xprtrdma: Allocate MRs on demand
[deliverable/linux.git] / net / sunrpc / xprtrdma / xprt_rdma.h
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 #define RPCRDMA_BIND_TO (60U * HZ)
59 #define RPCRDMA_INIT_REEST_TO (5U * HZ)
60 #define RPCRDMA_MAX_REEST_TO (30U * HZ)
61 #define RPCRDMA_IDLE_DISC_TO (5U * 60 * HZ)
62
63 /*
64 * Interface Adapter -- one per transport instance
65 */
66 struct rpcrdma_ia {
67 const struct rpcrdma_memreg_ops *ri_ops;
68 struct ib_device *ri_device;
69 struct rdma_cm_id *ri_id;
70 struct ib_pd *ri_pd;
71 struct completion ri_done;
72 int ri_async_rc;
73 unsigned int ri_max_frmr_depth;
74 unsigned int ri_max_inline_write;
75 unsigned int ri_max_inline_read;
76 struct ib_qp_attr ri_qp_attr;
77 struct ib_qp_init_attr ri_qp_init_attr;
78 };
79
80 /*
81 * RDMA Endpoint -- one per transport instance
82 */
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 };
94
95 #define INIT_CQCOUNT(ep) atomic_set(&(ep)->rep_cqcount, (ep)->rep_cqinit)
96 #define DECR_CQCOUNT(ep) atomic_sub_return(1, &(ep)->rep_cqcount)
97
98 /* Pre-allocate extra Work Requests for handling backward receives
99 * and sends. This is a fixed value because the Work Queues are
100 * allocated when the forward channel is set up.
101 */
102 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
103 #define RPCRDMA_BACKWARD_WRS (8)
104 #else
105 #define RPCRDMA_BACKWARD_WRS (0)
106 #endif
107
108 /* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
109 *
110 * The below structure appears at the front of a large region of kmalloc'd
111 * memory, which always starts on a good alignment boundary.
112 */
113
114 struct rpcrdma_regbuf {
115 size_t rg_size;
116 struct rpcrdma_req *rg_owner;
117 struct ib_sge rg_iov;
118 __be32 rg_base[0] __attribute__ ((aligned(256)));
119 };
120
121 static inline u64
122 rdmab_addr(struct rpcrdma_regbuf *rb)
123 {
124 return rb->rg_iov.addr;
125 }
126
127 static inline u32
128 rdmab_length(struct rpcrdma_regbuf *rb)
129 {
130 return rb->rg_iov.length;
131 }
132
133 static inline u32
134 rdmab_lkey(struct rpcrdma_regbuf *rb)
135 {
136 return rb->rg_iov.lkey;
137 }
138
139 static inline struct rpcrdma_msg *
140 rdmab_to_msg(struct rpcrdma_regbuf *rb)
141 {
142 return (struct rpcrdma_msg *)rb->rg_base;
143 }
144
145 #define RPCRDMA_DEF_GFP (GFP_NOIO | __GFP_NOWARN)
146
147 /* To ensure a transport can always make forward progress,
148 * the number of RDMA segments allowed in header chunk lists
149 * is capped at 8. This prevents less-capable devices and
150 * memory registrations from overrunning the Send buffer
151 * while building chunk lists.
152 *
153 * Elements of the Read list take up more room than the
154 * Write list or Reply chunk. 8 read segments means the Read
155 * list (or Write list or Reply chunk) cannot consume more
156 * than
157 *
158 * ((8 + 2) * read segment size) + 1 XDR words, or 244 bytes.
159 *
160 * And the fixed part of the header is another 24 bytes.
161 *
162 * The smallest inline threshold is 1024 bytes, ensuring that
163 * at least 750 bytes are available for RPC messages.
164 */
165 #define RPCRDMA_MAX_HDR_SEGS (8)
166
167 /*
168 * struct rpcrdma_rep -- this structure encapsulates state required to recv
169 * and complete a reply, asychronously. It needs several pieces of
170 * state:
171 * o recv buffer (posted to provider)
172 * o ib_sge (also donated to provider)
173 * o status of reply (length, success or not)
174 * o bookkeeping state to get run by tasklet (list, etc)
175 *
176 * These are allocated during initialization, per-transport instance;
177 * however, the tasklet execution list itself is global, as it should
178 * always be pretty short.
179 *
180 * N of these are associated with a transport instance, and stored in
181 * struct rpcrdma_buffer. N is the max number of outstanding requests.
182 */
183
184 #define RPCRDMA_MAX_DATA_SEGS ((1 * 1024 * 1024) / PAGE_SIZE)
185
186 /* data segments + head/tail for Call + head/tail for Reply */
187 #define RPCRDMA_MAX_SEGS (RPCRDMA_MAX_DATA_SEGS + 4)
188
189 struct rpcrdma_buffer;
190
191 struct rpcrdma_rep {
192 struct ib_cqe rr_cqe;
193 unsigned int rr_len;
194 struct ib_device *rr_device;
195 struct rpcrdma_xprt *rr_rxprt;
196 struct work_struct rr_work;
197 struct list_head rr_list;
198 struct rpcrdma_regbuf *rr_rdmabuf;
199 };
200
201 #define RPCRDMA_BAD_LEN (~0U)
202
203 /*
204 * struct rpcrdma_mw - external memory region metadata
205 *
206 * An external memory region is any buffer or page that is registered
207 * on the fly (ie, not pre-registered).
208 *
209 * Each rpcrdma_buffer has a list of free MWs anchored in rb_mws. During
210 * call_allocate, rpcrdma_buffer_get() assigns one to each segment in
211 * an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep
212 * track of registration metadata while each RPC is pending.
213 * rpcrdma_deregister_external() uses this metadata to unmap and
214 * release these resources when an RPC is complete.
215 */
216 enum rpcrdma_frmr_state {
217 FRMR_IS_INVALID, /* ready to be used */
218 FRMR_IS_VALID, /* in use */
219 FRMR_IS_STALE, /* failed completion */
220 };
221
222 struct rpcrdma_frmr {
223 struct ib_mr *fr_mr;
224 struct ib_cqe fr_cqe;
225 enum rpcrdma_frmr_state fr_state;
226 struct completion fr_linv_done;
227 union {
228 struct ib_reg_wr fr_regwr;
229 struct ib_send_wr fr_invwr;
230 };
231 };
232
233 struct rpcrdma_fmr {
234 struct ib_fmr *fm_mr;
235 u64 *fm_physaddrs;
236 };
237
238 struct rpcrdma_mw {
239 struct list_head mw_list;
240 struct scatterlist *mw_sg;
241 int mw_nents;
242 enum dma_data_direction mw_dir;
243 union {
244 struct rpcrdma_fmr fmr;
245 struct rpcrdma_frmr frmr;
246 };
247 struct rpcrdma_xprt *mw_xprt;
248 struct list_head mw_all;
249 };
250
251 /*
252 * struct rpcrdma_req -- structure central to the request/reply sequence.
253 *
254 * N of these are associated with a transport instance, and stored in
255 * struct rpcrdma_buffer. N is the max number of outstanding requests.
256 *
257 * It includes pre-registered buffer memory for send AND recv.
258 * The recv buffer, however, is not owned by this structure, and
259 * is "donated" to the hardware when a recv is posted. When a
260 * reply is handled, the recv buffer used is given back to the
261 * struct rpcrdma_req associated with the request.
262 *
263 * In addition to the basic memory, this structure includes an array
264 * of iovs for send operations. The reason is that the iovs passed to
265 * ib_post_{send,recv} must not be modified until the work request
266 * completes.
267 *
268 * NOTES:
269 * o RPCRDMA_MAX_SEGS is the max number of addressible chunk elements we
270 * marshal. The number needed varies depending on the iov lists that
271 * are passed to us and the memory registration mode we are in.
272 */
273
274 struct rpcrdma_mr_seg { /* chunk descriptors */
275 struct rpcrdma_mw *rl_mw; /* registered MR */
276 u64 mr_base; /* registration result */
277 u32 mr_rkey; /* registration result */
278 u32 mr_len; /* length of chunk or segment */
279 int mr_nsegs; /* number of segments in chunk or 0 */
280 struct page *mr_page; /* owning page, if any */
281 char *mr_offset; /* kva if no page, else offset */
282 };
283
284 #define RPCRDMA_MAX_IOVS (2)
285
286 struct rpcrdma_req {
287 struct list_head rl_free;
288 unsigned int rl_niovs;
289 unsigned int rl_nchunks;
290 unsigned int rl_connect_cookie;
291 struct rpc_task *rl_task;
292 struct rpcrdma_buffer *rl_buffer;
293 struct rpcrdma_rep *rl_reply;/* holder for reply buffer */
294 struct ib_sge rl_send_iov[RPCRDMA_MAX_IOVS];
295 struct rpcrdma_regbuf *rl_rdmabuf;
296 struct rpcrdma_regbuf *rl_sendbuf;
297 struct rpcrdma_mr_seg rl_segments[RPCRDMA_MAX_SEGS];
298 struct rpcrdma_mr_seg *rl_nextseg;
299
300 struct ib_cqe rl_cqe;
301 struct list_head rl_all;
302 bool rl_backchannel;
303 };
304
305 static inline struct rpcrdma_req *
306 rpcr_to_rdmar(struct rpc_rqst *rqst)
307 {
308 void *buffer = rqst->rq_buffer;
309 struct rpcrdma_regbuf *rb;
310
311 rb = container_of(buffer, struct rpcrdma_regbuf, rg_base);
312 return rb->rg_owner;
313 }
314
315 /*
316 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
317 * inline requests/replies, and client/server credits.
318 *
319 * One of these is associated with a transport instance
320 */
321 struct rpcrdma_buffer {
322 spinlock_t rb_mwlock; /* protect rb_mws list */
323 struct list_head rb_mws;
324 struct list_head rb_all;
325 char *rb_pool;
326
327 spinlock_t rb_lock; /* protect buf lists */
328 struct list_head rb_send_bufs;
329 struct list_head rb_recv_bufs;
330 u32 rb_max_requests;
331 atomic_t rb_credits; /* most recent credit grant */
332
333 u32 rb_bc_srv_max_requests;
334 spinlock_t rb_reqslock; /* protect rb_allreqs */
335 struct list_head rb_allreqs;
336
337 u32 rb_bc_max_requests;
338
339 spinlock_t rb_recovery_lock; /* protect rb_stale_mrs */
340 struct list_head rb_stale_mrs;
341 struct delayed_work rb_recovery_worker;
342 struct delayed_work rb_refresh_worker;
343 };
344 #define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia)
345
346 /*
347 * Internal structure for transport instance creation. This
348 * exists primarily for modularity.
349 *
350 * This data should be set with mount options
351 */
352 struct rpcrdma_create_data_internal {
353 struct sockaddr_storage addr; /* RDMA server address */
354 unsigned int max_requests; /* max requests (slots) in flight */
355 unsigned int rsize; /* mount rsize - max read hdr+data */
356 unsigned int wsize; /* mount wsize - max write hdr+data */
357 unsigned int inline_rsize; /* max non-rdma read data payload */
358 unsigned int inline_wsize; /* max non-rdma write data payload */
359 unsigned int padding; /* non-rdma write header padding */
360 };
361
362 #define RPCRDMA_INLINE_READ_THRESHOLD(rq) \
363 (rpcx_to_rdmad(rq->rq_xprt).inline_rsize)
364
365 #define RPCRDMA_INLINE_WRITE_THRESHOLD(rq)\
366 (rpcx_to_rdmad(rq->rq_xprt).inline_wsize)
367
368 #define RPCRDMA_INLINE_PAD_VALUE(rq)\
369 rpcx_to_rdmad(rq->rq_xprt).padding
370
371 /*
372 * Statistics for RPCRDMA
373 */
374 struct rpcrdma_stats {
375 unsigned long read_chunk_count;
376 unsigned long write_chunk_count;
377 unsigned long reply_chunk_count;
378
379 unsigned long long total_rdma_request;
380 unsigned long long total_rdma_reply;
381
382 unsigned long long pullup_copy_count;
383 unsigned long long fixup_copy_count;
384 unsigned long hardway_register_count;
385 unsigned long failed_marshal_count;
386 unsigned long bad_reply_count;
387 unsigned long nomsg_call_count;
388 unsigned long bcall_count;
389 unsigned long mrs_recovered;
390 unsigned long mrs_orphaned;
391 unsigned long mrs_allocated;
392 };
393
394 /*
395 * Per-registration mode operations
396 */
397 struct rpcrdma_xprt;
398 struct rpcrdma_memreg_ops {
399 int (*ro_map)(struct rpcrdma_xprt *,
400 struct rpcrdma_mr_seg *, int, bool);
401 void (*ro_unmap_sync)(struct rpcrdma_xprt *,
402 struct rpcrdma_req *);
403 void (*ro_unmap_safe)(struct rpcrdma_xprt *,
404 struct rpcrdma_req *, bool);
405 void (*ro_recover_mr)(struct rpcrdma_mw *);
406 int (*ro_open)(struct rpcrdma_ia *,
407 struct rpcrdma_ep *,
408 struct rpcrdma_create_data_internal *);
409 size_t (*ro_maxpages)(struct rpcrdma_xprt *);
410 int (*ro_init_mr)(struct rpcrdma_ia *,
411 struct rpcrdma_mw *);
412 void (*ro_release_mr)(struct rpcrdma_mw *);
413 const char *ro_displayname;
414 };
415
416 extern const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops;
417 extern const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops;
418
419 /*
420 * RPCRDMA transport -- encapsulates the structures above for
421 * integration with RPC.
422 *
423 * The contained structures are embedded, not pointers,
424 * for convenience. This structure need not be visible externally.
425 *
426 * It is allocated and initialized during mount, and released
427 * during unmount.
428 */
429 struct rpcrdma_xprt {
430 struct rpc_xprt rx_xprt;
431 struct rpcrdma_ia rx_ia;
432 struct rpcrdma_ep rx_ep;
433 struct rpcrdma_buffer rx_buf;
434 struct rpcrdma_create_data_internal rx_data;
435 struct delayed_work rx_connect_worker;
436 struct rpcrdma_stats rx_stats;
437 };
438
439 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
440 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data)
441
442 /* Setting this to 0 ensures interoperability with early servers.
443 * Setting this to 1 enhances certain unaligned read/write performance.
444 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
445 extern int xprt_rdma_pad_optimize;
446
447 /*
448 * Interface Adapter calls - xprtrdma/verbs.c
449 */
450 int rpcrdma_ia_open(struct rpcrdma_xprt *, struct sockaddr *, int);
451 void rpcrdma_ia_close(struct rpcrdma_ia *);
452 bool frwr_is_supported(struct rpcrdma_ia *);
453 bool fmr_is_supported(struct rpcrdma_ia *);
454
455 /*
456 * Endpoint calls - xprtrdma/verbs.c
457 */
458 int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *,
459 struct rpcrdma_create_data_internal *);
460 void rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *);
461 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *);
462 void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *);
463
464 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
465 struct rpcrdma_req *);
466 int rpcrdma_ep_post_recv(struct rpcrdma_ia *, struct rpcrdma_ep *,
467 struct rpcrdma_rep *);
468
469 /*
470 * Buffer calls - xprtrdma/verbs.c
471 */
472 struct rpcrdma_req *rpcrdma_create_req(struct rpcrdma_xprt *);
473 struct rpcrdma_rep *rpcrdma_create_rep(struct rpcrdma_xprt *);
474 void rpcrdma_destroy_req(struct rpcrdma_ia *, struct rpcrdma_req *);
475 int rpcrdma_buffer_create(struct rpcrdma_xprt *);
476 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
477
478 struct rpcrdma_mw *rpcrdma_get_mw(struct rpcrdma_xprt *);
479 void rpcrdma_put_mw(struct rpcrdma_xprt *, struct rpcrdma_mw *);
480 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
481 void rpcrdma_buffer_put(struct rpcrdma_req *);
482 void rpcrdma_recv_buffer_get(struct rpcrdma_req *);
483 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
484
485 void rpcrdma_defer_mr_recovery(struct rpcrdma_mw *);
486
487 struct rpcrdma_regbuf *rpcrdma_alloc_regbuf(struct rpcrdma_ia *,
488 size_t, gfp_t);
489 void rpcrdma_free_regbuf(struct rpcrdma_ia *,
490 struct rpcrdma_regbuf *);
491
492 int rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt *, unsigned int);
493
494 int rpcrdma_alloc_wq(void);
495 void rpcrdma_destroy_wq(void);
496
497 /*
498 * Wrappers for chunk registration, shared by read/write chunk code.
499 */
500
501 static inline enum dma_data_direction
502 rpcrdma_data_dir(bool writing)
503 {
504 return writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
505 }
506
507 /*
508 * RPC/RDMA connection management calls - xprtrdma/rpc_rdma.c
509 */
510 void rpcrdma_connect_worker(struct work_struct *);
511 void rpcrdma_conn_func(struct rpcrdma_ep *);
512 void rpcrdma_reply_handler(struct rpcrdma_rep *);
513
514 /*
515 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
516 */
517 int rpcrdma_marshal_req(struct rpc_rqst *);
518 void rpcrdma_set_max_header_sizes(struct rpcrdma_ia *,
519 struct rpcrdma_create_data_internal *,
520 unsigned int);
521
522 /* RPC/RDMA module init - xprtrdma/transport.c
523 */
524 extern unsigned int xprt_rdma_max_inline_read;
525 void xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap);
526 void xprt_rdma_free_addresses(struct rpc_xprt *xprt);
527 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq);
528 int xprt_rdma_init(void);
529 void xprt_rdma_cleanup(void);
530
531 /* Backchannel calls - xprtrdma/backchannel.c
532 */
533 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
534 int xprt_rdma_bc_setup(struct rpc_xprt *, unsigned int);
535 int xprt_rdma_bc_up(struct svc_serv *, struct net *);
536 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *);
537 int rpcrdma_bc_post_recv(struct rpcrdma_xprt *, unsigned int);
538 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *, struct rpcrdma_rep *);
539 int rpcrdma_bc_marshal_reply(struct rpc_rqst *);
540 void xprt_rdma_bc_free_rqst(struct rpc_rqst *);
541 void xprt_rdma_bc_destroy(struct rpc_xprt *, unsigned int);
542 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
543
544 extern struct xprt_class xprt_rdma_bc;
545
546 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */
This page took 0.042401 seconds and 5 git commands to generate.