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 #include <linux/sunrpc/debug.h> 43 #include <linux/sunrpc/rpc_rdma.h> 44 #include <linux/spinlock.h> 45 #include <asm/unaligned.h> 46 #include <rdma/ib_verbs.h> 47 #include <rdma/rdma_cm.h> 48 #include <linux/sunrpc/svc_rdma.h> 49 50 #define RPCDBG_FACILITY RPCDBG_SVCXPRT 51 52 /* Encode an XDR as an array of IB SGE 53 * 54 * Assumptions: 55 * - head[0] is physically contiguous. 56 * - tail[0] is physically contiguous. 57 * - pages[] is not physically or virtually contigous and consists of 58 * PAGE_SIZE elements. 59 * 60 * Output: 61 * SGE[0] reserved for RCPRDMA header 62 * SGE[1] data from xdr->head[] 63 * SGE[2..sge_count-2] data from xdr->pages[] 64 * SGE[sge_count-1] data from xdr->tail. 65 * 66 */ 67 static struct ib_sge *xdr_to_sge(struct svcxprt_rdma *xprt, 68 struct xdr_buf *xdr, 69 struct ib_sge *sge, 70 int *sge_count) 71 { 72 /* Max we need is the length of the XDR / pagesize + one for 73 * head + one for tail + one for RPCRDMA header 74 */ 75 int sge_max = (xdr->len+PAGE_SIZE-1) / PAGE_SIZE + 3; 76 int sge_no; 77 u32 byte_count = xdr->len; 78 u32 sge_bytes; 79 u32 page_bytes; 80 int page_off; 81 int page_no; 82 83 /* Skip the first sge, this is for the RPCRDMA header */ 84 sge_no = 1; 85 86 /* Head SGE */ 87 sge[sge_no].addr = ib_dma_map_single(xprt->sc_cm_id->device, 88 xdr->head[0].iov_base, 89 xdr->head[0].iov_len, 90 DMA_TO_DEVICE); 91 sge_bytes = min_t(u32, byte_count, xdr->head[0].iov_len); 92 byte_count -= sge_bytes; 93 sge[sge_no].length = sge_bytes; 94 sge[sge_no].lkey = xprt->sc_phys_mr->lkey; 95 sge_no++; 96 97 /* pages SGE */ 98 page_no = 0; 99 page_bytes = xdr->page_len; 100 page_off = xdr->page_base; 101 while (byte_count && page_bytes) { 102 sge_bytes = min_t(u32, byte_count, (PAGE_SIZE-page_off)); 103 sge[sge_no].addr = 104 ib_dma_map_page(xprt->sc_cm_id->device, 105 xdr->pages[page_no], page_off, 106 sge_bytes, DMA_TO_DEVICE); 107 sge_bytes = min(sge_bytes, page_bytes); 108 byte_count -= sge_bytes; 109 page_bytes -= sge_bytes; 110 sge[sge_no].length = sge_bytes; 111 sge[sge_no].lkey = xprt->sc_phys_mr->lkey; 112 113 sge_no++; 114 page_no++; 115 page_off = 0; /* reset for next time through loop */ 116 } 117 118 /* Tail SGE */ 119 if (byte_count && xdr->tail[0].iov_len) { 120 sge[sge_no].addr = 121 ib_dma_map_single(xprt->sc_cm_id->device, 122 xdr->tail[0].iov_base, 123 xdr->tail[0].iov_len, 124 DMA_TO_DEVICE); 125 sge_bytes = min_t(u32, byte_count, xdr->tail[0].iov_len); 126 byte_count -= sge_bytes; 127 sge[sge_no].length = sge_bytes; 128 sge[sge_no].lkey = xprt->sc_phys_mr->lkey; 129 sge_no++; 130 } 131 132 BUG_ON(sge_no > sge_max); 133 BUG_ON(byte_count != 0); 134 135 *sge_count = sge_no; 136 return sge; 137 } 138 139 140 /* Assumptions: 141 * - The specified write_len can be represented in sc_max_sge * PAGE_SIZE 142 */ 143 static int send_write(struct svcxprt_rdma *xprt, struct svc_rqst *rqstp, 144 u32 rmr, u64 to, 145 u32 xdr_off, int write_len, 146 struct ib_sge *xdr_sge, int sge_count) 147 { 148 struct svc_rdma_op_ctxt *tmp_sge_ctxt; 149 struct ib_send_wr write_wr; 150 struct ib_sge *sge; 151 int xdr_sge_no; 152 int sge_no; 153 int sge_bytes; 154 int sge_off; 155 int bc; 156 struct svc_rdma_op_ctxt *ctxt; 157 int ret = 0; 158 159 BUG_ON(sge_count >= 32); 160 dprintk("svcrdma: RDMA_WRITE rmr=%x, to=%llx, xdr_off=%d, " 161 "write_len=%d, xdr_sge=%p, sge_count=%d\n", 162 rmr, (unsigned long long)to, xdr_off, 163 write_len, xdr_sge, sge_count); 164 165 ctxt = svc_rdma_get_context(xprt); 166 ctxt->count = 0; 167 tmp_sge_ctxt = svc_rdma_get_context(xprt); 168 sge = tmp_sge_ctxt->sge; 169 170 /* Find the SGE associated with xdr_off */ 171 for (bc = xdr_off, xdr_sge_no = 1; bc && xdr_sge_no < sge_count; 172 xdr_sge_no++) { 173 if (xdr_sge[xdr_sge_no].length > bc) 174 break; 175 bc -= xdr_sge[xdr_sge_no].length; 176 } 177 178 sge_off = bc; 179 bc = write_len; 180 sge_no = 0; 181 182 /* Copy the remaining SGE */ 183 while (bc != 0 && xdr_sge_no < sge_count) { 184 sge[sge_no].addr = xdr_sge[xdr_sge_no].addr + sge_off; 185 sge[sge_no].lkey = xdr_sge[xdr_sge_no].lkey; 186 sge_bytes = min((size_t)bc, 187 (size_t)(xdr_sge[xdr_sge_no].length-sge_off)); 188 sge[sge_no].length = sge_bytes; 189 190 sge_off = 0; 191 sge_no++; 192 xdr_sge_no++; 193 bc -= sge_bytes; 194 } 195 196 BUG_ON(bc != 0); 197 BUG_ON(xdr_sge_no > sge_count); 198 199 /* Prepare WRITE WR */ 200 memset(&write_wr, 0, sizeof write_wr); 201 ctxt->wr_op = IB_WR_RDMA_WRITE; 202 write_wr.wr_id = (unsigned long)ctxt; 203 write_wr.sg_list = &sge[0]; 204 write_wr.num_sge = sge_no; 205 write_wr.opcode = IB_WR_RDMA_WRITE; 206 write_wr.send_flags = IB_SEND_SIGNALED; 207 write_wr.wr.rdma.rkey = rmr; 208 write_wr.wr.rdma.remote_addr = to; 209 210 /* Post It */ 211 atomic_inc(&rdma_stat_write); 212 if (svc_rdma_send(xprt, &write_wr)) { 213 svc_rdma_put_context(ctxt, 1); 214 /* Fatal error, close transport */ 215 ret = -EIO; 216 } 217 svc_rdma_put_context(tmp_sge_ctxt, 0); 218 return ret; 219 } 220 221 static int send_write_chunks(struct svcxprt_rdma *xprt, 222 struct rpcrdma_msg *rdma_argp, 223 struct rpcrdma_msg *rdma_resp, 224 struct svc_rqst *rqstp, 225 struct ib_sge *sge, 226 int sge_count) 227 { 228 u32 xfer_len = rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len; 229 int write_len; 230 int max_write; 231 u32 xdr_off; 232 int chunk_off; 233 int chunk_no; 234 struct rpcrdma_write_array *arg_ary; 235 struct rpcrdma_write_array *res_ary; 236 int ret; 237 238 arg_ary = svc_rdma_get_write_array(rdma_argp); 239 if (!arg_ary) 240 return 0; 241 res_ary = (struct rpcrdma_write_array *) 242 &rdma_resp->rm_body.rm_chunks[1]; 243 244 max_write = xprt->sc_max_sge * PAGE_SIZE; 245 246 /* Write chunks start at the pagelist */ 247 for (xdr_off = rqstp->rq_res.head[0].iov_len, chunk_no = 0; 248 xfer_len && chunk_no < arg_ary->wc_nchunks; 249 chunk_no++) { 250 struct rpcrdma_segment *arg_ch; 251 u64 rs_offset; 252 253 arg_ch = &arg_ary->wc_array[chunk_no].wc_target; 254 write_len = min(xfer_len, arg_ch->rs_length); 255 256 /* Prepare the response chunk given the length actually 257 * written */ 258 rs_offset = get_unaligned(&(arg_ch->rs_offset)); 259 svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no, 260 arg_ch->rs_handle, 261 rs_offset, 262 write_len); 263 chunk_off = 0; 264 while (write_len) { 265 int this_write; 266 this_write = min(write_len, max_write); 267 ret = send_write(xprt, rqstp, 268 arg_ch->rs_handle, 269 rs_offset + chunk_off, 270 xdr_off, 271 this_write, 272 sge, 273 sge_count); 274 if (ret) { 275 dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n", 276 ret); 277 return -EIO; 278 } 279 chunk_off += this_write; 280 xdr_off += this_write; 281 xfer_len -= this_write; 282 write_len -= this_write; 283 } 284 } 285 /* Update the req with the number of chunks actually used */ 286 svc_rdma_xdr_encode_write_list(rdma_resp, chunk_no); 287 288 return rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len; 289 } 290 291 static int send_reply_chunks(struct svcxprt_rdma *xprt, 292 struct rpcrdma_msg *rdma_argp, 293 struct rpcrdma_msg *rdma_resp, 294 struct svc_rqst *rqstp, 295 struct ib_sge *sge, 296 int sge_count) 297 { 298 u32 xfer_len = rqstp->rq_res.len; 299 int write_len; 300 int max_write; 301 u32 xdr_off; 302 int chunk_no; 303 int chunk_off; 304 struct rpcrdma_segment *ch; 305 struct rpcrdma_write_array *arg_ary; 306 struct rpcrdma_write_array *res_ary; 307 int ret; 308 309 arg_ary = svc_rdma_get_reply_array(rdma_argp); 310 if (!arg_ary) 311 return 0; 312 /* XXX: need to fix when reply lists occur with read-list and or 313 * write-list */ 314 res_ary = (struct rpcrdma_write_array *) 315 &rdma_resp->rm_body.rm_chunks[2]; 316 317 max_write = xprt->sc_max_sge * PAGE_SIZE; 318 319 /* xdr offset starts at RPC message */ 320 for (xdr_off = 0, chunk_no = 0; 321 xfer_len && chunk_no < arg_ary->wc_nchunks; 322 chunk_no++) { 323 u64 rs_offset; 324 ch = &arg_ary->wc_array[chunk_no].wc_target; 325 write_len = min(xfer_len, ch->rs_length); 326 327 328 /* Prepare the reply chunk given the length actually 329 * written */ 330 rs_offset = get_unaligned(&(ch->rs_offset)); 331 svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no, 332 ch->rs_handle, rs_offset, 333 write_len); 334 chunk_off = 0; 335 while (write_len) { 336 int this_write; 337 338 this_write = min(write_len, max_write); 339 ret = send_write(xprt, rqstp, 340 ch->rs_handle, 341 rs_offset + chunk_off, 342 xdr_off, 343 this_write, 344 sge, 345 sge_count); 346 if (ret) { 347 dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n", 348 ret); 349 return -EIO; 350 } 351 chunk_off += this_write; 352 xdr_off += this_write; 353 xfer_len -= this_write; 354 write_len -= this_write; 355 } 356 } 357 /* Update the req with the number of chunks actually used */ 358 svc_rdma_xdr_encode_reply_array(res_ary, chunk_no); 359 360 return rqstp->rq_res.len; 361 } 362 363 /* This function prepares the portion of the RPCRDMA message to be 364 * sent in the RDMA_SEND. This function is called after data sent via 365 * RDMA has already been transmitted. There are three cases: 366 * - The RPCRDMA header, RPC header, and payload are all sent in a 367 * single RDMA_SEND. This is the "inline" case. 368 * - The RPCRDMA header and some portion of the RPC header and data 369 * are sent via this RDMA_SEND and another portion of the data is 370 * sent via RDMA. 371 * - The RPCRDMA header [NOMSG] is sent in this RDMA_SEND and the RPC 372 * header and data are all transmitted via RDMA. 373 * In all three cases, this function prepares the RPCRDMA header in 374 * sge[0], the 'type' parameter indicates the type to place in the 375 * RPCRDMA header, and the 'byte_count' field indicates how much of 376 * the XDR to include in this RDMA_SEND. 377 */ 378 static int send_reply(struct svcxprt_rdma *rdma, 379 struct svc_rqst *rqstp, 380 struct page *page, 381 struct rpcrdma_msg *rdma_resp, 382 struct svc_rdma_op_ctxt *ctxt, 383 int sge_count, 384 int byte_count) 385 { 386 struct ib_send_wr send_wr; 387 int sge_no; 388 int sge_bytes; 389 int page_no; 390 int ret; 391 392 /* Prepare the context */ 393 ctxt->pages[0] = page; 394 ctxt->count = 1; 395 396 /* Prepare the SGE for the RPCRDMA Header */ 397 ctxt->sge[0].addr = 398 ib_dma_map_page(rdma->sc_cm_id->device, 399 page, 0, PAGE_SIZE, DMA_TO_DEVICE); 400 ctxt->direction = DMA_TO_DEVICE; 401 ctxt->sge[0].length = svc_rdma_xdr_get_reply_hdr_len(rdma_resp); 402 ctxt->sge[0].lkey = rdma->sc_phys_mr->lkey; 403 404 /* Determine how many of our SGE are to be transmitted */ 405 for (sge_no = 1; byte_count && sge_no < sge_count; sge_no++) { 406 sge_bytes = min((size_t)ctxt->sge[sge_no].length, 407 (size_t)byte_count); 408 byte_count -= sge_bytes; 409 } 410 BUG_ON(byte_count != 0); 411 412 /* Save all respages in the ctxt and remove them from the 413 * respages array. They are our pages until the I/O 414 * completes. 415 */ 416 for (page_no = 0; page_no < rqstp->rq_resused; page_no++) { 417 ctxt->pages[page_no+1] = rqstp->rq_respages[page_no]; 418 ctxt->count++; 419 rqstp->rq_respages[page_no] = NULL; 420 } 421 422 BUG_ON(sge_no > rdma->sc_max_sge); 423 memset(&send_wr, 0, sizeof send_wr); 424 ctxt->wr_op = IB_WR_SEND; 425 send_wr.wr_id = (unsigned long)ctxt; 426 send_wr.sg_list = ctxt->sge; 427 send_wr.num_sge = sge_no; 428 send_wr.opcode = IB_WR_SEND; 429 send_wr.send_flags = IB_SEND_SIGNALED; 430 431 ret = svc_rdma_send(rdma, &send_wr); 432 if (ret) 433 svc_rdma_put_context(ctxt, 1); 434 435 return ret; 436 } 437 438 void svc_rdma_prep_reply_hdr(struct svc_rqst *rqstp) 439 { 440 } 441 442 /* 443 * Return the start of an xdr buffer. 444 */ 445 static void *xdr_start(struct xdr_buf *xdr) 446 { 447 return xdr->head[0].iov_base - 448 (xdr->len - 449 xdr->page_len - 450 xdr->tail[0].iov_len - 451 xdr->head[0].iov_len); 452 } 453 454 int svc_rdma_sendto(struct svc_rqst *rqstp) 455 { 456 struct svc_xprt *xprt = rqstp->rq_xprt; 457 struct svcxprt_rdma *rdma = 458 container_of(xprt, struct svcxprt_rdma, sc_xprt); 459 struct rpcrdma_msg *rdma_argp; 460 struct rpcrdma_msg *rdma_resp; 461 struct rpcrdma_write_array *reply_ary; 462 enum rpcrdma_proc reply_type; 463 int ret; 464 int inline_bytes; 465 struct ib_sge *sge; 466 int sge_count = 0; 467 struct page *res_page; 468 struct svc_rdma_op_ctxt *ctxt; 469 470 dprintk("svcrdma: sending response for rqstp=%p\n", rqstp); 471 472 /* Get the RDMA request header. */ 473 rdma_argp = xdr_start(&rqstp->rq_arg); 474 475 /* Build an SGE for the XDR */ 476 ctxt = svc_rdma_get_context(rdma); 477 ctxt->direction = DMA_TO_DEVICE; 478 sge = xdr_to_sge(rdma, &rqstp->rq_res, ctxt->sge, &sge_count); 479 480 inline_bytes = rqstp->rq_res.len; 481 482 /* Create the RDMA response header */ 483 res_page = svc_rdma_get_page(); 484 rdma_resp = page_address(res_page); 485 reply_ary = svc_rdma_get_reply_array(rdma_argp); 486 if (reply_ary) 487 reply_type = RDMA_NOMSG; 488 else 489 reply_type = RDMA_MSG; 490 svc_rdma_xdr_encode_reply_header(rdma, rdma_argp, 491 rdma_resp, reply_type); 492 493 /* Send any write-chunk data and build resp write-list */ 494 ret = send_write_chunks(rdma, rdma_argp, rdma_resp, 495 rqstp, sge, sge_count); 496 if (ret < 0) { 497 printk(KERN_ERR "svcrdma: failed to send write chunks, rc=%d\n", 498 ret); 499 goto error; 500 } 501 inline_bytes -= ret; 502 503 /* Send any reply-list data and update resp reply-list */ 504 ret = send_reply_chunks(rdma, rdma_argp, rdma_resp, 505 rqstp, sge, sge_count); 506 if (ret < 0) { 507 printk(KERN_ERR "svcrdma: failed to send reply chunks, rc=%d\n", 508 ret); 509 goto error; 510 } 511 inline_bytes -= ret; 512 513 ret = send_reply(rdma, rqstp, res_page, rdma_resp, ctxt, sge_count, 514 inline_bytes); 515 dprintk("svcrdma: send_reply returns %d\n", ret); 516 return ret; 517 error: 518 svc_rdma_put_context(ctxt, 0); 519 put_page(res_page); 520 return ret; 521 } 522