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 ib_device *ri_device;
70 struct rdma_cm_id *ri_id;
72 struct completion ri_done;
73 struct completion ri_remove_done;
75 unsigned int ri_max_segs;
76 unsigned int ri_max_frwr_depth;
77 unsigned int ri_max_inline_write;
78 unsigned int ri_max_inline_read;
79 unsigned int ri_max_send_sges;
80 bool ri_implicit_roundup;
81 enum ib_mr_type ri_mrtype;
82 unsigned long ri_flags;
86 RPCRDMA_IAF_REMOVING = 0,
90 * RDMA Endpoint -- one per transport instance
94 unsigned int rep_send_count;
95 unsigned int rep_send_batch;
97 struct ib_qp_init_attr rep_attr;
98 wait_queue_head_t rep_connect_wait;
99 struct rpcrdma_connect_private rep_cm_private;
100 struct rdma_conn_param rep_remote_cma;
101 int rep_receive_count;
104 /* Pre-allocate extra Work Requests for handling backward receives
105 * and sends. This is a fixed value because the Work Queues are
106 * allocated when the forward channel is set up.
108 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
109 #define RPCRDMA_BACKWARD_WRS (8)
111 #define RPCRDMA_BACKWARD_WRS (0)
114 /* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV
116 * The below structure appears at the front of a large region of kmalloc'd
117 * memory, which always starts on a good alignment boundary.
120 struct rpcrdma_regbuf {
121 struct ib_sge rg_iov;
122 struct ib_device *rg_device;
123 enum dma_data_direction rg_direction;
124 __be32 rg_base[0] __attribute__ ((aligned(256)));
128 rdmab_addr(struct rpcrdma_regbuf *rb)
130 return rb->rg_iov.addr;
134 rdmab_length(struct rpcrdma_regbuf *rb)
136 return rb->rg_iov.length;
140 rdmab_lkey(struct rpcrdma_regbuf *rb)
142 return rb->rg_iov.lkey;
145 static inline struct ib_device *
146 rdmab_device(struct rpcrdma_regbuf *rb)
148 return rb->rg_device;
151 #define RPCRDMA_DEF_GFP (GFP_NOIO | __GFP_NOWARN)
153 /* To ensure a transport can always make forward progress,
154 * the number of RDMA segments allowed in header chunk lists
155 * is capped at 8. This prevents less-capable devices and
156 * memory registrations from overrunning the Send buffer
157 * while building chunk lists.
159 * Elements of the Read list take up more room than the
160 * Write list or Reply chunk. 8 read segments means the Read
161 * list (or Write list or Reply chunk) cannot consume more
164 * ((8 + 2) * read segment size) + 1 XDR words, or 244 bytes.
166 * And the fixed part of the header is another 24 bytes.
168 * The smallest inline threshold is 1024 bytes, ensuring that
169 * at least 750 bytes are available for RPC messages.
172 RPCRDMA_MAX_HDR_SEGS = 8,
173 RPCRDMA_HDRBUF_SIZE = 256,
177 * struct rpcrdma_rep -- this structure encapsulates state required
178 * to receive and complete an RPC Reply, asychronously. It needs
179 * several pieces of state:
181 * o receive buffer and ib_sge (donated to provider)
182 * o status of receive (success or not, length, inv rkey)
183 * o bookkeeping state to get run by reply handler (XDR stream)
185 * These structures are allocated during transport initialization.
186 * N of these are associated with a transport instance, managed by
187 * struct rpcrdma_buffer. N is the max number of outstanding RPCs.
191 struct ib_cqe rr_cqe;
198 struct rpcrdma_regbuf *rr_rdmabuf;
199 struct rpcrdma_xprt *rr_rxprt;
200 struct work_struct rr_work;
201 struct xdr_buf rr_hdrbuf;
202 struct xdr_stream rr_stream;
203 struct rpc_rqst *rr_rqst;
204 struct list_head rr_list;
205 struct ib_recv_wr rr_recv_wr;
208 /* struct rpcrdma_sendctx - DMA mapped SGEs to unmap after Send completes
212 struct rpcrdma_sendctx {
213 struct ib_send_wr sc_wr;
214 struct ib_cqe sc_cqe;
215 struct rpcrdma_xprt *sc_xprt;
216 struct rpcrdma_req *sc_req;
217 unsigned int sc_unmap_count;
218 struct ib_sge sc_sges[];
221 /* Limit the number of SGEs that can be unmapped during one
222 * Send completion. This caps the amount of work a single
223 * completion can do before returning to the provider.
225 * Setting this to zero disables Send completion batching.
228 RPCRDMA_MAX_SEND_BATCH = 7,
232 * struct rpcrdma_mr - external memory region metadata
234 * An external memory region is any buffer or page that is registered
235 * on the fly (ie, not pre-registered).
237 * Each rpcrdma_buffer has a list of free MWs anchored in rb_mrs. During
238 * call_allocate, rpcrdma_buffer_get() assigns one to each segment in
239 * an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep
240 * track of registration metadata while each RPC is pending.
241 * rpcrdma_deregister_external() uses this metadata to unmap and
242 * release these resources when an RPC is complete.
244 enum rpcrdma_frwr_state {
245 FRWR_IS_INVALID, /* ready to be used */
246 FRWR_IS_VALID, /* in use */
247 FRWR_FLUSHED_FR, /* flushed FASTREG WR */
248 FRWR_FLUSHED_LI, /* flushed LOCALINV WR */
251 struct rpcrdma_frwr {
253 struct ib_cqe fr_cqe;
254 enum rpcrdma_frwr_state fr_state;
255 struct completion fr_linv_done;
257 struct ib_reg_wr fr_regwr;
258 struct ib_send_wr fr_invwr;
263 struct list_head mr_list;
264 struct scatterlist *mr_sg;
266 enum dma_data_direction mr_dir;
267 struct rpcrdma_frwr frwr;
268 struct rpcrdma_xprt *mr_xprt;
272 struct work_struct mr_recycle;
273 struct list_head mr_all;
277 * struct rpcrdma_req -- structure central to the request/reply sequence.
279 * N of these are associated with a transport instance, and stored in
280 * struct rpcrdma_buffer. N is the max number of outstanding requests.
282 * It includes pre-registered buffer memory for send AND recv.
283 * The recv buffer, however, is not owned by this structure, and
284 * is "donated" to the hardware when a recv is posted. When a
285 * reply is handled, the recv buffer used is given back to the
286 * struct rpcrdma_req associated with the request.
288 * In addition to the basic memory, this structure includes an array
289 * of iovs for send operations. The reason is that the iovs passed to
290 * ib_post_{send,recv} must not be modified until the work request
294 /* Maximum number of page-sized "segments" per chunk list to be
295 * registered or invalidated. Must handle a Reply chunk:
298 RPCRDMA_MAX_IOV_SEGS = 3,
299 RPCRDMA_MAX_DATA_SEGS = ((1 * 1024 * 1024) / PAGE_SIZE) + 1,
300 RPCRDMA_MAX_SEGS = RPCRDMA_MAX_DATA_SEGS +
301 RPCRDMA_MAX_IOV_SEGS,
304 struct rpcrdma_mr_seg { /* chunk descriptors */
305 u32 mr_len; /* length of chunk or segment */
306 struct page *mr_page; /* owning page, if any */
307 char *mr_offset; /* kva if no page, else offset */
310 /* The Send SGE array is provisioned to send a maximum size
312 * - RPC-over-RDMA header
313 * - xdr_buf head iovec
314 * - RPCRDMA_MAX_INLINE bytes, in pages
315 * - xdr_buf tail iovec
317 * The actual number of array elements consumed by each RPC
318 * depends on the device's max_sge limit.
321 RPCRDMA_MIN_SEND_SGES = 3,
322 RPCRDMA_MAX_PAGE_SGES = RPCRDMA_MAX_INLINE >> PAGE_SHIFT,
323 RPCRDMA_MAX_SEND_SGES = 1 + 1 + RPCRDMA_MAX_PAGE_SGES + 1,
326 struct rpcrdma_buffer;
328 struct list_head rl_list;
329 struct rpc_rqst rl_slot;
330 struct rpcrdma_buffer *rl_buffer;
331 struct rpcrdma_rep *rl_reply;
332 struct xdr_stream rl_stream;
333 struct xdr_buf rl_hdrbuf;
334 struct rpcrdma_sendctx *rl_sendctx;
335 struct rpcrdma_regbuf *rl_rdmabuf; /* xprt header */
336 struct rpcrdma_regbuf *rl_sendbuf; /* rq_snd_buf */
337 struct rpcrdma_regbuf *rl_recvbuf; /* rq_rcv_buf */
339 struct list_head rl_all;
340 unsigned long rl_flags;
342 struct list_head rl_registered; /* registered segments */
343 struct rpcrdma_mr_seg rl_segments[RPCRDMA_MAX_SEGS];
348 RPCRDMA_REQ_F_PENDING = 0,
349 RPCRDMA_REQ_F_TX_RESOURCES,
352 static inline struct rpcrdma_req *
353 rpcr_to_rdmar(const struct rpc_rqst *rqst)
355 return container_of(rqst, struct rpcrdma_req, rl_slot);
359 rpcrdma_mr_push(struct rpcrdma_mr *mr, struct list_head *list)
361 list_add_tail(&mr->mr_list, list);
364 static inline struct rpcrdma_mr *
365 rpcrdma_mr_pop(struct list_head *list)
367 struct rpcrdma_mr *mr;
369 mr = list_first_entry(list, struct rpcrdma_mr, mr_list);
370 list_del_init(&mr->mr_list);
375 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for
376 * inline requests/replies, and client/server credits.
378 * One of these is associated with a transport instance
380 struct rpcrdma_buffer {
381 spinlock_t rb_mrlock; /* protect rb_mrs list */
382 struct list_head rb_mrs;
383 struct list_head rb_all;
385 unsigned long rb_sc_head;
386 unsigned long rb_sc_tail;
387 unsigned long rb_sc_last;
388 struct rpcrdma_sendctx **rb_sc_ctxs;
390 spinlock_t rb_lock; /* protect buf lists */
391 struct list_head rb_send_bufs;
392 struct list_head rb_recv_bufs;
393 struct list_head rb_allreqs;
395 unsigned long rb_flags;
397 u32 rb_credits; /* most recent credit grant */
399 u32 rb_bc_srv_max_requests;
400 u32 rb_bc_max_requests;
402 struct workqueue_struct *rb_completion_wq;
403 struct delayed_work rb_refresh_worker;
408 RPCRDMA_BUF_F_EMPTY_SCQ = 0,
412 * Internal structure for transport instance creation. This
413 * exists primarily for modularity.
415 * This data should be set with mount options
417 struct rpcrdma_create_data_internal {
418 unsigned int max_requests; /* max requests (slots) in flight */
419 unsigned int rsize; /* mount rsize - max read hdr+data */
420 unsigned int wsize; /* mount wsize - max write hdr+data */
421 unsigned int inline_rsize; /* max non-rdma read data payload */
422 unsigned int inline_wsize; /* max non-rdma write data payload */
426 * Statistics for RPCRDMA
428 struct rpcrdma_stats {
429 /* accessed when sending a call */
430 unsigned long read_chunk_count;
431 unsigned long write_chunk_count;
432 unsigned long reply_chunk_count;
433 unsigned long long total_rdma_request;
435 /* rarely accessed error counters */
436 unsigned long long pullup_copy_count;
437 unsigned long hardway_register_count;
438 unsigned long failed_marshal_count;
439 unsigned long bad_reply_count;
440 unsigned long mrs_recycled;
441 unsigned long mrs_orphaned;
442 unsigned long mrs_allocated;
443 unsigned long empty_sendctx_q;
445 /* accessed when receiving a reply */
446 unsigned long long total_rdma_reply;
447 unsigned long long fixup_copy_count;
448 unsigned long reply_waits_for_send;
449 unsigned long local_inv_needed;
450 unsigned long nomsg_call_count;
451 unsigned long bcall_count;
455 * RPCRDMA transport -- encapsulates the structures above for
456 * integration with RPC.
458 * The contained structures are embedded, not pointers,
459 * for convenience. This structure need not be visible externally.
461 * It is allocated and initialized during mount, and released
464 struct rpcrdma_xprt {
465 struct rpc_xprt rx_xprt;
466 struct rpcrdma_ia rx_ia;
467 struct rpcrdma_ep rx_ep;
468 struct rpcrdma_buffer rx_buf;
469 struct rpcrdma_create_data_internal rx_data;
470 struct delayed_work rx_connect_worker;
471 struct rpcrdma_stats rx_stats;
474 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt)
475 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data)
477 static inline const char *
478 rpcrdma_addrstr(const struct rpcrdma_xprt *r_xprt)
480 return r_xprt->rx_xprt.address_strings[RPC_DISPLAY_ADDR];
483 static inline const char *
484 rpcrdma_portstr(const struct rpcrdma_xprt *r_xprt)
486 return r_xprt->rx_xprt.address_strings[RPC_DISPLAY_PORT];
489 /* Setting this to 0 ensures interoperability with early servers.
490 * Setting this to 1 enhances certain unaligned read/write performance.
491 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */
492 extern int xprt_rdma_pad_optimize;
494 /* This setting controls the hunt for a supported memory
495 * registration strategy.
497 extern unsigned int xprt_rdma_memreg_strategy;
500 * Interface Adapter calls - xprtrdma/verbs.c
502 int rpcrdma_ia_open(struct rpcrdma_xprt *xprt);
503 void rpcrdma_ia_remove(struct rpcrdma_ia *ia);
504 void rpcrdma_ia_close(struct rpcrdma_ia *);
507 * Endpoint calls - xprtrdma/verbs.c
509 int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *,
510 struct rpcrdma_create_data_internal *);
511 void rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *);
512 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *);
513 void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *);
515 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *,
516 struct rpcrdma_req *);
519 * Buffer calls - xprtrdma/verbs.c
521 struct rpcrdma_req *rpcrdma_create_req(struct rpcrdma_xprt *);
522 void rpcrdma_req_destroy(struct rpcrdma_req *req);
523 int rpcrdma_buffer_create(struct rpcrdma_xprt *);
524 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *);
525 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_buffer *buf);
527 struct rpcrdma_mr *rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt);
528 void rpcrdma_mr_put(struct rpcrdma_mr *mr);
529 void rpcrdma_mr_unmap_and_put(struct rpcrdma_mr *mr);
532 rpcrdma_mr_recycle(struct rpcrdma_mr *mr)
534 schedule_work(&mr->mr_recycle);
537 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *);
538 void rpcrdma_buffer_put(struct rpcrdma_req *);
539 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *);
541 struct rpcrdma_regbuf *rpcrdma_alloc_regbuf(size_t, enum dma_data_direction,
543 bool __rpcrdma_dma_map_regbuf(struct rpcrdma_ia *, struct rpcrdma_regbuf *);
544 void rpcrdma_free_regbuf(struct rpcrdma_regbuf *);
547 rpcrdma_regbuf_is_mapped(struct rpcrdma_regbuf *rb)
549 return rb->rg_device != NULL;
553 rpcrdma_dma_map_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
555 if (likely(rpcrdma_regbuf_is_mapped(rb)))
557 return __rpcrdma_dma_map_regbuf(ia, rb);
561 * Wrappers for chunk registration, shared by read/write chunk code.
564 static inline enum dma_data_direction
565 rpcrdma_data_dir(bool writing)
567 return writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
570 /* Memory registration calls xprtrdma/frwr_ops.c
572 bool frwr_is_supported(struct rpcrdma_ia *);
573 int frwr_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep,
574 struct rpcrdma_create_data_internal *cdata);
575 int frwr_init_mr(struct rpcrdma_ia *ia, struct rpcrdma_mr *mr);
576 void frwr_release_mr(struct rpcrdma_mr *mr);
577 size_t frwr_maxpages(struct rpcrdma_xprt *r_xprt);
578 struct rpcrdma_mr_seg *frwr_map(struct rpcrdma_xprt *r_xprt,
579 struct rpcrdma_mr_seg *seg,
580 int nsegs, bool writing, u32 xid,
581 struct rpcrdma_mr **mr);
582 int frwr_send(struct rpcrdma_ia *ia, struct rpcrdma_req *req);
583 void frwr_reminv(struct rpcrdma_rep *rep, struct list_head *mrs);
584 void frwr_unmap_sync(struct rpcrdma_xprt *r_xprt,
585 struct list_head *mrs);
588 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c
591 enum rpcrdma_chunktype {
599 int rpcrdma_prepare_send_sges(struct rpcrdma_xprt *r_xprt,
600 struct rpcrdma_req *req, u32 hdrlen,
602 enum rpcrdma_chunktype rtype);
603 void rpcrdma_unmap_sendctx(struct rpcrdma_sendctx *sc);
604 int rpcrdma_marshal_req(struct rpcrdma_xprt *r_xprt, struct rpc_rqst *rqst);
605 void rpcrdma_set_max_header_sizes(struct rpcrdma_xprt *);
606 void rpcrdma_complete_rqst(struct rpcrdma_rep *rep);
607 void rpcrdma_reply_handler(struct rpcrdma_rep *rep);
608 void rpcrdma_release_rqst(struct rpcrdma_xprt *r_xprt,
609 struct rpcrdma_req *req);
610 void rpcrdma_deferred_completion(struct work_struct *work);
612 static inline void rpcrdma_set_xdrlen(struct xdr_buf *xdr, size_t len)
614 xdr->head[0].iov_len = len;
618 /* RPC/RDMA module init - xprtrdma/transport.c
620 extern unsigned int xprt_rdma_max_inline_read;
621 void xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap);
622 void xprt_rdma_free_addresses(struct rpc_xprt *xprt);
623 void xprt_rdma_close(struct rpc_xprt *xprt);
624 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq);
625 int xprt_rdma_init(void);
626 void xprt_rdma_cleanup(void);
628 /* Backchannel calls - xprtrdma/backchannel.c
630 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
631 int xprt_rdma_bc_setup(struct rpc_xprt *, unsigned int);
632 size_t xprt_rdma_bc_maxpayload(struct rpc_xprt *);
633 int rpcrdma_bc_post_recv(struct rpcrdma_xprt *, unsigned int);
634 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *, struct rpcrdma_rep *);
635 int xprt_rdma_bc_send_reply(struct rpc_rqst *rqst);
636 void xprt_rdma_bc_free_rqst(struct rpc_rqst *);
637 void xprt_rdma_bc_destroy(struct rpc_xprt *, unsigned int);
638 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
640 extern struct xprt_class xprt_rdma_bc;
642 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */