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.h> /* RPC/RDMA protocol */
57 #include <linux/sunrpc/xprtrdma.h> /* xprt parameters */
59 #define RDMA_RESOLVE_TIMEOUT (5000) /* 5 seconds */
60 #define RDMA_CONNECT_RETRY_MAX (2) /* retries if no listener backlog */
62 #define RPCRDMA_BIND_TO (60U * HZ)
63 #define RPCRDMA_INIT_REEST_TO (5U * HZ)
64 #define RPCRDMA_MAX_REEST_TO (30U * HZ)
65 #define RPCRDMA_IDLE_DISC_TO (5U * 60 * HZ)
68 * Interface Adapter -- one per transport instance
71 struct rdma_cm_id *ri_id;
74 unsigned int ri_max_segs;
75 unsigned int ri_max_frwr_depth;
76 unsigned int ri_max_send_sges;
77 bool ri_implicit_roundup;
78 enum ib_mr_type ri_mrtype;
79 unsigned long ri_flags;
80 struct completion ri_done;
81 struct completion ri_remove_done;
85 RPCRDMA_IAF_REMOVING = 0,
89 * RDMA Endpoint -- one per transport instance
93 unsigned int rep_send_count;
94 unsigned int rep_send_batch;
95 unsigned int rep_max_inline_send;
96 unsigned int rep_max_inline_recv;
98 struct ib_qp_init_attr rep_attr;
99 wait_queue_head_t rep_connect_wait;
100 struct rpcrdma_connect_private rep_cm_private;
101 struct rdma_conn_param rep_remote_cma;
102 unsigned int rep_max_requests; /* set by /proc */
103 unsigned int rep_inline_send; /* negotiated */
104 unsigned int rep_inline_recv; /* negotiated */
105 int rep_receive_count;
108 /* Pre-allocate extra Work Requests for handling backward receives
109 * and sends. This is a fixed value because the Work Queues are
110 * allocated when the forward channel is set up, long before the
111 * backchannel is provisioned. This value is two times
112 * NFS4_DEF_CB_SLOT_TABLE_SIZE.
114 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
115 #define RPCRDMA_BACKWARD_WRS (32)
117 #define RPCRDMA_BACKWARD_WRS (0)
120 /* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
123 struct rpcrdma_regbuf {
124 struct ib_sge rg_iov;
125 struct ib_device *rg_device;
126 enum dma_data_direction rg_direction;
130 static inline u64 rdmab_addr(struct rpcrdma_regbuf *rb)
132 return rb->rg_iov.addr;
135 static inline u32 rdmab_length(struct rpcrdma_regbuf *rb)
137 return rb->rg_iov.length;
140 static inline u32 rdmab_lkey(struct rpcrdma_regbuf *rb)
142 return rb->rg_iov.lkey;
145 static inline struct ib_device *rdmab_device(struct rpcrdma_regbuf *rb)
147 return rb->rg_device;
150 static inline void *rdmab_data(const struct rpcrdma_regbuf *rb)
155 #define RPCRDMA_DEF_GFP (GFP_NOIO | __GFP_NOWARN)
157 /* To ensure a transport can always make forward progress,
158 * the number of RDMA segments allowed in header chunk lists
159 * is capped at 16. This prevents less-capable devices from
160 * overrunning the Send buffer while building chunk lists.
162 * Elements of the Read list take up more room than the
163 * Write list or Reply chunk. 16 read segments means the
164 * chunk lists cannot consume more than
166 * ((16 + 2) * read segment size) + 1 XDR words,
168 * or about 400 bytes. The fixed part of the header is
169 * another 24 bytes. Thus when the inline threshold is
170 * 1024 bytes, at least 600 bytes are available for RPC
174 RPCRDMA_MAX_HDR_SEGS = 16,
178 * struct rpcrdma_rep -- this structure encapsulates state required
179 * to receive and complete an RPC Reply, asychronously. It needs
180 * several pieces of state:
182 * o receive buffer and ib_sge (donated to provider)
183 * o status of receive (success or not, length, inv rkey)
184 * o bookkeeping state to get run by reply handler (XDR stream)
186 * These structures are allocated during transport initialization.
187 * N of these are associated with a transport instance, managed by
188 * struct rpcrdma_buffer. N is the max number of outstanding RPCs.
192 struct ib_cqe rr_cqe;
199 struct rpcrdma_regbuf *rr_rdmabuf;
200 struct rpcrdma_xprt *rr_rxprt;
201 struct rpc_rqst *rr_rqst;
202 struct xdr_buf rr_hdrbuf;
203 struct xdr_stream rr_stream;
204 struct llist_node rr_node;
205 struct ib_recv_wr rr_recv_wr;
208 /* To reduce the rate at which a transport invokes ib_post_recv
209 * (and thus the hardware doorbell rate), xprtrdma posts Receive
212 * Setting this to zero disables Receive post batching.
215 RPCRDMA_MAX_RECV_BATCH = 7,
218 /* struct rpcrdma_sendctx - DMA mapped SGEs to unmap after Send completes
222 struct rpcrdma_sendctx {
223 struct ib_send_wr sc_wr;
224 struct ib_cqe sc_cqe;
225 struct ib_device *sc_device;
226 struct rpcrdma_xprt *sc_xprt;
227 struct rpcrdma_req *sc_req;
228 unsigned int sc_unmap_count;
229 struct ib_sge sc_sges[];
233 * struct rpcrdma_mr - external memory region metadata
235 * An external memory region is any buffer or page that is registered
236 * on the fly (ie, not pre-registered).
238 struct rpcrdma_frwr {
240 struct ib_cqe fr_cqe;
241 struct completion fr_linv_done;
243 struct ib_reg_wr fr_regwr;
244 struct ib_send_wr fr_invwr;
250 struct list_head mr_list;
251 struct rpcrdma_req *mr_req;
252 struct scatterlist *mr_sg;
254 enum dma_data_direction mr_dir;
255 struct rpcrdma_frwr frwr;
256 struct rpcrdma_xprt *mr_xprt;
260 struct work_struct mr_recycle;
261 struct list_head mr_all;
265 * struct rpcrdma_req -- structure central to the request/reply sequence.
267 * N of these are associated with a transport instance, and stored in
268 * struct rpcrdma_buffer. N is the max number of outstanding requests.
270 * It includes pre-registered buffer memory for send AND recv.
271 * The recv buffer, however, is not owned by this structure, and
272 * is "donated" to the hardware when a recv is posted. When a
273 * reply is handled, the recv buffer used is given back to the
274 * struct rpcrdma_req associated with the request.
276 * In addition to the basic memory, this structure includes an array
277 * of iovs for send operations. The reason is that the iovs passed to
278 * ib_post_{send,recv} must not be modified until the work request
282 /* Maximum number of page-sized "segments" per chunk list to be
283 * registered or invalidated. Must handle a Reply chunk:
286 RPCRDMA_MAX_IOV_SEGS = 3,
287 RPCRDMA_MAX_DATA_SEGS = ((1 * 1024 * 1024) / PAGE_SIZE) + 1,
288 RPCRDMA_MAX_SEGS = RPCRDMA_MAX_DATA_SEGS +
289 RPCRDMA_MAX_IOV_SEGS,
292 struct rpcrdma_mr_seg { /* chunk descriptors */
293 u32 mr_len; /* length of chunk or segment */
294 struct page *mr_page; /* owning page, if any */
295 char *mr_offset; /* kva if no page, else offset */
298 /* The Send SGE array is provisioned to send a maximum size
300 * - RPC-over-RDMA header
301 * - xdr_buf head iovec
302 * - RPCRDMA_MAX_INLINE bytes, in pages
303 * - xdr_buf tail iovec
305 * The actual number of array elements consumed by each RPC
306 * depends on the device's max_sge limit.
309 RPCRDMA_MIN_SEND_SGES = 3,
310 RPCRDMA_MAX_PAGE_SGES = RPCRDMA_MAX_INLINE >> PAGE_SHIFT,
311 RPCRDMA_MAX_SEND_SGES = 1 + 1 + RPCRDMA_MAX_PAGE_SGES + 1,
314 struct rpcrdma_buffer;
316 struct list_head rl_list;
317 struct rpc_rqst rl_slot;
318 struct rpcrdma_rep *rl_reply;
319 struct xdr_stream rl_stream;
320 struct xdr_buf rl_hdrbuf;
321 struct rpcrdma_sendctx *rl_sendctx;
322 struct rpcrdma_regbuf *rl_rdmabuf; /* xprt header */
323 struct rpcrdma_regbuf *rl_sendbuf; /* rq_snd_buf */
324 struct rpcrdma_regbuf *rl_recvbuf; /* rq_rcv_buf */
326 struct list_head rl_all;
329 struct list_head rl_free_mrs;
330 struct list_head rl_registered;
331 struct rpcrdma_mr_seg rl_segments[RPCRDMA_MAX_SEGS];
334 static inline struct rpcrdma_req *
335 rpcr_to_rdmar(const struct rpc_rqst *rqst)
337 return container_of(rqst, struct rpcrdma_req, rl_slot);
341 rpcrdma_mr_push(struct rpcrdma_mr *mr, struct list_head *list)
343 list_add(&mr->mr_list, list);
346 static inline struct rpcrdma_mr *
347 rpcrdma_mr_pop(struct list_head *list)
349 struct rpcrdma_mr *mr;
351 mr = list_first_entry_or_null(list, struct rpcrdma_mr, mr_list);
353 list_del_init(&mr->mr_list);
358 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
359 * inline requests/replies, and client/server credits.
361 * One of these is associated with a transport instance
363 struct rpcrdma_buffer {
365 struct list_head rb_send_bufs;
366 struct list_head rb_mrs;
368 unsigned long rb_sc_head;
369 unsigned long rb_sc_tail;
370 unsigned long rb_sc_last;
371 struct rpcrdma_sendctx **rb_sc_ctxs;
373 struct list_head rb_allreqs;
374 struct list_head rb_all_mrs;
376 struct llist_head rb_free_reps;
379 u32 rb_credits; /* most recent credit grant */
381 u32 rb_bc_srv_max_requests;
382 u32 rb_bc_max_requests;
384 struct work_struct rb_refresh_worker;
388 * Statistics for RPCRDMA
390 struct rpcrdma_stats {
391 /* accessed when sending a call */
392 unsigned long read_chunk_count;
393 unsigned long write_chunk_count;
394 unsigned long reply_chunk_count;
395 unsigned long long total_rdma_request;
397 /* rarely accessed error counters */
398 unsigned long long pullup_copy_count;
399 unsigned long hardway_register_count;
400 unsigned long failed_marshal_count;
401 unsigned long bad_reply_count;
402 unsigned long mrs_recycled;
403 unsigned long mrs_orphaned;
404 unsigned long mrs_allocated;
405 unsigned long empty_sendctx_q;
407 /* accessed when receiving a reply */
408 unsigned long long total_rdma_reply;
409 unsigned long long fixup_copy_count;
410 unsigned long reply_waits_for_send;
411 unsigned long local_inv_needed;
412 unsigned long nomsg_call_count;
413 unsigned long bcall_count;
417 * RPCRDMA transport -- encapsulates the structures above for
418 * integration with RPC.
420 * The contained structures are embedded, not pointers,
421 * for convenience. This structure need not be visible externally.
423 * It is allocated and initialized during mount, and released
426 struct rpcrdma_xprt {
427 struct rpc_xprt rx_xprt;
428 struct rpcrdma_ia rx_ia;
429 struct rpcrdma_ep rx_ep;
430 struct rpcrdma_buffer rx_buf;
431 struct delayed_work rx_connect_worker;
432 struct rpc_timeout rx_timeout;
433 struct rpcrdma_stats rx_stats;
436 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
438 static inline const char *
439 rpcrdma_addrstr(const struct rpcrdma_xprt *r_xprt)
441 return r_xprt->rx_xprt.address_strings[RPC_DISPLAY_ADDR];
444 static inline const char *
445 rpcrdma_portstr(const struct rpcrdma_xprt *r_xprt)
447 return r_xprt->rx_xprt.address_strings[RPC_DISPLAY_PORT];
450 /* Setting this to 0 ensures interoperability with early servers.
451 * Setting this to 1 enhances certain unaligned read/write performance.
452 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
453 extern int xprt_rdma_pad_optimize;
455 /* This setting controls the hunt for a supported memory
456 * registration strategy.
458 extern unsigned int xprt_rdma_memreg_strategy;
461 * Interface Adapter calls - xprtrdma/verbs.c
463 int rpcrdma_ia_open(struct rpcrdma_xprt *xprt);
464 void rpcrdma_ia_remove(struct rpcrdma_ia *ia);
465 void rpcrdma_ia_close(struct rpcrdma_ia *);
468 * Endpoint calls - xprtrdma/verbs.c
470 int rpcrdma_ep_create(struct rpcrdma_xprt *r_xprt);
471 void rpcrdma_ep_destroy(struct rpcrdma_xprt *r_xprt);
472 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *);
473 void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *);
475 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
476 struct rpcrdma_req *);
479 * Buffer calls - xprtrdma/verbs.c
481 struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
483 void rpcrdma_req_destroy(struct rpcrdma_req *req);
484 int rpcrdma_buffer_create(struct rpcrdma_xprt *);
485 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
486 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt);
488 struct rpcrdma_mr *rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt);
489 void rpcrdma_mr_put(struct rpcrdma_mr *mr);
492 rpcrdma_mr_recycle(struct rpcrdma_mr *mr)
494 schedule_work(&mr->mr_recycle);
497 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
498 void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers,
499 struct rpcrdma_req *req);
500 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
502 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size,
504 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
505 struct rpcrdma_regbuf *rb);
508 * rpcrdma_regbuf_is_mapped - check if buffer is DMA mapped
510 * Returns true if the buffer is now mapped to rb->rg_device.
512 static inline bool rpcrdma_regbuf_is_mapped(struct rpcrdma_regbuf *rb)
514 return rb->rg_device != NULL;
518 * rpcrdma_regbuf_dma_map - DMA-map a regbuf
519 * @r_xprt: controlling transport instance
520 * @rb: regbuf to be mapped
522 * Returns true if the buffer is currently DMA mapped.
524 static inline bool rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
525 struct rpcrdma_regbuf *rb)
527 if (likely(rpcrdma_regbuf_is_mapped(rb)))
529 return __rpcrdma_regbuf_dma_map(r_xprt, rb);
533 * Wrappers for chunk registration, shared by read/write chunk code.
536 static inline enum dma_data_direction
537 rpcrdma_data_dir(bool writing)
539 return writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
542 /* Memory registration calls xprtrdma/frwr_ops.c
544 bool frwr_is_supported(struct ib_device *device);
545 void frwr_reset(struct rpcrdma_req *req);
546 int frwr_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep);
547 int frwr_init_mr(struct rpcrdma_ia *ia, struct rpcrdma_mr *mr);
548 void frwr_release_mr(struct rpcrdma_mr *mr);
549 size_t frwr_maxpages(struct rpcrdma_xprt *r_xprt);
550 struct rpcrdma_mr_seg *frwr_map(struct rpcrdma_xprt *r_xprt,
551 struct rpcrdma_mr_seg *seg,
552 int nsegs, bool writing, __be32 xid,
553 struct rpcrdma_mr *mr);
554 int frwr_send(struct rpcrdma_ia *ia, struct rpcrdma_req *req);
555 void frwr_reminv(struct rpcrdma_rep *rep, struct list_head *mrs);
556 void frwr_unmap_sync(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req);
557 void frwr_unmap_async(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req);
560 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
563 enum rpcrdma_chunktype {
571 int rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
572 struct rpcrdma_req *req, u32 hdrlen,
574 enum rpcrdma_chunktype rtype);
575 void rpcrdma_sendctx_unmap(struct rpcrdma_sendctx *sc);
576 int rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst);
577 void rpcrdma_set_max_header_sizes(struct rpcrdma_xprt *);
578 void rpcrdma_complete_rqst(struct rpcrdma_rep *rep);
579 void rpcrdma_reply_handler(struct rpcrdma_rep *rep);
581 static inline void rpcrdma_set_xdrlen(struct xdr_buf *xdr, size_t len)
583 xdr->head[0].iov_len = len;
587 /* RPC/RDMA module init - xprtrdma/transport.c
589 extern unsigned int xprt_rdma_slot_table_entries;
590 extern unsigned int xprt_rdma_max_inline_read;
591 extern unsigned int xprt_rdma_max_inline_write;
592 void xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap);
593 void xprt_rdma_free_addresses(struct rpc_xprt *xprt);
594 void xprt_rdma_close(struct rpc_xprt *xprt);
595 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq);
596 int xprt_rdma_init(void);
597 void xprt_rdma_cleanup(void);
599 /* Backchannel calls - xprtrdma/backchannel.c
601 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
602 int xprt_rdma_bc_setup(struct rpc_xprt *, unsigned int);
603 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *);
604 unsigned int xprt_rdma_bc_max_slots(struct rpc_xprt *);
605 int rpcrdma_bc_post_recv(struct rpcrdma_xprt *, unsigned int);
606 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *, struct rpcrdma_rep *);
607 int xprt_rdma_bc_send_reply(struct rpc_rqst *rqst);
608 void xprt_rdma_bc_free_rqst(struct rpc_rqst *);
609 void xprt_rdma_bc_destroy(struct rpc_xprt *, unsigned int);
610 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
612 extern struct xprt_class xprt_rdma_bc;
614 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */