1 /* SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause */
3 * Copyright (c) 2014-2017 Oracle. All rights reserved.
4 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
6 * This software is available to you under a choice of one of two
7 * licenses. You may choose to be licensed under the terms of the GNU
8 * General Public License (GPL) Version 2, available from the file
9 * COPYING in the main directory of this source tree, or the BSD-type
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
16 * Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
19 * Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials provided
22 * with the distribution.
24 * Neither the name of the Network Appliance, Inc. nor the names of
25 * its contributors may be used to endorse or promote products
26 * derived from this software without specific prior written
29 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
30 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
31 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
32 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
33 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
34 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
35 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
36 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
37 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
38 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
39 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
42 #ifndef _LINUX_SUNRPC_XPRT_RDMA_H
43 #define _LINUX_SUNRPC_XPRT_RDMA_H
45 #include <linux/wait.h> /* wait_queue_head_t, etc */
46 #include <linux/spinlock.h> /* spinlock_t, etc */
47 #include <linux/atomic.h> /* atomic_t, etc */
48 #include <linux/workqueue.h> /* struct work_struct */
50 #include <rdma/rdma_cm.h> /* RDMA connection api */
51 #include <rdma/ib_verbs.h> /* RDMA verbs api */
53 #include <linux/sunrpc/clnt.h> /* rpc_xprt */
54 #include <linux/sunrpc/rpc_rdma.h> /* RPC/RDMA protocol */
55 #include <linux/sunrpc/xprtrdma.h> /* xprt parameters */
57 #define RDMA_RESOLVE_TIMEOUT (5000) /* 5 seconds */
58 #define RDMA_CONNECT_RETRY_MAX (2) /* retries if no listener backlog */
60 #define RPCRDMA_BIND_TO (60U * HZ)
61 #define RPCRDMA_INIT_REEST_TO (5U * HZ)
62 #define RPCRDMA_MAX_REEST_TO (30U * HZ)
63 #define RPCRDMA_IDLE_DISC_TO (5U * 60 * HZ)
66 * Interface Adapter -- one per transport instance
69 struct rdma_cm_id *ri_id;
72 unsigned int ri_max_segs;
73 unsigned int ri_max_frwr_depth;
74 unsigned int ri_max_send_sges;
75 bool ri_implicit_roundup;
76 enum ib_mr_type ri_mrtype;
77 unsigned long ri_flags;
78 struct completion ri_done;
79 struct completion ri_remove_done;
83 RPCRDMA_IAF_REMOVING = 0,
87 * RDMA Endpoint -- one per transport instance
91 unsigned int rep_send_count;
92 unsigned int rep_send_batch;
93 unsigned int rep_max_inline_send;
94 unsigned int rep_max_inline_recv;
96 struct ib_qp_init_attr rep_attr;
97 wait_queue_head_t rep_connect_wait;
98 struct rpcrdma_connect_private rep_cm_private;
99 struct rdma_conn_param rep_remote_cma;
100 unsigned int rep_max_requests; /* set by /proc */
101 unsigned int rep_inline_send; /* negotiated */
102 unsigned int rep_inline_recv; /* negotiated */
103 int rep_receive_count;
106 /* Pre-allocate extra Work Requests for handling backward receives
107 * and sends. This is a fixed value because the Work Queues are
108 * allocated when the forward channel is set up, long before the
109 * backchannel is provisioned. This value is two times
110 * NFS4_DEF_CB_SLOT_TABLE_SIZE.
112 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
113 #define RPCRDMA_BACKWARD_WRS (32)
115 #define RPCRDMA_BACKWARD_WRS (0)
118 /* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
120 * The below structure appears at the front of a large region of kmalloc'd
121 * memory, which always starts on a good alignment boundary.
124 struct rpcrdma_regbuf {
125 struct ib_sge rg_iov;
126 struct ib_device *rg_device;
127 enum dma_data_direction rg_direction;
131 static inline u64 rdmab_addr(struct rpcrdma_regbuf *rb)
133 return rb->rg_iov.addr;
136 static inline u32 rdmab_length(struct rpcrdma_regbuf *rb)
138 return rb->rg_iov.length;
141 static inline u32 rdmab_lkey(struct rpcrdma_regbuf *rb)
143 return rb->rg_iov.lkey;
146 static inline struct ib_device *rdmab_device(struct rpcrdma_regbuf *rb)
148 return rb->rg_device;
151 static inline void *rdmab_data(const struct rpcrdma_regbuf *rb)
156 #define RPCRDMA_DEF_GFP (GFP_NOIO | __GFP_NOWARN)
158 /* To ensure a transport can always make forward progress,
159 * the number of RDMA segments allowed in header chunk lists
160 * is capped at 8. This prevents less-capable devices and
161 * memory registrations from overrunning the Send buffer
162 * while building chunk lists.
164 * Elements of the Read list take up more room than the
165 * Write list or Reply chunk. 8 read segments means the Read
166 * list (or Write list or Reply chunk) cannot consume more
169 * ((8 + 2) * read segment size) + 1 XDR words, or 244 bytes.
171 * And the fixed part of the header is another 24 bytes.
173 * The smallest inline threshold is 1024 bytes, ensuring that
174 * at least 750 bytes are available for RPC messages.
177 RPCRDMA_MAX_HDR_SEGS = 8,
178 RPCRDMA_HDRBUF_SIZE = 256,
182 * struct rpcrdma_rep -- this structure encapsulates state required
183 * to receive and complete an RPC Reply, asychronously. It needs
184 * several pieces of state:
186 * o receive buffer and ib_sge (donated to provider)
187 * o status of receive (success or not, length, inv rkey)
188 * o bookkeeping state to get run by reply handler (XDR stream)
190 * These structures are allocated during transport initialization.
191 * N of these are associated with a transport instance, managed by
192 * struct rpcrdma_buffer. N is the max number of outstanding RPCs.
196 struct ib_cqe rr_cqe;
203 struct rpcrdma_regbuf *rr_rdmabuf;
204 struct rpcrdma_xprt *rr_rxprt;
205 struct work_struct rr_work;
206 struct xdr_buf rr_hdrbuf;
207 struct xdr_stream rr_stream;
208 struct rpc_rqst *rr_rqst;
209 struct list_head rr_list;
210 struct ib_recv_wr rr_recv_wr;
213 /* To reduce the rate at which a transport invokes ib_post_recv
214 * (and thus the hardware doorbell rate), xprtrdma posts Receive
217 * Setting this to zero disables Receive post batching.
220 RPCRDMA_MAX_RECV_BATCH = 7,
223 /* struct rpcrdma_sendctx - DMA mapped SGEs to unmap after Send completes
227 struct rpcrdma_sendctx {
228 struct ib_send_wr sc_wr;
229 struct ib_cqe sc_cqe;
230 struct ib_device *sc_device;
231 struct rpcrdma_xprt *sc_xprt;
232 struct rpcrdma_req *sc_req;
233 unsigned int sc_unmap_count;
234 struct ib_sge sc_sges[];
238 * struct rpcrdma_mr - external memory region metadata
240 * An external memory region is any buffer or page that is registered
241 * on the fly (ie, not pre-registered).
243 enum rpcrdma_frwr_state {
244 FRWR_IS_INVALID, /* ready to be used */
245 FRWR_IS_VALID, /* in use */
246 FRWR_FLUSHED_FR, /* flushed FASTREG WR */
247 FRWR_FLUSHED_LI, /* flushed LOCALINV WR */
250 struct rpcrdma_frwr {
252 struct ib_cqe fr_cqe;
253 enum rpcrdma_frwr_state fr_state;
254 struct completion fr_linv_done;
256 struct ib_reg_wr fr_regwr;
257 struct ib_send_wr fr_invwr;
262 struct list_head mr_list;
263 struct scatterlist *mr_sg;
265 enum dma_data_direction mr_dir;
266 struct rpcrdma_frwr frwr;
267 struct rpcrdma_xprt *mr_xprt;
271 struct work_struct mr_recycle;
272 struct list_head mr_all;
276 * struct rpcrdma_req -- structure central to the request/reply sequence.
278 * N of these are associated with a transport instance, and stored in
279 * struct rpcrdma_buffer. N is the max number of outstanding requests.
281 * It includes pre-registered buffer memory for send AND recv.
282 * The recv buffer, however, is not owned by this structure, and
283 * is "donated" to the hardware when a recv is posted. When a
284 * reply is handled, the recv buffer used is given back to the
285 * struct rpcrdma_req associated with the request.
287 * In addition to the basic memory, this structure includes an array
288 * of iovs for send operations. The reason is that the iovs passed to
289 * ib_post_{send,recv} must not be modified until the work request
293 /* Maximum number of page-sized "segments" per chunk list to be
294 * registered or invalidated. Must handle a Reply chunk:
297 RPCRDMA_MAX_IOV_SEGS = 3,
298 RPCRDMA_MAX_DATA_SEGS = ((1 * 1024 * 1024) / PAGE_SIZE) + 1,
299 RPCRDMA_MAX_SEGS = RPCRDMA_MAX_DATA_SEGS +
300 RPCRDMA_MAX_IOV_SEGS,
303 struct rpcrdma_mr_seg { /* chunk descriptors */
304 u32 mr_len; /* length of chunk or segment */
305 struct page *mr_page; /* owning page, if any */
306 char *mr_offset; /* kva if no page, else offset */
309 /* The Send SGE array is provisioned to send a maximum size
311 * - RPC-over-RDMA header
312 * - xdr_buf head iovec
313 * - RPCRDMA_MAX_INLINE bytes, in pages
314 * - xdr_buf tail iovec
316 * The actual number of array elements consumed by each RPC
317 * depends on the device's max_sge limit.
320 RPCRDMA_MIN_SEND_SGES = 3,
321 RPCRDMA_MAX_PAGE_SGES = RPCRDMA_MAX_INLINE >> PAGE_SHIFT,
322 RPCRDMA_MAX_SEND_SGES = 1 + 1 + RPCRDMA_MAX_PAGE_SGES + 1,
325 struct rpcrdma_buffer;
327 struct list_head rl_list;
328 struct rpc_rqst rl_slot;
329 struct rpcrdma_buffer *rl_buffer;
330 struct rpcrdma_rep *rl_reply;
331 struct xdr_stream rl_stream;
332 struct xdr_buf rl_hdrbuf;
333 struct rpcrdma_sendctx *rl_sendctx;
334 struct rpcrdma_regbuf *rl_rdmabuf; /* xprt header */
335 struct rpcrdma_regbuf *rl_sendbuf; /* rq_snd_buf */
336 struct rpcrdma_regbuf *rl_recvbuf; /* rq_rcv_buf */
338 struct list_head rl_all;
339 unsigned long rl_flags;
341 struct list_head rl_registered; /* registered segments */
342 struct rpcrdma_mr_seg rl_segments[RPCRDMA_MAX_SEGS];
347 RPCRDMA_REQ_F_PENDING = 0,
348 RPCRDMA_REQ_F_TX_RESOURCES,
351 static inline struct rpcrdma_req *
352 rpcr_to_rdmar(const struct rpc_rqst *rqst)
354 return container_of(rqst, struct rpcrdma_req, rl_slot);
358 rpcrdma_mr_push(struct rpcrdma_mr *mr, struct list_head *list)
360 list_add_tail(&mr->mr_list, list);
363 static inline struct rpcrdma_mr *
364 rpcrdma_mr_pop(struct list_head *list)
366 struct rpcrdma_mr *mr;
368 mr = list_first_entry(list, struct rpcrdma_mr, mr_list);
369 list_del_init(&mr->mr_list);
374 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
375 * inline requests/replies, and client/server credits.
377 * One of these is associated with a transport instance
379 struct rpcrdma_buffer {
380 spinlock_t rb_mrlock; /* protect rb_mrs list */
381 struct list_head rb_mrs;
382 struct list_head rb_all;
384 unsigned long rb_sc_head;
385 unsigned long rb_sc_tail;
386 unsigned long rb_sc_last;
387 struct rpcrdma_sendctx **rb_sc_ctxs;
389 spinlock_t rb_lock; /* protect buf lists */
390 struct list_head rb_send_bufs;
391 struct list_head rb_recv_bufs;
392 struct list_head rb_allreqs;
394 unsigned long rb_flags;
396 u32 rb_credits; /* most recent credit grant */
398 u32 rb_bc_srv_max_requests;
399 u32 rb_bc_max_requests;
401 struct workqueue_struct *rb_completion_wq;
402 struct delayed_work rb_refresh_worker;
407 RPCRDMA_BUF_F_EMPTY_SCQ = 0,
411 * Statistics for RPCRDMA
413 struct rpcrdma_stats {
414 /* accessed when sending a call */
415 unsigned long read_chunk_count;
416 unsigned long write_chunk_count;
417 unsigned long reply_chunk_count;
418 unsigned long long total_rdma_request;
420 /* rarely accessed error counters */
421 unsigned long long pullup_copy_count;
422 unsigned long hardway_register_count;
423 unsigned long failed_marshal_count;
424 unsigned long bad_reply_count;
425 unsigned long mrs_recycled;
426 unsigned long mrs_orphaned;
427 unsigned long mrs_allocated;
428 unsigned long empty_sendctx_q;
430 /* accessed when receiving a reply */
431 unsigned long long total_rdma_reply;
432 unsigned long long fixup_copy_count;
433 unsigned long reply_waits_for_send;
434 unsigned long local_inv_needed;
435 unsigned long nomsg_call_count;
436 unsigned long bcall_count;
440 * RPCRDMA transport -- encapsulates the structures above for
441 * integration with RPC.
443 * The contained structures are embedded, not pointers,
444 * for convenience. This structure need not be visible externally.
446 * It is allocated and initialized during mount, and released
449 struct rpcrdma_xprt {
450 struct rpc_xprt rx_xprt;
451 struct rpcrdma_ia rx_ia;
452 struct rpcrdma_ep rx_ep;
453 struct rpcrdma_buffer rx_buf;
454 struct delayed_work rx_connect_worker;
455 struct rpcrdma_stats rx_stats;
458 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
460 static inline const char *
461 rpcrdma_addrstr(const struct rpcrdma_xprt *r_xprt)
463 return r_xprt->rx_xprt.address_strings[RPC_DISPLAY_ADDR];
466 static inline const char *
467 rpcrdma_portstr(const struct rpcrdma_xprt *r_xprt)
469 return r_xprt->rx_xprt.address_strings[RPC_DISPLAY_PORT];
472 /* Setting this to 0 ensures interoperability with early servers.
473 * Setting this to 1 enhances certain unaligned read/write performance.
474 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
475 extern int xprt_rdma_pad_optimize;
477 /* This setting controls the hunt for a supported memory
478 * registration strategy.
480 extern unsigned int xprt_rdma_memreg_strategy;
483 * Interface Adapter calls - xprtrdma/verbs.c
485 int rpcrdma_ia_open(struct rpcrdma_xprt *xprt);
486 void rpcrdma_ia_remove(struct rpcrdma_ia *ia);
487 void rpcrdma_ia_close(struct rpcrdma_ia *);
490 * Endpoint calls - xprtrdma/verbs.c
492 int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt);
493 void rpcrdma_ep_destroy(struct rpcrdma_xprt *r_xprt);
494 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *);
495 void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *);
497 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
498 struct rpcrdma_req *);
501 * Buffer calls - xprtrdma/verbs.c
503 struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
505 void rpcrdma_req_destroy(struct rpcrdma_req *req);
506 int rpcrdma_buffer_create(struct rpcrdma_xprt *);
507 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
508 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt);
510 struct rpcrdma_mr *rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt);
511 void rpcrdma_mr_put(struct rpcrdma_mr *mr);
512 void rpcrdma_mr_unmap_and_put(struct rpcrdma_mr *mr);
515 rpcrdma_mr_recycle(struct rpcrdma_mr *mr)
517 schedule_work(&mr->mr_recycle);
520 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
521 void rpcrdma_buffer_put(struct rpcrdma_req *);
522 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
524 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size,
526 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
527 struct rpcrdma_regbuf *rb);
530 * rpcrdma_regbuf_is_mapped - check if buffer is DMA mapped
532 * Returns true if the buffer is now mapped to rb->rg_device.
