1 /* 2 * Copyright (c) 2005-2006 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 * Author: Tom Tucker <tom@opengridcomputing.com> 40 */ 41 42 #ifndef SVC_RDMA_H 43 #define SVC_RDMA_H 44 #include <linux/sunrpc/xdr.h> 45 #include <linux/sunrpc/svcsock.h> 46 #include <linux/sunrpc/rpc_rdma.h> 47 #include <rdma/ib_verbs.h> 48 #include <rdma/rdma_cm.h> 49 #define SVCRDMA_DEBUG 50 51 /* RPC/RDMA parameters and stats */ 52 extern unsigned int svcrdma_ord; 53 extern unsigned int svcrdma_max_requests; 54 extern unsigned int svcrdma_max_req_size; 55 56 extern atomic_t rdma_stat_recv; 57 extern atomic_t rdma_stat_read; 58 extern atomic_t rdma_stat_write; 59 extern atomic_t rdma_stat_sq_starve; 60 extern atomic_t rdma_stat_rq_starve; 61 extern atomic_t rdma_stat_rq_poll; 62 extern atomic_t rdma_stat_rq_prod; 63 extern atomic_t rdma_stat_sq_poll; 64 extern atomic_t rdma_stat_sq_prod; 65 66 /* 67 * Contexts are built when an RDMA request is created and are a 68 * record of the resources that can be recovered when the request 69 * completes. 70 */ 71 struct svc_rdma_op_ctxt { 72 struct svc_rdma_op_ctxt *read_hdr; 73 struct svc_rdma_fastreg_mr *frmr; 74 int hdr_count; 75 struct xdr_buf arg; 76 struct list_head dto_q; 77 enum ib_wr_opcode wr_op; 78 enum ib_wc_status wc_status; 79 u32 byte_len; 80 u32 position; 81 struct svcxprt_rdma *xprt; 82 unsigned long flags; 83 enum dma_data_direction direction; 84 int count; 85 struct ib_sge sge[RPCSVC_MAXPAGES]; 86 struct page *pages[RPCSVC_MAXPAGES]; 87 }; 88 89 /* 90 * NFS_ requests are mapped on the client side by the chunk lists in 91 * the RPCRDMA header. During the fetching of the RPC from the client 92 * and the writing of the reply to the client, the memory in the 93 * client and the memory in the server must be mapped as contiguous 94 * vaddr/len for access by the hardware. These data strucures keep 95 * these mappings. 96 * 97 * For an RDMA_WRITE, the 'sge' maps the RPC REPLY. For RDMA_READ, the 98 * 'sge' in the svc_rdma_req_map maps the server side RPC reply and the 99 * 'ch' field maps the read-list of the RPCRDMA header to the 'sge' 100 * mapping of the reply. 101 */ 102 struct svc_rdma_chunk_sge { 103 int start; /* sge no for this chunk */ 104 int count; /* sge count for this chunk */ 105 }; 106 struct svc_rdma_fastreg_mr { 107 struct ib_mr *mr; 108 void *kva; 109 struct ib_fast_reg_page_list *page_list; 110 int page_list_len; 111 unsigned long access_flags; 112 unsigned long map_len; 113 enum dma_data_direction direction; 114 struct list_head frmr_list; 115 }; 116 struct svc_rdma_req_map { 117 unsigned long count; 118 union { 119 struct kvec sge[RPCSVC_MAXPAGES]; 120 struct svc_rdma_chunk_sge ch[RPCSVC_MAXPAGES]; 121 unsigned long lkey[RPCSVC_MAXPAGES]; 122 }; 123 }; 124 #define RDMACTXT_F_LAST_CTXT 2 125 126 #define SVCRDMA_DEVCAP_FAST_REG 1 /* fast mr registration */ 127 #define SVCRDMA_DEVCAP_READ_W_INV 2 /* read w/ invalidate */ 128 129 struct svcxprt_rdma { 130 struct svc_xprt sc_xprt; /* SVC transport structure */ 131 struct rdma_cm_id *sc_cm_id; /* RDMA connection id */ 132 struct list_head sc_accept_q; /* Conn. waiting accept */ 133 int sc_ord; /* RDMA read limit */ 134 int sc_max_sge; 135 136 int sc_sq_depth; /* Depth of SQ */ 137 atomic_t sc_sq_count; /* Number of SQ WR on queue */ 138 139 int sc_max_requests; /* Depth of RQ */ 140 int sc_max_req_size; /* Size of each RQ WR buf */ 141 142 struct ib_pd *sc_pd; 143 144 atomic_t sc_dma_used; 145 atomic_t sc_ctxt_used; 146 struct list_head sc_rq_dto_q; 147 spinlock_t sc_rq_dto_lock; 148 struct ib_qp *sc_qp; 149 struct ib_cq *sc_rq_cq; 150 struct ib_cq *sc_sq_cq; 151 struct ib_mr *sc_phys_mr; /* MR for server memory */ 152 int (*sc_reader)(struct svcxprt_rdma *, 153 struct svc_rqst *, 154 struct svc_rdma_op_ctxt *, 155 int *, u32 *, u32, u32, u64, bool); 156 u32 sc_dev_caps; /* distilled device caps */ 157 u32 sc_dma_lkey; /* local dma key */ 158 unsigned int sc_frmr_pg_list_len; 159 struct list_head sc_frmr_q; 160 spinlock_t sc_frmr_q_lock; 161 162 spinlock_t sc_lock; /* transport lock */ 163 164 wait_queue_head_t sc_send_wait; /* SQ exhaustion waitlist */ 165 unsigned long sc_flags; 166 struct list_head sc_dto_q; /* DTO tasklet I/O pending Q */ 167 struct list_head sc_read_complete_q; 168 struct work_struct sc_work; 169 }; 170 /* sc_flags */ 171 #define RDMAXPRT_RQ_PENDING 1 172 #define RDMAXPRT_SQ_PENDING 2 173 #define RDMAXPRT_CONN_PENDING 3 174 175 #define RPCRDMA_MAX_SVC_SEGS (64) /* server max scatter/gather */ 176 #if RPCSVC_MAXPAYLOAD < (RPCRDMA_MAX_SVC_SEGS << PAGE_SHIFT) 177 #define RPCRDMA_MAXPAYLOAD RPCSVC_MAXPAYLOAD 178 #else 179 #define RPCRDMA_MAXPAYLOAD (RPCRDMA_MAX_SVC_SEGS << PAGE_SHIFT) 180 #endif 181 182 #define RPCRDMA_LISTEN_BACKLOG 10 183 /* The default ORD value is based on two outstanding full-size writes with a 184 * page size of 4k, or 32k * 2 ops / 4k = 16 outstanding RDMA_READ. */ 185 #define RPCRDMA_ORD (64/4) 186 #define RPCRDMA_SQ_DEPTH_MULT 8 187 #define RPCRDMA_MAX_REQUESTS 32 188 #define RPCRDMA_MAX_REQ_SIZE 4096 189 190 /* svc_rdma_marshal.c */ 191 extern int svc_rdma_xdr_decode_req(struct rpcrdma_msg **, struct svc_rqst *); 192 extern int svc_rdma_xdr_encode_error(struct svcxprt_rdma *, 193 struct rpcrdma_msg *, 194 enum rpcrdma_errcode, __be32 *); 195 extern void svc_rdma_xdr_encode_write_list(struct rpcrdma_msg *, int); 196 extern void svc_rdma_xdr_encode_reply_array(struct rpcrdma_write_array *, int); 197 extern void svc_rdma_xdr_encode_array_chunk(struct rpcrdma_write_array *, int, 198 __be32, __be64, u32); 199 extern void svc_rdma_xdr_encode_reply_header(struct svcxprt_rdma *, 200 struct rpcrdma_msg *, 201 struct rpcrdma_msg *, 202 enum rpcrdma_proc); 203 extern int svc_rdma_xdr_get_reply_hdr_len(struct rpcrdma_msg *); 204 205 /* svc_rdma_recvfrom.c */ 206 extern int svc_rdma_recvfrom(struct svc_rqst *); 207 extern int rdma_read_chunk_lcl(struct svcxprt_rdma *, struct svc_rqst *, 208 struct svc_rdma_op_ctxt *, int *, u32 *, 209 u32, u32, u64, bool); 210 extern int rdma_read_chunk_frmr(struct svcxprt_rdma *, struct svc_rqst *, 211 struct svc_rdma_op_ctxt *, int *, u32 *, 212 u32, u32, u64, bool); 213 214 /* svc_rdma_sendto.c */ 215 extern int svc_rdma_sendto(struct svc_rqst *); 216 217 /* svc_rdma_transport.c */ 218 extern int svc_rdma_send(struct svcxprt_rdma *, struct ib_send_wr *); 219 extern void svc_rdma_send_error(struct svcxprt_rdma *, struct rpcrdma_msg *, 220 enum rpcrdma_errcode); 221 extern int svc_rdma_post_recv(struct svcxprt_rdma *); 222 extern int svc_rdma_create_listen(struct svc_serv *, int, struct sockaddr *); 223 extern struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *); 224 extern void svc_rdma_put_context(struct svc_rdma_op_ctxt *, int); 225 extern void svc_rdma_unmap_dma(struct svc_rdma_op_ctxt *ctxt); 226 extern struct svc_rdma_req_map *svc_rdma_get_req_map(void); 227 extern void svc_rdma_put_req_map(struct svc_rdma_req_map *); 228 extern int svc_rdma_fastreg(struct svcxprt_rdma *, struct svc_rdma_fastreg_mr *); 229 extern struct svc_rdma_fastreg_mr *svc_rdma_get_frmr(struct svcxprt_rdma *); 230 extern void svc_rdma_put_frmr(struct svcxprt_rdma *, 231 struct svc_rdma_fastreg_mr *); 232 extern void svc_sq_reap(struct svcxprt_rdma *); 233 extern void svc_rq_reap(struct svcxprt_rdma *); 234 extern struct svc_xprt_class svc_rdma_class; 235 extern void svc_rdma_prep_reply_hdr(struct svc_rqst *); 236 237 /* svc_rdma.c */ 238 extern int svc_rdma_init(void); 239 extern void svc_rdma_cleanup(void); 240 241 /* 242 * Returns the address of the first read chunk or <nul> if no read chunk is 243 * present 244 */ 245 static inline struct rpcrdma_read_chunk * 246 svc_rdma_get_read_chunk(struct rpcrdma_msg *rmsgp) 247 { 248 struct rpcrdma_read_chunk *ch = 249 (struct rpcrdma_read_chunk *)&rmsgp->rm_body.rm_chunks[0]; 250 251 if (ch->rc_discrim == 0) 252 return NULL; 253 254 return ch; 255 } 256 257 /* 258 * Returns the address of the first read write array element or <nul> if no 259 * write array list is present 260 */ 261 static inline struct rpcrdma_write_array * 262 svc_rdma_get_write_array(struct rpcrdma_msg *rmsgp) 263 { 264 if (rmsgp->rm_body.rm_chunks[0] != 0 265 || rmsgp->rm_body.rm_chunks[1] == 0) 266 return NULL; 267 268 return (struct rpcrdma_write_array *)&rmsgp->rm_body.rm_chunks[1]; 269 } 270 271 /* 272 * Returns the address of the first reply array element or <nul> if no 273 * reply array is present 274 */ 275 static inline struct rpcrdma_write_array * 276 svc_rdma_get_reply_array(struct rpcrdma_msg *rmsgp) 277 { 278 struct rpcrdma_read_chunk *rch; 279 struct rpcrdma_write_array *wr_ary; 280 struct rpcrdma_write_array *rp_ary; 281 282 /* XXX: Need to fix when reply list may occur with read-list and/or 283 * write list */ 284 if (rmsgp->rm_body.rm_chunks[0] != 0 || 285 rmsgp->rm_body.rm_chunks[1] != 0) 286 return NULL; 287 288 rch = svc_rdma_get_read_chunk(rmsgp); 289 if (rch) { 290 while (rch->rc_discrim) 291 rch++; 292 293 /* The reply list follows an empty write array located 294 * at 'rc_position' here. The reply array is at rc_target. 295 */ 296 rp_ary = (struct rpcrdma_write_array *)&rch->rc_target; 297 298 goto found_it; 299 } 300 301 wr_ary = svc_rdma_get_write_array(rmsgp); 302 if (wr_ary) { 303 rp_ary = (struct rpcrdma_write_array *) 304 &wr_ary-> 305 wc_array[ntohl(wr_ary->wc_nchunks)].wc_target.rs_length; 306 307 goto found_it; 308 } 309 310 /* No read list, no write list */ 311 rp_ary = (struct rpcrdma_write_array *) 312 &rmsgp->rm_body.rm_chunks[2]; 313 314 found_it: 315 if (rp_ary->wc_discrim == 0) 316 return NULL; 317 318 return rp_ary; 319 } 320 #endif 321