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 53 /* 54 * Replace the pages in the rq_argpages array with the pages from the SGE in 55 * the RDMA_RECV completion. The SGL should contain full pages up until the 56 * last one. 57 */ 58 static void rdma_build_arg_xdr(struct svc_rqst *rqstp, 59 struct svc_rdma_op_ctxt *ctxt, 60 u32 byte_count) 61 { 62 struct rpcrdma_msg *rmsgp; 63 struct page *page; 64 u32 bc; 65 int sge_no; 66 67 /* Swap the page in the SGE with the page in argpages */ 68 page = ctxt->pages[0]; 69 put_page(rqstp->rq_pages[0]); 70 rqstp->rq_pages[0] = page; 71 72 /* Set up the XDR head */ 73 rqstp->rq_arg.head[0].iov_base = page_address(page); 74 rqstp->rq_arg.head[0].iov_len = 75 min_t(size_t, byte_count, ctxt->sge[0].length); 76 rqstp->rq_arg.len = byte_count; 77 rqstp->rq_arg.buflen = byte_count; 78 79 /* Compute bytes past head in the SGL */ 80 bc = byte_count - rqstp->rq_arg.head[0].iov_len; 81 82 /* If data remains, store it in the pagelist */ 83 rqstp->rq_arg.page_len = bc; 84 rqstp->rq_arg.page_base = 0; 85 86 /* RDMA_NOMSG: RDMA READ data should land just after RDMA RECV data */ 87 rmsgp = (struct rpcrdma_msg *)rqstp->rq_arg.head[0].iov_base; 88 if (rmsgp->rm_type == rdma_nomsg) 89 rqstp->rq_arg.pages = &rqstp->rq_pages[0]; 90 else 91 rqstp->rq_arg.pages = &rqstp->rq_pages[1]; 92 93 sge_no = 1; 94 while (bc && sge_no < ctxt->count) { 95 page = ctxt->pages[sge_no]; 96 put_page(rqstp->rq_pages[sge_no]); 97 rqstp->rq_pages[sge_no] = page; 98 bc -= min_t(u32, bc, ctxt->sge[sge_no].length); 99 rqstp->rq_arg.buflen += ctxt->sge[sge_no].length; 100 sge_no++; 101 } 102 rqstp->rq_respages = &rqstp->rq_pages[sge_no]; 103 rqstp->rq_next_page = rqstp->rq_respages + 1; 104 105 /* If not all pages were used from the SGL, free the remaining ones */ 106 bc = sge_no; 107 while (sge_no < ctxt->count) { 108 page = ctxt->pages[sge_no++]; 109 put_page(page); 110 } 111 ctxt->count = bc; 112 113 /* Set up tail */ 114 rqstp->rq_arg.tail[0].iov_base = NULL; 115 rqstp->rq_arg.tail[0].iov_len = 0; 116 } 117 118 /* Issue an RDMA_READ using the local lkey to map the data sink */ 119 int rdma_read_chunk_lcl(struct svcxprt_rdma *xprt, 120 struct svc_rqst *rqstp, 121 struct svc_rdma_op_ctxt *head, 122 int *page_no, 123 u32 *page_offset, 124 u32 rs_handle, 125 u32 rs_length, 126 u64 rs_offset, 127 bool last) 128 { 129 struct ib_rdma_wr read_wr; 130 int pages_needed = PAGE_ALIGN(*page_offset + rs_length) >> PAGE_SHIFT; 131 struct svc_rdma_op_ctxt *ctxt = svc_rdma_get_context(xprt); 132 int ret, read, pno; 133 u32 pg_off = *page_offset; 134 u32 pg_no = *page_no; 135 136 ctxt->direction = DMA_FROM_DEVICE; 137 ctxt->read_hdr = head; 138 pages_needed = min_t(int, pages_needed, xprt->sc_max_sge_rd); 139 read = min_t(int, (pages_needed << PAGE_SHIFT) - *page_offset, 140 rs_length); 141 142 for (pno = 0; pno < pages_needed; pno++) { 143 int len = min_t(int, rs_length, PAGE_SIZE - pg_off); 144 145 head->arg.pages[pg_no] = rqstp->rq_arg.pages[pg_no]; 146 head->arg.page_len += len; 147 148 head->arg.len += len; 149 if (!pg_off) 150 head->count++; 151 rqstp->rq_respages = &rqstp->rq_arg.pages[pg_no+1]; 152 rqstp->rq_next_page = rqstp->rq_respages + 1; 153 ctxt->sge[pno].addr = 154 ib_dma_map_page(xprt->sc_cm_id->device, 155 head->arg.pages[pg_no], pg_off, 156 PAGE_SIZE - pg_off, 157 DMA_FROM_DEVICE); 158 ret = ib_dma_mapping_error(xprt->sc_cm_id->device, 159 ctxt->sge[pno].addr); 160 if (ret) 161 goto err; 162 svc_rdma_count_mappings(xprt, ctxt); 163 164 ctxt->sge[pno].lkey = xprt->sc_pd->local_dma_lkey; 165 ctxt->sge[pno].length = len; 166 ctxt->count++; 167 168 /* adjust offset and wrap to next page if needed */ 169 pg_off += len; 170 if (pg_off == PAGE_SIZE) { 171 pg_off = 0; 172 pg_no++; 173 } 174 rs_length -= len; 175 } 176 177 if (last && rs_length == 0) 178 set_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags); 179 else 180 clear_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags); 181 182 memset(&read_wr, 0, sizeof(read_wr)); 183 ctxt->cqe.done = svc_rdma_wc_read; 184 read_wr.wr.wr_cqe = &ctxt->cqe; 185 read_wr.wr.opcode = IB_WR_RDMA_READ; 186 read_wr.wr.send_flags = IB_SEND_SIGNALED; 187 read_wr.rkey = rs_handle; 188 read_wr.remote_addr = rs_offset; 189 read_wr.wr.sg_list = ctxt->sge; 190 read_wr.wr.num_sge = pages_needed; 191 192 ret = svc_rdma_send(xprt, &read_wr.wr); 193 if (ret) { 194 pr_err("svcrdma: Error %d posting RDMA_READ\n", ret); 195 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); 196 goto err; 197 } 198 199 /* return current location in page array */ 200 *page_no = pg_no; 201 *page_offset = pg_off; 202 ret = read; 203 atomic_inc(&rdma_stat_read); 204 return ret; 205 err: 206 svc_rdma_unmap_dma(ctxt); 207 svc_rdma_put_context(ctxt, 0); 208 return ret; 209 } 210 211 /* Issue an RDMA_READ using an FRMR to map the data sink */ 212 int rdma_read_chunk_frmr(struct svcxprt_rdma *xprt, 213 struct svc_rqst *rqstp, 214 struct svc_rdma_op_ctxt *head, 215 int *page_no, 216 u32 *page_offset, 217 u32 rs_handle, 218 u32 rs_length, 219 u64 rs_offset, 220 bool last) 221 { 222 struct ib_rdma_wr read_wr; 223 struct ib_send_wr inv_wr; 224 struct ib_reg_wr reg_wr; 225 u8 key; 226 int nents = PAGE_ALIGN(*page_offset + rs_length) >> PAGE_SHIFT; 227 struct svc_rdma_op_ctxt *ctxt = svc_rdma_get_context(xprt); 228 struct svc_rdma_fastreg_mr *frmr = svc_rdma_get_frmr(xprt); 229 int ret, read, pno, dma_nents, n; 230 u32 pg_off = *page_offset; 231 u32 pg_no = *page_no; 232 233 if (IS_ERR(frmr)) 234 return -ENOMEM; 235 236 ctxt->direction = DMA_FROM_DEVICE; 237 ctxt->frmr = frmr; 238 nents = min_t(unsigned int, nents, xprt->sc_frmr_pg_list_len); 239 read = min_t(int, (nents << PAGE_SHIFT) - *page_offset, rs_length); 240 241 frmr->direction = DMA_FROM_DEVICE; 242 frmr->access_flags = (IB_ACCESS_LOCAL_WRITE|IB_ACCESS_REMOTE_WRITE); 243 frmr->sg_nents = nents; 244 245 for (pno = 0; pno < nents; pno++) { 246 int len = min_t(int, rs_length, PAGE_SIZE - pg_off); 247 248 head->arg.pages[pg_no] = rqstp->rq_arg.pages[pg_no]; 249 head->arg.page_len += len; 250 head->arg.len += len; 251 if (!pg_off) 252 head->count++; 253 254 sg_set_page(&frmr->sg[pno], rqstp->rq_arg.pages[pg_no], 255 len, pg_off); 256 257 rqstp->rq_respages = &rqstp->rq_arg.pages[pg_no+1]; 258 rqstp->rq_next_page = rqstp->rq_respages + 1; 259 260 /* adjust offset and wrap to next page if needed */ 261 pg_off += len; 262 if (pg_off == PAGE_SIZE) { 263 pg_off = 0; 264 pg_no++; 265 } 266 rs_length -= len; 267 } 268 269 if (last && rs_length == 0) 270 set_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags); 271 else 272 clear_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags); 273 274 dma_nents = ib_dma_map_sg(xprt->sc_cm_id->device, 275 frmr->sg, frmr->sg_nents, 276 frmr->direction); 277 if (!dma_nents) { 278 pr_err("svcrdma: failed to dma map sg %p\n", 279 frmr->sg); 280 return -ENOMEM; 281 } 282 283 n = ib_map_mr_sg(frmr->mr, frmr->sg, frmr->sg_nents, NULL, PAGE_SIZE); 284 if (unlikely(n != frmr->sg_nents)) { 285 pr_err("svcrdma: failed to map mr %p (%d/%d elements)\n", 286 frmr->mr, n, frmr->sg_nents); 287 return n < 0 ? n : -EINVAL; 288 } 289 290 /* Bump the key */ 291 key = (u8)(frmr->mr->lkey & 0x000000FF); 292 ib_update_fast_reg_key(frmr->mr, ++key); 293 294 ctxt->sge[0].addr = frmr->mr->iova; 295 ctxt->sge[0].lkey = frmr->mr->lkey; 296 ctxt->sge[0].length = frmr->mr->length; 297 ctxt->count = 1; 298 ctxt->read_hdr = head; 299 300 /* Prepare REG WR */ 301 ctxt->reg_cqe.done = svc_rdma_wc_reg; 302 reg_wr.wr.wr_cqe = &ctxt->reg_cqe; 303 reg_wr.wr.opcode = IB_WR_REG_MR; 304 reg_wr.wr.send_flags = IB_SEND_SIGNALED; 305 reg_wr.wr.num_sge = 0; 306 reg_wr.mr = frmr->mr; 307 reg_wr.key = frmr->mr->lkey; 308 reg_wr.access = frmr->access_flags; 309 reg_wr.wr.next = &read_wr.wr; 310 311 /* Prepare RDMA_READ */ 312 memset(&read_wr, 0, sizeof(read_wr)); 313 ctxt->cqe.done = svc_rdma_wc_read; 314 read_wr.wr.wr_cqe = &ctxt->cqe; 315 read_wr.wr.send_flags = IB_SEND_SIGNALED; 316 read_wr.rkey = rs_handle; 317 read_wr.remote_addr = rs_offset; 318 read_wr.wr.sg_list = ctxt->sge; 319 read_wr.wr.num_sge = 1; 320 if (xprt->sc_dev_caps & SVCRDMA_DEVCAP_READ_W_INV) { 321 read_wr.wr.opcode = IB_WR_RDMA_READ_WITH_INV; 322 read_wr.wr.ex.invalidate_rkey = ctxt->frmr->mr->lkey; 323 } else { 324 read_wr.wr.opcode = IB_WR_RDMA_READ; 325 read_wr.wr.next = &inv_wr; 326 /* Prepare invalidate */ 327 memset(&inv_wr, 0, sizeof(inv_wr)); 328 ctxt->inv_cqe.done = svc_rdma_wc_inv; 329 inv_wr.wr_cqe = &ctxt->inv_cqe; 330 inv_wr.opcode = IB_WR_LOCAL_INV; 331 inv_wr.send_flags = IB_SEND_SIGNALED | IB_SEND_FENCE; 332 inv_wr.ex.invalidate_rkey = frmr->mr->lkey; 333 } 334 335 /* Post the chain */ 336 ret = svc_rdma_send(xprt, ®_wr.wr); 337 if (ret) { 338 pr_err("svcrdma: Error %d posting RDMA_READ\n", ret); 339 set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags); 340 goto err; 341 } 342 343 /* return current location in page array */ 344 *page_no = pg_no; 345 *page_offset = pg_off; 346 ret = read; 347 atomic_inc(&rdma_stat_read); 348 return ret; 349 err: 350 ib_dma_unmap_sg(xprt->sc_cm_id->device, 351 frmr->sg, frmr->sg_nents, frmr->direction); 352 svc_rdma_put_context(ctxt, 0); 353 svc_rdma_put_frmr(xprt, frmr); 354 return ret; 355 } 356 357 static unsigned int 358 rdma_rcl_chunk_count(struct rpcrdma_read_chunk *ch) 359 { 360 unsigned int count; 361 362 for (count = 0; ch->rc_discrim != xdr_zero; ch++) 363 count++; 364 return count; 365 } 366 367 /* If there was additional inline content, append it to the end of arg.pages. 368 * Tail copy has to be done after the reader function has determined how many 369 * pages are needed for RDMA READ. 370 */ 371 static int 372 rdma_copy_tail(struct svc_rqst *rqstp, struct svc_rdma_op_ctxt *head, 373 u32 position, u32 byte_count, u32 page_offset, int page_no) 374 { 375 char *srcp, *destp; 376 377 srcp = head->arg.head[0].iov_base + position; 378 byte_count = head->arg.head[0].iov_len - position; 379 if (byte_count > PAGE_SIZE) { 380 dprintk("svcrdma: large tail unsupported\n"); 381 return 0; 382 } 383 384 /* Fit as much of the tail on the current page as possible */ 385 if (page_offset != PAGE_SIZE) { 386 destp = page_address(rqstp->rq_arg.pages[page_no]); 387 destp += page_offset; 388 while (byte_count--) { 389 *destp++ = *srcp++; 390 page_offset++; 391 if (page_offset == PAGE_SIZE && byte_count) 392 goto more; 393 } 394 goto done; 395 } 396 397 more: 398 /* Fit the rest on the next page */ 399 page_no++; 400 destp = page_address(rqstp->rq_arg.pages[page_no]); 401 while (byte_count--) 402 *destp++ = *srcp++; 403 404 rqstp->rq_respages = &rqstp->rq_arg.pages[page_no+1]; 405 rqstp->rq_next_page = rqstp->rq_respages + 1; 406 407 done: 408 byte_count = head->arg.head[0].iov_len - position; 409 head->arg.page_len += byte_count; 410 head->arg.len += byte_count; 411 head->arg.buflen += byte_count; 412 return 1; 413 } 414 415 /* Returns the address of the first read chunk or <nul> if no read chunk 416 * is present 417 */ 418 static struct rpcrdma_read_chunk * 419 svc_rdma_get_read_chunk(struct rpcrdma_msg *rmsgp) 420 { 421 struct rpcrdma_read_chunk *ch = 422 (struct rpcrdma_read_chunk *)&rmsgp->rm_body.rm_chunks[0]; 423 424 if (ch->rc_discrim == xdr_zero) 425 return NULL; 426 return ch; 427 } 428 429 static int rdma_read_chunks(struct svcxprt_rdma *xprt, 430 struct rpcrdma_msg *rmsgp, 431 struct svc_rqst *rqstp, 432 struct svc_rdma_op_ctxt *head) 433 { 434 int page_no, ret; 435 struct rpcrdma_read_chunk *ch; 436 u32 handle, page_offset, byte_count; 437 u32 position; 438 u64 rs_offset; 439 bool last; 440 441 /* If no read list is present, return 0 */ 442 ch = svc_rdma_get_read_chunk(rmsgp); 443 if (!ch) 444 return 0; 445 446 if (rdma_rcl_chunk_count(ch) > RPCSVC_MAXPAGES) 447 return -EINVAL; 448 449 /* The request is completed when the RDMA_READs complete. The 450 * head context keeps all the pages that comprise the 451 * request. 452 */ 453 head->arg.head[0] = rqstp->rq_arg.head[0]; 454 head->arg.tail[0] = rqstp->rq_arg.tail[0]; 455 head->hdr_count = head->count; 456 head->arg.page_base = 0; 457 head->arg.page_len = 0; 458 head->arg.len = rqstp->rq_arg.len; 459 head->arg.buflen = rqstp->rq_arg.buflen; 460 461 /* RDMA_NOMSG: RDMA READ data should land just after RDMA RECV data */ 462 position = be32_to_cpu(ch->rc_position); 463 if (position == 0) { 464 head->arg.pages = &head->pages[0]; 465 page_offset = head->byte_len; 466 } else { 467 head->arg.pages = &head->pages[head->count]; 468 page_offset = 0; 469 } 470 471 ret = 0; 472 page_no = 0; 473 for (; ch->rc_discrim != xdr_zero; ch++) { 474 if (be32_to_cpu(ch->rc_position) != position) 475 goto err; 476 477 handle = be32_to_cpu(ch->rc_target.rs_handle), 478 byte_count = be32_to_cpu(ch->rc_target.rs_length); 479 xdr_decode_hyper((__be32 *)&ch->rc_target.rs_offset, 480 &rs_offset); 481 482 while (byte_count > 0) { 483 last = (ch + 1)->rc_discrim == xdr_zero; 484 ret = xprt->sc_reader(xprt, rqstp, head, 485 &page_no, &page_offset, 486 handle, byte_count, 487 rs_offset, last); 488 if (ret < 0) 489 goto err; 490 byte_count -= ret; 491 rs_offset += ret; 492 head->arg.buflen += ret; 493 } 494 } 495 496 /* Read list may need XDR round-up (see RFC 5666, s. 3.7) */ 497 if (page_offset & 3) { 498 u32 pad = 4 - (page_offset & 3); 499 500 head->arg.tail[0].iov_len += pad; 501 head->arg.len += pad; 502 head->arg.buflen += pad; 503 page_offset += pad; 504 } 505 506 ret = 1; 507 if (position && position < head->arg.head[0].iov_len) 508 ret = rdma_copy_tail(rqstp, head, position, 509 byte_count, page_offset, page_no); 510 head->arg.head[0].iov_len = position; 511 head->position = position; 512 513 err: 514 /* Detach arg pages. svc_recv will replenish them */ 515 for (page_no = 0; 516 &rqstp->rq_pages[page_no] < rqstp->rq_respages; page_no++) 517 rqstp->rq_pages[page_no] = NULL; 518 519 return ret; 520 } 521 522 static void rdma_read_complete(struct svc_rqst *rqstp, 523 struct svc_rdma_op_ctxt *head) 524 { 525 int page_no; 526 527 /* Copy RPC pages */ 528 for (page_no = 0; page_no < head->count; page_no++) { 529 put_page(rqstp->rq_pages[page_no]); 530 rqstp->rq_pages[page_no] = head->pages[page_no]; 531 } 532 533 /* Adjustments made for RDMA_NOMSG type requests */ 534 if (head->position == 0) { 535 if (head->arg.len <= head->sge[0].length) { 536 head->arg.head[0].iov_len = head->arg.len - 537 head->byte_len; 538 head->arg.page_len = 0; 539 } else { 540 head->arg.head[0].iov_len = head->sge[0].length - 541 head->byte_len; 542 head->arg.page_len = head->arg.len - 543 head->sge[0].length; 544 } 545 } 546 547 /* Point rq_arg.pages past header */ 548 rqstp->rq_arg.pages = &rqstp->rq_pages[head->hdr_count]; 549 rqstp->rq_arg.page_len = head->arg.page_len; 550 rqstp->rq_arg.page_base = head->arg.page_base; 551 552 /* rq_respages starts after the last arg page */ 553 rqstp->rq_respages = &rqstp->rq_pages[page_no]; 554 rqstp->rq_next_page = rqstp->rq_respages + 1; 555 556 /* Rebuild rq_arg head and tail. */ 557 rqstp->rq_arg.head[0] = head->arg.head[0]; 558 rqstp->rq_arg.tail[0] = head->arg.tail[0]; 559 rqstp->rq_arg.len = head->arg.len; 560 rqstp->rq_arg.buflen = head->arg.buflen; 561 } 562 563 /* By convention, backchannel calls arrive via rdma_msg type 564 * messages, and never populate the chunk lists. This makes 565 * the RPC/RDMA header small and fixed in size, so it is 566 * straightforward to check the RPC header's direction field. 567 */ 568 static bool 569 svc_rdma_is_backchannel_reply(struct svc_xprt *xprt, struct rpcrdma_msg *rmsgp) 570 { 571 __be32 *p = (__be32 *)rmsgp; 572 573 if (!xprt->xpt_bc_xprt) 574 return false; 575 576 if (rmsgp->rm_type != rdma_msg) 577 return false; 578 if (rmsgp->rm_body.rm_chunks[0] != xdr_zero) 579 return false; 580 if (rmsgp->rm_body.rm_chunks[1] != xdr_zero) 581 return false; 582 if (rmsgp->rm_body.rm_chunks[2] != xdr_zero) 583 return false; 584 585 /* sanity */ 586 if (p[7] != rmsgp->rm_xid) 587 return false; 588 /* call direction */ 589 if (p[8] == cpu_to_be32(RPC_CALL)) 590 return false; 591 592 return true; 593 } 594 595 /* 596 * Set up the rqstp thread context to point to the RQ buffer. If 597 * necessary, pull additional data from the client with an RDMA_READ 598 * request. 599 */ 600 int svc_rdma_recvfrom(struct svc_rqst *rqstp) 601 { 602 struct svc_xprt *xprt = rqstp->rq_xprt; 603 struct svcxprt_rdma *rdma_xprt = 604 container_of(xprt, struct svcxprt_rdma, sc_xprt); 605 struct svc_rdma_op_ctxt *ctxt = NULL; 606 struct rpcrdma_msg *rmsgp; 607 int ret = 0; 608 609 dprintk("svcrdma: rqstp=%p\n", rqstp); 610 611 spin_lock_bh(&rdma_xprt->sc_rq_dto_lock); 612 if (!list_empty(&rdma_xprt->sc_read_complete_q)) { 613 ctxt = list_entry(rdma_xprt->sc_read_complete_q.next, 614 struct svc_rdma_op_ctxt, 615 dto_q); 616 list_del_init(&ctxt->dto_q); 617 spin_unlock_bh(&rdma_xprt->sc_rq_dto_lock); 618 rdma_read_complete(rqstp, ctxt); 619 goto complete; 620 } else if (!list_empty(&rdma_xprt->sc_rq_dto_q)) { 621 ctxt = list_entry(rdma_xprt->sc_rq_dto_q.next, 622 struct svc_rdma_op_ctxt, 623 dto_q); 624 list_del_init(&ctxt->dto_q); 625 } else { 626 atomic_inc(&rdma_stat_rq_starve); 627 clear_bit(XPT_DATA, &xprt->xpt_flags); 628 ctxt = NULL; 629 } 630 spin_unlock_bh(&rdma_xprt->sc_rq_dto_lock); 631 if (!ctxt) { 632 /* This is the EAGAIN path. The svc_recv routine will 633 * return -EAGAIN, the nfsd thread will go to call into 634 * svc_recv again and we shouldn't be on the active 635 * transport list 636 */ 637 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) 638 goto defer; 639 goto out; 640 } 641 dprintk("svcrdma: processing ctxt=%p on xprt=%p, rqstp=%p\n", 642 ctxt, rdma_xprt, rqstp); 643 atomic_inc(&rdma_stat_recv); 644 645 /* Build up the XDR from the receive buffers. */ 646 rdma_build_arg_xdr(rqstp, ctxt, ctxt->byte_len); 647 648 /* Decode the RDMA header. */ 649 rmsgp = (struct rpcrdma_msg *)rqstp->rq_arg.head[0].iov_base; 650 ret = svc_rdma_xdr_decode_req(&rqstp->rq_arg); 651 if (ret < 0) 652 goto out_err; 653 if (ret == 0) 654 goto out_drop; 655 rqstp->rq_xprt_hlen = ret; 656 657 if (svc_rdma_is_backchannel_reply(xprt, rmsgp)) { 658 ret = svc_rdma_handle_bc_reply(xprt->xpt_bc_xprt, rmsgp, 659 &rqstp->rq_arg); 660 svc_rdma_put_context(ctxt, 0); 661 if (ret) 662 goto repost; 663 return ret; 664 } 665 666 /* Read read-list data. */ 667 ret = rdma_read_chunks(rdma_xprt, rmsgp, rqstp, ctxt); 668 if (ret > 0) { 669 /* read-list posted, defer until data received from client. */ 670 goto defer; 671 } else if (ret < 0) { 672 /* Post of read-list failed, free context. */ 673 svc_rdma_put_context(ctxt, 1); 674 return 0; 675 } 676 677 complete: 678 ret = rqstp->rq_arg.head[0].iov_len 679 + rqstp->rq_arg.page_len 680 + rqstp->rq_arg.tail[0].iov_len; 681 svc_rdma_put_context(ctxt, 0); 682 out: 683 dprintk("svcrdma: ret=%d, rq_arg.len=%u, " 684 "rq_arg.head[0].iov_base=%p, rq_arg.head[0].iov_len=%zd\n", 685 ret, rqstp->rq_arg.len, 686 rqstp->rq_arg.head[0].iov_base, 687 rqstp->rq_arg.head[0].iov_len); 688 rqstp->rq_prot = IPPROTO_MAX; 689 svc_xprt_copy_addrs(rqstp, xprt); 690 return ret; 691 692 out_err: 693 svc_rdma_send_error(rdma_xprt, rmsgp, ret); 694 svc_rdma_put_context(ctxt, 0); 695 return 0; 696 697 defer: 698 return 0; 699 700 out_drop: 701 svc_rdma_put_context(ctxt, 1); 702 repost: 703 return svc_rdma_repost_recv(rdma_xprt, GFP_KERNEL); 704 } 705