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 /* 101 * Force a signaled SEND Work Request every so often, 102 * in case the provider needs to do some housekeeping. 103 */ 104 #define RPCRDMA_MAX_UNSIGNALED_SENDS (32) 105 106 #define INIT_CQCOUNT(ep) atomic_set(&(ep)->rep_cqcount, (ep)->rep_cqinit) 107 #define DECR_CQCOUNT(ep) atomic_sub_return(1, &(ep)->rep_cqcount) 108 109 enum rpcrdma_chunktype { 110 rpcrdma_noch = 0, 111 rpcrdma_readch, 112 rpcrdma_areadch, 113 rpcrdma_writech, 114 rpcrdma_replych 115 }; 116 117 /* 118 * struct rpcrdma_rep -- this structure encapsulates state required to recv 119 * and complete a reply, asychronously. It needs several pieces of 120 * state: 121 * o recv buffer (posted to provider) 122 * o ib_sge (also donated to provider) 123 * o status of reply (length, success or not) 124 * o bookkeeping state to get run by tasklet (list, etc) 125 * 126 * These are allocated during initialization, per-transport instance; 127 * however, the tasklet execution list itself is global, as it should 128 * always be pretty short. 129 * 130 * N of these are associated with a transport instance, and stored in 131 * struct rpcrdma_buffer. N is the max number of outstanding requests. 132 */ 133 134 /* temporary static scatter/gather max */ 135 #define RPCRDMA_MAX_DATA_SEGS (64) /* max scatter/gather */ 136 #define RPCRDMA_MAX_SEGS (RPCRDMA_MAX_DATA_SEGS + 2) /* head+tail = 2 */ 137 #define MAX_RPCRDMAHDR (\ 138 /* max supported RPC/RDMA header */ \ 139 sizeof(struct rpcrdma_msg) + (2 * sizeof(u32)) + \ 140 (sizeof(struct rpcrdma_read_chunk) * RPCRDMA_MAX_SEGS) + sizeof(u32)) 141 142 struct rpcrdma_buffer; 143 144 struct rpcrdma_rep { 145 unsigned int rr_len; /* actual received reply length */ 146 struct rpcrdma_buffer *rr_buffer; /* home base for this structure */ 147 struct rpc_xprt *rr_xprt; /* needed for request/reply matching */ 148 void (*rr_func)(struct rpcrdma_rep *);/* called by tasklet in softint */ 149 struct list_head rr_list; /* tasklet list */ 150 struct ib_sge rr_iov; /* for posting */ 151 struct ib_mr *rr_handle; /* handle for mem in rr_iov */ 152 char rr_base[MAX_RPCRDMAHDR]; /* minimal inline receive buffer */ 153 }; 154 155 /* 156 * struct rpcrdma_mw - external memory region metadata 157 * 158 * An external memory region is any buffer or page that is registered 159 * on the fly (ie, not pre-registered). 160 * 161 * Each rpcrdma_buffer has a list of free MWs anchored in rb_mws. During 162 * call_allocate, rpcrdma_buffer_get() assigns one to each segment in 163 * an rpcrdma_req. Then rpcrdma_register_external() grabs these to keep 164 * track of registration metadata while each RPC is pending. 165 * rpcrdma_deregister_external() uses this metadata to unmap and 166 * release these resources when an RPC is complete. 167 */ 168 enum rpcrdma_frmr_state { 169 FRMR_IS_INVALID, /* ready to be used */ 170 FRMR_IS_VALID, /* in use */ 171 FRMR_IS_STALE, /* failed completion */ 172 }; 173 174 struct rpcrdma_frmr { 175 struct ib_fast_reg_page_list *fr_pgl; 176 struct ib_mr *fr_mr; 177 enum rpcrdma_frmr_state fr_state; 178 }; 179 180 struct rpcrdma_mw { 181 union { 182 struct ib_fmr *fmr; 183 struct rpcrdma_frmr frmr; 184 } r; 185 struct list_head mw_list; 186 struct list_head mw_all; 187 }; 188 189 /* 190 * struct rpcrdma_req -- structure central to the request/reply sequence. 191 * 192 * N of these are associated with a transport instance, and stored in 193 * struct rpcrdma_buffer. N is the max number of outstanding requests. 194 * 195 * It includes pre-registered buffer memory for send AND recv. 196 * The recv buffer, however, is not owned by this structure, and 197 * is "donated" to the hardware when a recv is posted. When a 198 * reply is handled, the recv buffer used is given back to the 199 * struct rpcrdma_req associated with the request. 200 * 201 * In addition to the basic memory, this structure includes an array 202 * of iovs for send operations. The reason is that the iovs passed to 203 * ib_post_{send,recv} must not be modified until the work request 204 * completes. 205 * 206 * NOTES: 207 * o RPCRDMA_MAX_SEGS is the max number of addressible chunk elements we 208 * marshal. The number needed varies depending on the iov lists that 209 * are passed to us, the memory registration mode we are in, and if 210 * physical addressing is used, the layout. 211 */ 212 213 struct rpcrdma_mr_seg { /* chunk descriptors */ 214 union { /* chunk memory handles */ 215 struct ib_mr *rl_mr; /* if registered directly */ 216 struct rpcrdma_mw *rl_mw; /* if registered from region */ 217 } mr_chunk; 218 u64 mr_base; /* registration result */ 219 u32 mr_rkey; /* registration result */ 220 u32 mr_len; /* length of chunk or segment */ 221 int mr_nsegs; /* number of segments in chunk or 0 */ 222 enum dma_data_direction mr_dir; /* segment mapping direction */ 223 dma_addr_t mr_dma; /* segment mapping address */ 224 size_t mr_dmalen; /* segment mapping length */ 225 struct page *mr_page; /* owning page, if any */ 226 char *mr_offset; /* kva if no page, else offset */ 227 }; 228 229 struct rpcrdma_req { 230 size_t rl_size; /* actual length of buffer */ 231 unsigned int rl_niovs; /* 0, 2 or 4 */ 232 unsigned int rl_nchunks; /* non-zero if chunks */ 233 unsigned int rl_connect_cookie; /* retry detection */ 234 enum rpcrdma_chunktype rl_rtype, rl_wtype; 235 struct rpcrdma_buffer *rl_buffer; /* home base for this structure */ 236 struct rpcrdma_rep *rl_reply;/* holder for reply buffer */ 237 struct rpcrdma_mr_seg rl_segments[RPCRDMA_MAX_SEGS];/* chunk segments */ 238 struct ib_sge rl_send_iov[4]; /* for active requests */ 239 struct ib_sge rl_iov; /* for posting */ 240 struct ib_mr *rl_handle; /* handle for mem in rl_iov */ 241 char rl_base[MAX_RPCRDMAHDR]; /* start of actual buffer */ 242 __u32 rl_xdr_buf[0]; /* start of returned rpc rq_buffer */ 243 }; 244 #define rpcr_to_rdmar(r) \ 245 container_of((r)->rq_buffer, struct rpcrdma_req, rl_xdr_buf[0]) 246 247 /* 248 * struct rpcrdma_buffer -- holds list/queue of pre-registered memory for 249 * inline requests/replies, and client/server credits. 250 * 251 * One of these is associated with a transport instance 252 */ 253 struct rpcrdma_buffer { 254 spinlock_t rb_lock; /* protects indexes */ 255 atomic_t rb_credits; /* most recent server credits */ 256 int rb_max_requests;/* client max requests */ 257 struct list_head rb_mws; /* optional memory windows/fmrs/frmrs */ 258 struct list_head rb_all; 259 int rb_send_index; 260 struct rpcrdma_req **rb_send_bufs; 261 int rb_recv_index; 262 struct rpcrdma_rep **rb_recv_bufs; 263 char *rb_pool; 264 }; 265 #define rdmab_to_ia(b) (&container_of((b), struct rpcrdma_xprt, rx_buf)->rx_ia) 266 267 /* 268 * Internal structure for transport instance creation. This 269 * exists primarily for modularity. 270 * 271 * This data should be set with mount options 272 */ 273 struct rpcrdma_create_data_internal { 274 struct sockaddr_storage addr; /* RDMA server address */ 275 unsigned int max_requests; /* max requests (slots) in flight */ 276 unsigned int rsize; /* mount rsize - max read hdr+data */ 277 unsigned int wsize; /* mount wsize - max write hdr+data */ 278 unsigned int inline_rsize; /* max non-rdma read data payload */ 279 unsigned int inline_wsize; /* max non-rdma write data payload */ 280 unsigned int padding; /* non-rdma write header padding */ 281 }; 282 283 #define RPCRDMA_INLINE_READ_THRESHOLD(rq) \ 284 (rpcx_to_rdmad(rq->rq_xprt).inline_rsize) 285 286 #define RPCRDMA_INLINE_WRITE_THRESHOLD(rq)\ 287 (rpcx_to_rdmad(rq->rq_xprt).inline_wsize) 288 289 #define RPCRDMA_INLINE_PAD_VALUE(rq)\ 290 rpcx_to_rdmad(rq->rq_xprt).padding 291 292 /* 293 * Statistics for RPCRDMA 294 */ 295 struct rpcrdma_stats { 296 unsigned long read_chunk_count; 297 unsigned long write_chunk_count; 298 unsigned long reply_chunk_count; 299 300 unsigned long long total_rdma_request; 301 unsigned long long total_rdma_reply; 302 303 unsigned long long pullup_copy_count; 304 unsigned long long fixup_copy_count; 305 unsigned long hardway_register_count; 306 unsigned long failed_marshal_count; 307 unsigned long bad_reply_count; 308 }; 309 310 /* 311 * RPCRDMA transport -- encapsulates the structures above for 312 * integration with RPC. 313 * 314 * The contained structures are embedded, not pointers, 315 * for convenience. This structure need not be visible externally. 316 * 317 * It is allocated and initialized during mount, and released 318 * during unmount. 319 */ 320 struct rpcrdma_xprt { 321 struct rpc_xprt xprt; 322 struct rpcrdma_ia rx_ia; 323 struct rpcrdma_ep rx_ep; 324 struct rpcrdma_buffer rx_buf; 325 struct rpcrdma_create_data_internal rx_data; 326 struct delayed_work rdma_connect; 327 struct rpcrdma_stats rx_stats; 328 }; 329 330 #define rpcx_to_rdmax(x) container_of(x, struct rpcrdma_xprt, xprt) 331 #define rpcx_to_rdmad(x) (rpcx_to_rdmax(x)->rx_data) 332 333 /* Setting this to 0 ensures interoperability with early servers. 334 * Setting this to 1 enhances certain unaligned read/write performance. 335 * Default is 0, see sysctl entry and rpc_rdma.c rpcrdma_convert_iovs() */ 336 extern int xprt_rdma_pad_optimize; 337 338 /* 339 * Interface Adapter calls - xprtrdma/verbs.c 340 */ 341 int rpcrdma_ia_open(struct rpcrdma_xprt *, struct sockaddr *, int); 342 void rpcrdma_ia_close(struct rpcrdma_ia *); 343 344 /* 345 * Endpoint calls - xprtrdma/verbs.c 346 */ 347 int rpcrdma_ep_create(struct rpcrdma_ep *, struct rpcrdma_ia *, 348 struct rpcrdma_create_data_internal *); 349 void rpcrdma_ep_destroy(struct rpcrdma_ep *, struct rpcrdma_ia *); 350 int rpcrdma_ep_connect(struct rpcrdma_ep *, struct rpcrdma_ia *); 351 void rpcrdma_ep_disconnect(struct rpcrdma_ep *, struct rpcrdma_ia *); 352 353 int rpcrdma_ep_post(struct rpcrdma_ia *, struct rpcrdma_ep *, 354 struct rpcrdma_req *); 355 int rpcrdma_ep_post_recv(struct rpcrdma_ia *, struct rpcrdma_ep *, 356 struct rpcrdma_rep *); 357 358 /* 359 * Buffer calls - xprtrdma/verbs.c 360 */ 361 int rpcrdma_buffer_create(struct rpcrdma_buffer *, struct rpcrdma_ep *, 362 struct rpcrdma_ia *, 363 struct rpcrdma_create_data_internal *); 364 void rpcrdma_buffer_destroy(struct rpcrdma_buffer *); 365 366 struct rpcrdma_req *rpcrdma_buffer_get(struct rpcrdma_buffer *); 367 void rpcrdma_buffer_put(struct rpcrdma_req *); 368 void rpcrdma_recv_buffer_get(struct rpcrdma_req *); 369 void rpcrdma_recv_buffer_put(struct rpcrdma_rep *); 370 371 int rpcrdma_register_internal(struct rpcrdma_ia *, void *, int, 372 struct ib_mr **, struct ib_sge *); 373 int rpcrdma_deregister_internal(struct rpcrdma_ia *, 374 struct ib_mr *, struct ib_sge *); 375 376 int rpcrdma_register_external(struct rpcrdma_mr_seg *, 377 int, int, struct rpcrdma_xprt *); 378 int rpcrdma_deregister_external(struct rpcrdma_mr_seg *, 379 struct rpcrdma_xprt *); 380 381 /* 382 * RPC/RDMA connection management calls - xprtrdma/rpc_rdma.c 383 */ 384 void rpcrdma_connect_worker(struct work_struct *); 385 void rpcrdma_conn_func(struct rpcrdma_ep *); 386 void rpcrdma_reply_handler(struct rpcrdma_rep *); 387 388 /* 389 * RPC/RDMA protocol calls - xprtrdma/rpc_rdma.c 390 */ 391 ssize_t rpcrdma_marshal_chunks(struct rpc_rqst *, ssize_t); 392 int rpcrdma_marshal_req(struct rpc_rqst *); 393 size_t rpcrdma_max_payload(struct rpcrdma_xprt *); 394 395 /* Temporary NFS request map cache. Created in svc_rdma.c */ 396 extern struct kmem_cache *svc_rdma_map_cachep; 397 /* WR context cache. Created in svc_rdma.c */ 398 extern struct kmem_cache *svc_rdma_ctxt_cachep; 399 /* Workqueue created in svc_rdma.c */ 400 extern struct workqueue_struct *svc_rdma_wq; 401 402 #if RPCSVC_MAXPAYLOAD < (RPCRDMA_MAX_DATA_SEGS << PAGE_SHIFT) 403 #define RPCSVC_MAXPAYLOAD_RDMA RPCSVC_MAXPAYLOAD 404 #else 405 #define RPCSVC_MAXPAYLOAD_RDMA (RPCRDMA_MAX_DATA_SEGS << PAGE_SHIFT) 406 #endif 407 408 #endif /* _LINUX_SUNRPC_XPRT_RDMA_H */ 409