534 static inline bool rpcrdma_regbuf_is_mapped(struct rpcrdma_regbuf *rb)
536 return rb->rg_device != NULL;
540 * rpcrdma_regbuf_dma_map - DMA-map a regbuf
541 * @r_xprt: controlling transport instance
542 * @rb: regbuf to be mapped
544 * Returns true if the buffer is currently DMA mapped.
546 static inline bool rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
547 struct rpcrdma_regbuf *rb)
549 if (likely(rpcrdma_regbuf_is_mapped(rb)))
551 return __rpcrdma_regbuf_dma_map(r_xprt, rb);
555 * Wrappers for chunk registration, shared by read/write chunk code.
558 static inline enum dma_data_direction
559 rpcrdma_data_dir(bool writing)
561 return writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
564 /* Memory registration calls xprtrdma/frwr_ops.c
566 bool frwr_is_supported(struct ib_device *device);
567 int frwr_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep);
568 int frwr_init_mr(struct rpcrdma_ia *ia, struct rpcrdma_mr *mr);
569 void frwr_release_mr(struct rpcrdma_mr *mr);
570 size_t frwr_maxpages(struct rpcrdma_xprt *r_xprt);
571 struct rpcrdma_mr_seg *frwr_map(struct rpcrdma_xprt *r_xprt,
572 struct rpcrdma_mr_seg *seg,
573 int nsegs, bool writing, __be32 xid,
574 struct rpcrdma_mr **mr);
575 int frwr_send(struct rpcrdma_ia *ia, struct rpcrdma_req *req);
576 void frwr_reminv(struct rpcrdma_rep *rep, struct list_head *mrs);
577 void frwr_unmap_sync(struct rpcrdma_xprt *r_xprt,
578 struct list_head *mrs);
581 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
584 enum rpcrdma_chunktype {
592 int rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
593 struct rpcrdma_req *req, u32 hdrlen,
595 enum rpcrdma_chunktype rtype);
596 void rpcrdma_sendctx_unmap(struct rpcrdma_sendctx *sc);
597 int rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst);
598 void rpcrdma_set_max_header_sizes(struct rpcrdma_xprt *);
599 void rpcrdma_complete_rqst(struct rpcrdma_rep *rep);
600 void rpcrdma_reply_handler(struct rpcrdma_rep *rep);
601 void rpcrdma_release_rqst(struct rpcrdma_xprt *r_xprt,
602 struct rpcrdma_req *req);
603 void rpcrdma_deferred_completion(struct work_struct *work);
605 static inline void rpcrdma_set_xdrlen(struct xdr_buf *xdr, size_t len)
607 xdr->head[0].iov_len = len;
611 /* RPC/RDMA module init - xprtrdma/transport.c
613 extern unsigned int xprt_rdma_slot_table_entries;
614 extern unsigned int xprt_rdma_max_inline_read;
615 extern unsigned int xprt_rdma_max_inline_write;
616 void xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap);
617 void xprt_rdma_free_addresses(struct rpc_xprt *xprt);
618 void xprt_rdma_close(struct rpc_xprt *xprt);
619 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq);
620 int xprt_rdma_init(void);
621 void xprt_rdma_cleanup(void);
623 /* Backchannel calls - xprtrdma/backchannel.c
625 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
626 int xprt_rdma_bc_setup(struct rpc_xprt *, unsigned int);
627 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *);
628 int rpcrdma_bc_post_recv(struct rpcrdma_xprt *, unsigned int);
629 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *, struct rpcrdma_rep *);
630 int xprt_rdma_bc_send_reply(struct rpc_rqst *rqst);
631 void xprt_rdma_bc_free_rqst(struct rpc_rqst *);
632 void xprt_rdma_bc_destroy(struct rpc_xprt *, unsigned int);
633 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
635 extern struct xprt_class xprt_rdma_bc;
637 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */