1 /* 2 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved. 3 * 4 * This software is available to you under a choice of one of two 5 * licenses. You may choose to be licensed under the terms of the GNU 6 * General Public License (GPL) Version 2, available from the file 7 * COPYING in the main directory of this source tree, or the BSD-type 8 * license below: 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 14 * Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 17 * Redistributions in binary form must reproduce the above 18 * copyright notice, this list of conditions and the following 19 * disclaimer in the documentation and/or other materials provided 20 * with the distribution. 21 * 22 * Neither the name of the Network Appliance, Inc. nor the names of 23 * its contributors may be used to endorse or promote products 24 * derived from this software without specific prior written 25 * permission. 26 * 27 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 28 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 29 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 30 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 31 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 32 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 33 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 34 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 35 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 36 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 37 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 38 */ 39 40 #ifndef _LINUX_SUNRPC_XPRT_RDMA_H 41 #define _LINUX_SUNRPC_XPRT_RDMA_H 42 43 #include <linux/wait.h> /* wait_queue_head_t, etc */ 44 #include <linux/spinlock.h> /* spinlock_t, etc */ 45 #include <linux/atomic.h> /* atomic_t, etc */ 46 #include <linux/workqueue.h> /* struct work_struct */ 47 48 #include <rdma/rdma_cm.h> /* RDMA connection api */ 49 #include <rdma/ib_verbs.h> /* RDMA verbs api */ 50 51 #include <linux/sunrpc/clnt.h> /* rpc_xprt */ 52 #include <linux/sunrpc/rpc_rdma.h> /* RPC/RDMA protocol */ 53 #include <linux/sunrpc/xprtrdma.h> /* xprt parameters */ 54 55 #define RDMA_RESOLVE_TIMEOUT (5000) /* 5 seconds */ 56 #define RDMA_CONNECT_RETRY_MAX (2) /* retries if no listener backlog */ 57 58 /* 59 * Interface Adapter -- one per transport instance 60 */ 61 struct rpcrdma_ia { 62 const struct rpcrdma_memreg_ops *ri_ops; 63 rwlock_t ri_qplock; 64 struct ib_device *ri_device; 65 struct rdma_cm_id *ri_id; 66 struct ib_pd *ri_pd; 67 struct ib_mr *ri_dma_mr; 68 struct completion ri_done; 69 int ri_async_rc; 70 unsigned int ri_max_frmr_depth; 71 struct ib_device_attr ri_devattr; 72 struct ib_qp_attr ri_qp_attr; 73 struct ib_qp_init_attr ri_qp_init_attr; 74 }; 75 76 /* 77 * RDMA Endpoint -- one per transport instance 78 */ 79 80 struct rpcrdma_ep { 81 atomic_t rep_cqcount; 82 int rep_cqinit; 83 int rep_connected; 84 struct ib_qp_init_attr rep_attr; 85 wait_queue_head_t rep_connect_wait; 86 struct rdma_conn_param rep_remote_cma; 87 struct sockaddr_storage rep_remote_addr; 88 struct delayed_work rep_connect_worker; 89 }; 90 91 /* 92 * Force a signaled SEND Work Request every so often, 93 * in case the provider needs to do some housekeeping. 94 */ 95 #define RPCRDMA_MAX_UNSIGNALED_SENDS (32) 96 97 #define INIT_CQCOUNT(ep) atomic_set(&(ep)->rep_cqcount, (ep)->rep_cqinit) 98 #define DECR_CQCOUNT(ep) atomic_sub_return(1, &(ep)->rep_cqcount) 99 100 /* Force completion handler to ignore the signal 101 */ 102 #define RPCRDMA_IGNORE_COMPLETION (0ULL) 103 104 /* 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. 107 */ 108 #if defined(CONFIG_SUNRPC_BACKCHANNEL) 109 #define RPCRDMA_BACKWARD_WRS (8) 110 #else 111 #define RPCRDMA_BACKWARD_WRS (0) 112 #endif 113 114 /* Registered buffer -- registered kmalloc'd memory for RDMA SEND/RECV 115 * 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. 118 */ 119 120 struct rpcrdma_regbuf { 121 size_t rg_size; 122 struct rpcrdma_req *rg_owner; 123 struct ib_sge rg_iov; 124 __be32 rg_base[0] __attribute__ ((aligned(256))); 125 }; 126 127 static inline u64 128 rdmab_addr(struct rpcrdma_regbuf *rb) 129 { 130 return rb->rg_iov.addr; 131 } 132 133 static inline u32 134 rdmab_length(struct rpcrdma_regbuf *rb) 135 { 136 return rb->rg_iov.length; 137 } 138 139 static inline u32 140 rdmab_lkey(struct rpcrdma_regbuf *rb) 141 { 142 return rb->rg_iov.lkey; 143 } 144 145 static inline struct rpcrdma_msg * 146 rdmab_to_msg(struct rpcrdma_regbuf *rb) 147 { 148 return (struct rpcrdma_msg *)rb->rg_base; 149 } 150 151 /* 152 * struct rpcrdma_rep -- this structure encapsulates state required to recv 153 * and complete a reply, asychronously. It needs several pieces of 154 * state: 155 * o recv buffer (posted to provider) 156 * o ib_sge (also donated to provider) 157 * o status of reply (length, success or not) 158 * o bookkeeping state to get run by tasklet (list, etc) 159 * 160 * These are allocated during initialization, per-transport instance; 161 * however, the tasklet execution list itself is global, as it should 162 * always be pretty short. 163 * 164 * N of these are associated with a transport instance, and stored in 165 * struct rpcrdma_buffer. N is the max number of outstanding requests. 166 */ 167 168 #define RPCRDMA_MAX_DATA_SEGS ((1 * 1024 * 1024) / PAGE_SIZE) 169 #define RPCRDMA_MAX_SEGS (RPCRDMA_MAX_DATA_SEGS + 2) /* head+tail = 2 */ 170 171 struct rpcrdma_buffer; 172 173 struct rpcrdma_rep { 174 unsigned int rr_len; 175 struct ib_device *rr_device; 176 struct rpcrdma_xprt *rr_rxprt; 177 struct work_struct rr_work; 178 struct list_head rr_list; 179 struct rpcrdma_regbuf *rr_rdmabuf; 180 }; 181 182 #define RPCRDMA_BAD_LEN (~0U) 183 184 /* 185 * struct rpcrdma_mw - external memory region metadata 186 * 187 * An external memory region is any buffer or page that is registered 188 * on the fly (ie, not pre-registered). 189 * 190 * Each rpcrdma_buffer has a list of free MWs anchored in rb_mws. During 191 * call_allocate, rpcrdma_buffer_get() assigns one to each segment in 192 * an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep 193 * track of registration metadata while each RPC is pending. 194 * rpcrdma_deregister_external() uses this metadata to unmap and 195 * release these resources when an RPC is complete. 196 */ 197 enum rpcrdma_frmr_state { 198 FRMR_IS_INVALID, /* ready to be used */ 199 FRMR_IS_VALID, /* in use */ 200 FRMR_IS_STALE, /* failed completion */ 201 }; 202 203 struct rpcrdma_frmr { 204 struct scatterlist *sg; 205 int sg_nents; 206 struct ib_mr *fr_mr; 207 enum rpcrdma_frmr_state fr_state; 208 struct work_struct fr_work; 209 struct rpcrdma_xprt *fr_xprt; 210 }; 211 212 struct rpcrdma_fmr { 213 struct ib_fmr *fmr; 214 u64 *physaddrs; 215 }; 216 217 struct rpcrdma_mw { 218 union { 219 struct rpcrdma_fmr fmr; 220 struct rpcrdma_frmr frmr; 221 } r; 222 void (*mw_sendcompletion)(struct ib_wc *); 223 struct list_head mw_list; 224 struct list_head mw_all; 225 }; 226 227 /* 228 * struct rpcrdma_req -- structure central to the request/reply sequence. 229 * 230 * N of these are associated with a transport instance, and stored in 231 * struct rpcrdma_buffer. N is the max number of outstanding requests. 232 * 233 * It includes pre-registered buffer memory for send AND recv. 234 * The recv buffer, however, is not owned by this structure, and 235 * is "donated" to the hardware when a recv is posted. When a 236 * reply is handled, the recv buffer used is given back to the 237 * struct rpcrdma_req associated with the request. 238 * 239 * In addition to the basic memory, this structure includes an array 240 * of iovs for send operations. The reason is that the iovs passed to 241 * ib_post_{send,recv} must not be modified until the work request 242 * completes. 243 * 244 * NOTES: 245 * o RPCRDMA_MAX_SEGS is the max number of addressible chunk elements we 246 * marshal. The number needed varies depending on the iov lists that 247 * are passed to us, the memory registration mode we are in, and if 248 * physical addressing is used, the layout. 249 */ 250 251 struct rpcrdma_mr_seg { /* chunk descriptors */ 252 struct rpcrdma_mw *rl_mw; /* registered MR */ 253 u64 mr_base; /* registration result */ 254 u32 mr_rkey; /* registration result */ 255 u32 mr_len; /* length of chunk or segment */ 256 int mr_nsegs; /* number of segments in chunk or 0 */ 257 enum dma_data_direction mr_dir; /* segment mapping direction */ 258 dma_addr_t mr_dma; /* segment mapping address */ 259 size_t mr_dmalen; /* segment mapping length */ 260 struct page *mr_page; /* owning page, if any */ 261 char *mr_offset; /* kva if no page, else offset */ 262 }; 263 264 #define RPCRDMA_MAX_IOVS (2) 265 266 struct rpcrdma_req { 267 struct list_head rl_free; 268 unsigned int rl_niovs; 269 unsigned int rl_nchunks; 270 unsigned int rl_connect_cookie; 271 struct rpcrdma_buffer *rl_buffer; 272 struct rpcrdma_rep *rl_reply;/* holder for reply buffer */ 273 struct ib_sge rl_send_iov[RPCRDMA_MAX_IOVS]; 274 struct rpcrdma_regbuf *rl_rdmabuf; 275 struct rpcrdma_regbuf *rl_sendbuf; 276 struct rpcrdma_mr_seg rl_segments[RPCRDMA_MAX_SEGS]; 277 278 struct list_head rl_all; 279 bool rl_backchannel; 280 }; 281 282 static inline struct rpcrdma_req * 283 rpcr_to_rdmar(struct rpc_rqst *rqst) 284 { 285 void *buffer = rqst->rq_buffer; 286 struct rpcrdma_regbuf *rb; 287 288 rb = container_of(buffer, struct rpcrdma_regbuf, rg_base); 289 return rb->rg_owner; 290 } 291 292 /* 293 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for 294 * inline requests/replies, and client/server credits. 295 * 296 * One of these is associated with a transport instance 297 */ 298 struct rpcrdma_buffer { 299 spinlock_t rb_mwlock; /* protect rb_mws list */ 300 struct list_head rb_mws; 301 struct list_head rb_all; 302 char *rb_pool; 303 304 spinlock_t rb_lock; /* protect buf lists */ 305 struct list_head rb_send_bufs; 306 struct list_head rb_recv_bufs; 307 u32 rb_max_requests; 308 309 u32 rb_bc_srv_max_requests; 310 spinlock_t rb_reqslock; /* protect rb_allreqs */ 311 struct list_head rb_allreqs; 312 }; 313 #define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia) 314 315 /* 316 * Internal structure for transport instance creation. This 317 * exists primarily for modularity. 318 * 319 * This data should be set with mount options 320 */ 321 struct rpcrdma_create_data_internal { 322 struct sockaddr_storage addr; /* RDMA server address */ 323 unsigned int max_requests; /* max requests (slots) in flight */ 324 unsigned int rsize; /* mount rsize - max read hdr+data */ 325 unsigned int wsize; /* mount wsize - max write hdr+data */ 326 unsigned int inline_rsize; /* max non-rdma read data payload */ 327 unsigned int inline_wsize; /* max non-rdma write data payload */ 328 unsigned int padding; /* non-rdma write header padding */ 329 }; 330 331 #define RPCRDMA_INLINE_READ_THRESHOLD(rq) \ 332 (rpcx_to_rdmad(rq->rq_xprt).inline_rsize) 333 334 #define RPCRDMA_INLINE_WRITE_THRESHOLD(rq)\ 335 (rpcx_to_rdmad(rq->rq_xprt).inline_wsize) 336 337 #define RPCRDMA_INLINE_PAD_VALUE(rq)\ 338 rpcx_to_rdmad(rq->rq_xprt).padding 339 340 /* 341 * Statistics for RPCRDMA 342 */ 343 struct rpcrdma_stats { 344 unsigned long read_chunk_count; 345 unsigned long write_chunk_count; 346 unsigned long reply_chunk_count; 347 348 unsigned long long total_rdma_request; 349 unsigned long long total_rdma_reply; 350 351 unsigned long long pullup_copy_count; 352 unsigned long long fixup_copy_count; 353 unsigned long hardway_register_count; 354 unsigned long failed_marshal_count; 355 unsigned long bad_reply_count; 356 unsigned long nomsg_call_count; 357 unsigned long bcall_count; 358 }; 359 360 /* 361 * Per-registration mode operations 362 */ 363 struct rpcrdma_xprt; 364 struct rpcrdma_memreg_ops { 365 int (*ro_map)(struct rpcrdma_xprt *, 366 struct rpcrdma_mr_seg *, int, bool); 367 int (*ro_unmap)(struct rpcrdma_xprt *, 368 struct rpcrdma_mr_seg *); 369 int (*ro_open)(struct rpcrdma_ia *, 370 struct rpcrdma_ep *, 371 struct rpcrdma_create_data_internal *); 372 size_t (*ro_maxpages)(struct rpcrdma_xprt *); 373 int (*ro_init)(struct rpcrdma_xprt *); 374 void (*ro_destroy)(struct rpcrdma_buffer *); 375 const char *ro_displayname; 376 }; 377 378 extern const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops; 379 extern const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops; 380 extern const struct rpcrdma_memreg_ops rpcrdma_physical_memreg_ops; 381 382 /* 383 * RPCRDMA transport -- encapsulates the structures above for 384 * integration with RPC. 385 * 386 * The contained structures are embedded, not pointers, 387 * for convenience. This structure need not be visible externally. 388 * 389 * It is allocated and initialized during mount, and released 390 * during unmount. 391 */ 392 struct rpcrdma_xprt { 393 struct rpc_xprt rx_xprt; 394 struct rpcrdma_ia rx_ia; 395 struct rpcrdma_ep rx_ep; 396 struct rpcrdma_buffer rx_buf; 397 struct rpcrdma_create_data_internal rx_data; 398 struct delayed_work rx_connect_worker; 399 struct rpcrdma_stats rx_stats; 400 }; 401 402 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, rx_xprt) 403 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data) 404 405 /* Setting this to 0 ensures interoperability with early servers. 406 * Setting this to 1 enhances certain unaligned read/write performance. 407 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */ 408 extern int xprt_rdma_pad_optimize; 409 410 /* 411 * Interface Adapter calls - xprtrdma/verbs.c 412 */ 413 int rpcrdma_ia_open(struct rpcrdma_xprt *, struct sockaddr *, int); 414 void rpcrdma_ia_close(struct rpcrdma_ia *); 415 416 /* 417 * Endpoint calls - xprtrdma/verbs.c 418 */ 419 int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *, 420 struct rpcrdma_create_data_internal *); 421 void rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *); 422 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *); 423 void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *); 424 425 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *, 426 struct rpcrdma_req *); 427 int rpcrdma_ep_post_recv(struct rpcrdma_ia *, struct rpcrdma_ep *, 428 struct rpcrdma_rep *); 429 430 /* 431 * Buffer calls - xprtrdma/verbs.c 432 */ 433 struct rpcrdma_req *rpcrdma_create_req(struct rpcrdma_xprt *); 434 struct rpcrdma_rep *rpcrdma_create_rep(struct rpcrdma_xprt *); 435 void rpcrdma_destroy_req(struct rpcrdma_ia *, struct rpcrdma_req *); 436 int rpcrdma_buffer_create(struct rpcrdma_xprt *); 437 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *); 438 439 struct rpcrdma_mw *rpcrdma_get_mw(struct rpcrdma_xprt *); 440 void rpcrdma_put_mw(struct rpcrdma_xprt *, struct rpcrdma_mw *); 441 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *); 442 void rpcrdma_buffer_put(struct rpcrdma_req *); 443 void rpcrdma_recv_buffer_get(struct rpcrdma_req *); 444 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *); 445 446 struct rpcrdma_regbuf *rpcrdma_alloc_regbuf(struct rpcrdma_ia *, 447 size_t, gfp_t); 448 void rpcrdma_free_regbuf(struct rpcrdma_ia *, 449 struct rpcrdma_regbuf *); 450 451 unsigned int rpcrdma_max_segments(struct rpcrdma_xprt *); 452 int rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt *, unsigned int); 453 454 int frwr_alloc_recovery_wq(void); 455 void frwr_destroy_recovery_wq(void); 456 457 int rpcrdma_alloc_wq(void); 458 void rpcrdma_destroy_wq(void); 459 460 /* 461 * Wrappers for chunk registration, shared by read/write chunk code. 462 */ 463 464 void rpcrdma_mapping_error(struct rpcrdma_mr_seg *); 465 466 static inline enum dma_data_direction 467 rpcrdma_data_dir(bool writing) 468 { 469 return writing ? DMA_FROM_DEVICE : DMA_TO_DEVICE; 470 } 471 472 static inline void 473 rpcrdma_map_one(struct ib_device *device, struct rpcrdma_mr_seg *seg, 474 enum dma_data_direction direction) 475 { 476 seg->mr_dir = direction; 477 seg->mr_dmalen = seg->mr_len; 478 479 if (seg->mr_page) 480 seg->mr_dma = ib_dma_map_page(device, 481 seg->mr_page, offset_in_page(seg->mr_offset), 482 seg->mr_dmalen, seg->mr_dir); 483 else 484 seg->mr_dma = ib_dma_map_single(device, 485 seg->mr_offset, 486 seg->mr_dmalen, seg->mr_dir); 487 488 if (ib_dma_mapping_error(device, seg->mr_dma)) 489 rpcrdma_mapping_error(seg); 490 } 491 492 static inline void 493 rpcrdma_unmap_one(struct ib_device *device, struct rpcrdma_mr_seg *seg) 494 { 495 if (seg->mr_page) 496 ib_dma_unmap_page(device, 497 seg->mr_dma, seg->mr_dmalen, seg->mr_dir); 498 else 499 ib_dma_unmap_single(device, 500 seg->mr_dma, seg->mr_dmalen, seg->mr_dir); 501 } 502 503 /* 504 * RPC/RDMA connection management calls - xprtrdma/rpc_rdma.c 505 */ 506 void rpcrdma_connect_worker(struct work_struct *); 507 void rpcrdma_conn_func(struct rpcrdma_ep *); 508 void rpcrdma_reply_handler(struct rpcrdma_rep *); 509 510 /* 511 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c 512 */ 513 int rpcrdma_marshal_req(struct rpc_rqst *); 514 515 /* RPC/RDMA module init - xprtrdma/transport.c 516 */ 517 int xprt_rdma_init(void); 518 void xprt_rdma_cleanup(void); 519 520 /* Backchannel calls - xprtrdma/backchannel.c 521 */ 522 #if defined(CONFIG_SUNRPC_BACKCHANNEL) 523 int xprt_rdma_bc_setup(struct rpc_xprt *, unsigned int); 524 int xprt_rdma_bc_up(struct svc_serv *, struct net *); 525 int rpcrdma_bc_post_recv(struct rpcrdma_xprt *, unsigned int); 526 void rpcrdma_bc_receive_call(struct rpcrdma_xprt *, struct rpcrdma_rep *); 527 int rpcrdma_bc_marshal_reply(struct rpc_rqst *); 528 void xprt_rdma_bc_free_rqst(struct rpc_rqst *); 529 void xprt_rdma_bc_destroy(struct rpc_xprt *, unsigned int); 530 #endif /* CONFIG_SUNRPC_BACKCHANNEL */ 531 532 /* Temporary NFS request map cache. Created in svc_rdma.c */ 533 extern struct kmem_cache *svc_rdma_map_cachep; 534 /* WR context cache. Created in svc_rdma.c */ 535 extern struct kmem_cache *svc_rdma_ctxt_cachep; 536 /* Workqueue created in svc_rdma.c */ 537 extern struct workqueue_struct *svc_rdma_wq; 538 539 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */ 540