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 rwlock_t ri_qplock; 63 struct rdma_cm_id *ri_id; 64 struct ib_pd *ri_pd; 65 struct ib_mr *ri_bind_mem; 66 u32 ri_dma_lkey; 67 int ri_have_dma_lkey; 68 struct completion ri_done; 69 int ri_async_rc; 70 enum rpcrdma_memreg ri_memreg_strategy; 71 unsigned int ri_max_frmr_depth; 72 }; 73 74 /* 75 * RDMA Endpoint -- one per transport instance 76 */ 77 78 #define RPCRDMA_WC_BUDGET (128) 79 #define RPCRDMA_POLLSIZE (16) 80 81 struct rpcrdma_ep { 82 atomic_t rep_cqcount; 83 int rep_cqinit; 84 int rep_connected; 85 struct rpcrdma_ia *rep_ia; 86 struct ib_qp_init_attr rep_attr; 87 wait_queue_head_t rep_connect_wait; 88 struct ib_sge rep_pad; /* holds zeroed pad */ 89 struct ib_mr *rep_pad_mr; /* holds zeroed pad */ 90 void (*rep_func)(struct rpcrdma_ep *); 91 struct rpc_xprt *rep_xprt; /* for rep_func */ 92 struct rdma_conn_param rep_remote_cma; 93 struct sockaddr_storage rep_remote_addr; 94 struct delayed_work rep_connect_worker; 95 struct ib_wc rep_send_wcs[RPCRDMA_POLLSIZE]; 96 struct ib_wc rep_recv_wcs[RPCRDMA_POLLSIZE]; 97 }; 98 99 #define INIT_CQCOUNT(ep) atomic_set(&(ep)->rep_cqcount, (ep)->rep_cqinit) 100 #define DECR_CQCOUNT(ep) atomic_sub_return(1, &(ep)->rep_cqcount) 101 102 enum rpcrdma_chunktype { 103 rpcrdma_noch = 0, 104 rpcrdma_readch, 105 rpcrdma_areadch, 106 rpcrdma_writech, 107 rpcrdma_replych 108 }; 109 110 /* 111 * struct rpcrdma_rep -- this structure encapsulates state required to recv 112 * and complete a reply, asychronously. It needs several pieces of 113 * state: 114 * o recv buffer (posted to provider) 115 * o ib_sge (also donated to provider) 116 * o status of reply (length, success or not) 117 * o bookkeeping state to get run by tasklet (list, etc) 118 * 119 * These are allocated during initialization, per-transport instance; 120 * however, the tasklet execution list itself is global, as it should 121 * always be pretty short. 122 * 123 * N of these are associated with a transport instance, and stored in 124 * struct rpcrdma_buffer. N is the max number of outstanding requests. 125 */ 126 127 /* temporary static scatter/gather max */ 128 #define RPCRDMA_MAX_DATA_SEGS (64) /* max scatter/gather */ 129 #define RPCRDMA_MAX_SEGS (RPCRDMA_MAX_DATA_SEGS + 2) /* head+tail = 2 */ 130 #define MAX_RPCRDMAHDR (\ 131 /* max supported RPC/RDMA header */ \ 132 sizeof(struct rpcrdma_msg) + (2 * sizeof(u32)) + \ 133 (sizeof(struct rpcrdma_read_chunk) * RPCRDMA_MAX_SEGS) + sizeof(u32)) 134 135 struct rpcrdma_buffer; 136 137 struct rpcrdma_rep { 138 unsigned int rr_len; /* actual received reply length */ 139 struct rpcrdma_buffer *rr_buffer; /* home base for this structure */ 140 struct rpc_xprt *rr_xprt; /* needed for request/reply matching */ 141 void (*rr_func)(struct rpcrdma_rep *);/* called by tasklet in softint */ 142 struct list_head rr_list; /* tasklet list */ 143 struct ib_sge rr_iov; /* for posting */ 144 struct ib_mr *rr_handle; /* handle for mem in rr_iov */ 145 char rr_base[MAX_RPCRDMAHDR]; /* minimal inline receive buffer */ 146 }; 147 148 /* 149 * struct rpcrdma_mw - external memory region metadata 150 * 151 * An external memory region is any buffer or page that is registered 152 * on the fly (ie, not pre-registered). 153 * 154 * Each rpcrdma_buffer has a list of free MWs anchored in rb_mws. During 155 * call_allocate, rpcrdma_buffer_get() assigns one to each segment in 156 * an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep 157 * track of registration metadata while each RPC is pending. 158 * rpcrdma_deregister_external() uses this metadata to unmap and 159 * release these resources when an RPC is complete. 160 */ 161 enum rpcrdma_frmr_state { 162 FRMR_IS_INVALID, /* ready to be used */ 163 FRMR_IS_VALID, /* in use */ 164 FRMR_IS_STALE, /* failed completion */ 165 }; 166 167 struct rpcrdma_frmr { 168 struct ib_fast_reg_page_list *fr_pgl; 169 struct ib_mr *fr_mr; 170 enum rpcrdma_frmr_state fr_state; 171 }; 172 173 struct rpcrdma_mw { 174 union { 175 struct ib_fmr *fmr; 176 struct rpcrdma_frmr frmr; 177 } r; 178 struct list_head mw_list; 179 struct list_head mw_all; 180 }; 181 182 /* 183 * struct rpcrdma_req -- structure central to the request/reply sequence. 184 * 185 * N of these are associated with a transport instance, and stored in 186 * struct rpcrdma_buffer. N is the max number of outstanding requests. 187 * 188 * It includes pre-registered buffer memory for send AND recv. 189 * The recv buffer, however, is not owned by this structure, and 190 * is "donated" to the hardware when a recv is posted. When a 191 * reply is handled, the recv buffer used is given back to the 192 * struct rpcrdma_req associated with the request. 193 * 194 * In addition to the basic memory, this structure includes an array 195 * of iovs for send operations. The reason is that the iovs passed to 196 * ib_post_{send,recv} must not be modified until the work request 197 * completes. 198 * 199 * NOTES: 200 * o RPCRDMA_MAX_SEGS is the max number of addressible chunk elements we 201 * marshal. The number needed varies depending on the iov lists that 202 * are passed to us, the memory registration mode we are in, and if 203 * physical addressing is used, the layout. 204 */ 205 206 struct rpcrdma_mr_seg { /* chunk descriptors */ 207 union { /* chunk memory handles */ 208 struct ib_mr *rl_mr; /* if registered directly */ 209 struct rpcrdma_mw *rl_mw; /* if registered from region */ 210 } mr_chunk; 211 u64 mr_base; /* registration result */ 212 u32 mr_rkey; /* registration result */ 213 u32 mr_len; /* length of chunk or segment */ 214 int mr_nsegs; /* number of segments in chunk or 0 */ 215 enum dma_data_direction mr_dir; /* segment mapping direction */ 216 dma_addr_t mr_dma; /* segment mapping address */ 217 size_t mr_dmalen; /* segment mapping length */ 218 struct page *mr_page; /* owning page, if any */ 219 char *mr_offset; /* kva if no page, else offset */ 220 }; 221 222 struct rpcrdma_req { 223 size_t rl_size; /* actual length of buffer */ 224 unsigned int rl_niovs; /* 0, 2 or 4 */ 225 unsigned int rl_nchunks; /* non-zero if chunks */ 226 unsigned int rl_connect_cookie; /* retry detection */ 227 enum rpcrdma_chunktype rl_rtype, rl_wtype; 228 struct rpcrdma_buffer *rl_buffer; /* home base for this structure */ 229 struct rpcrdma_rep *rl_reply;/* holder for reply buffer */ 230 struct rpcrdma_mr_seg rl_segments[RPCRDMA_MAX_SEGS];/* chunk segments */ 231 struct ib_sge rl_send_iov[4]; /* for active requests */ 232 struct ib_sge rl_iov; /* for posting */ 233 struct ib_mr *rl_handle; /* handle for mem in rl_iov */ 234 char rl_base[MAX_RPCRDMAHDR]; /* start of actual buffer */ 235 __u32 rl_xdr_buf[0]; /* start of returned rpc rq_buffer */ 236 }; 237 #define rpcr_to_rdmar(r) \ 238 container_of((r)->rq_buffer, struct rpcrdma_req, rl_xdr_buf[0]) 239 240 /* 241 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for 242 * inline requests/replies, and client/server credits. 243 * 244 * One of these is associated with a transport instance 245 */ 246 struct rpcrdma_buffer { 247 spinlock_t rb_lock; /* protects indexes */ 248 atomic_t rb_credits; /* most recent server credits */ 249 int rb_max_requests;/* client max requests */ 250 struct list_head rb_mws; /* optional memory windows/fmrs/frmrs */ 251 struct list_head rb_all; 252 int rb_send_index; 253 struct rpcrdma_req **rb_send_bufs; 254 int rb_recv_index; 255 struct rpcrdma_rep **rb_recv_bufs; 256 char *rb_pool; 257 }; 258 #define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia) 259 260 /* 261 * Internal structure for transport instance creation. This 262 * exists primarily for modularity. 263 * 264 * This data should be set with mount options 265 */ 266 struct rpcrdma_create_data_internal { 267 struct sockaddr_storage addr; /* RDMA server address */ 268 unsigned int max_requests; /* max requests (slots) in flight */ 269 unsigned int rsize; /* mount rsize - max read hdr+data */ 270 unsigned int wsize; /* mount wsize - max write hdr+data */ 271 unsigned int inline_rsize; /* max non-rdma read data payload */ 272 unsigned int inline_wsize; /* max non-rdma write data payload */ 273 unsigned int padding; /* non-rdma write header padding */ 274 }; 275 276 #define RPCRDMA_INLINE_READ_THRESHOLD(rq) \ 277 (rpcx_to_rdmad(rq->rq_xprt).inline_rsize) 278 279 #define RPCRDMA_INLINE_WRITE_THRESHOLD(rq)\ 280 (rpcx_to_rdmad(rq->rq_xprt).inline_wsize) 281 282 #define RPCRDMA_INLINE_PAD_VALUE(rq)\ 283 rpcx_to_rdmad(rq->rq_xprt).padding 284 285 /* 286 * Statistics for RPCRDMA 287 */ 288 struct rpcrdma_stats { 289 unsigned long read_chunk_count; 290 unsigned long write_chunk_count; 291 unsigned long reply_chunk_count; 292 293 unsigned long long total_rdma_request; 294 unsigned long long total_rdma_reply; 295 296 unsigned long long pullup_copy_count; 297 unsigned long long fixup_copy_count; 298 unsigned long hardway_register_count; 299 unsigned long failed_marshal_count; 300 unsigned long bad_reply_count; 301 }; 302 303 /* 304 * RPCRDMA transport -- encapsulates the structures above for 305 * integration with RPC. 306 * 307 * The contained structures are embedded, not pointers, 308 * for convenience. This structure need not be visible externally. 309 * 310 * It is allocated and initialized during mount, and released 311 * during unmount. 312 */ 313 struct rpcrdma_xprt { 314 struct rpc_xprt xprt; 315 struct rpcrdma_ia rx_ia; 316 struct rpcrdma_ep rx_ep; 317 struct rpcrdma_buffer rx_buf; 318 struct rpcrdma_create_data_internal rx_data; 319 struct delayed_work rdma_connect; 320 struct rpcrdma_stats rx_stats; 321 }; 322 323 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, xprt) 324 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data) 325 326 /* Setting this to 0 ensures interoperability with early servers. 327 * Setting this to 1 enhances certain unaligned read/write performance. 328 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */ 329 extern int xprt_rdma_pad_optimize; 330 331 /* 332 * Interface Adapter calls - xprtrdma/verbs.c 333 */ 334 int rpcrdma_ia_open(struct rpcrdma_xprt *, struct sockaddr *, int); 335 void rpcrdma_ia_close(struct rpcrdma_ia *); 336 337 /* 338 * Endpoint calls - xprtrdma/verbs.c 339 */ 340 int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *, 341 struct rpcrdma_create_data_internal *); 342 void rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *); 343 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *); 344 void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *); 345 346 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *, 347 struct rpcrdma_req *); 348 int rpcrdma_ep_post_recv(struct rpcrdma_ia *, struct rpcrdma_ep *, 349 struct rpcrdma_rep *); 350 351 /* 352 * Buffer calls - xprtrdma/verbs.c 353 */ 354 int rpcrdma_buffer_create(struct rpcrdma_buffer *, struct rpcrdma_ep *, 355 struct rpcrdma_ia *, 356 struct rpcrdma_create_data_internal *); 357 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *); 358 359 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *); 360 void rpcrdma_buffer_put(struct rpcrdma_req *); 361 void rpcrdma_recv_buffer_get(struct rpcrdma_req *); 362 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *); 363 364 int rpcrdma_register_internal(struct rpcrdma_ia *, void *, int, 365 struct ib_mr **, struct ib_sge *); 366 int rpcrdma_deregister_internal(struct rpcrdma_ia *, 367 struct ib_mr *, struct ib_sge *); 368 369 int rpcrdma_register_external(struct rpcrdma_mr_seg *, 370 int, int, struct rpcrdma_xprt *); 371 int rpcrdma_deregister_external(struct rpcrdma_mr_seg *, 372 struct rpcrdma_xprt *); 373 374 /* 375 * RPC/RDMA connection management calls - xprtrdma/rpc_rdma.c 376 */ 377 void rpcrdma_connect_worker(struct work_struct *); 378 void rpcrdma_conn_func(struct rpcrdma_ep *); 379 void rpcrdma_reply_handler(struct rpcrdma_rep *); 380 381 /* 382 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c 383 */ 384 ssize_t rpcrdma_marshal_chunks(struct rpc_rqst *, ssize_t); 385 int rpcrdma_marshal_req(struct rpc_rqst *); 386 size_t rpcrdma_max_payload(struct rpcrdma_xprt *); 387 388 /* Temporary NFS request map cache. Created in svc_rdma.c */ 389 extern struct kmem_cache *svc_rdma_map_cachep; 390 /* WR context cache. Created in svc_rdma.c */ 391 extern struct kmem_cache *svc_rdma_ctxt_cachep; 392 /* Workqueue created in svc_rdma.c */ 393 extern struct workqueue_struct *svc_rdma_wq; 394 395 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */ 396