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 */
55 #define RDMA_RESOLVE_TIMEOUT (5000) /* 5 seconds */
56 #define RDMA_CONNECT_RETRY_MAX (2) /* retries if no listener backlog */
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)
64 * Interface Adapter -- one per transport instance
67 const struct rpcrdma_memreg_ops *ri_ops;
68 struct ib_device *ri_device;
69 struct rdma_cm_id *ri_id;
71 struct completion ri_done;
73 unsigned int ri_max_segs;
74 unsigned int ri_max_frmr_depth;
75 unsigned int ri_max_inline_write;
76 unsigned int ri_max_inline_read;
77 bool ri_reminv_expected;
78 struct ib_qp_attr ri_qp_attr;
79 struct ib_qp_init_attr ri_qp_init_attr;
83 * RDMA Endpoint -- one per transport instance
90 struct ib_qp_init_attr rep_attr;
91 wait_queue_head_t rep_connect_wait;
92 struct rpcrdma_connect_private rep_cm_private;
93 struct rdma_conn_param rep_remote_cma;
94 struct sockaddr_storage rep_remote_addr;
95 struct delayed_work rep_connect_worker;
98 #define INIT_CQCOUNT(ep) atomic_set(&(ep)->rep_cqcount, (ep)->rep_cqinit)
99 #define DECR_CQCOUNT(ep) atomic_sub_return(1, &(ep)->rep_cqcount)
101 /* Pre-allocate extra Work Requests for handling backward receives
102 * and sends. This is a fixed value because the Work Queues are
103 * allocated when the forward channel is set up.
105 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
106 #define RPCRDMA_BACKWARD_WRS (8)
108 #define RPCRDMA_BACKWARD_WRS (0)
111 /* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
113 * The below structure appears at the front of a large region of kmalloc'd
114 * memory, which always starts on a good alignment boundary.
117 struct rpcrdma_regbuf {
118 struct ib_sge rg_iov;
119 struct ib_device *rg_device;
120 enum dma_data_direction rg_direction;
121 __be32 rg_base[0] __attribute__ ((aligned(256)));
125 rdmab_addr(struct rpcrdma_regbuf *rb)
127 return rb->rg_iov.addr;
131 rdmab_length(struct rpcrdma_regbuf *rb)
133 return rb->rg_iov.length;
137 rdmab_lkey(struct rpcrdma_regbuf *rb)
139 return rb->rg_iov.lkey;
142 static inline struct rpcrdma_msg *
143 rdmab_to_msg(struct rpcrdma_regbuf *rb)
145 return (struct rpcrdma_msg *)rb->rg_base;
148 #define RPCRDMA_DEF_GFP (GFP_NOIO | __GFP_NOWARN)
150 /* To ensure a transport can always make forward progress,
151 * the number of RDMA segments allowed in header chunk lists
152 * is capped at 8. This prevents less-capable devices and
153 * memory registrations from overrunning the Send buffer
154 * while building chunk lists.
156 * Elements of the Read list take up more room than the
157 * Write list or Reply chunk. 8 read segments means the Read
158 * list (or Write list or Reply chunk) cannot consume more
161 * ((8 + 2) * read segment size) + 1 XDR words, or 244 bytes.
163 * And the fixed part of the header is another 24 bytes.
165 * The smallest inline threshold is 1024 bytes, ensuring that
166 * at least 750 bytes are available for RPC messages.
169 RPCRDMA_MAX_HDR_SEGS = 8,
170 RPCRDMA_HDRBUF_SIZE = 256,
174 * struct rpcrdma_rep -- this structure encapsulates state required to recv
175 * and complete a reply, asychronously. It needs several pieces of
177 * o recv buffer (posted to provider)
178 * o ib_sge (also donated to provider)
179 * o status of reply (length, success or not)
180 * o bookkeeping state to get run by reply handler (list, etc)
182 * These are allocated during initialization, per-transport instance.
184 * N of these are associated with a transport instance, and stored in
185 * struct rpcrdma_buffer. N is the max number of outstanding requests.
189 struct ib_cqe rr_cqe;
193 struct ib_device *rr_device;
194 struct rpcrdma_xprt *rr_rxprt;
195 struct work_struct rr_work;
196 struct list_head rr_list;
197 struct ib_recv_wr rr_recv_wr;
198 struct rpcrdma_regbuf *rr_rdmabuf;
201 #define RPCRDMA_BAD_LEN (~0U)
204 * struct rpcrdma_mw - external memory region metadata
206 * An external memory region is any buffer or page that is registered
207 * on the fly (ie, not pre-registered).
