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/kref.h> /* struct kref */
49 #include <linux/workqueue.h> /* struct work_struct */
50 #include <linux/llist.h>
52 #include <rdma/rdma_cm.h> /* RDMA connection api */
53 #include <rdma/ib_verbs.h> /* RDMA verbs api */
55 #include <linux/sunrpc/clnt.h> /* rpc_xprt */
56 #include <linux/sunrpc/rpc_rdma_cid.h> /* completion IDs */
57 #include <linux/sunrpc/rpc_rdma.h> /* RPC/RDMA protocol */
58 #include <linux/sunrpc/xprtrdma.h> /* xprt parameters */
60 #define RDMA_RESOLVE_TIMEOUT (5000) /* 5 seconds */
61 #define RDMA_CONNECT_RETRY_MAX (2) /* retries if no listener backlog */
63 #define RPCRDMA_BIND_TO (60U * HZ)
64 #define RPCRDMA_INIT_REEST_TO (5U * HZ)
65 #define RPCRDMA_MAX_REEST_TO (30U * HZ)
66 #define RPCRDMA_IDLE_DISC_TO (5U * 60 * HZ)
69 * RDMA Endpoint -- connection endpoint details
73 struct rdma_cm_id *re_id;
75 unsigned int re_max_rdma_segs;
76 unsigned int re_max_fr_depth;
77 bool re_implicit_roundup;
78 enum ib_mr_type re_mrtype;
79 struct completion re_done;
80 unsigned int re_send_count;
81 unsigned int re_send_batch;
82 unsigned int re_max_inline_send;
83 unsigned int re_max_inline_recv;
85 int re_connect_status;
86 atomic_t re_force_disconnect;
87 struct ib_qp_init_attr re_attr;
88 wait_queue_head_t re_connect_wait;
89 struct rpc_xprt *re_xprt;
90 struct rpcrdma_connect_private
92 struct rdma_conn_param re_remote_cma;
94 unsigned int re_max_requests; /* depends on device */
95 unsigned int re_inline_send; /* negotiated */
96 unsigned int re_inline_recv; /* negotiated */
98 atomic_t re_completion_ids;
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, long before the
104 * backchannel is provisioned. This value is two times
105 * NFS4_DEF_CB_SLOT_TABLE_SIZE.
107 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
108 #define RPCRDMA_BACKWARD_WRS (32)
110 #define RPCRDMA_BACKWARD_WRS (0)
113 /* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
116 struct rpcrdma_regbuf {
117 struct ib_sge rg_iov;
118 struct ib_device *rg_device;
119 enum dma_data_direction rg_direction;
123 static inline u64 rdmab_addr(struct rpcrdma_regbuf *rb)
125 return rb->rg_iov.addr;
128 static inline u32 rdmab_length(struct rpcrdma_regbuf *rb)
130 return rb->rg_iov.length;
133 static inline u32 rdmab_lkey(struct rpcrdma_regbuf *rb)
135 return rb->rg_iov.lkey;
138 static inline struct ib_device *rdmab_device(struct rpcrdma_regbuf *rb)
140 return rb->rg_device;
143 static inline void *rdmab_data(const struct rpcrdma_regbuf *rb)
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 16. This prevents less-capable devices from
153 * overrunning the Send buffer while building chunk lists.
155 * Elements of the Read list take up more room than the
156 * Write list or Reply chunk. 16 read segments means the
157 * chunk lists cannot consume more than
159 * ((16 + 2) * read segment size) + 1 XDR words,
161 * or about 400 bytes. The fixed part of the header is
162 * another 24 bytes. Thus when the inline threshold is
163 * 1024 bytes, at least 600 bytes are available for RPC
167 RPCRDMA_MAX_HDR_SEGS = 16,
171 * struct rpcrdma_rep -- this structure encapsulates state required
172 * to receive and complete an RPC Reply, asychronously. It needs
173 * several pieces of state:
175 * o receive buffer and ib_sge (donated to provider)
176 * o status of receive (success or not, length, inv rkey)
177 * o bookkeeping state to get run by reply handler (XDR stream)
179 * These structures are allocated during transport initialization.
180 * N of these are associated with a transport instance, managed by
181 * struct rpcrdma_buffer. N is the max number of outstanding RPCs.
185 struct ib_cqe rr_cqe;
186 struct rpc_rdma_cid rr_cid;
194 struct rpcrdma_regbuf *rr_rdmabuf;
195 struct rpcrdma_xprt *rr_rxprt;
196 struct rpc_rqst *rr_rqst;
197 struct xdr_buf rr_hdrbuf;
198 struct xdr_stream rr_stream;
199 struct llist_node rr_node;
200 struct ib_recv_wr rr_recv_wr;
201 struct list_head rr_all;
204 /* To reduce the rate at which a transport invokes ib_post_recv
205 * (and thus the hardware doorbell rate), xprtrdma posts Receive
208 * Setting this to zero disables Receive post batching.
211 RPCRDMA_MAX_RECV_BATCH = 7,
214 /* struct rpcrdma_sendctx - DMA mapped SGEs to unmap after Send completes
217 struct rpcrdma_sendctx {
218 struct ib_cqe sc_cqe;
219 struct rpc_rdma_cid sc_cid;
220 struct rpcrdma_req *sc_req;
221 unsigned int sc_unmap_count;
222 struct ib_sge sc_sges[];
226 * struct rpcrdma_mr - external memory region metadata
228 * An external memory region is any buffer or page that is registered
229 * on the fly (ie, not pre-registered).
231 struct rpcrdma_frwr {
233 struct ib_cqe fr_cqe;
234 struct rpc_rdma_cid fr_cid;
235 struct completion fr_linv_done;
237 struct ib_reg_wr fr_regwr;
238 struct ib_send_wr fr_invwr;
244 struct list_head mr_list;
245 struct rpcrdma_req *mr_req;
246 struct ib_device *mr_device;
247 struct scatterlist *mr_sg;
249 enum dma_data_direction mr_dir;
250 struct rpcrdma_frwr frwr;
251 struct rpcrdma_xprt *mr_xprt;
255 struct list_head mr_all;
259 * struct rpcrdma_req -- structure central to the request/reply sequence.
261 * N of these are associated with a transport instance, and stored in
262 * struct rpcrdma_buffer. N is the max number of outstanding requests.
264 * It includes pre-registered buffer memory for send AND recv.
265 * The recv buffer, however, is not owned by this structure, and
266 * is "donated" to the hardware when a recv is posted. When a
267 * reply is handled, the recv buffer used is given back to the
268 * struct rpcrdma_req associated with the request.
270 * In addition to the basic memory, this structure includes an array
271 * of iovs for send operations. The reason is that the iovs passed to
272 * ib_post_{send,recv} must not be modified until the work request
276 /* Maximum number of page-sized "segments" per chunk list to be
277 * registered or invalidated. Must handle a Reply chunk:
280 RPCRDMA_MAX_IOV_SEGS = 3,
281 RPCRDMA_MAX_DATA_SEGS = ((1 * 1024 * 1024) / PAGE_SIZE) + 1,
282 RPCRDMA_MAX_SEGS = RPCRDMA_MAX_DATA_SEGS +
283 RPCRDMA_MAX_IOV_SEGS,
286 struct rpcrdma_mr_seg { /* chunk descriptors */
287 u32 mr_len; /* length of chunk or segment */
288 struct page *mr_page; /* owning page, if any */
289 char *mr_offset; /* kva if no page, else offset */
292 /* The Send SGE array is provisioned to send a maximum size
294 * - RPC-over-RDMA header
295 * - xdr_buf head iovec
296 * - RPCRDMA_MAX_INLINE bytes, in pages
297 * - xdr_buf tail iovec
299 * The actual number of array elements consumed by each RPC
300 * depends on the device's max_sge limit.
303 RPCRDMA_MIN_SEND_SGES = 3,
304 RPCRDMA_MAX_PAGE_SGES = RPCRDMA_MAX_INLINE >> PAGE_SHIFT,
305 RPCRDMA_MAX_SEND_SGES = 1 + 1 + RPCRDMA_MAX_PAGE_SGES + 1,
308 struct rpcrdma_buffer;
310 struct list_head rl_list;
311 struct rpc_rqst rl_slot;
312 struct rpcrdma_rep *rl_reply;
313 struct xdr_stream rl_stream;
314 struct xdr_buf rl_hdrbuf;
315 struct ib_send_wr rl_wr;
316 struct rpcrdma_sendctx *rl_sendctx;
317 struct rpcrdma_regbuf *rl_rdmabuf; /* xprt header */
318 struct rpcrdma_regbuf *rl_sendbuf; /* rq_snd_buf */
319 struct rpcrdma_regbuf *rl_recvbuf; /* rq_rcv_buf */
321 struct list_head rl_all;
324 struct list_head rl_free_mrs;
325 struct list_head rl_registered;
326 struct rpcrdma_mr_seg rl_segments[RPCRDMA_MAX_SEGS];
329 static inline struct rpcrdma_req *
330 rpcr_to_rdmar(const struct rpc_rqst *rqst)
332 return container_of(rqst, struct rpcrdma_req, rl_slot);
336 rpcrdma_mr_push(struct rpcrdma_mr *mr, struct list_head *list)
338 list_add(&mr->mr_list, list);
341 static inline struct rpcrdma_mr *
342 rpcrdma_mr_pop(struct list_head *list)
344 struct rpcrdma_mr *mr;
346 mr = list_first_entry_or_null(list, struct rpcrdma_mr, mr_list);
348 list_del_init(&mr->mr_list);
353 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
354 * inline requests/replies, and client/server credits.
356 * One of these is associated with a transport instance
358 struct rpcrdma_buffer {
360 struct list_head rb_send_bufs;
361 struct list_head rb_mrs;
363 unsigned long rb_sc_head;
364 unsigned long rb_sc_tail;
365 unsigned long rb_sc_last;
366 struct rpcrdma_sendctx **rb_sc_ctxs;
368 struct list_head rb_allreqs;
369 struct list_head rb_all_mrs;
370 struct list_head rb_all_reps;
372 struct llist_head rb_free_reps;
374 __be32 rb_max_requests;
375 u32 rb_credits; /* most recent credit grant */
377 u32 rb_bc_srv_max_requests;
378 u32 rb_bc_max_requests;
380 struct work_struct rb_refresh_worker;
384 * Statistics for RPCRDMA
386 struct rpcrdma_stats {
387 /* accessed when sending a call */
388 unsigned long read_chunk_count;
389 unsigned long write_chunk_count;
390 unsigned long reply_chunk_count;
391 unsigned long long total_rdma_request;
393 /* rarely accessed error counters */
394 unsigned long long pullup_copy_count;
395 unsigned long hardway_register_count;
396 unsigned long failed_marshal_count;
397 unsigned long bad_reply_count;
398 unsigned long mrs_recycled;
399 unsigned long mrs_orphaned;
400 unsigned long mrs_allocated;
401 unsigned long empty_sendctx_q;
403 /* accessed when receiving a reply */
404 unsigned long long total_rdma_reply;
405 unsigned long long fixup_copy_count;
406 unsigned long reply_waits_for_send;
407 unsigned long local_inv_needed;
408 unsigned long nomsg_call_count;
409 unsigned long bcall_count;
413 * RPCRDMA transport -- encapsulates the structures above for
414 * integration with RPC.
416 * The contained structures are embedded, not pointers,
417 * for convenience. This structure need not be visible externally.
419 * It is allocated and initialized during mount, and released
422 struct rpcrdma_xprt {
423 struct rpc_xprt rx_xprt;
424 struct rpcrdma_ep *rx_ep;
425 struct rpcrdma_buffer rx_buf;
426 struct delayed_work rx_connect_worker;
427 struct rpc_timeout rx_timeout;
428 struct rpcrdma_stats rx_stats;
431 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
433 static inline const char *
434 rpcrdma_addrstr(const struct rpcrdma_xprt *r_xprt)
436 return r_xprt->rx_xprt.address_strings[RPC_DISPLAY_ADDR];
439 static inline const char *
440 rpcrdma_portstr(const struct rpcrdma_xprt *r_xprt)
442 return r_xprt->rx_xprt.address_strings[RPC_DISPLAY_PORT];
445 /* Setting this to 0 ensures interoperability with early servers.
446 * Setting this to 1 enhances certain unaligned read/write performance.
447 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
448 extern int xprt_rdma_pad_optimize;
450 /* This setting controls the hunt for a supported memory
451 * registration strategy.
453 extern unsigned int xprt_rdma_memreg_strategy;
456 * Endpoint calls - xprtrdma/verbs.c
458 void rpcrdma_flush_disconnect(struct rpcrdma_xprt *r_xprt, struct ib_wc *wc);
459 int rpcrdma_xprt_connect(struct rpcrdma_xprt *r_xprt);
460 void rpcrdma_xprt_disconnect(struct rpcrdma_xprt *r_xprt);
462 int rpcrdma_post_sends(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req);
463 void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp);
466 * Buffer calls - xprtrdma/verbs.c
468 struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
470 int rpcrdma_req_setup(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req);
471 void rpcrdma_req_destroy(struct rpcrdma_req *req);
472 int rpcrdma_buffer_create(struct rpcrdma_xprt *);
473 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
474 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt);
476 struct rpcrdma_mr *rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt);
477 void rpcrdma_mrs_refresh(struct rpcrdma_xprt *r_xprt);
479 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
480 void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers,
481 struct rpcrdma_req *req);
482 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
484 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size,
486 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
487 struct rpcrdma_regbuf *rb);
490 * rpcrdma_regbuf_is_mapped - check if buffer is DMA mapped
492 * Returns true if the buffer is now mapped to rb->rg_device.
494 static inline bool rpcrdma_regbuf_is_mapped(struct rpcrdma_regbuf *rb)
496 return rb->rg_device != NULL;
500 * rpcrdma_regbuf_dma_map - DMA-map a regbuf
501 * @r_xprt: controlling transport instance
502 * @rb: regbuf to be mapped
504 * Returns true if the buffer is currently DMA mapped.
506 static inline bool rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
507 struct rpcrdma_regbuf *rb)
509 if (likely(rpcrdma_regbuf_is_mapped(rb)))
511 return __rpcrdma_regbuf_dma_map(r_xprt, rb);
515 * Wrappers for chunk registration, shared by read/write chunk code.
518 static inline enum dma_data_direction
519 rpcrdma_data_dir(bool writing)
521 return writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
524 /* Memory registration calls xprtrdma/frwr_ops.c
526 void frwr_reset(struct rpcrdma_req *req);
527 int frwr_query_device(struct rpcrdma_ep *ep, const struct ib_device *device);
528 int frwr_mr_init(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr *mr);
529 void frwr_release_mr(struct rpcrdma_mr *mr);
530 struct rpcrdma_mr_seg *frwr_map(struct rpcrdma_xprt *r_xprt,
531 struct rpcrdma_mr_seg *seg,
532 int nsegs, bool writing, __be32 xid,
533 struct rpcrdma_mr *mr);
534 int frwr_send(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req);
535 void frwr_reminv(struct rpcrdma_rep *rep, struct list_head *mrs);
536 void frwr_unmap_sync(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req);
537 void frwr_unmap_async(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req);
540 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
543 enum rpcrdma_chunktype {
553 int rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
554 struct rpcrdma_req *req, u32 hdrlen,
556 enum rpcrdma_chunktype rtype);
557 void rpcrdma_sendctx_unmap(struct rpcrdma_sendctx *sc);
558 int rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst);
559 void rpcrdma_set_max_header_sizes(struct rpcrdma_ep *ep);
560 void rpcrdma_reset_cwnd(struct rpcrdma_xprt *r_xprt);
561 void rpcrdma_complete_rqst(struct rpcrdma_rep *rep);
562 void rpcrdma_reply_handler(struct rpcrdma_rep *rep);
564 static inline void rpcrdma_set_xdrlen(struct xdr_buf *xdr, size_t len)
566 xdr->head[0].iov_len = len;
570 /* RPC/RDMA module init - xprtrdma/transport.c
572 extern unsigned int xprt_rdma_max_inline_read;
573 extern unsigned int xprt_rdma_max_inline_write;
574 void xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap);
575 void xprt_rdma_free_addresses(struct rpc_xprt *xprt);
576 void xprt_rdma_close(struct rpc_xprt *xprt);
577 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq);
578 int xprt_rdma_init(void);
579 void xprt_rdma_cleanup(void);
581 /* Backchannel calls - xprtrdma/backchannel.c
583 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
584 int xprt_rdma_bc_setup(struct rpc_xprt *, unsigned int);
585 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *);
586 unsigned int xprt_rdma_bc_max_slots(struct rpc_xprt *);
587 int rpcrdma_bc_post_recv(struct rpcrdma_xprt *, unsigned int);
588 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *, struct rpcrdma_rep *);
589 int xprt_rdma_bc_send_reply(struct rpc_rqst *rqst);
590 void xprt_rdma_bc_free_rqst(struct rpc_rqst *);
591 void xprt_rdma_bc_destroy(struct rpc_xprt *, unsigned int);
592 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
594 extern struct xprt_class xprt_rdma_bc;
596 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */