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 * RDMA Endpoint -- connection endpoint details
72 struct rdma_cm_id *re_id;
74 unsigned int re_max_rdma_segs;
75 unsigned int re_max_fr_depth;
76 bool re_implicit_roundup;
77 enum ib_mr_type re_mrtype;
78 struct completion re_done;
79 unsigned int re_send_count;
80 unsigned int re_send_batch;
81 unsigned int re_max_inline_send;
82 unsigned int re_max_inline_recv;
84 int re_connect_status;
85 atomic_t re_force_disconnect;
86 struct ib_qp_init_attr re_attr;
87 wait_queue_head_t re_connect_wait;
88 struct rpc_xprt *re_xprt;
89 struct rpcrdma_connect_private
91 struct rdma_conn_param re_remote_cma;
93 unsigned int re_max_requests; /* depends on device */
94 unsigned int re_inline_send; /* negotiated */
95 unsigned int re_inline_recv; /* negotiated */
98 /* Pre-allocate extra Work Requests for handling backward receives
99 * and sends. This is a fixed value because the Work Queues are
100 * allocated when the forward channel is set up, long before the
101 * backchannel is provisioned. This value is two times
102 * NFS4_DEF_CB_SLOT_TABLE_SIZE.
104 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
105 #define RPCRDMA_BACKWARD_WRS (32)
107 #define RPCRDMA_BACKWARD_WRS (0)
110 /* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
113 struct rpcrdma_regbuf {
114 struct ib_sge rg_iov;
115 struct ib_device *rg_device;
116 enum dma_data_direction rg_direction;
120 static inline u64 rdmab_addr(struct rpcrdma_regbuf *rb)
122 return rb->rg_iov.addr;
125 static inline u32 rdmab_length(struct rpcrdma_regbuf *rb)
127 return rb->rg_iov.length;
130 static inline u32 rdmab_lkey(struct rpcrdma_regbuf *rb)
132 return rb->rg_iov.lkey;
135 static inline struct ib_device *rdmab_device(struct rpcrdma_regbuf *rb)
137 return rb->rg_device;
140 static inline void *rdmab_data(const struct rpcrdma_regbuf *rb)
145 #define RPCRDMA_DEF_GFP (GFP_NOIO | __GFP_NOWARN)
147 /* To ensure a transport can always make forward progress,
148 * the number of RDMA segments allowed in header chunk lists
149 * is capped at 16. This prevents less-capable devices from
150 * overrunning the Send buffer while building chunk lists.
152 * Elements of the Read list take up more room than the
153 * Write list or Reply chunk. 16 read segments means the
154 * chunk lists cannot consume more than
156 * ((16 + 2) * read segment size) + 1 XDR words,
158 * or about 400 bytes. The fixed part of the header is
159 * another 24 bytes. Thus when the inline threshold is
160 * 1024 bytes, at least 600 bytes are available for RPC
164 RPCRDMA_MAX_HDR_SEGS = 16,
168 * struct rpcrdma_rep -- this structure encapsulates state required
169 * to receive and complete an RPC Reply, asychronously. It needs
170 * several pieces of state:
172 * o receive buffer and ib_sge (donated to provider)
173 * o status of receive (success or not, length, inv rkey)
174 * o bookkeeping state to get run by reply handler (XDR stream)
176 * These structures are allocated during transport initialization.
177 * N of these are associated with a transport instance, managed by
178 * struct rpcrdma_buffer. N is the max number of outstanding RPCs.
182 struct ib_cqe rr_cqe;
189 struct rpcrdma_regbuf *rr_rdmabuf;
190 struct rpcrdma_xprt *rr_rxprt;
191 struct rpc_rqst *rr_rqst;
192 struct xdr_buf rr_hdrbuf;
193 struct xdr_stream rr_stream;
194 struct llist_node rr_node;
195 struct ib_recv_wr rr_recv_wr;
196 struct list_head rr_all;
199 /* To reduce the rate at which a transport invokes ib_post_recv
200 * (and thus the hardware doorbell rate), xprtrdma posts Receive
203 * Setting this to zero disables Receive post batching.
