1/* 2 * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved. 3 * Copyright (c) 2005-2006 Network Appliance, Inc. All rights reserved. 4 * 5 * This software is available to you under a choice of one of two 6 * licenses. You may choose to be licensed under the terms of the GNU 7 * General Public License (GPL) Version 2, available from the file 8 * COPYING in the main directory of this source tree, or the BSD-type 9 * license below: 10 * 11 * Redistribution and use in source and binary forms, with or without 12 * modification, are permitted provided that the following conditions 13 * are met: 14 * 15 * Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 18 * Redistributions in binary form must reproduce the above 19 * copyright notice, this list of conditions and the following 20 * disclaimer in the documentation and/or other materials provided 21 * with the distribution. 22 * 23 * Neither the name of the Network Appliance, Inc. nor the names of 24 * its contributors may be used to endorse or promote products 25 * derived from this software without specific prior written 26 * permission. 27 * 28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 29 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 31 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 32 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 33 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 34 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 35 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 36 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 37 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 38 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 39 * 40 * Author: Tom Tucker <tom@opengridcomputing.com> 41 */ 42 43#include <linux/sunrpc/debug.h> 44#include <linux/sunrpc/rpc_rdma.h> 45#include <linux/spinlock.h> 46#include <asm/unaligned.h> 47#include <rdma/ib_verbs.h> 48#include <rdma/rdma_cm.h> 49#include <linux/sunrpc/svc_rdma.h> 50 51#define RPCDBG_FACILITY RPCDBG_SVCXPRT 52 53static int map_xdr(struct svcxprt_rdma *xprt, 54 struct xdr_buf *xdr, 55 struct svc_rdma_req_map *vec) 56{ 57 int sge_no; 58 u32 sge_bytes; 59 u32 page_bytes; 60 u32 page_off; 61 int page_no; 62 63 BUG_ON(xdr->len != 64 (xdr->head[0].iov_len + xdr->page_len + xdr->tail[0].iov_len)); 65 66 /* Skip the first sge, this is for the RPCRDMA header */ 67 sge_no = 1; 68 69 /* Head SGE */ 70 vec->sge[sge_no].iov_base = xdr->head[0].iov_base; 71 vec->sge[sge_no].iov_len = xdr->head[0].iov_len; 72 sge_no++; 73 74 /* pages SGE */ 75 page_no = 0; 76 page_bytes = xdr->page_len; 77 page_off = xdr->page_base; 78 while (page_bytes) { 79 vec->sge[sge_no].iov_base = 80 page_address(xdr->pages[page_no]) + page_off; 81 sge_bytes = min_t(u32, page_bytes, (PAGE_SIZE - page_off)); 82 page_bytes -= sge_bytes; 83 vec->sge[sge_no].iov_len = sge_bytes; 84 85 sge_no++; 86 page_no++; 87 page_off = 0; /* reset for next time through loop */ 88 } 89 90 /* Tail SGE */ 91 if (xdr->tail[0].iov_len) { 92 vec->sge[sge_no].iov_base = xdr->tail[0].iov_base; 93 vec->sge[sge_no].iov_len = xdr->tail[0].iov_len; 94 sge_no++; 95 } 96 97 dprintk("svcrdma: map_xdr: sge_no %d page_no %d " 98 "page_base %u page_len %u head_len %zu tail_len %zu\n", 99 sge_no, page_no, xdr->page_base, xdr->page_len, 100 xdr->head[0].iov_len, xdr->tail[0].iov_len); 101 102 vec->count = sge_no; 103 return 0; 104} 105 106static dma_addr_t dma_map_xdr(struct svcxprt_rdma *xprt, 107 struct xdr_buf *xdr, 108 u32 xdr_off, size_t len, int dir) 109{ 110 struct page *page; 111 dma_addr_t dma_addr; 112 if (xdr_off < xdr->head[0].iov_len) { 113 /* This offset is in the head */ 114 xdr_off += (unsigned long)xdr->head[0].iov_base & ~PAGE_MASK; 115 page = virt_to_page(xdr->head[0].iov_base); 116 } else { 117 xdr_off -= xdr->head[0].iov_len; 118 if (xdr_off < xdr->page_len) { 119 /* This offset is in the page list */ 120 xdr_off += xdr->page_base; 121 page = xdr->pages[xdr_off >> PAGE_SHIFT]; 122 xdr_off &= ~PAGE_MASK; 123 } else { 124 /* This offset is in the tail */ 125 xdr_off -= xdr->page_len; 126 xdr_off += (unsigned long) 127 xdr->tail[0].iov_base & ~PAGE_MASK; 128 page = virt_to_page(xdr->tail[0].iov_base); 129 } 130 } 131 dma_addr = ib_dma_map_page(xprt->sc_cm_id->device, page, xdr_off, 132 min_t(size_t, PAGE_SIZE, len), dir); 133 return dma_addr; 134} 135 136/* Assumptions: 137 * - The specified write_len can be represented in sc_max_sge * PAGE_SIZE 138 */ 139static int send_write(struct svcxprt_rdma *xprt, struct svc_rqst *rqstp, 140 u32 rmr, u64 to, 141 u32 xdr_off, int write_len, 142 struct svc_rdma_req_map *vec) 143{ 144 struct ib_send_wr write_wr; 145 struct ib_sge *sge; 146 int xdr_sge_no; 147 int sge_no; 148 int sge_bytes; 149 int sge_off; 150 int bc; 151 struct svc_rdma_op_ctxt *ctxt; 152 153 BUG_ON(vec->count > RPCSVC_MAXPAGES); 154 dprintk("svcrdma: RDMA_WRITE rmr=%x, to=%llx, xdr_off=%d, " 155 "write_len=%d, vec->sge=%p, vec->count=%lu\n", 156 rmr, (unsigned long long)to, xdr_off, 157 write_len, vec->sge, vec->count); 158 159 ctxt = svc_rdma_get_context(xprt); 160 ctxt->direction = DMA_TO_DEVICE; 161 sge = ctxt->sge; 162 163 /* Find the SGE associated with xdr_off */ 164 for (bc = xdr_off, xdr_sge_no = 1; bc && xdr_sge_no < vec->count; 165 xdr_sge_no++) { 166 if (vec->sge[xdr_sge_no].iov_len > bc) 167 break; 168 bc -= vec->sge[xdr_sge_no].iov_len; 169 } 170 171 sge_off = bc; 172 bc = write_len; 173 sge_no = 0; 174 175 /* Copy the remaining SGE */ 176 while (bc != 0) { 177 sge_bytes = min_t(size_t, 178 bc, vec->sge[xdr_sge_no].iov_len-sge_off); 179 sge[sge_no].length = sge_bytes; 180 sge[sge_no].addr = 181 dma_map_xdr(xprt, &rqstp->rq_res, xdr_off, 182 sge_bytes, DMA_TO_DEVICE); 183 xdr_off += sge_bytes; 184 if (ib_dma_mapping_error(xprt->sc_cm_id->device, 185 sge[sge_no].addr)) 186 goto err; 187 atomic_inc(&xprt->sc_dma_used); 188 sge[sge_no].lkey = xprt->sc_dma_lkey; 189 ctxt->count++; 190 sge_off = 0; 191 sge_no++; 192 xdr_sge_no++; 193 BUG_ON(xdr_sge_no > vec->count); 194 bc -= sge_bytes; 195 if (sge_no == xprt->sc_max_sge) 196 break; 197 } 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 goto err; 214 return write_len - bc; 215 err: 216 svc_rdma_unmap_dma(ctxt); 217 svc_rdma_put_context(ctxt, 0); 218 /* Fatal error, close transport */ 219 return -EIO; 220} 221 222static int send_write_chunks(struct svcxprt_rdma *xprt, 223 struct rpcrdma_msg *rdma_argp, 224 struct rpcrdma_msg *rdma_resp, 225 struct svc_rqst *rqstp, 226 struct svc_rdma_req_map *vec) 227{ 228 u32 xfer_len = rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len; 229 int write_len; 230 u32 xdr_off; 231 int chunk_off; 232 int chunk_no; 233 struct rpcrdma_write_array *arg_ary; 234 struct rpcrdma_write_array *res_ary; 235 int ret; 236 237 arg_ary = svc_rdma_get_write_array(rdma_argp); 238 if (!arg_ary) 239 return 0; 240 res_ary = (struct rpcrdma_write_array *) 241 &rdma_resp->rm_body.rm_chunks[1]; 242 243 /* Write chunks start at the pagelist */ 244 for (xdr_off = rqstp->rq_res.head[0].iov_len, chunk_no = 0; 245 xfer_len && chunk_no < arg_ary->wc_nchunks; 246 chunk_no++) { 247 struct rpcrdma_segment *arg_ch; 248 u64 rs_offset; 249 250 arg_ch = &arg_ary->wc_array[chunk_no].wc_target; 251 write_len = min(xfer_len, ntohl(arg_ch->rs_length)); 252 253 /* Prepare the response chunk given the length actually 254 * written */ 255 xdr_decode_hyper((__be32 *)&arg_ch->rs_offset, &rs_offset); 256 svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no, 257 arg_ch->rs_handle, 258 arg_ch->rs_offset, 259 write_len); 260 chunk_off = 0; 261 while (write_len) { 262 ret = send_write(xprt, rqstp, 263 ntohl(arg_ch->rs_handle), 264 rs_offset + chunk_off, 265 xdr_off, 266 write_len, 267 vec); 268 if (ret <= 0) { 269 dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n", 270 ret); 271 return -EIO; 272 } 273 chunk_off += ret; 274 xdr_off += ret; 275 xfer_len -= ret; 276 write_len -= ret; 277 } 278 } 279 /* Update the req with the number of chunks actually used */ 280 svc_rdma_xdr_encode_write_list(rdma_resp, chunk_no); 281 282 return rqstp->rq_res.page_len + rqstp->rq_res.tail[0].iov_len; 283} 284 285static int send_reply_chunks(struct svcxprt_rdma *xprt, 286 struct rpcrdma_msg *rdma_argp, 287 struct rpcrdma_msg *rdma_resp, 288 struct svc_rqst *rqstp, 289 struct svc_rdma_req_map *vec) 290{ 291 u32 xfer_len = rqstp->rq_res.len; 292 int write_len; 293 u32 xdr_off; 294 int chunk_no; 295 int chunk_off; 296 int nchunks; 297 struct rpcrdma_segment *ch; 298 struct rpcrdma_write_array *arg_ary; 299 struct rpcrdma_write_array *res_ary; 300 int ret; 301 302 arg_ary = svc_rdma_get_reply_array(rdma_argp); 303 if (!arg_ary) 304 return 0; 305 /* XXX: need to fix when reply lists occur with read-list and or 306 * write-list */ 307 res_ary = (struct rpcrdma_write_array *) 308 &rdma_resp->rm_body.rm_chunks[2]; 309 310 /* xdr offset starts at RPC message */ 311 nchunks = ntohl(arg_ary->wc_nchunks); 312 for (xdr_off = 0, chunk_no = 0; 313 xfer_len && chunk_no < nchunks; 314 chunk_no++) { 315 u64 rs_offset; 316 ch = &arg_ary->wc_array[chunk_no].wc_target; 317 write_len = min(xfer_len, htonl(ch->rs_length)); 318 319 /* Prepare the reply chunk given the length actually 320 * written */ 321 xdr_decode_hyper((__be32 *)&ch->rs_offset, &rs_offset); 322 svc_rdma_xdr_encode_array_chunk(res_ary, chunk_no, 323 ch->rs_handle, ch->rs_offset, 324 write_len); 325 chunk_off = 0; 326 while (write_len) { 327 ret = send_write(xprt, rqstp, 328 ntohl(ch->rs_handle), 329 rs_offset + chunk_off, 330 xdr_off, 331 write_len, 332 vec); 333 if (ret <= 0) { 334 dprintk("svcrdma: RDMA_WRITE failed, ret=%d\n", 335 ret); 336 return -EIO; 337 } 338 chunk_off += ret; 339 xdr_off += ret; 340 xfer_len -= ret; 341 write_len -= ret; 342 } 343 } 344 /* Update the req with the number of chunks actually used */ 345 svc_rdma_xdr_encode_reply_array(res_ary, chunk_no); 346 347 return rqstp->rq_res.len; 348} 349 350/* This function prepares the portion of the RPCRDMA message to be 351 * sent in the RDMA_SEND. This function is called after data sent via 352 * RDMA has already been transmitted. There are three cases: 353 * - The RPCRDMA header, RPC header, and payload are all sent in a 354 * single RDMA_SEND. This is the "inline" case. 355 * - The RPCRDMA header and some portion of the RPC header and data 356 * are sent via this RDMA_SEND and another portion of the data is 357 * sent via RDMA. 358 * - The RPCRDMA header [NOMSG] is sent in this RDMA_SEND and the RPC 359 * header and data are all transmitted via RDMA. 360 * In all three cases, this function prepares the RPCRDMA header in 361 * sge[0], the 'type' parameter indicates the type to place in the 362 * RPCRDMA header, and the 'byte_count' field indicates how much of 363 * the XDR to include in this RDMA_SEND. NB: The offset of the payload 364 * to send is zero in the XDR. 365 */ 366static int send_reply(struct svcxprt_rdma *rdma, 367 struct svc_rqst *rqstp, 368 struct page *page, 369 struct rpcrdma_msg *rdma_resp, 370 struct svc_rdma_op_ctxt *ctxt, 371 struct svc_rdma_req_map *vec, 372 int byte_count) 373{ 374 struct ib_send_wr send_wr; 375 int sge_no; 376 int sge_bytes; 377 int page_no; 378 int pages; 379 int ret; 380 381 /* Post a recv buffer to handle another request. */ 382 ret = svc_rdma_post_recv(rdma); 383 if (ret) { 384 printk(KERN_INFO 385 "svcrdma: could not post a receive buffer, err=%d." 386 "Closing transport %p.\n", ret, rdma); 387 set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags); 388 svc_rdma_put_context(ctxt, 0); 389 return -ENOTCONN; 390 } 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].lkey = rdma->sc_dma_lkey; 398 ctxt->sge[0].length = svc_rdma_xdr_get_reply_hdr_len(rdma_resp); 399 ctxt->sge[0].addr = 400 ib_dma_map_page(rdma->sc_cm_id->device, page, 0, 401 ctxt->sge[0].length, DMA_TO_DEVICE); 402 if (ib_dma_mapping_error(rdma->sc_cm_id->device, ctxt->sge[0].addr)) 403 goto err; 404 atomic_inc(&rdma->sc_dma_used); 405 406 ctxt->direction = DMA_TO_DEVICE; 407 408 /* Map the payload indicated by 'byte_count' */ 409 for (sge_no = 1; byte_count && sge_no < vec->count; sge_no++) { 410 int xdr_off = 0; 411 sge_bytes = min_t(size_t, vec->sge[sge_no].iov_len, byte_count); 412 byte_count -= sge_bytes; 413 ctxt->sge[sge_no].addr = 414 dma_map_xdr(rdma, &rqstp->rq_res, xdr_off, 415 sge_bytes, DMA_TO_DEVICE); 416 xdr_off += sge_bytes; 417 if (ib_dma_mapping_error(rdma->sc_cm_id->device, 418 ctxt->sge[sge_no].addr)) 419 goto err; 420 atomic_inc(&rdma->sc_dma_used); 421 ctxt->sge[sge_no].lkey = rdma->sc_dma_lkey; 422 ctxt->sge[sge_no].length = sge_bytes; 423 } 424 BUG_ON(byte_count != 0); 425 426 /* Save all respages in the ctxt and remove them from the 427 * respages array. They are our pages until the I/O 428 * completes. 429 */ 430 pages = rqstp->rq_next_page - rqstp->rq_respages; 431 for (page_no = 0; page_no < pages; page_no++) { 432 ctxt->pages[page_no+1] = rqstp->rq_respages[page_no]; 433 ctxt->count++; 434 rqstp->rq_respages[page_no] = NULL; 435 /* 436 * If there are more pages than SGE, terminate SGE 437 * list so that svc_rdma_unmap_dma doesn't attempt to 438 * unmap garbage. 439 */ 440 if (page_no+1 >= sge_no) 441 ctxt->sge[page_no+1].length = 0; 442 } 443 rqstp->rq_next_page = rqstp->rq_respages + 1; 444 445 BUG_ON(sge_no > rdma->sc_max_sge); 446 memset(&send_wr, 0, sizeof send_wr); 447 ctxt->wr_op = IB_WR_SEND; 448 send_wr.wr_id = (unsigned long)ctxt; 449 send_wr.sg_list = ctxt->sge; 450 send_wr.num_sge = sge_no; 451 send_wr.opcode = IB_WR_SEND; 452 send_wr.send_flags = IB_SEND_SIGNALED; 453 454 ret = svc_rdma_send(rdma, &send_wr); 455 if (ret) 456 goto err; 457 458 return 0; 459 460 err: 461 svc_rdma_unmap_dma(ctxt); 462 svc_rdma_put_context(ctxt, 1); 463 return -EIO; 464} 465 466void svc_rdma_prep_reply_hdr(struct svc_rqst *rqstp) 467{ 468} 469 470/* 471 * Return the start of an xdr buffer. 472 */ 473static void *xdr_start(struct xdr_buf *xdr) 474{ 475 return xdr->head[0].iov_base - 476 (xdr->len - 477 xdr->page_len - 478 xdr->tail[0].iov_len - 479 xdr->head[0].iov_len); 480} 481 482int svc_rdma_sendto(struct svc_rqst *rqstp) 483{ 484 struct svc_xprt *xprt = rqstp->rq_xprt; 485 struct svcxprt_rdma *rdma = 486 container_of(xprt, struct svcxprt_rdma, sc_xprt); 487 struct rpcrdma_msg *rdma_argp; 488 struct rpcrdma_msg *rdma_resp; 489 struct rpcrdma_write_array *reply_ary; 490 enum rpcrdma_proc reply_type; 491 int ret; 492 int inline_bytes; 493 struct page *res_page; 494 struct svc_rdma_op_ctxt *ctxt; 495 struct svc_rdma_req_map *vec; 496 497 dprintk("svcrdma: sending response for rqstp=%p\n", rqstp); 498 499 /* Get the RDMA request header. */ 500 rdma_argp = xdr_start(&rqstp->rq_arg); 501 502 /* Build an req vec for the XDR */ 503 ctxt = svc_rdma_get_context(rdma); 504 ctxt->direction = DMA_TO_DEVICE; 505 vec = svc_rdma_get_req_map(); 506 ret = map_xdr(rdma, &rqstp->rq_res, vec); 507 if (ret) 508 goto err0; 509 inline_bytes = rqstp->rq_res.len; 510 511 /* Create the RDMA response header */ 512 res_page = svc_rdma_get_page(); 513 rdma_resp = page_address(res_page); 514 reply_ary = svc_rdma_get_reply_array(rdma_argp); 515 if (reply_ary) 516 reply_type = RDMA_NOMSG; 517 else 518 reply_type = RDMA_MSG; 519 svc_rdma_xdr_encode_reply_header(rdma, rdma_argp, 520 rdma_resp, reply_type); 521 522 /* Send any write-chunk data and build resp write-list */ 523 ret = send_write_chunks(rdma, rdma_argp, rdma_resp, 524 rqstp, vec); 525 if (ret < 0) { 526 printk(KERN_ERR "svcrdma: failed to send write chunks, rc=%d\n", 527 ret); 528 goto err1; 529 } 530 inline_bytes -= ret; 531 532 /* Send any reply-list data and update resp reply-list */ 533 ret = send_reply_chunks(rdma, rdma_argp, rdma_resp, 534 rqstp, vec); 535 if (ret < 0) { 536 printk(KERN_ERR "svcrdma: failed to send reply chunks, rc=%d\n", 537 ret); 538 goto err1; 539 } 540 inline_bytes -= ret; 541 542 ret = send_reply(rdma, rqstp, res_page, rdma_resp, ctxt, vec, 543 inline_bytes); 544 svc_rdma_put_req_map(vec); 545 dprintk("svcrdma: send_reply returns %d\n", ret); 546 return ret; 547 548 err1: 549 put_page(res_page); 550 err0: 551 svc_rdma_put_req_map(vec); 552 svc_rdma_put_context(ctxt, 0); 553 return ret; 554} 555