xref: /openbmc/linux/net/sunrpc/xprtrdma/xprt_rdma.h (revision 8b235f2f)
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 	u32			ri_dma_lkey;
69 	struct completion	ri_done;
70 	int			ri_async_rc;
71 	unsigned int		ri_max_frmr_depth;
72 	struct ib_device_attr	ri_devattr;
73 	struct ib_qp_attr	ri_qp_attr;
74 	struct ib_qp_init_attr	ri_qp_init_attr;
75 };
76 
77 /*
78  * RDMA Endpoint -- one per transport instance
79  */
80 
81 #define RPCRDMA_WC_BUDGET	(128)
82 #define RPCRDMA_POLLSIZE	(16)
83 
84 struct rpcrdma_ep {
85 	atomic_t		rep_cqcount;
86 	int			rep_cqinit;
87 	int			rep_connected;
88 	struct ib_qp_init_attr	rep_attr;
89 	wait_queue_head_t 	rep_connect_wait;
90 	struct rdma_conn_param	rep_remote_cma;
91 	struct sockaddr_storage	rep_remote_addr;
92 	struct delayed_work	rep_connect_worker;
93 	struct ib_wc		rep_send_wcs[RPCRDMA_POLLSIZE];
94 	struct ib_wc		rep_recv_wcs[RPCRDMA_POLLSIZE];
95 };
96 
97 /*
98  * Force a signaled SEND Work Request every so often,
99  * in case the provider needs to do some housekeeping.
100  */
101 #define RPCRDMA_MAX_UNSIGNALED_SENDS	(32)
102 
103 #define INIT_CQCOUNT(ep) atomic_set(&(ep)->rep_cqcount, (ep)->rep_cqinit)
104 #define DECR_CQCOUNT(ep) atomic_sub_return(1, &(ep)->rep_cqcount)
105 
106 /* Force completion handler to ignore the signal
107  */
108 #define RPCRDMA_IGNORE_COMPLETION	(0ULL)
109 
110 /* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
111  *
112  * The below structure appears at the front of a large region of kmalloc'd
113  * memory, which always starts on a good alignment boundary.
114  */
115 
116 struct rpcrdma_regbuf {
117 	size_t			rg_size;
118 	struct rpcrdma_req	*rg_owner;
119 	struct ib_sge		rg_iov;
120 	__be32			rg_base[0] __attribute__ ((aligned(256)));
121 };
122 
123 static inline u64
124 rdmab_addr(struct rpcrdma_regbuf *rb)
125 {
126 	return rb->rg_iov.addr;
127 }
128 
129 static inline u32
130 rdmab_length(struct rpcrdma_regbuf *rb)
131 {
132 	return rb->rg_iov.length;
133 }
134 
135 static inline u32
136 rdmab_lkey(struct rpcrdma_regbuf *rb)
137 {
138 	return rb->rg_iov.lkey;
139 }
140 
141 static inline struct rpcrdma_msg *
142 rdmab_to_msg(struct rpcrdma_regbuf *rb)
143 {
144 	return (struct rpcrdma_msg *)rb->rg_base;
145 }
146 
147 /*
148  * struct rpcrdma_rep -- this structure encapsulates state required to recv
149  * and complete a reply, asychronously. It needs several pieces of
150  * state:
151  *   o recv buffer (posted to provider)
152  *   o ib_sge (also donated to provider)
153  *   o status of reply (length, success or not)
154  *   o bookkeeping state to get run by tasklet (list, etc)
155  *
156  * These are allocated during initialization, per-transport instance;
157  * however, the tasklet execution list itself is global, as it should
158  * always be pretty short.
159  *
160  * N of these are associated with a transport instance, and stored in
161  * struct rpcrdma_buffer. N is the max number of outstanding requests.
162  */
163 
164 #define RPCRDMA_MAX_DATA_SEGS	((1 * 1024 * 1024) / PAGE_SIZE)
165 #define RPCRDMA_MAX_SEGS 	(RPCRDMA_MAX_DATA_SEGS + 2) /* head+tail = 2 */
166 
167 struct rpcrdma_buffer;
168 
169 struct rpcrdma_rep {
170 	unsigned int		rr_len;
171 	struct ib_device	*rr_device;
172 	struct rpcrdma_xprt	*rr_rxprt;
173 	struct list_head	rr_list;
174 	struct rpcrdma_regbuf	*rr_rdmabuf;
175 };
176 
177 /*
178  * struct rpcrdma_mw - external memory region metadata
179  *
180  * An external memory region is any buffer or page that is registered
181  * on the fly (ie, not pre-registered).
182  *
183  * Each rpcrdma_buffer has a list of free MWs anchored in rb_mws. During
184  * call_allocate, rpcrdma_buffer_get() assigns one to each segment in
185  * an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep
186  * track of registration metadata while each RPC is pending.
187  * rpcrdma_deregister_external() uses this metadata to unmap and
188  * release these resources when an RPC is complete.
189  */
190 enum rpcrdma_frmr_state {
191 	FRMR_IS_INVALID,	/* ready to be used */
192 	FRMR_IS_VALID,		/* in use */
193 	FRMR_IS_STALE,		/* failed completion */
194 };
195 
196 struct rpcrdma_frmr {
197 	struct ib_fast_reg_page_list	*fr_pgl;
198 	struct ib_mr			*fr_mr;
199 	enum rpcrdma_frmr_state		fr_state;
200 	struct work_struct		fr_work;
201 	struct rpcrdma_xprt		*fr_xprt;
202 };
203 
204 struct rpcrdma_fmr {
205 	struct ib_fmr		*fmr;
206 	u64			*physaddrs;
207 };
208 
209 struct rpcrdma_mw {
210 	union {
211 		struct rpcrdma_fmr	fmr;
212 		struct rpcrdma_frmr	frmr;
213 	} r;
214 	void			(*mw_sendcompletion)(struct ib_wc *);
215 	struct list_head	mw_list;
216 	struct list_head	mw_all;
217 };
218 
219 /*
220  * struct rpcrdma_req -- structure central to the request/reply sequence.
221  *
222  * N of these are associated with a transport instance, and stored in
223  * struct rpcrdma_buffer. N is the max number of outstanding requests.
224  *
225  * It includes pre-registered buffer memory for send AND recv.
226  * The recv buffer, however, is not owned by this structure, and
227  * is "donated" to the hardware when a recv is posted. When a
228  * reply is handled, the recv buffer used is given back to the
229  * struct rpcrdma_req associated with the request.
230  *
231  * In addition to the basic memory, this structure includes an array
232  * of iovs for send operations. The reason is that the iovs passed to
233  * ib_post_{send,recv} must not be modified until the work request
234  * completes.
235  *
236  * NOTES:
237  *   o RPCRDMA_MAX_SEGS is the max number of addressible chunk elements we
238  *     marshal. The number needed varies depending on the iov lists that
239  *     are passed to us, the memory registration mode we are in, and if
240  *     physical addressing is used, the layout.
241  */
242 
243 struct rpcrdma_mr_seg {		/* chunk descriptors */
244 	struct rpcrdma_mw *rl_mw;	/* registered MR */
245 	u64		mr_base;	/* registration result */
246 	u32		mr_rkey;	/* registration result */
247 	u32		mr_len;		/* length of chunk or segment */
248 	int		mr_nsegs;	/* number of segments in chunk or 0 */
249 	enum dma_data_direction	mr_dir;	/* segment mapping direction */
250 	dma_addr_t	mr_dma;		/* segment mapping address */
251 	size_t		mr_dmalen;	/* segment mapping length */
252 	struct page	*mr_page;	/* owning page, if any */
253 	char		*mr_offset;	/* kva if no page, else offset */
254 };
255 
256 #define RPCRDMA_MAX_IOVS	(2)
257 
258 struct rpcrdma_req {
259 	unsigned int		rl_niovs;
260 	unsigned int		rl_nchunks;
261 	unsigned int		rl_connect_cookie;
262 	struct rpcrdma_buffer	*rl_buffer;
263 	struct rpcrdma_rep	*rl_reply;/* holder for reply buffer */
264 	struct ib_sge		rl_send_iov[RPCRDMA_MAX_IOVS];
265 	struct rpcrdma_regbuf	*rl_rdmabuf;
266 	struct rpcrdma_regbuf	*rl_sendbuf;
267 	struct rpcrdma_mr_seg	rl_segments[RPCRDMA_MAX_SEGS];
268 };
269 
270 static inline struct rpcrdma_req *
271 rpcr_to_rdmar(struct rpc_rqst *rqst)
272 {
273 	void *buffer = rqst->rq_buffer;
274 	struct rpcrdma_regbuf *rb;
275 
276 	rb = container_of(buffer, struct rpcrdma_regbuf, rg_base);
277 	return rb->rg_owner;
278 }
279 
280 /*
281  * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
282  * inline requests/replies, and client/server credits.
283  *
284  * One of these is associated with a transport instance
285  */
286 struct rpcrdma_buffer {
287 	spinlock_t		rb_mwlock;	/* protect rb_mws list */
288 	struct list_head	rb_mws;
289 	struct list_head	rb_all;
290 	char			*rb_pool;
291 
292 	spinlock_t		rb_lock;	/* protect buf arrays */
293 	u32			rb_max_requests;
294 	int			rb_send_index;
295 	int			rb_recv_index;
296 	struct rpcrdma_req	**rb_send_bufs;
297 	struct rpcrdma_rep	**rb_recv_bufs;
298 };
299 #define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia)
300 
301 /*
302  * Internal structure for transport instance creation. This
303  * exists primarily for modularity.
304  *
305  * This data should be set with mount options
306  */
307 struct rpcrdma_create_data_internal {
308 	struct sockaddr_storage	addr;	/* RDMA server address */
309 	unsigned int	max_requests;	/* max requests (slots) in flight */
310 	unsigned int	rsize;		/* mount rsize - max read hdr+data */
311 	unsigned int	wsize;		/* mount wsize - max write hdr+data */
312 	unsigned int	inline_rsize;	/* max non-rdma read data payload */
313 	unsigned int	inline_wsize;	/* max non-rdma write data payload */
314 	unsigned int	padding;	/* non-rdma write header padding */
315 };
316 
317 #define RPCRDMA_INLINE_READ_THRESHOLD(rq) \
318 	(rpcx_to_rdmad(rq->rq_xprt).inline_rsize)
319 
320 #define RPCRDMA_INLINE_WRITE_THRESHOLD(rq)\
321 	(rpcx_to_rdmad(rq->rq_xprt).inline_wsize)
322 
323 #define RPCRDMA_INLINE_PAD_VALUE(rq)\
324 	rpcx_to_rdmad(rq->rq_xprt).padding
325 
326 /*
327  * Statistics for RPCRDMA
328  */
329 struct rpcrdma_stats {
330 	unsigned long		read_chunk_count;
331 	unsigned long		write_chunk_count;
332 	unsigned long		reply_chunk_count;
333 
334 	unsigned long long	total_rdma_request;
335 	unsigned long long	total_rdma_reply;
336 
337 	unsigned long long	pullup_copy_count;
338 	unsigned long long	fixup_copy_count;
339 	unsigned long		hardway_register_count;
340 	unsigned long		failed_marshal_count;
341 	unsigned long		bad_reply_count;
342 	unsigned long		nomsg_call_count;
343 };
344 
345 /*
346  * Per-registration mode operations
347  */
348 struct rpcrdma_xprt;
349 struct rpcrdma_memreg_ops {
350 	int		(*ro_map)(struct rpcrdma_xprt *,
351 				  struct rpcrdma_mr_seg *, int, bool);
352 	int		(*ro_unmap)(struct rpcrdma_xprt *,
353 				    struct rpcrdma_mr_seg *);
354 	int		(*ro_open)(struct rpcrdma_ia *,
355 				   struct rpcrdma_ep *,
356 				   struct rpcrdma_create_data_internal *);
357 	size_t		(*ro_maxpages)(struct rpcrdma_xprt *);
358 	int		(*ro_init)(struct rpcrdma_xprt *);
359 	void		(*ro_destroy)(struct rpcrdma_buffer *);
360 	const char	*ro_displayname;
361 };
362 
363 extern const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops;
364 extern const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops;
365 extern const struct rpcrdma_memreg_ops rpcrdma_physical_memreg_ops;
366 
367 /*
368  * RPCRDMA transport -- encapsulates the structures above for
369  * integration with RPC.
370  *
371  * The contained structures are embedded, not pointers,
372  * for convenience. This structure need not be visible externally.
373  *
374  * It is allocated and initialized during mount, and released
375  * during unmount.
376  */
377 struct rpcrdma_xprt {
378 	struct rpc_xprt		rx_xprt;
379 	struct rpcrdma_ia	rx_ia;
380 	struct rpcrdma_ep	rx_ep;
381 	struct rpcrdma_buffer	rx_buf;
382 	struct rpcrdma_create_data_internal rx_data;
383 	struct delayed_work	rx_connect_worker;
384 	struct rpcrdma_stats	rx_stats;
385 };
386 
387 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
388 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data)
389 
390 /* Setting this to 0 ensures interoperability with early servers.
391  * Setting this to 1 enhances certain unaligned read/write performance.
392  * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
393 extern int xprt_rdma_pad_optimize;
394 
395 /*
396  * Interface Adapter calls - xprtrdma/verbs.c
397  */
398 int rpcrdma_ia_open(struct rpcrdma_xprt *, struct sockaddr *, int);
399 void rpcrdma_ia_close(struct rpcrdma_ia *);
400 
401 /*
402  * Endpoint calls - xprtrdma/verbs.c
403  */
404 int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *,
405 				struct rpcrdma_create_data_internal *);
406 void rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *);
407 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *);
408 void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *);
409 
410 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
411 				struct rpcrdma_req *);
412 int rpcrdma_ep_post_recv(struct rpcrdma_ia *, struct rpcrdma_ep *,
413 				struct rpcrdma_rep *);
414 
415 /*
416  * Buffer calls - xprtrdma/verbs.c
417  */
418 int rpcrdma_buffer_create(struct rpcrdma_xprt *);
419 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
420 
421 struct rpcrdma_mw *rpcrdma_get_mw(struct rpcrdma_xprt *);
422 void rpcrdma_put_mw(struct rpcrdma_xprt *, struct rpcrdma_mw *);
423 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
424 void rpcrdma_buffer_put(struct rpcrdma_req *);
425 void rpcrdma_recv_buffer_get(struct rpcrdma_req *);
426 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
427 
428 struct rpcrdma_regbuf *rpcrdma_alloc_regbuf(struct rpcrdma_ia *,
429 					    size_t, gfp_t);
430 void rpcrdma_free_regbuf(struct rpcrdma_ia *,
431 			 struct rpcrdma_regbuf *);
432 
433 unsigned int rpcrdma_max_segments(struct rpcrdma_xprt *);
434 
435 int frwr_alloc_recovery_wq(void);
436 void frwr_destroy_recovery_wq(void);
437 
438 /*
439  * Wrappers for chunk registration, shared by read/write chunk code.
440  */
441 
442 void rpcrdma_mapping_error(struct rpcrdma_mr_seg *);
443 
444 static inline enum dma_data_direction
445 rpcrdma_data_dir(bool writing)
446 {
447 	return writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
448 }
449 
450 static inline void
451 rpcrdma_map_one(struct ib_device *device, struct rpcrdma_mr_seg *seg,
452 		enum dma_data_direction direction)
453 {
454 	seg->mr_dir = direction;
455 	seg->mr_dmalen = seg->mr_len;
456 
457 	if (seg->mr_page)
458 		seg->mr_dma = ib_dma_map_page(device,
459 				seg->mr_page, offset_in_page(seg->mr_offset),
460 				seg->mr_dmalen, seg->mr_dir);
461 	else
462 		seg->mr_dma = ib_dma_map_single(device,
463 				seg->mr_offset,
464 				seg->mr_dmalen, seg->mr_dir);
465 
466 	if (ib_dma_mapping_error(device, seg->mr_dma))
467 		rpcrdma_mapping_error(seg);
468 }
469 
470 static inline void
471 rpcrdma_unmap_one(struct ib_device *device, struct rpcrdma_mr_seg *seg)
472 {
473 	if (seg->mr_page)
474 		ib_dma_unmap_page(device,
475 				  seg->mr_dma, seg->mr_dmalen, seg->mr_dir);
476 	else
477 		ib_dma_unmap_single(device,
478 				    seg->mr_dma, seg->mr_dmalen, seg->mr_dir);
479 }
480 
481 /*
482  * RPC/RDMA connection management calls - xprtrdma/rpc_rdma.c
483  */
484 void rpcrdma_connect_worker(struct work_struct *);
485 void rpcrdma_conn_func(struct rpcrdma_ep *);
486 void rpcrdma_reply_handler(struct rpcrdma_rep *);
487 
488 /*
489  * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
490  */
491 int rpcrdma_marshal_req(struct rpc_rqst *);
492 
493 /* RPC/RDMA module init - xprtrdma/transport.c
494  */
495 int xprt_rdma_init(void);
496 void xprt_rdma_cleanup(void);
497 
498 /* Temporary NFS request map cache. Created in svc_rdma.c  */
499 extern struct kmem_cache *svc_rdma_map_cachep;
500 /* WR context cache. Created in svc_rdma.c  */
501 extern struct kmem_cache *svc_rdma_ctxt_cachep;
502 /* Workqueue created in svc_rdma.c */
503 extern struct workqueue_struct *svc_rdma_wq;
504 
505 #endif				/* _LINUX_SUNRPC_XPRT_RDMA_H */
506