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.
216 enum rpcrdma_frmr_state {
217 FRMR_IS_INVALID, /* ready to be used */
218 FRMR_IS_VALID, /* in use */
219 FRMR_FLUSHED_FR, /* flushed FASTREG WR */
220 FRMR_FLUSHED_LI, /* flushed LOCALINV WR */
223 struct rpcrdma_frmr {
225 struct ib_cqe fr_cqe;
226 enum rpcrdma_frmr_state fr_state;
227 struct completion fr_linv_done;
229 struct ib_reg_wr fr_regwr;
230 struct ib_send_wr fr_invwr;
235 struct ib_fmr *fm_mr;
240 struct list_head mw_list;
241 struct scatterlist *mw_sg;
243 enum dma_data_direction mw_dir;
245 struct rpcrdma_fmr fmr;
246 struct rpcrdma_frmr frmr;
248 struct rpcrdma_xprt *mw_xprt;
252 struct list_head mw_all;
256 * struct rpcrdma_req -- structure central to the request/reply sequence.
258 * N of these are associated with a transport instance, and stored in
259 * struct rpcrdma_buffer. N is the max number of outstanding requests.
261 * It includes pre-registered buffer memory for send AND recv.
262 * The recv buffer, however, is not owned by this structure, and
263 * is "donated" to the hardware when a recv is posted. When a
264 * reply is handled, the recv buffer used is given back to the
265 * struct rpcrdma_req associated with the request.
267 * In addition to the basic memory, this structure includes an array
268 * of iovs for send operations. The reason is that the iovs passed to
269 * ib_post_{send,recv} must not be modified until the work request
273 /* Maximum number of page-sized "segments" per chunk list to be
274 * registered or invalidated. Must handle a Reply chunk:
277 RPCRDMA_MAX_IOV_SEGS = 3,
278 RPCRDMA_MAX_DATA_SEGS = ((1 * 1024 * 1024) / PAGE_SIZE) + 1,
279 RPCRDMA_MAX_SEGS = RPCRDMA_MAX_DATA_SEGS +
280 RPCRDMA_MAX_IOV_SEGS,
283 struct rpcrdma_mr_seg { /* chunk descriptors */
284 u32 mr_len; /* length of chunk or segment */
285 struct page *mr_page; /* owning page, if any */
286 char *mr_offset; /* kva if no page, else offset */
289 /* Reserve enough Send SGEs to send a maximum size inline request:
290 * - RPC-over-RDMA header
291 * - xdr_buf head iovec
292 * - RPCRDMA_MAX_INLINE bytes, possibly unaligned, in pages
293 * - xdr_buf tail iovec
296 RPCRDMA_MAX_SEND_PAGES = PAGE_SIZE + RPCRDMA_MAX_INLINE - 1,
297 RPCRDMA_MAX_PAGE_SGES = (RPCRDMA_MAX_SEND_PAGES >> PAGE_SHIFT) + 1,
298 RPCRDMA_MAX_SEND_SGES = 1 + 1 + RPCRDMA_MAX_PAGE_SGES + 1,
301 struct rpcrdma_buffer;
303 struct list_head rl_free;
304 unsigned int rl_mapped_sges;
305 unsigned int rl_connect_cookie;
306 struct rpcrdma_buffer *rl_buffer;
307 struct rpcrdma_rep *rl_reply;
308 struct ib_send_wr rl_send_wr;
309 struct ib_sge rl_send_sge[RPCRDMA_MAX_SEND_SGES];
310 struct rpcrdma_regbuf *rl_rdmabuf; /* xprt header */
311 struct rpcrdma_regbuf *rl_sendbuf; /* rq_snd_buf */
312 struct rpcrdma_regbuf *rl_recvbuf; /* rq_rcv_buf */
314 struct ib_cqe rl_cqe;
315 struct list_head rl_all;
318 struct list_head rl_registered; /* registered segments */
319 struct rpcrdma_mr_seg rl_segments[RPCRDMA_MAX_SEGS];
323 rpcrdma_set_xprtdata(struct rpc_rqst *rqst, struct rpcrdma_req *req)
325 rqst->rq_xprtdata = req;
328 static inline struct rpcrdma_req *
329 rpcr_to_rdmar(struct rpc_rqst *rqst)
331 return rqst->rq_xprtdata;
335 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
336 * inline requests/replies, and client/server credits.
338 * One of these is associated with a transport instance
340 struct rpcrdma_buffer {
341 spinlock_t rb_mwlock; /* protect rb_mws list */
342 struct list_head rb_mws;
343 struct list_head rb_all;
346 spinlock_t rb_lock; /* protect buf lists */
347 int rb_send_count, rb_recv_count;
348 struct list_head rb_send_bufs;
349 struct list_head rb_recv_bufs;
351 atomic_t rb_credits; /* most recent credit grant */
353 u32 rb_bc_srv_max_requests;
354 spinlock_t rb_reqslock; /* protect rb_allreqs */
355 struct list_head rb_allreqs;
357 u32 rb_bc_max_requests;
359 spinlock_t rb_recovery_lock; /* protect rb_stale_mrs */
360 struct list_head rb_stale_mrs;
361 struct delayed_work rb_recovery_worker;
362 struct delayed_work rb_refresh_worker;
364 #define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia)
367 * Internal structure for transport instance creation. This
368 * exists primarily for modularity.
370 * This data should be set with mount options
372 struct rpcrdma_create_data_internal {
373 struct sockaddr_storage addr; /* RDMA server address */
374 unsigned int max_requests; /* max requests (slots) in flight */
375 unsigned int rsize; /* mount rsize - max read hdr+data */
376 unsigned int wsize; /* mount wsize - max write hdr+data */
377 unsigned int inline_rsize; /* max non-rdma read data payload */
378 unsigned int inline_wsize; /* max non-rdma write data payload */
379 unsigned int padding; /* non-rdma write header padding */
383 * Statistics for RPCRDMA
385 struct rpcrdma_stats {
386 unsigned long read_chunk_count;
387 unsigned long write_chunk_count;
388 unsigned long reply_chunk_count;
390 unsigned long long total_rdma_request;
391 unsigned long long total_rdma_reply;
393 unsigned long long pullup_copy_count;
394 unsigned long long fixup_copy_count;
395 unsigned long hardway_register_count;
396 unsigned long failed_marshal_count;
397 unsigned long bad_reply_count;
398 unsigned long nomsg_call_count;
399 unsigned long bcall_count;
400 unsigned long mrs_recovered;
401 unsigned long mrs_orphaned;
402 unsigned long mrs_allocated;
403 unsigned long local_inv_needed;
407 * Per-registration mode operations
410 struct rpcrdma_memreg_ops {
411 int (*ro_map)(struct rpcrdma_xprt *,
412 struct rpcrdma_mr_seg *, int, bool,
413 struct rpcrdma_mw **);
414 void (*ro_unmap_sync)(struct rpcrdma_xprt *,
415 struct rpcrdma_req *);
416 void (*ro_unmap_safe)(struct rpcrdma_xprt *,
417 struct rpcrdma_req *, bool);
418 void (*ro_recover_mr)(struct rpcrdma_mw *);
419 int (*ro_open)(struct rpcrdma_ia *,
421 struct rpcrdma_create_data_internal *);
422 size_t (*ro_maxpages)(struct rpcrdma_xprt *);
423 int (*ro_init_mr)(struct rpcrdma_ia *,
424 struct rpcrdma_mw *);
425 void (*ro_release_mr)(struct rpcrdma_mw *);
426 const char *ro_displayname;
427 const int ro_send_w_inv_ok;
430 extern const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops;
431 extern const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops;
434 * RPCRDMA transport -- encapsulates the structures above for
435 * integration with RPC.
437 * The contained structures are embedded, not pointers,
438 * for convenience. This structure need not be visible externally.
440 * It is allocated and initialized during mount, and released
443 struct rpcrdma_xprt {
444 struct rpc_xprt rx_xprt;
445 struct rpcrdma_ia rx_ia;
446 struct rpcrdma_ep rx_ep;
447 struct rpcrdma_buffer rx_buf;
448 struct rpcrdma_create_data_internal rx_data;
449 struct delayed_work rx_connect_worker;
450 struct rpcrdma_stats rx_stats;
453 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
454 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data)
456 /* Setting this to 0 ensures interoperability with early servers.
457 * Setting this to 1 enhances certain unaligned read/write performance.
458 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
459 extern int xprt_rdma_pad_optimize;
462 * Interface Adapter calls - xprtrdma/verbs.c
464 int rpcrdma_ia_open(struct rpcrdma_xprt *, struct sockaddr *, int);
465 void rpcrdma_ia_close(struct rpcrdma_ia *);
466 bool frwr_is_supported(struct rpcrdma_ia *);
467 bool fmr_is_supported(struct rpcrdma_ia *);
470 * Endpoint calls - xprtrdma/verbs.c
472 int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *,
473 struct rpcrdma_create_data_internal *);
474 void rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *);
475 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *);
476 void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *);
478 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
479 struct rpcrdma_req *);
480 int rpcrdma_ep_post_recv(struct rpcrdma_ia *, struct rpcrdma_rep *);
483 * Buffer calls - xprtrdma/verbs.c
485 struct rpcrdma_req *rpcrdma_create_req(struct rpcrdma_xprt *);
486 struct rpcrdma_rep *rpcrdma_create_rep(struct rpcrdma_xprt *);
487 void rpcrdma_destroy_req(struct rpcrdma_req *);
488 int rpcrdma_buffer_create(struct rpcrdma_xprt *);
489 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
491 struct rpcrdma_mw *rpcrdma_get_mw(struct rpcrdma_xprt *);
492 void rpcrdma_put_mw(struct rpcrdma_xprt *, struct rpcrdma_mw *);
493 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
494 void rpcrdma_buffer_put(struct rpcrdma_req *);
495 void rpcrdma_recv_buffer_get(struct rpcrdma_req *);
496 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
498 void rpcrdma_defer_mr_recovery(struct rpcrdma_mw *);
500 struct rpcrdma_regbuf *rpcrdma_alloc_regbuf(size_t, enum dma_data_direction,
502 bool __rpcrdma_dma_map_regbuf(struct rpcrdma_ia *, struct rpcrdma_regbuf *);
503 void rpcrdma_free_regbuf(struct rpcrdma_regbuf *);
506 rpcrdma_regbuf_is_mapped(struct rpcrdma_regbuf *rb)
508 return rb->rg_device != NULL;
512 rpcrdma_dma_map_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
514 if (likely(rpcrdma_regbuf_is_mapped(rb)))
516 return __rpcrdma_dma_map_regbuf(ia, rb);
519 int rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt *, unsigned int);
521 int rpcrdma_alloc_wq(void);
522 void rpcrdma_destroy_wq(void);
525 * Wrappers for chunk registration, shared by read/write chunk code.
528 static inline enum dma_data_direction
529 rpcrdma_data_dir(bool writing)
531 return writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
535 * RPC/RDMA connection management calls - xprtrdma/rpc_rdma.c
537 void rpcrdma_connect_worker(struct work_struct *);
538 void rpcrdma_conn_func(struct rpcrdma_ep *);
539 void rpcrdma_reply_handler(struct work_struct *);
542 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
545 enum rpcrdma_chunktype {
553 bool rpcrdma_prepare_send_sges(struct rpcrdma_ia *, struct rpcrdma_req *,
554 u32, struct xdr_buf *, enum rpcrdma_chunktype);
555 void rpcrdma_unmap_sges(struct rpcrdma_ia *, struct rpcrdma_req *);
556 int rpcrdma_marshal_req(struct rpc_rqst *);
557 void rpcrdma_set_max_header_sizes(struct rpcrdma_xprt *);
559 /* RPC/RDMA module init - xprtrdma/transport.c
561 extern unsigned int xprt_rdma_max_inline_read;
562 void xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap);
563 void xprt_rdma_free_addresses(struct rpc_xprt *xprt);
564 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq);
565 int xprt_rdma_init(void);
566 void xprt_rdma_cleanup(void);
568 /* Backchannel calls - xprtrdma/backchannel.c
570 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
571 int xprt_rdma_bc_setup(struct rpc_xprt *, unsigned int);
572 int xprt_rdma_bc_up(struct svc_serv *, struct net *);
573 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *);
574 int rpcrdma_bc_post_recv(struct rpcrdma_xprt *, unsigned int);
575 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *, struct rpcrdma_rep *);
576 int rpcrdma_bc_marshal_reply(struct rpc_rqst *);
577 void xprt_rdma_bc_free_rqst(struct rpc_rqst *);
578 void xprt_rdma_bc_destroy(struct rpc_xprt *, unsigned int);
579 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
581 extern struct xprt_class xprt_rdma_bc;
583 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */