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 #include <linux/sunrpc/svc.h> /* RPCSVC_MAXPAYLOAD */ 55 56 #define RDMA_RESOLVE_TIMEOUT (5000) /* 5 seconds */ 57 #define RDMA_CONNECT_RETRY_MAX (2) /* retries if no listener backlog */ 58 59 /* 60 * Interface Adapter -- one per transport instance 61 */ 62 struct rpcrdma_ia { 63 rwlock_t ri_qplock; 64 struct rdma_cm_id *ri_id; 65 struct ib_pd *ri_pd; 66 struct ib_mr *ri_bind_mem; 67 u32 ri_dma_lkey; 68 int ri_have_dma_lkey; 69 struct completion ri_done; 70 int ri_async_rc; 71 enum rpcrdma_memreg ri_memreg_strategy; 72 unsigned int ri_max_frmr_depth; 73 }; 74 75 /* 76 * RDMA Endpoint -- one per transport instance 77 */ 78 79 #define RPCRDMA_WC_BUDGET (128) 80 #define RPCRDMA_POLLSIZE (16) 81 82 struct rpcrdma_ep { 83 atomic_t rep_cqcount; 84 int rep_cqinit; 85 int rep_connected; 86 struct rpcrdma_ia *rep_ia; 87 struct ib_qp_init_attr rep_attr; 88 wait_queue_head_t rep_connect_wait; 89 struct ib_sge rep_pad; /* holds zeroed pad */ 90 struct ib_mr *rep_pad_mr; /* holds zeroed pad */ 91 void (*rep_func)(struct rpcrdma_ep *); 92 struct rpc_xprt *rep_xprt; /* for rep_func */ 93 struct rdma_conn_param rep_remote_cma; 94 struct sockaddr_storage rep_remote_addr; 95 struct delayed_work rep_connect_worker; 96 struct ib_wc rep_send_wcs[RPCRDMA_POLLSIZE]; 97 struct ib_wc rep_recv_wcs[RPCRDMA_POLLSIZE]; 98 }; 99 100 #define INIT_CQCOUNT(ep) atomic_set(&(ep)->rep_cqcount, (ep)->rep_cqinit) 101 #define DECR_CQCOUNT(ep) atomic_sub_return(1, &(ep)->rep_cqcount) 102 103 enum rpcrdma_chunktype { 104 rpcrdma_noch = 0, 105 rpcrdma_readch, 106 rpcrdma_areadch, 107 rpcrdma_writech, 108 rpcrdma_replych 109 }; 110 111 /* 112 * struct rpcrdma_rep -- this structure encapsulates state required to recv 113 * and complete a reply, asychronously. It needs several pieces of 114 * state: 115 * o recv buffer (posted to provider) 116 * o ib_sge (also donated to provider) 117 * o status of reply (length, success or not) 118 * o bookkeeping state to get run by tasklet (list, etc) 119 * 120 * These are allocated during initialization, per-transport instance; 121 * however, the tasklet execution list itself is global, as it should 122 * always be pretty short. 123 * 124 * N of these are associated with a transport instance, and stored in 125 * struct rpcrdma_buffer. N is the max number of outstanding requests. 126 */ 127 128 /* temporary static scatter/gather max */ 129 #define RPCRDMA_MAX_DATA_SEGS (64) /* max scatter/gather */ 130 #define RPCRDMA_MAX_SEGS (RPCRDMA_MAX_DATA_SEGS + 2) /* head+tail = 2 */ 131 #define MAX_RPCRDMAHDR (\ 132 /* max supported RPC/RDMA header */ \ 133 sizeof(struct rpcrdma_msg) + (2 * sizeof(u32)) + \ 134 (sizeof(struct rpcrdma_read_chunk) * RPCRDMA_MAX_SEGS) + sizeof(u32)) 135 136 struct rpcrdma_buffer; 137 138 struct rpcrdma_rep { 139 unsigned int rr_len; /* actual received reply length */ 140 struct rpcrdma_buffer *rr_buffer; /* home base for this structure */ 141 struct rpc_xprt *rr_xprt; /* needed for request/reply matching */ 142 void (*rr_func)(struct rpcrdma_rep *);/* called by tasklet in softint */ 143 struct list_head rr_list; /* tasklet list */ 144 struct ib_sge rr_iov; /* for posting */ 145 struct ib_mr *rr_handle; /* handle for mem in rr_iov */ 146 char rr_base[MAX_RPCRDMAHDR]; /* minimal inline receive buffer */ 147 }; 148 149 /* 150 * struct rpcrdma_mw - external memory region metadata 151 * 152 * An external memory region is any buffer or page that is registered 153 * on the fly (ie, not pre-registered). 154 * 155 * Each rpcrdma_buffer has a list of free MWs anchored in rb_mws. During 156 * call_allocate, rpcrdma_buffer_get() assigns one to each segment in 157 * an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep 158 * track of registration metadata while each RPC is pending. 159 * rpcrdma_deregister_external() uses this metadata to unmap and 160 * release these resources when an RPC is complete. 161 */ 162 enum rpcrdma_frmr_state { 163 FRMR_IS_INVALID, /* ready to be used */ 164 FRMR_IS_VALID, /* in use */ 165 FRMR_IS_STALE, /* failed completion */ 166 }; 167 168 struct rpcrdma_frmr { 169 struct ib_fast_reg_page_list *fr_pgl; 170 struct ib_mr *fr_mr; 171 enum rpcrdma_frmr_state fr_state; 172 }; 173 174 struct rpcrdma_mw { 175 union { 176 struct ib_fmr *fmr; 177 struct rpcrdma_frmr frmr; 178 } r; 179 struct list_head mw_list; 180 struct list_head mw_all; 181 }; 182 183 /* 184 * struct rpcrdma_req -- structure central to the request/reply sequence. 185 * 186 * N of these are associated with a transport instance, and stored in 187 * struct rpcrdma_buffer. N is the max number of outstanding requests. 188 * 189 * It includes pre-registered buffer memory for send AND recv. 190 * The recv buffer, however, is not owned by this structure, and 191 * is "donated" to the hardware when a recv is posted. When a 192 * reply is handled, the recv buffer used is given back to the 193 * struct rpcrdma_req associated with the request. 194 * 195 * In addition to the basic memory, this structure includes an array 196 * of iovs for send operations. The reason is that the iovs passed to 197 * ib_post_{send,recv} must not be modified until the work request 198 * completes. 199 * 200 * NOTES: 201 * o RPCRDMA_MAX_SEGS is the max number of addressible chunk elements we 202 * marshal. The number needed varies depending on the iov lists that 203 * are passed to us, the memory registration mode we are in, and if 204 * physical addressing is used, the layout. 205 */ 206 207 struct rpcrdma_mr_seg { /* chunk descriptors */ 208 union { /* chunk memory handles */ 209 struct ib_mr *rl_mr; /* if registered directly */ 210 struct rpcrdma_mw *rl_mw; /* if registered from region */ 211 } mr_chunk; 212 u64 mr_base; /* registration result */ 213 u32 mr_rkey; /* registration result */ 214 u32 mr_len; /* length of chunk or segment */ 215 int mr_nsegs; /* number of segments in chunk or 0 */ 216 enum dma_data_direction mr_dir; /* segment mapping direction */ 217 dma_addr_t mr_dma; /* segment mapping address */ 218 size_t mr_dmalen; /* segment mapping length */ 219 struct page *mr_page; /* owning page, if any */ 220 char *mr_offset; /* kva if no page, else offset */ 221 }; 222 223 struct rpcrdma_req { 224 size_t rl_size; /* actual length of buffer */ 225 unsigned int rl_niovs; /* 0, 2 or 4 */ 226 unsigned int rl_nchunks; /* non-zero if chunks */ 227 unsigned int rl_connect_cookie; /* retry detection */ 228 enum rpcrdma_chunktype rl_rtype, rl_wtype; 229 struct rpcrdma_buffer *rl_buffer; /* home base for this structure */ 230 struct rpcrdma_rep *rl_reply;/* holder for reply buffer */ 231 struct rpcrdma_mr_seg rl_segments[RPCRDMA_MAX_SEGS];/* chunk segments */ 232 struct ib_sge rl_send_iov[4]; /* for active requests */ 233 struct ib_sge rl_iov; /* for posting */ 234 struct ib_mr *rl_handle; /* handle for mem in rl_iov */ 235 char rl_base[MAX_RPCRDMAHDR]; /* start of actual buffer */ 236 __u32 rl_xdr_buf[0]; /* start of returned rpc rq_buffer */ 237 }; 238 #define rpcr_to_rdmar(r) \ 239 container_of((r)->rq_buffer, struct rpcrdma_req, rl_xdr_buf[0]) 240 241 /* 242 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for 243 * inline requests/replies, and client/server credits. 244 * 245 * One of these is associated with a transport instance 246 */ 247 struct rpcrdma_buffer { 248 spinlock_t rb_lock; /* protects indexes */ 249 atomic_t rb_credits; /* most recent server credits */ 250 int rb_max_requests;/* client max requests */ 251 struct list_head rb_mws; /* optional memory windows/fmrs/frmrs */ 252 struct list_head rb_all; 253 int rb_send_index; 254 struct rpcrdma_req **rb_send_bufs; 255 int rb_recv_index; 256 struct rpcrdma_rep **rb_recv_bufs; 257 char *rb_pool; 258 }; 259 #define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia) 260 261 /* 262 * Internal structure for transport instance creation. This 263 * exists primarily for modularity. 264 * 265 * This data should be set with mount options 266 */ 267 struct rpcrdma_create_data_internal { 268 struct sockaddr_storage addr; /* RDMA server address */ 269 unsigned int max_requests; /* max requests (slots) in flight */ 270 unsigned int rsize; /* mount rsize - max read hdr+data */ 271 unsigned int wsize; /* mount wsize - max write hdr+data */ 272 unsigned int inline_rsize; /* max non-rdma read data payload */ 273 unsigned int inline_wsize; /* max non-rdma write data payload */ 274 unsigned int padding; /* non-rdma write header padding */ 275 }; 276 277 #define RPCRDMA_INLINE_READ_THRESHOLD(rq) \ 278 (rpcx_to_rdmad(rq->rq_xprt).inline_rsize) 279 280 #define RPCRDMA_INLINE_WRITE_THRESHOLD(rq)\ 281 (rpcx_to_rdmad(rq->rq_xprt).inline_wsize) 282 283 #define RPCRDMA_INLINE_PAD_VALUE(rq)\ 284 rpcx_to_rdmad(rq->rq_xprt).padding 285 286 /* 287 * Statistics for RPCRDMA 288 */ 289 struct rpcrdma_stats { 290 unsigned long read_chunk_count; 291 unsigned long write_chunk_count; 292 unsigned long reply_chunk_count; 293 294 unsigned long long total_rdma_request; 295 unsigned long long total_rdma_reply; 296 297 unsigned long long pullup_copy_count; 298 unsigned long long fixup_copy_count; 299 unsigned long hardway_register_count; 300 unsigned long failed_marshal_count; 301 unsigned long bad_reply_count; 302 }; 303 304 /* 305 * RPCRDMA transport -- encapsulates the structures above for 306 * integration with RPC. 307 * 308 * The contained structures are embedded, not pointers, 309 * for convenience. This structure need not be visible externally. 310 * 311 * It is allocated and initialized during mount, and released 312 * during unmount. 313 */ 314 struct rpcrdma_xprt { 315 struct rpc_xprt xprt; 316 struct rpcrdma_ia rx_ia; 317 struct rpcrdma_ep rx_ep; 318 struct rpcrdma_buffer rx_buf; 319 struct rpcrdma_create_data_internal rx_data; 320 struct delayed_work rdma_connect; 321 struct rpcrdma_stats rx_stats; 322 }; 323 324 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, xprt) 325 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data) 326 327 /* Setting this to 0 ensures interoperability with early servers. 328 * Setting this to 1 enhances certain unaligned read/write performance. 329 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */ 330 extern int xprt_rdma_pad_optimize; 331 332 /* 333 * Interface Adapter calls - xprtrdma/verbs.c 334 */ 335 int rpcrdma_ia_open(struct rpcrdma_xprt *, struct sockaddr *, int); 336 void rpcrdma_ia_close(struct rpcrdma_ia *); 337 338 /* 339 * Endpoint calls - xprtrdma/verbs.c 340 */ 341 int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *, 342 struct rpcrdma_create_data_internal *); 343 void rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *); 344 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *); 345 void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *); 346 347 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *, 348 struct rpcrdma_req *); 349 int rpcrdma_ep_post_recv(struct rpcrdma_ia *, struct rpcrdma_ep *, 350 struct rpcrdma_rep *); 351 352 /* 353 * Buffer calls - xprtrdma/verbs.c 354 */ 355 int rpcrdma_buffer_create(struct rpcrdma_buffer *, struct rpcrdma_ep *, 356 struct rpcrdma_ia *, 357 struct rpcrdma_create_data_internal *); 358 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *); 359 360 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *); 361 void rpcrdma_buffer_put(struct rpcrdma_req *); 362 void rpcrdma_recv_buffer_get(struct rpcrdma_req *); 363 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *); 364 365 int rpcrdma_register_internal(struct rpcrdma_ia *, void *, int, 366 struct ib_mr **, struct ib_sge *); 367 int rpcrdma_deregister_internal(struct rpcrdma_ia *, 368 struct ib_mr *, struct ib_sge *); 369 370 int rpcrdma_register_external(struct rpcrdma_mr_seg *, 371 int, int, struct rpcrdma_xprt *); 372 int rpcrdma_deregister_external(struct rpcrdma_mr_seg *, 373 struct rpcrdma_xprt *); 374 375 /* 376 * RPC/RDMA connection management calls - xprtrdma/rpc_rdma.c 377 */ 378 void rpcrdma_connect_worker(struct work_struct *); 379 void rpcrdma_conn_func(struct rpcrdma_ep *); 380 void rpcrdma_reply_handler(struct rpcrdma_rep *); 381 382 /* 383 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c 384 */ 385 ssize_t rpcrdma_marshal_chunks(struct rpc_rqst *, ssize_t); 386 int rpcrdma_marshal_req(struct rpc_rqst *); 387 size_t rpcrdma_max_payload(struct rpcrdma_xprt *); 388 389 /* Temporary NFS request map cache. Created in svc_rdma.c */ 390 extern struct kmem_cache *svc_rdma_map_cachep; 391 /* WR context cache. Created in svc_rdma.c */ 392 extern struct kmem_cache *svc_rdma_ctxt_cachep; 393 /* Workqueue created in svc_rdma.c */ 394 extern struct workqueue_struct *svc_rdma_wq; 395 396 #if RPCSVC_MAXPAYLOAD < (RPCRDMA_MAX_DATA_SEGS << PAGE_SHIFT) 397 #define RPCSVC_MAXPAYLOAD_RDMA RPCSVC_MAXPAYLOAD 398 #else 399 #define RPCSVC_MAXPAYLOAD_RDMA (RPCRDMA_MAX_DATA_SEGS << PAGE_SHIFT) 400 #endif 401 402 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */ 403