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_inline_write;
75 unsigned int ri_max_inline_read;
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;
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 int rep_receive_count;
103 /* Pre-allocate extra Work Requests for handling backward receives
104 * and sends. This is a fixed value because the Work Queues are
105 * allocated when the forward channel is set up, long before the
106 * backchannel is provisioned. This value is two times
107 * NFS4_DEF_CB_SLOT_TABLE_SIZE.
109 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
110 #define RPCRDMA_BACKWARD_WRS (32)
112 #define RPCRDMA_BACKWARD_WRS (0)
115 /* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
117 * The below structure appears at the front of a large region of kmalloc'd
118 * memory, which always starts on a good alignment boundary.
121 struct rpcrdma_regbuf {
122 struct ib_sge rg_iov;
123 struct ib_device *rg_device;
124 enum dma_data_direction rg_direction;
128 static inline u64 rdmab_addr(struct rpcrdma_regbuf *rb)
130 return rb->rg_iov.addr;
133 static inline u32 rdmab_length(struct rpcrdma_regbuf *rb)
135 return rb->rg_iov.length;
138 static inline u32 rdmab_lkey(struct rpcrdma_regbuf *rb)
140 return rb->rg_iov.lkey;
143 static inline struct ib_device *rdmab_device(struct rpcrdma_regbuf *rb)
145 return rb->rg_device;
148 static inline void *rdmab_data(const struct rpcrdma_regbuf *rb)
153 #define RPCRDMA_DEF_GFP (GFP_NOIO | __GFP_NOWARN)
155 /* To ensure a transport can always make forward progress,
156 * the number of RDMA segments allowed in header chunk lists
157 * is capped at 8. This prevents less-capable devices and
158 * memory registrations from overrunning the Send buffer
159 * while building chunk lists.
161 * Elements of the Read list take up more room than the
162 * Write list or Reply chunk. 8 read segments means the Read
163 * list (or Write list or Reply chunk) cannot consume more
166 * ((8 + 2) * read segment size) + 1 XDR words, or 244 bytes.
168 * And the fixed part of the header is another 24 bytes.
170 * The smallest inline threshold is 1024 bytes, ensuring that
171 * at least 750 bytes are available for RPC messages.
174 RPCRDMA_MAX_HDR_SEGS = 8,
175 RPCRDMA_HDRBUF_SIZE = 256,
179 * struct rpcrdma_rep -- this structure encapsulates state required
180 * to receive and complete an RPC Reply, asychronously. It needs
181 * several pieces of state:
183 * o receive buffer and ib_sge (donated to provider)
184 * o status of receive (success or not, length, inv rkey)
185 * o bookkeeping state to get run by reply handler (XDR stream)
187 * These structures are allocated during transport initialization.
188 * N of these are associated with a transport instance, managed by
189 * struct rpcrdma_buffer. N is the max number of outstanding RPCs.
193 struct ib_cqe rr_cqe;
200 struct rpcrdma_regbuf *rr_rdmabuf;
201 struct rpcrdma_xprt *rr_rxprt;
202 struct work_struct rr_work;
203 struct xdr_buf rr_hdrbuf;
204 struct xdr_stream rr_stream;
205 struct rpc_rqst *rr_rqst;
206 struct list_head rr_list;
207 struct ib_recv_wr rr_recv_wr;
210 /* To reduce the rate at which a transport invokes ib_post_recv
211 * (and thus the hardware doorbell rate), xprtrdma posts Receive
214 * Setting this to zero disables Receive post batching.
217 RPCRDMA_MAX_RECV_BATCH = 7,
220 /* struct rpcrdma_sendctx - DMA mapped SGEs to unmap after Send completes
224 struct rpcrdma_sendctx {
225 struct ib_send_wr sc_wr;
226 struct ib_cqe sc_cqe;
227 struct ib_device *sc_device;
228 struct rpcrdma_xprt *sc_xprt;
229 struct rpcrdma_req *sc_req;
230 unsigned int sc_unmap_count;
231 struct ib_sge sc_sges[];
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 struct list_head rb_allreqs;
398 unsigned long rb_flags;
400 u32 rb_credits; /* most recent credit grant */
402 u32 rb_bc_srv_max_requests;
403 u32 rb_bc_max_requests;
405 struct workqueue_struct *rb_completion_wq;
406 struct delayed_work rb_refresh_worker;
411 RPCRDMA_BUF_F_EMPTY_SCQ = 0,
415 * Internal structure for transport instance creation. This
416 * exists primarily for modularity.
418 * This data should be set with mount options
420 struct rpcrdma_create_data_internal {
421 unsigned int max_requests; /* max requests (slots) in flight */
422 unsigned int rsize; /* mount rsize - max read hdr+data */
423 unsigned int wsize; /* mount wsize - max write hdr+data */
424 unsigned int inline_rsize; /* max non-rdma read data payload */
425 unsigned int inline_wsize; /* max non-rdma write data payload */
429 * Statistics for RPCRDMA
431 struct rpcrdma_stats {
432 /* accessed when sending a call */
433 unsigned long read_chunk_count;
434 unsigned long write_chunk_count;
435 unsigned long reply_chunk_count;
436 unsigned long long total_rdma_request;
438 /* rarely accessed error counters */
439 unsigned long long pullup_copy_count;
440 unsigned long hardway_register_count;
441 unsigned long failed_marshal_count;
442 unsigned long bad_reply_count;
443 unsigned long mrs_recycled;
444 unsigned long mrs_orphaned;
445 unsigned long mrs_allocated;
446 unsigned long empty_sendctx_q;
448 /* accessed when receiving a reply */
449 unsigned long long total_rdma_reply;
450 unsigned long long fixup_copy_count;
451 unsigned long reply_waits_for_send;
452 unsigned long local_inv_needed;
453 unsigned long nomsg_call_count;
454 unsigned long bcall_count;
458 * RPCRDMA transport -- encapsulates the structures above for
459 * integration with RPC.
461 * The contained structures are embedded, not pointers,
462 * for convenience. This structure need not be visible externally.
464 * It is allocated and initialized during mount, and released
467 struct rpcrdma_xprt {
468 struct rpc_xprt rx_xprt;
469 struct rpcrdma_ia rx_ia;
470 struct rpcrdma_ep rx_ep;
471 struct rpcrdma_buffer rx_buf;
472 struct rpcrdma_create_data_internal rx_data;
473 struct delayed_work rx_connect_worker;
474 struct rpcrdma_stats rx_stats;
477 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
478 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data)
480 static inline const char *
481 rpcrdma_addrstr(const struct rpcrdma_xprt *r_xprt)
483 return r_xprt->rx_xprt.address_strings[RPC_DISPLAY_ADDR];
486 static inline const char *
487 rpcrdma_portstr(const struct rpcrdma_xprt *r_xprt)
489 return r_xprt->rx_xprt.address_strings[RPC_DISPLAY_PORT];
492 /* Setting this to 0 ensures interoperability with early servers.
493 * Setting this to 1 enhances certain unaligned read/write performance.
494 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
495 extern int xprt_rdma_pad_optimize;
497 /* This setting controls the hunt for a supported memory
498 * registration strategy.
500 extern unsigned int xprt_rdma_memreg_strategy;
503 * Interface Adapter calls - xprtrdma/verbs.c
505 int rpcrdma_ia_open(struct rpcrdma_xprt *xprt);
506 void rpcrdma_ia_remove(struct rpcrdma_ia *ia);
507 void rpcrdma_ia_close(struct rpcrdma_ia *);
510 * Endpoint calls - xprtrdma/verbs.c
512 int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *,
513 struct rpcrdma_create_data_internal *);
514 void rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *);
515 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *);
516 void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *);
518 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
519 struct rpcrdma_req *);
522 * Buffer calls - xprtrdma/verbs.c
524 struct rpcrdma_req *rpcrdma_req_create(struct rpcrdma_xprt *r_xprt, size_t size,
526 void rpcrdma_req_destroy(struct rpcrdma_req *req);
527 int rpcrdma_buffer_create(struct rpcrdma_xprt *);
528 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
529 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_xprt *r_xprt);
531 struct rpcrdma_mr *rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt);
532 void rpcrdma_mr_put(struct rpcrdma_mr *mr);
533 void rpcrdma_mr_unmap_and_put(struct rpcrdma_mr *mr);
536 rpcrdma_mr_recycle(struct rpcrdma_mr *mr)
538 schedule_work(&mr->mr_recycle);
541 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
542 void rpcrdma_buffer_put(struct rpcrdma_req *);
543 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
545 bool rpcrdma_regbuf_realloc(struct rpcrdma_regbuf *rb, size_t size,
547 bool __rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
548 struct rpcrdma_regbuf *rb);
551 * rpcrdma_regbuf_is_mapped - check if buffer is DMA mapped
553 * Returns true if the buffer is now mapped to rb->rg_device.
555 static inline bool rpcrdma_regbuf_is_mapped(struct rpcrdma_regbuf *rb)
557 return rb->rg_device != NULL;
561 * rpcrdma_regbuf_dma_map - DMA-map a regbuf
562 * @r_xprt: controlling transport instance
563 * @rb: regbuf to be mapped
565 * Returns true if the buffer is currently DMA mapped.
567 static inline bool rpcrdma_regbuf_dma_map(struct rpcrdma_xprt *r_xprt,
568 struct rpcrdma_regbuf *rb)
570 if (likely(rpcrdma_regbuf_is_mapped(rb)))
572 return __rpcrdma_regbuf_dma_map(r_xprt, rb);
576 * Wrappers for chunk registration, shared by read/write chunk code.
579 static inline enum dma_data_direction
580 rpcrdma_data_dir(bool writing)
582 return writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
585 /* Memory registration calls xprtrdma/frwr_ops.c
587 bool frwr_is_supported(struct ib_device *device);
588 int frwr_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep,
589 struct rpcrdma_create_data_internal *cdata);
590 int frwr_init_mr(struct rpcrdma_ia *ia, struct rpcrdma_mr *mr);
591 void frwr_release_mr(struct rpcrdma_mr *mr);
592 size_t frwr_maxpages(struct rpcrdma_xprt *r_xprt);
593 struct rpcrdma_mr_seg *frwr_map(struct rpcrdma_xprt *r_xprt,
594 struct rpcrdma_mr_seg *seg,
595 int nsegs, bool writing, __be32 xid,
596 struct rpcrdma_mr **mr);
597 int frwr_send(struct rpcrdma_ia *ia, struct rpcrdma_req *req);
598 void frwr_reminv(struct rpcrdma_rep *rep, struct list_head *mrs);
599 void frwr_unmap_sync(struct rpcrdma_xprt *r_xprt,
600 struct list_head *mrs);
603 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
606 enum rpcrdma_chunktype {
614 int rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
615 struct rpcrdma_req *req, u32 hdrlen,
617 enum rpcrdma_chunktype rtype);
618 void rpcrdma_sendctx_unmap(struct rpcrdma_sendctx *sc);
619 int rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst);
620 void rpcrdma_set_max_header_sizes(struct rpcrdma_xprt *);
621 void rpcrdma_complete_rqst(struct rpcrdma_rep *rep);
622 void rpcrdma_reply_handler(struct rpcrdma_rep *rep);
623 void rpcrdma_release_rqst(struct rpcrdma_xprt *r_xprt,
624 struct rpcrdma_req *req);
625 void rpcrdma_deferred_completion(struct work_struct *work);
627 static inline void rpcrdma_set_xdrlen(struct xdr_buf *xdr, size_t len)
629 xdr->head[0].iov_len = len;
633 /* RPC/RDMA module init - xprtrdma/transport.c
635 extern unsigned int xprt_rdma_max_inline_read;
636 void xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap);
637 void xprt_rdma_free_addresses(struct rpc_xprt *xprt);
638 void xprt_rdma_close(struct rpc_xprt *xprt);
639 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq);
640 int xprt_rdma_init(void);
641 void xprt_rdma_cleanup(void);
643 /* Backchannel calls - xprtrdma/backchannel.c
645 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
646 int xprt_rdma_bc_setup(struct rpc_xprt *, unsigned int);
647 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *);
648 int rpcrdma_bc_post_recv(struct rpcrdma_xprt *, unsigned int);
649 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *, struct rpcrdma_rep *);
650 int xprt_rdma_bc_send_reply(struct rpc_rqst *rqst);
651 void xprt_rdma_bc_free_rqst(struct rpc_rqst *);
652 void xprt_rdma_bc_destroy(struct rpc_xprt *, unsigned int);
653 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
655 extern struct xprt_class xprt_rdma_bc;
657 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */