpcrdma_reply_handler(struct rpcrdma_rep *);
486 
487 /*
488  * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
489  */
490 int rpcrdma_marshal_req(struct rpc_rqst *);
491 
492 /* RPC/RDMA module init - xprtrdma/transport.c
493  */
494 int xprt_rdma_init(void);
495 void xprt_rdma_cleanup(void);
496 
497 /* Temporary NFS request map cache. Created in svc_rdma.c  */
498 extern struct kmem_cache *svc_rdma_map_cachep;
499 /* WR context cache. Created in svc_rdma.c  */
500 extern struct kmem_cache *svc_rdma_ctxt_cachep;
501 /* Workqueue created in svc_rdma.c */
502 extern struct workqueue_struct *svc_rdma_wq;
503 
504 #endif				/* _LINUX_SUNRPC_XPRT_RDMA_H */
505