206 RPCRDMA_MAX_RECV_BATCH = 7,
209 /* struct rpcrdma_sendctx - DMA mapped SGEs to unmap after Send completes
212 struct rpcrdma_sendctx {
213 struct ib_cqe sc_cqe;
214 struct rpcrdma_req *sc_req;
215 unsigned int sc_unmap_count;
216 struct ib_sge sc_sges[];
220 * struct rpcrdma_mr - external memory region metadata
222 * An external memory region is any buffer or page that is registered
223 * on the fly (ie, not pre-registered).
225 struct rpcrdma_frwr {
227 struct ib_cqe fr_cqe;
228 struct completion fr_linv_done;
230 struct ib_reg_wr fr_regwr;
231 struct ib_send_wr fr_invwr;
237 struct list_head mr_list;
238 struct rpcrdma_req *mr_req;
239 struct scatterlist *mr_sg;
241 enum dma_data_direction mr_dir;
242 struct rpcrdma_frwr frwr;
243 struct rpcrdma_xprt *mr_xprt;
247 struct list_head mr_all;
251 * struct rpcrdma_req -- structure central to the request/reply sequence.
253 * N of these are associated with a transport instance, and stored in
254 * struct rpcrdma_buffer. N is the max number of outstanding requests.
256 * It includes pre-registered buffer memory for send AND recv.
257 * The recv buffer, however, is not owned by this structure, and
258 * is "donated" to the hardware when a recv is posted. When a
259 * reply is handled, the recv buffer used is given back to the
260 * struct rpcrdma_req associated with the request.
262 * In addition to the basic memory, this structure includes an array
263 * of iovs for send operations. The reason is that the iovs passed to
264 * ib_post_{send,recv} must not be modified until the work request
268 /* Maximum number of page-sized "segments" per chunk list to be
269 * registered or invalidated. Must handle a Reply chunk:
272 RPCRDMA_MAX_IOV_SEGS = 3,
273 RPCRDMA_MAX_DATA_SEGS = ((1 * 1024 * 1024) / PAGE_SIZE) + 1,
274 RPCRDMA_MAX_SEGS = RPCRDMA_MAX_DATA_SEGS +
275 RPCRDMA_MAX_IOV_SEGS,
278 struct rpcrdma_mr_seg { /* chunk descriptors */
279 u32 mr_len; /* length of chunk or segment */
280 struct page *mr_page; /* owning page, if any */
281 char *mr_offset; /* kva if no page, else offset */
284 /* The Send SGE array is provisioned to send a maximum size
286 * - RPC-over-RDMA header
287 * - xdr_buf head iovec
288 * - RPCRDMA_MAX_INLINE bytes, in pages
289 * - xdr_buf tail iovec
291 * The actual number of array elements consumed by each RPC
292 * depends on the device's max_sge limit.
295 RPCRDMA_MIN_SEND_SGES = 3,
296 RPCRDMA_MAX_PAGE_SGES = RPCRDMA_MAX_INLINE >> PAGE_SHIFT,
297 RPCRDMA_MAX_SEND_SGES = 1 + 1 + RPCRDMA_MAX_PAGE_SGES + 1,
300 struct rpcrdma_buffer;
302 struct list_head rl_list;
303 struct rpc_rqst rl_slot;
304 struct rpcrdma_rep *rl_reply;
305 struct xdr_stream rl_stream;
306 struct xdr_buf rl_hdrbuf;
307 struct ib_send_wr rl_wr;
308 struct rpcrdma_sendctx *rl_sendctx;
309 struct rpcrdma_regbuf *rl_rdmabuf; /* xprt header */
310 struct rpcrdma_regbuf *rl_sendbuf; /* rq_snd_buf */
311 struct rpcrdma_regbuf *rl_recvbuf; /* rq_rcv_buf */
313 struct list_head rl_all;
316 struct list_head rl_free_mrs;
317 struct list_head rl_registered;
318 struct rpcrdma_mr_seg rl_segments[RPCRDMA_MAX_SEGS];
321 static inline struct rpcrdma_req *
322 rpcr_to_rdmar(const struct rpc_rqst *rqst)
324 return container_of(rqst, struct rpcrdma_req, rl_slot);
328 rpcrdma_mr_push(struct rpcrdma_mr *mr, struct list_head *list)
330 list_add(&mr->mr_list, list);
333 static inline struct rpcrdma_mr *
334 rpcrdma_mr_pop(struct list_head *list)
336 struct rpcrdma_mr *mr;
338 mr = list_first_entry_or_null(list, struct rpcrdma_mr, mr_list);
340 list_del_init(&mr->mr_list);
345 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
346 * inline requests/replies, and client/server credits.
348 * One of these is associated with a transport instance
350 struct rpcrdma_buffer {
352 struct list_head rb_send_bufs;
353 struct list_head rb_mrs;
355 unsigned long rb_sc_head;
356 unsigned long rb_sc_tail;
357 unsigned long rb_sc_last;
358 struct rpcrdma_sendctx **rb_sc_ctxs;
360 struct list_head rb_allreqs;
361 struct list_head rb_all_mrs;
362 struct list_head rb_all_reps;
364 struct llist_head rb_free_reps;
366 __be32 rb_max_requests;
367 u32 rb_credits; /* most recent credit grant */
369 u32 rb_bc_srv_max_requests;
370 u32 rb_bc_max_requests;
372 struct work_struct rb_refresh_worker;
376 * Statistics for RPCRDMA
378 struct rpcrdma_stats {
379 /* accessed when sending a call */
380 unsigned long read_chunk_count;
381 unsigned long write_chunk_count;
382 unsigned long reply_chunk_count;
383 unsigned long long total_rdma_request;
385 /* rarely accessed error counters */
386 unsigned long long pullup_copy_count;
387 unsigned long hardway_register_count;
388 unsigned long failed_marshal_count;
389 unsigned long bad_reply_count;
390 unsigned long mrs_recycled;
391 unsigned long mrs_orphaned;
392 unsigned long mrs_allocated;
393 unsigned long empty_sendctx_q;
395 /* accessed when receiving a reply */
396 unsigned long long total_rdma_reply;
397 unsigned long long fixup_copy_count;
398 unsigned long reply_waits_for_send;
399 unsigned long local_inv_needed;
400 unsigned long nomsg_call_count;
401 unsigned long bcall_count;
405 * RPCRDMA transport -- encapsulates the structures above for
406 * integration with RPC.
408 * The contained structures are embedded, not pointers,
409 * for convenience. This structure need not be visible externally.
411 * It is allocated and initialized during mount, and released
414 struct rpcrdma_xprt {
415 struct rpc_xprt rx_xprt;
416 struct rpcrdma_ep *rx_ep;
417 struct rpcrdma_buffer rx_buf;
418 struct delayed_work rx_connect_worker;
419 struct rpc_timeout rx_timeout;
420 struct rpcrdma_stats rx_stats;
423 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
425 static inline const char *
426 rpcrdma_addrstr(const struct rpcrdma_xprt *r_xprt)
428 return r_xprt->rx_xprt.address_strings[RPC_DISPLAY_ADDR];
431 static inline const char *
432 rpcrdma_portstr(const struct rpcrdma_xprt *r_xprt)
434 return r_xprt->rx_xprt.address_strings[RPC_DISPLAY_PORT];
437 /* Setting this to 0 ensures interoperability with early servers.
438 * Setting this to 1 enhances certain unaligned read/write performance.
439 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
440 extern int xprt_rdma_pad_optimize;
442 /* This setting controls the hunt for a supported memory
443 * registration strategy.
445 extern unsigned int xprt_rdma_memreg_strategy;
448 * Endpoint calls - xprtrdma/verbs.c
450 void rpcrdma_flush_disconnect(struct rpcrdma_xprt *r_xprt, struct ib_wc *wc);
451 int rpcrdma_xprt_connect(struct rpcrdma_xprt *r_xprt);
452 void rpcrdma_xprt_disconnect(struct rpcrdma_xprt *r_xprt);
454 int rpcrdma_post_sends(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req);
455 void rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, int needed, bool temp);
458 * Buffer calls - xprtrdma/verbs.c
460 struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
462 int rpcrdma_req_setup(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req);
463 void rpcrdma_req_destroy(struct rpcrdma_req *req);
464 int rpcrdma_buffer_create(struct rpcrdma_xprt *);
465 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
466 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt);
468 struct rpcrdma_mr *rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt);
469 void rpcrdma_mr_put(struct rpcrdma_mr *mr);
470 void rpcrdma_mrs_refresh(struct rpcrdma_xprt *r_xprt);
472 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
473 void rpcrdma_buffer_put(struct rpcrdma_buffer *buffers,
474 struct rpcrdma_req *req);
475 void rpcrdma_reply_put(struct rpcrdma_buffer *buffers, struct rpcrdma_req *req);
476 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
478 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size,
480 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
481 struct rpcrdma_regbuf *rb);
484 * rpcrdma_regbuf_is_mapped - check if buffer is DMA mapped
486 * Returns true if the buffer is now mapped to rb->rg_device.
488 static inline bool rpcrdma_regbuf_is_mapped(struct rpcrdma_regbuf *rb)
490 return rb->rg_device != NULL;
494 * rpcrdma_regbuf_dma_map - DMA-map a regbuf
495 * @r_xprt: controlling transport instance
496 * @rb: regbuf to be mapped
498 * Returns true if the buffer is currently DMA mapped.
500 static inline bool rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
501 struct rpcrdma_regbuf *rb)
503 if (likely(rpcrdma_regbuf_is_mapped(rb)))
505 return __rpcrdma_regbuf_dma_map(r_xprt, rb);
509 * Wrappers for chunk registration, shared by read/write chunk code.
512 static inline enum dma_data_direction
513 rpcrdma_data_dir(bool writing)
515 return writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
518 /* Memory registration calls xprtrdma/frwr_ops.c
520 void frwr_reset(struct rpcrdma_req *req);
521 int frwr_query_device(struct rpcrdma_ep *ep, const struct ib_device *device);
522 int frwr_mr_init(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr *mr);
523 void frwr_release_mr(struct rpcrdma_mr *mr);
524 struct rpcrdma_mr_seg *frwr_map(struct rpcrdma_xprt *r_xprt,
525 struct rpcrdma_mr_seg *seg,
526 int nsegs, bool writing, __be32 xid,
527 struct rpcrdma_mr *mr);
528 int frwr_send(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req);
529 void frwr_reminv(struct rpcrdma_rep *rep, struct list_head *mrs);
530 void frwr_unmap_sync(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req);
531 void frwr_unmap_async(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req);
534 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
537 enum rpcrdma_chunktype {
547 int rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
548 struct rpcrdma_req *req, u32 hdrlen,
550 enum rpcrdma_chunktype rtype);
551 void rpcrdma_sendctx_unmap(struct rpcrdma_sendctx *sc);
552 int rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst);
553 void rpcrdma_set_max_header_sizes(struct rpcrdma_ep *ep);
554 void rpcrdma_reset_cwnd(struct rpcrdma_xprt *r_xprt);
555 void rpcrdma_complete_rqst(struct rpcrdma_rep *rep);
556 void rpcrdma_reply_handler(struct rpcrdma_rep *rep);
558 static inline void rpcrdma_set_xdrlen(struct xdr_buf *xdr, size_t len)
560 xdr->head[0].iov_len = len;
564 /* RPC/RDMA module init - xprtrdma/transport.c
566 extern unsigned int xprt_rdma_max_inline_read;
567 extern unsigned int xprt_rdma_max_inline_write;
568 void xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap);
569 void xprt_rdma_free_addresses(struct rpc_xprt *xprt);
570 void xprt_rdma_close(struct rpc_xprt *xprt);
571 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq);
572 int xprt_rdma_init(void);
573 void xprt_rdma_cleanup(void);
575 /* Backchannel calls - xprtrdma/backchannel.c
577 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
578 int xprt_rdma_bc_setup(struct rpc_xprt *, unsigned int);
579 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *);
580 unsigned int xprt_rdma_bc_max_slots(struct rpc_xprt *);
581 int rpcrdma_bc_post_recv(struct rpcrdma_xprt *, unsigned int);
582 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *, struct rpcrdma_rep *);
583 int xprt_rdma_bc_send_reply(struct rpc_rqst *rqst);
584 void xprt_rdma_bc_free_rqst(struct rpc_rqst *);
585 void xprt_rdma_bc_destroy(struct rpc_xprt *, unsigned int);
586 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
588 extern struct xprt_class xprt_rdma_bc;
590 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */