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 const struct rpcrdma_memreg_ops *ri_ops;
70 struct ib_device *ri_device;
71 struct rdma_cm_id *ri_id;
73 struct completion ri_done;
74 struct completion ri_remove_done;
76 unsigned int ri_max_segs;
77 unsigned int ri_max_frwr_depth;
78 unsigned int ri_max_inline_write;
79 unsigned int ri_max_inline_read;
80 unsigned int ri_max_send_sges;
81 bool ri_implicit_roundup;
82 enum ib_mr_type ri_mrtype;
83 unsigned long ri_flags;
84 struct ib_qp_attr ri_qp_attr;
85 struct ib_qp_init_attr ri_qp_init_attr;
89 RPCRDMA_IAF_REMOVING = 0,
93 * RDMA Endpoint -- one per transport instance
97 unsigned int rep_send_count;
98 unsigned int rep_send_batch;
100 struct ib_qp_init_attr rep_attr;
101 wait_queue_head_t rep_connect_wait;
102 struct rpcrdma_connect_private rep_cm_private;
103 struct rdma_conn_param rep_remote_cma;
104 int rep_receive_count;
107 /* Pre-allocate extra Work Requests for handling backward receives
108 * and sends. This is a fixed value because the Work Queues are
109 * allocated when the forward channel is set up.
111 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
112 #define RPCRDMA_BACKWARD_WRS (8)
114 #define RPCRDMA_BACKWARD_WRS (0)
117 /* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
119 * The below structure appears at the front of a large region of kmalloc'd
120 * memory, which always starts on a good alignment boundary.
123 struct rpcrdma_regbuf {
124 struct ib_sge rg_iov;
125 struct ib_device *rg_device;
126 enum dma_data_direction rg_direction;
127 __be32 rg_base[0] __attribute__ ((aligned(256)));
131 rdmab_addr(struct rpcrdma_regbuf *rb)
133 return rb->rg_iov.addr;
137 rdmab_length(struct rpcrdma_regbuf *rb)
139 return rb->rg_iov.length;
143 rdmab_lkey(struct rpcrdma_regbuf *rb)
145 return rb->rg_iov.lkey;
148 static inline struct ib_device *
149 rdmab_device(struct rpcrdma_regbuf *rb)
151 return rb->rg_device;
154 #define RPCRDMA_DEF_GFP (GFP_NOIO | __GFP_NOWARN)
156 /* To ensure a transport can always make forward progress,
157 * the number of RDMA segments allowed in header chunk lists
158 * is capped at 8. This prevents less-capable devices and
159 * memory registrations from overrunning the Send buffer
160 * while building chunk lists.
162 * Elements of the Read list take up more room than the
163 * Write list or Reply chunk. 8 read segments means the Read
164 * list (or Write list or Reply chunk) cannot consume more
167 * ((8 + 2) * read segment size) + 1 XDR words, or 244 bytes.
169 * And the fixed part of the header is another 24 bytes.
171 * The smallest inline threshold is 1024 bytes, ensuring that
172 * at least 750 bytes are available for RPC messages.
175 RPCRDMA_MAX_HDR_SEGS = 8,
176 RPCRDMA_HDRBUF_SIZE = 256,
180 * struct rpcrdma_rep -- this structure encapsulates state required
181 * to receive and complete an RPC Reply, asychronously. It needs
182 * several pieces of state:
184 * o receive buffer and ib_sge (donated to provider)
185 * o status of receive (success or not, length, inv rkey)
186 * o bookkeeping state to get run by reply handler (XDR stream)
188 * These structures are allocated during transport initialization.
189 * N of these are associated with a transport instance, managed by
190 * struct rpcrdma_buffer. N is the max number of outstanding RPCs.
194 struct ib_cqe rr_cqe;
201 struct rpcrdma_regbuf *rr_rdmabuf;
202 struct rpcrdma_xprt *rr_rxprt;
203 struct work_struct rr_work;
204 struct xdr_buf rr_hdrbuf;
205 struct xdr_stream rr_stream;
206 struct rpc_rqst *rr_rqst;
207 struct list_head rr_list;
208 struct ib_recv_wr rr_recv_wr;
211 /* struct rpcrdma_sendctx - DMA mapped SGEs to unmap after Send completes
215 struct rpcrdma_sendctx {
216 struct ib_send_wr sc_wr;
217 struct ib_cqe sc_cqe;
218 struct rpcrdma_xprt *sc_xprt;
219 struct rpcrdma_req *sc_req;
220 unsigned int sc_unmap_count;
221 struct ib_sge sc_sges[];
224 /* Limit the number of SGEs that can be unmapped during one
225 * Send completion. This caps the amount of work a single
226 * completion can do before returning to the provider.
228 * Setting this to zero disables Send completion batching.
231 RPCRDMA_MAX_SEND_BATCH = 7,
235 * struct rpcrdma_mr - external memory region metadata
237 * An external memory region is any buffer or page that is registered
238 * on the fly (ie, not pre-registered).
240 * Each rpcrdma_buffer has a list of free MWs anchored in rb_mrs. During
241 * call_allocate, rpcrdma_buffer_get() assigns one to each segment in
242 * an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep
243 * track of registration metadata while each RPC is pending.
244 * rpcrdma_deregister_external() uses this metadata to unmap and
245 * release these resources when an RPC is complete.
247 enum rpcrdma_frwr_state {
248 FRWR_IS_INVALID, /* ready to be used */
249 FRWR_IS_VALID, /* in use */
250 FRWR_FLUSHED_FR, /* flushed FASTREG WR */
251 FRWR_FLUSHED_LI, /* flushed LOCALINV WR */
254 struct rpcrdma_frwr {
256 struct ib_cqe fr_cqe;
257 enum rpcrdma_frwr_state fr_state;
258 struct completion fr_linv_done;
260 struct ib_reg_wr fr_regwr;
261 struct ib_send_wr fr_invwr;
266 struct list_head mr_list;
267 struct scatterlist *mr_sg;
269 enum dma_data_direction mr_dir;
270 struct rpcrdma_frwr frwr;
271 struct rpcrdma_xprt *mr_xprt;
275 struct work_struct mr_recycle;
276 struct list_head mr_all;
280 * struct rpcrdma_req -- structure central to the request/reply sequence.
282 * N of these are associated with a transport instance, and stored in
283 * struct rpcrdma_buffer. N is the max number of outstanding requests.
285 * It includes pre-registered buffer memory for send AND recv.
286 * The recv buffer, however, is not owned by this structure, and
287 * is "donated" to the hardware when a recv is posted. When a
288 * reply is handled, the recv buffer used is given back to the
289 * struct rpcrdma_req associated with the request.
291 * In addition to the basic memory, this structure includes an array
292 * of iovs for send operations. The reason is that the iovs passed to
293 * ib_post_{send,recv} must not be modified until the work request
297 /* Maximum number of page-sized "segments" per chunk list to be
298 * registered or invalidated. Must handle a Reply chunk:
301 RPCRDMA_MAX_IOV_SEGS = 3,
302 RPCRDMA_MAX_DATA_SEGS = ((1 * 1024 * 1024) / PAGE_SIZE) + 1,
303 RPCRDMA_MAX_SEGS = RPCRDMA_MAX_DATA_SEGS +
304 RPCRDMA_MAX_IOV_SEGS,
307 struct rpcrdma_mr_seg { /* chunk descriptors */
308 u32 mr_len; /* length of chunk or segment */
309 struct page *mr_page; /* owning page, if any */
310 char *mr_offset; /* kva if no page, else offset */
313 /* The Send SGE array is provisioned to send a maximum size
315 * - RPC-over-RDMA header
316 * - xdr_buf head iovec
317 * - RPCRDMA_MAX_INLINE bytes, in pages
318 * - xdr_buf tail iovec
320 * The actual number of array elements consumed by each RPC
321 * depends on the device's max_sge limit.
324 RPCRDMA_MIN_SEND_SGES = 3,
325 RPCRDMA_MAX_PAGE_SGES = RPCRDMA_MAX_INLINE >> PAGE_SHIFT,
326 RPCRDMA_MAX_SEND_SGES = 1 + 1 + RPCRDMA_MAX_PAGE_SGES + 1,
329 struct rpcrdma_buffer;
331 struct list_head rl_list;
332 struct rpc_rqst rl_slot;
333 struct rpcrdma_buffer *rl_buffer;
334 struct rpcrdma_rep *rl_reply;
335 struct xdr_stream rl_stream;
336 struct xdr_buf rl_hdrbuf;
337 struct rpcrdma_sendctx *rl_sendctx;
338 struct rpcrdma_regbuf *rl_rdmabuf; /* xprt header */
339 struct rpcrdma_regbuf *rl_sendbuf; /* rq_snd_buf */
340 struct rpcrdma_regbuf *rl_recvbuf; /* rq_rcv_buf */
342 struct list_head rl_all;
343 unsigned long rl_flags;
345 struct list_head rl_registered; /* registered segments */
346 struct rpcrdma_mr_seg rl_segments[RPCRDMA_MAX_SEGS];
351 RPCRDMA_REQ_F_PENDING = 0,
352 RPCRDMA_REQ_F_TX_RESOURCES,
355 static inline struct rpcrdma_req *
356 rpcr_to_rdmar(const struct rpc_rqst *rqst)
358 return container_of(rqst, struct rpcrdma_req, rl_slot);
362 rpcrdma_mr_push(struct rpcrdma_mr *mr, struct list_head *list)
364 list_add_tail(&mr->mr_list, list);
367 static inline struct rpcrdma_mr *
368 rpcrdma_mr_pop(struct list_head *list)
370 struct rpcrdma_mr *mr;
372 mr = list_first_entry(list, struct rpcrdma_mr, mr_list);
373 list_del_init(&mr->mr_list);
378 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
379 * inline requests/replies, and client/server credits.
381 * One of these is associated with a transport instance
383 struct rpcrdma_buffer {
384 spinlock_t rb_mrlock; /* protect rb_mrs list */
385 struct list_head rb_mrs;
386 struct list_head rb_all;
388 unsigned long rb_sc_head;
389 unsigned long rb_sc_tail;
390 unsigned long rb_sc_last;
391 struct rpcrdma_sendctx **rb_sc_ctxs;
393 spinlock_t rb_lock; /* protect buf lists */
394 struct list_head rb_send_bufs;
395 struct list_head rb_recv_bufs;
396 unsigned long rb_flags;
398 u32 rb_credits; /* most recent credit grant */
400 u32 rb_bc_srv_max_requests;
401 spinlock_t rb_reqslock; /* protect rb_allreqs */
402 struct list_head rb_allreqs;
404 u32 rb_bc_max_requests;
406 struct workqueue_struct *rb_completion_wq;
407 struct delayed_work rb_refresh_worker;
409 #define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia)
413 RPCRDMA_BUF_F_EMPTY_SCQ = 0,
417 * Internal structure for transport instance creation. This
418 * exists primarily for modularity.
420 * This data should be set with mount options
422 struct rpcrdma_create_data_internal {
423 unsigned int max_requests; /* max requests (slots) in flight */
424 unsigned int rsize; /* mount rsize - max read hdr+data */
425 unsigned int wsize; /* mount wsize - max write hdr+data */
426 unsigned int inline_rsize; /* max non-rdma read data payload */
427 unsigned int inline_wsize; /* max non-rdma write data payload */
431 * Statistics for RPCRDMA
433 struct rpcrdma_stats {
434 /* accessed when sending a call */
435 unsigned long read_chunk_count;
436 unsigned long write_chunk_count;
437 unsigned long reply_chunk_count;
438 unsigned long long total_rdma_request;
440 /* rarely accessed error counters */
441 unsigned long long pullup_copy_count;
442 unsigned long hardway_register_count;
443 unsigned long failed_marshal_count;
444 unsigned long bad_reply_count;
445 unsigned long mrs_recycled;
446 unsigned long mrs_orphaned;
447 unsigned long mrs_allocated;
448 unsigned long empty_sendctx_q;
450 /* accessed when receiving a reply */
451 unsigned long long total_rdma_reply;
452 unsigned long long fixup_copy_count;
453 unsigned long reply_waits_for_send;
454 unsigned long local_inv_needed;
455 unsigned long nomsg_call_count;
456 unsigned long bcall_count;
460 * Per-registration mode operations
463 struct rpcrdma_memreg_ops {
464 struct rpcrdma_mr_seg *
465 (*ro_map)(struct rpcrdma_xprt *,
466 struct rpcrdma_mr_seg *, int, bool,
467 struct rpcrdma_mr **);
468 int (*ro_send)(struct rpcrdma_ia *ia,
469 struct rpcrdma_req *req);
470 void (*ro_reminv)(struct rpcrdma_rep *rep,
471 struct list_head *mrs);
472 void (*ro_unmap_sync)(struct rpcrdma_xprt *,
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_mr *);
480 void (*ro_release_mr)(struct rpcrdma_mr *mr);
481 const char *ro_displayname;
482 const int ro_send_w_inv_ok;
485 extern const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops;
488 * RPCRDMA transport -- encapsulates the structures above for
489 * integration with RPC.
491 * The contained structures are embedded, not pointers,
492 * for convenience. This structure need not be visible externally.
494 * It is allocated and initialized during mount, and released
497 struct rpcrdma_xprt {
498 struct rpc_xprt rx_xprt;
499 struct rpcrdma_ia rx_ia;
500 struct rpcrdma_ep rx_ep;
501 struct rpcrdma_buffer rx_buf;
502 struct rpcrdma_create_data_internal rx_data;
503 struct delayed_work rx_connect_worker;
504 struct rpcrdma_stats rx_stats;
507 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
508 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data)
510 static inline const char *
511 rpcrdma_addrstr(const struct rpcrdma_xprt *r_xprt)
513 return r_xprt->rx_xprt.address_strings[RPC_DISPLAY_ADDR];
516 static inline const char *
517 rpcrdma_portstr(const struct rpcrdma_xprt *r_xprt)
519 return r_xprt->rx_xprt.address_strings[RPC_DISPLAY_PORT];
522 /* Setting this to 0 ensures interoperability with early servers.
523 * Setting this to 1 enhances certain unaligned read/write performance.
524 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
525 extern int xprt_rdma_pad_optimize;
527 /* This setting controls the hunt for a supported memory
528 * registration strategy.
530 extern unsigned int xprt_rdma_memreg_strategy;
533 * Interface Adapter calls - xprtrdma/verbs.c
535 int rpcrdma_ia_open(struct rpcrdma_xprt *xprt);
536 void rpcrdma_ia_remove(struct rpcrdma_ia *ia);
537 void rpcrdma_ia_close(struct rpcrdma_ia *);
538 bool frwr_is_supported(struct rpcrdma_ia *);
541 * Endpoint calls - xprtrdma/verbs.c
543 int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *,
544 struct rpcrdma_create_data_internal *);
545 void rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *);
546 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *);
547 void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *);
549 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
550 struct rpcrdma_req *);
553 * Buffer calls - xprtrdma/verbs.c
555 struct rpcrdma_req *rpcrdma_create_req(struct rpcrdma_xprt *);
556 void rpcrdma_destroy_req(struct rpcrdma_req *);
557 int rpcrdma_buffer_create(struct rpcrdma_xprt *);
558 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
559 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_buffer *buf);
561 struct rpcrdma_mr *rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt);
562 void rpcrdma_mr_put(struct rpcrdma_mr *mr);
563 void rpcrdma_mr_unmap_and_put(struct rpcrdma_mr *mr);
566 rpcrdma_mr_recycle(struct rpcrdma_mr *mr)
568 schedule_work(&mr->mr_recycle);
571 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
572 void rpcrdma_buffer_put(struct rpcrdma_req *);
573 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
575 struct rpcrdma_regbuf *rpcrdma_alloc_regbuf(size_t, enum dma_data_direction,
577 bool __rpcrdma_dma_map_regbuf(struct rpcrdma_ia *, struct rpcrdma_regbuf *);
578 void rpcrdma_free_regbuf(struct rpcrdma_regbuf *);
581 rpcrdma_regbuf_is_mapped(struct rpcrdma_regbuf *rb)
583 return rb->rg_device != NULL;
587 rpcrdma_dma_map_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
589 if (likely(rpcrdma_regbuf_is_mapped(rb)))
591 return __rpcrdma_dma_map_regbuf(ia, rb);
595 * Wrappers for chunk registration, shared by read/write chunk code.
598 static inline enum dma_data_direction
599 rpcrdma_data_dir(bool writing)
601 return writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
605 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
608 enum rpcrdma_chunktype {
616 int rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
617 struct rpcrdma_req *req, u32 hdrlen,
619 enum rpcrdma_chunktype rtype);
620 void rpcrdma_unmap_sendctx(struct rpcrdma_sendctx *sc);
621 int rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst);
622 void rpcrdma_set_max_header_sizes(struct rpcrdma_xprt *);
623 void rpcrdma_complete_rqst(struct rpcrdma_rep *rep);
624 void rpcrdma_reply_handler(struct rpcrdma_rep *rep);
625 void rpcrdma_release_rqst(struct rpcrdma_xprt *r_xprt,
626 struct rpcrdma_req *req);
627 void rpcrdma_deferred_completion(struct work_struct *work);
629 static inline void rpcrdma_set_xdrlen(struct xdr_buf *xdr, size_t len)
631 xdr->head[0].iov_len = len;
635 /* RPC/RDMA module init - xprtrdma/transport.c
637 extern unsigned int xprt_rdma_max_inline_read;
638 void xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap);
639 void xprt_rdma_free_addresses(struct rpc_xprt *xprt);
640 void xprt_rdma_close(struct rpc_xprt *xprt);
641 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq);
642 int xprt_rdma_init(void);
643 void xprt_rdma_cleanup(void);
645 /* Backchannel calls - xprtrdma/backchannel.c
647 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
648 int xprt_rdma_bc_setup(struct rpc_xprt *, unsigned int);
649 int xprt_rdma_bc_up(struct svc_serv *, struct net *);
650 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *);
651 int rpcrdma_bc_post_recv(struct rpcrdma_xprt *, unsigned int);
652 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *, struct rpcrdma_rep *);
653 int xprt_rdma_bc_send_reply(struct rpc_rqst *rqst);
654 void xprt_rdma_bc_free_rqst(struct rpc_rqst *);
655 void xprt_rdma_bc_destroy(struct rpc_xprt *, unsigned int);
656 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
658 extern struct xprt_class xprt_rdma_bc;
660 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */