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;
72 struct completion ri_remove_done;
74 unsigned int ri_max_segs;
75 unsigned int ri_max_frmr_depth;
76 unsigned int ri_max_inline_write;
77 unsigned int ri_max_inline_read;
78 unsigned int ri_max_send_sges;
79 bool ri_reminv_expected;
80 bool ri_implicit_roundup;
81 enum ib_mr_type ri_mrtype;
82 unsigned long ri_flags;
83 struct ib_qp_attr ri_qp_attr;
84 struct ib_qp_init_attr ri_qp_init_attr;
88 RPCRDMA_IAF_REMOVING = 0,
92 * RDMA Endpoint -- one per transport instance
99 struct ib_qp_init_attr rep_attr;
100 wait_queue_head_t rep_connect_wait;
101 struct rpcrdma_connect_private rep_cm_private;
102 struct rdma_conn_param rep_remote_cma;
103 struct sockaddr_storage rep_remote_addr;
104 struct delayed_work rep_connect_worker;
108 rpcrdma_init_cqcount(struct rpcrdma_ep *ep, int count)
110 atomic_set(&ep->rep_cqcount, ep->rep_cqinit - count);
113 /* To update send queue accounting, provider must take a
114 * send completion every now and then.
117 rpcrdma_set_signaled(struct rpcrdma_ep *ep, struct ib_send_wr *send_wr)
119 send_wr->send_flags = 0;
120 if (unlikely(atomic_sub_return(1, &ep->rep_cqcount) <= 0)) {
121 rpcrdma_init_cqcount(ep, 0);
122 send_wr->send_flags = IB_SEND_SIGNALED;
126 /* Pre-allocate extra Work Requests for handling backward receives
127 * and sends. This is a fixed value because the Work Queues are
128 * allocated when the forward channel is set up.
130 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
131 #define RPCRDMA_BACKWARD_WRS (8)
133 #define RPCRDMA_BACKWARD_WRS (0)
136 /* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
138 * The below structure appears at the front of a large region of kmalloc'd
139 * memory, which always starts on a good alignment boundary.
142 struct rpcrdma_regbuf {
143 struct ib_sge rg_iov;
144 struct ib_device *rg_device;
145 enum dma_data_direction rg_direction;
146 __be32 rg_base[0] __attribute__ ((aligned(256)));
150 rdmab_addr(struct rpcrdma_regbuf *rb)
152 return rb->rg_iov.addr;
156 rdmab_length(struct rpcrdma_regbuf *rb)
158 return rb->rg_iov.length;
162 rdmab_lkey(struct rpcrdma_regbuf *rb)
164 return rb->rg_iov.lkey;
167 static inline struct rpcrdma_msg *
168 rdmab_to_msg(struct rpcrdma_regbuf *rb)
170 return (struct rpcrdma_msg *)rb->rg_base;
173 static inline struct ib_device *
174 rdmab_device(struct rpcrdma_regbuf *rb)
176 return rb->rg_device;
179 #define RPCRDMA_DEF_GFP (GFP_NOIO | __GFP_NOWARN)
181 /* To ensure a transport can always make forward progress,
182 * the number of RDMA segments allowed in header chunk lists
183 * is capped at 8. This prevents less-capable devices and
184 * memory registrations from overrunning the Send buffer
185 * while building chunk lists.
187 * Elements of the Read list take up more room than the
188 * Write list or Reply chunk. 8 read segments means the Read
189 * list (or Write list or Reply chunk) cannot consume more
192 * ((8 + 2) * read segment size) + 1 XDR words, or 244 bytes.
194 * And the fixed part of the header is another 24 bytes.
196 * The smallest inline threshold is 1024 bytes, ensuring that
197 * at least 750 bytes are available for RPC messages.
200 RPCRDMA_MAX_HDR_SEGS = 8,
201 RPCRDMA_HDRBUF_SIZE = 256,
205 * struct rpcrdma_rep -- this structure encapsulates state required to recv
206 * and complete a reply, asychronously. It needs several pieces of
208 * o recv buffer (posted to provider)
209 * o ib_sge (also donated to provider)
210 * o status of reply (length, success or not)
211 * o bookkeeping state to get run by reply handler (list, etc)
213 * These are allocated during initialization, per-transport instance.
215 * N of these are associated with a transport instance, and stored in
216 * struct rpcrdma_buffer. N is the max number of outstanding requests.
220 struct ib_cqe rr_cqe;
224 struct rpcrdma_xprt *rr_rxprt;
225 struct work_struct rr_work;
226 struct list_head rr_list;
227 struct ib_recv_wr rr_recv_wr;
228 struct rpcrdma_regbuf *rr_rdmabuf;
231 #define RPCRDMA_BAD_LEN (~0U)
234 * struct rpcrdma_mw - external memory region metadata
236 * An external memory region is any buffer or page that is registered
237 * on the fly (ie, not pre-registered).
239 * Each rpcrdma_buffer has a list of free MWs anchored in rb_mws. During
240 * call_allocate, rpcrdma_buffer_get() assigns one to each segment in
241 * an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep
242 * track of registration metadata while each RPC is pending.
243 * rpcrdma_deregister_external() uses this metadata to unmap and
244 * release these resources when an RPC is complete.
246 enum rpcrdma_frmr_state {
247 FRMR_IS_INVALID, /* ready to be used */
248 FRMR_IS_VALID, /* in use */
249 FRMR_FLUSHED_FR, /* flushed FASTREG WR */
250 FRMR_FLUSHED_LI, /* flushed LOCALINV WR */
253 struct rpcrdma_frmr {
255 struct ib_cqe fr_cqe;
256 enum rpcrdma_frmr_state fr_state;
257 struct completion fr_linv_done;
259 struct ib_reg_wr fr_regwr;
260 struct ib_send_wr fr_invwr;
265 struct ib_fmr *fm_mr;
270 struct list_head mw_list;
271 struct scatterlist *mw_sg;
273 enum dma_data_direction mw_dir;
274 unsigned long mw_flags;
276 struct rpcrdma_fmr fmr;
277 struct rpcrdma_frmr frmr;
279 struct rpcrdma_xprt *mw_xprt;
283 struct list_head mw_all;
292 * struct rpcrdma_req -- structure central to the request/reply sequence.
294 * N of these are associated with a transport instance, and stored in
295 * struct rpcrdma_buffer. N is the max number of outstanding requests.
297 * It includes pre-registered buffer memory for send AND recv.
298 * The recv buffer, however, is not owned by this structure, and
299 * is "donated" to the hardware when a recv is posted. When a
300 * reply is handled, the recv buffer used is given back to the
301 * struct rpcrdma_req associated with the request.
303 * In addition to the basic memory, this structure includes an array
304 * of iovs for send operations. The reason is that the iovs passed to
305 * ib_post_{send,recv} must not be modified until the work request
309 /* Maximum number of page-sized "segments" per chunk list to be
310 * registered or invalidated. Must handle a Reply chunk:
313 RPCRDMA_MAX_IOV_SEGS = 3,
314 RPCRDMA_MAX_DATA_SEGS = ((1 * 1024 * 1024) / PAGE_SIZE) + 1,
315 RPCRDMA_MAX_SEGS = RPCRDMA_MAX_DATA_SEGS +
316 RPCRDMA_MAX_IOV_SEGS,
319 struct rpcrdma_mr_seg { /* chunk descriptors */
320 u32 mr_len; /* length of chunk or segment */
321 struct page *mr_page; /* owning page, if any */
322 char *mr_offset; /* kva if no page, else offset */
325 /* The Send SGE array is provisioned to send a maximum size
327 * - RPC-over-RDMA header
328 * - xdr_buf head iovec
329 * - RPCRDMA_MAX_INLINE bytes, in pages
330 * - xdr_buf tail iovec
332 * The actual number of array elements consumed by each RPC
333 * depends on the device's max_sge limit.
336 RPCRDMA_MIN_SEND_SGES = 3,
337 RPCRDMA_MAX_PAGE_SGES = RPCRDMA_MAX_INLINE >> PAGE_SHIFT,
338 RPCRDMA_MAX_SEND_SGES = 1 + 1 + RPCRDMA_MAX_PAGE_SGES + 1,
341 struct rpcrdma_buffer;
343 struct list_head rl_free;
344 unsigned int rl_mapped_sges;
345 unsigned int rl_connect_cookie;
346 struct rpcrdma_buffer *rl_buffer;
347 struct rpcrdma_rep *rl_reply;
348 struct ib_send_wr rl_send_wr;
349 struct ib_sge rl_send_sge[RPCRDMA_MAX_SEND_SGES];
350 struct rpcrdma_regbuf *rl_rdmabuf; /* xprt header */
351 struct rpcrdma_regbuf *rl_sendbuf; /* rq_snd_buf */
352 struct rpcrdma_regbuf *rl_recvbuf; /* rq_rcv_buf */
354 struct ib_cqe rl_cqe;
355 struct list_head rl_all;
358 struct list_head rl_registered; /* registered segments */
359 struct rpcrdma_mr_seg rl_segments[RPCRDMA_MAX_SEGS];
363 rpcrdma_set_xprtdata(struct rpc_rqst *rqst, struct rpcrdma_req *req)
365 rqst->rq_xprtdata = req;
368 static inline struct rpcrdma_req *
369 rpcr_to_rdmar(struct rpc_rqst *rqst)
371 return rqst->rq_xprtdata;
375 rpcrdma_push_mw(struct rpcrdma_mw *mw, struct list_head *list)
377 list_add_tail(&mw->mw_list, list);
380 static inline struct rpcrdma_mw *
381 rpcrdma_pop_mw(struct list_head *list)
383 struct rpcrdma_mw *mw;
385 mw = list_first_entry(list, struct rpcrdma_mw, mw_list);
386 list_del(&mw->mw_list);
391 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
392 * inline requests/replies, and client/server credits.
394 * One of these is associated with a transport instance
396 struct rpcrdma_buffer {
397 spinlock_t rb_mwlock; /* protect rb_mws list */
398 struct list_head rb_mws;
399 struct list_head rb_all;
401 spinlock_t rb_lock; /* protect buf lists */
402 int rb_send_count, rb_recv_count;
403 struct list_head rb_send_bufs;
404 struct list_head rb_recv_bufs;
406 atomic_t rb_credits; /* most recent credit grant */
408 u32 rb_bc_srv_max_requests;
409 spinlock_t rb_reqslock; /* protect rb_allreqs */
410 struct list_head rb_allreqs;
412 u32 rb_bc_max_requests;
414 spinlock_t rb_recovery_lock; /* protect rb_stale_mrs */
415 struct list_head rb_stale_mrs;
416 struct delayed_work rb_recovery_worker;
417 struct delayed_work rb_refresh_worker;
419 #define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia)
422 * Internal structure for transport instance creation. This
423 * exists primarily for modularity.
425 * This data should be set with mount options
427 struct rpcrdma_create_data_internal {
428 struct sockaddr_storage addr; /* RDMA server address */
429 unsigned int max_requests; /* max requests (slots) in flight */
430 unsigned int rsize; /* mount rsize - max read hdr+data */
431 unsigned int wsize; /* mount wsize - max write hdr+data */
432 unsigned int inline_rsize; /* max non-rdma read data payload */
433 unsigned int inline_wsize; /* max non-rdma write data payload */
434 unsigned int padding; /* non-rdma write header padding */
438 * Statistics for RPCRDMA
440 struct rpcrdma_stats {
441 unsigned long read_chunk_count;
442 unsigned long write_chunk_count;
443 unsigned long reply_chunk_count;
445 unsigned long long total_rdma_request;
446 unsigned long long total_rdma_reply;
448 unsigned long long pullup_copy_count;
449 unsigned long long fixup_copy_count;
450 unsigned long hardway_register_count;
451 unsigned long failed_marshal_count;
452 unsigned long bad_reply_count;
453 unsigned long nomsg_call_count;
454 unsigned long bcall_count;
455 unsigned long mrs_recovered;
456 unsigned long mrs_orphaned;
457 unsigned long mrs_allocated;
458 unsigned long local_inv_needed;
462 * Per-registration mode operations
465 struct rpcrdma_memreg_ops {
466 int (*ro_map)(struct rpcrdma_xprt *,
467 struct rpcrdma_mr_seg *, int, bool,
468 struct rpcrdma_mw **);
469 void (*ro_unmap_sync)(struct rpcrdma_xprt *,
470 struct rpcrdma_req *);
471 void (*ro_unmap_safe)(struct rpcrdma_xprt *,
472 struct rpcrdma_req *, bool);
473 void (*ro_recover_mr)(struct rpcrdma_mw *);
474 int (*ro_open)(struct rpcrdma_ia *,
476 struct rpcrdma_create_data_internal *);
477 size_t (*ro_maxpages)(struct rpcrdma_xprt *);
478 int (*ro_init_mr)(struct rpcrdma_ia *,
479 struct rpcrdma_mw *);
480 void (*ro_release_mr)(struct rpcrdma_mw *);
481 const char *ro_displayname;
482 const int ro_send_w_inv_ok;
485 extern const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops;
486 extern const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops;
489 * RPCRDMA transport -- encapsulates the structures above for
490 * integration with RPC.
492 * The contained structures are embedded, not pointers,
493 * for convenience. This structure need not be visible externally.
495 * It is allocated and initialized during mount, and released
498 struct rpcrdma_xprt {
499 struct rpc_xprt rx_xprt;
500 struct rpcrdma_ia rx_ia;
501 struct rpcrdma_ep rx_ep;
502 struct rpcrdma_buffer rx_buf;
503 struct rpcrdma_create_data_internal rx_data;
504 struct delayed_work rx_connect_worker;
505 struct rpcrdma_stats rx_stats;
508 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
509 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data)
511 /* Setting this to 0 ensures interoperability with early servers.
512 * Setting this to 1 enhances certain unaligned read/write performance.
513 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
514 extern int xprt_rdma_pad_optimize;
516 /* This setting controls the hunt for a supported memory
517 * registration strategy.
519 extern unsigned int xprt_rdma_memreg_strategy;
522 * Interface Adapter calls - xprtrdma/verbs.c
524 int rpcrdma_ia_open(struct rpcrdma_xprt *xprt, struct sockaddr *addr);
525 void rpcrdma_ia_remove(struct rpcrdma_ia *ia);
526 void rpcrdma_ia_close(struct rpcrdma_ia *);
527 bool frwr_is_supported(struct rpcrdma_ia *);
528 bool fmr_is_supported(struct rpcrdma_ia *);
531 * Endpoint calls - xprtrdma/verbs.c
533 int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *,
534 struct rpcrdma_create_data_internal *);
535 void rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *);
536 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *);
537 void rpcrdma_conn_func(struct rpcrdma_ep *ep);
538 void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *);
540 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
541 struct rpcrdma_req *);
542 int rpcrdma_ep_post_recv(struct rpcrdma_ia *, struct rpcrdma_rep *);
545 * Buffer calls - xprtrdma/verbs.c
547 struct rpcrdma_req *rpcrdma_create_req(struct rpcrdma_xprt *);
548 struct rpcrdma_rep *rpcrdma_create_rep(struct rpcrdma_xprt *);
549 void rpcrdma_destroy_req(struct rpcrdma_req *);
550 int rpcrdma_buffer_create(struct rpcrdma_xprt *);
551 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
553 struct rpcrdma_mw *rpcrdma_get_mw(struct rpcrdma_xprt *);
554 void rpcrdma_put_mw(struct rpcrdma_xprt *, struct rpcrdma_mw *);
555 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
556 void rpcrdma_buffer_put(struct rpcrdma_req *);
557 void rpcrdma_recv_buffer_get(struct rpcrdma_req *);
558 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
560 void rpcrdma_defer_mr_recovery(struct rpcrdma_mw *);
562 struct rpcrdma_regbuf *rpcrdma_alloc_regbuf(size_t, enum dma_data_direction,
564 bool __rpcrdma_dma_map_regbuf(struct rpcrdma_ia *, struct rpcrdma_regbuf *);
565 void rpcrdma_free_regbuf(struct rpcrdma_regbuf *);
568 rpcrdma_regbuf_is_mapped(struct rpcrdma_regbuf *rb)
570 return rb->rg_device != NULL;
574 rpcrdma_dma_map_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
576 if (likely(rpcrdma_regbuf_is_mapped(rb)))
578 return __rpcrdma_dma_map_regbuf(ia, rb);
581 int rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt *, unsigned int);
583 int rpcrdma_alloc_wq(void);
584 void rpcrdma_destroy_wq(void);
587 * Wrappers for chunk registration, shared by read/write chunk code.
590 static inline enum dma_data_direction
591 rpcrdma_data_dir(bool writing)
593 return writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
597 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
600 enum rpcrdma_chunktype {
608 bool rpcrdma_prepare_send_sges(struct rpcrdma_ia *, struct rpcrdma_req *,
609 u32, struct xdr_buf *, enum rpcrdma_chunktype);
610 void rpcrdma_unmap_sges(struct rpcrdma_ia *, struct rpcrdma_req *);
611 int rpcrdma_marshal_req(struct rpc_rqst *);
612 void rpcrdma_set_max_header_sizes(struct rpcrdma_xprt *);
613 void rpcrdma_reply_handler(struct work_struct *work);
615 /* RPC/RDMA module init - xprtrdma/transport.c
617 extern unsigned int xprt_rdma_max_inline_read;
618 void xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap);
619 void xprt_rdma_free_addresses(struct rpc_xprt *xprt);
620 void rpcrdma_connect_worker(struct work_struct *work);
621 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq);
622 int xprt_rdma_init(void);
623 void xprt_rdma_cleanup(void);
625 /* Backchannel calls - xprtrdma/backchannel.c
627 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
628 int xprt_rdma_bc_setup(struct rpc_xprt *, unsigned int);
629 int xprt_rdma_bc_up(struct svc_serv *, struct net *);
630 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *);
631 int rpcrdma_bc_post_recv(struct rpcrdma_xprt *, unsigned int);
632 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *, struct rpcrdma_rep *);
633 int rpcrdma_bc_marshal_reply(struct rpc_rqst *);
634 void xprt_rdma_bc_free_rqst(struct rpc_rqst *);
635 void xprt_rdma_bc_destroy(struct rpc_xprt *, unsigned int);
636 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
638 extern struct xprt_class xprt_rdma_bc;
640 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */