1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause 2 /* 3 * Copyright (c) 2016-2018 Oracle. All rights reserved. 4 * Copyright (c) 2014 Open Grid Computing, Inc. All rights reserved. 5 * Copyright (c) 2005-2006 Network Appliance, Inc. All rights reserved. 6 * 7 * This software is available to you under a choice of one of two 8 * licenses. You may choose to be licensed under the terms of the GNU 9 * General Public License (GPL) Version 2, available from the file 10 * COPYING in the main directory of this source tree, or the BSD-type 11 * license below: 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 17 * Redistributions of source code must retain the above copyright 18 * notice, this list of conditions and the following disclaimer. 19 * 20 * Redistributions in binary form must reproduce the above 21 * copyright notice, this list of conditions and the following 22 * disclaimer in the documentation and/or other materials provided 23 * with the distribution. 24 * 25 * Neither the name of the Network Appliance, Inc. nor the names of 26 * its contributors may be used to endorse or promote products 27 * derived from this software without specific prior written 28 * permission. 29 * 30 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 31 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 32 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 33 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 34 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 35 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 36 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 37 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 38 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 39 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 40 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 41 * 42 * Author: Tom Tucker <tom@opengridcomputing.com> 43 */ 44 45 /* Operation 46 * 47 * The main entry point is svc_rdma_sendto. This is called by the 48 * RPC server when an RPC Reply is ready to be transmitted to a client. 49 * 50 * The passed-in svc_rqst contains a struct xdr_buf which holds an 51 * XDR-encoded RPC Reply message. sendto must construct the RPC-over-RDMA 52 * transport header, post all Write WRs needed for this Reply, then post 53 * a Send WR conveying the transport header and the RPC message itself to 54 * the client. 55 * 56 * svc_rdma_sendto must fully transmit the Reply before returning, as 57 * the svc_rqst will be recycled as soon as sendto returns. Remaining 58 * resources referred to by the svc_rqst are also recycled at that time. 59 * Therefore any resources that must remain longer must be detached 60 * from the svc_rqst and released later. 61 * 62 * Page Management 63 * 64 * The I/O that performs Reply transmission is asynchronous, and may 65 * complete well after sendto returns. Thus pages under I/O must be 66 * removed from the svc_rqst before sendto returns. 67 * 68 * The logic here depends on Send Queue and completion ordering. Since 69 * the Send WR is always posted last, it will always complete last. Thus 70 * when it completes, it is guaranteed that all previous Write WRs have 71 * also completed. 72 * 73 * Write WRs are constructed and posted. Each Write segment gets its own 74 * svc_rdma_rw_ctxt, allowing the Write completion handler to find and 75 * DMA-unmap the pages under I/O for that Write segment. The Write 76 * completion handler does not release any pages. 77 * 78 * When the Send WR is constructed, it also gets its own svc_rdma_send_ctxt. 79 * The ownership of all of the Reply's pages are transferred into that 80 * ctxt, the Send WR is posted, and sendto returns. 81 * 82 * The svc_rdma_send_ctxt is presented when the Send WR completes. The 83 * Send completion handler finally releases the Reply's pages. 84 * 85 * This mechanism also assumes that completions on the transport's Send 86 * Completion Queue do not run in parallel. Otherwise a Write completion 87 * and Send completion running at the same time could release pages that 88 * are still DMA-mapped. 89 * 90 * Error Handling 91 * 92 * - If the Send WR is posted successfully, it will either complete 93 * successfully, or get flushed. Either way, the Send completion 94 * handler releases the Reply's pages. 95 * - If the Send WR cannot be not posted, the forward path releases 96 * the Reply's pages. 97 * 98 * This handles the case, without the use of page reference counting, 99 * where two different Write segments send portions of the same page. 100 */ 101 102 #include <linux/spinlock.h> 103 #include <asm/unaligned.h> 104 105 #include <rdma/ib_verbs.h> 106 #include <rdma/rdma_cm.h> 107 108 #include <linux/sunrpc/debug.h> 109 #include <linux/sunrpc/rpc_rdma.h> 110 #include <linux/sunrpc/svc_rdma.h> 111 112 #include "xprt_rdma.h" 113 #include <trace/events/rpcrdma.h> 114 115 #define RPCDBG_FACILITY RPCDBG_SVCXPRT 116 117 static void svc_rdma_wc_send(struct ib_cq *cq, struct ib_wc *wc); 118 119 static inline struct svc_rdma_send_ctxt * 120 svc_rdma_next_send_ctxt(struct list_head *list) 121 { 122 return list_first_entry_or_null(list, struct svc_rdma_send_ctxt, 123 sc_list); 124 } 125 126 static struct svc_rdma_send_ctxt * 127 svc_rdma_send_ctxt_alloc(struct svcxprt_rdma *rdma) 128 { 129 struct svc_rdma_send_ctxt *ctxt; 130 dma_addr_t addr; 131 void *buffer; 132 size_t size; 133 int i; 134 135 size = sizeof(*ctxt); 136 size += rdma->sc_max_send_sges * sizeof(struct ib_sge); 137 ctxt = kmalloc(size, GFP_KERNEL); 138 if (!ctxt) 139 goto fail0; 140 buffer = kmalloc(rdma->sc_max_req_size, GFP_KERNEL); 141 if (!buffer) 142 goto fail1; 143 addr = ib_dma_map_single(rdma->sc_pd->device, buffer, 144 rdma->sc_max_req_size, DMA_TO_DEVICE); 145 if (ib_dma_mapping_error(rdma->sc_pd->device, addr)) 146 goto fail2; 147 148 ctxt->sc_send_wr.next = NULL; 149 ctxt->sc_send_wr.wr_cqe = &ctxt->sc_cqe; 150 ctxt->sc_send_wr.sg_list = ctxt->sc_sges; 151 ctxt->sc_send_wr.send_flags = IB_SEND_SIGNALED; 152 ctxt->sc_cqe.done = svc_rdma_wc_send; 153 ctxt->sc_xprt_buf = buffer; 154 ctxt->sc_sges[0].addr = addr; 155 156 for (i = 0; i < rdma->sc_max_send_sges; i++) 157 ctxt->sc_sges[i].lkey = rdma->sc_pd->local_dma_lkey; 158 return ctxt; 159 160 fail2: 161 kfree(buffer); 162 fail1: 163 kfree(ctxt); 164 fail0: 165 return NULL; 166 } 167 168 /** 169 * svc_rdma_send_ctxts_destroy - Release all send_ctxt's for an xprt 170 * @rdma: svcxprt_rdma being torn down 171 * 172 */ 173 void svc_rdma_send_ctxts_destroy(struct svcxprt_rdma *rdma) 174 { 175 struct svc_rdma_send_ctxt *ctxt; 176 177 while ((ctxt = svc_rdma_next_send_ctxt(&rdma->sc_send_ctxts))) { 178 list_del(&ctxt->sc_list); 179 ib_dma_unmap_single(rdma->sc_pd->device, 180 ctxt->sc_sges[0].addr, 181 rdma->sc_max_req_size, 182 DMA_TO_DEVICE); 183 kfree(ctxt->sc_xprt_buf); 184 kfree(ctxt); 185 } 186 } 187 188 /** 189 * svc_rdma_send_ctxt_get - Get a free send_ctxt 190 * @rdma: controlling svcxprt_rdma 191 * 192 * Returns a ready-to-use send_ctxt, or NULL if none are 193 * available and a fresh one cannot be allocated. 194 */ 195 struct svc_rdma_send_ctxt *svc_rdma_send_ctxt_get(struct svcxprt_rdma *rdma) 196 { 197 struct svc_rdma_send_ctxt *ctxt; 198 199 spin_lock(&rdma->sc_send_lock); 200 ctxt = svc_rdma_next_send_ctxt(&rdma->sc_send_ctxts); 201 if (!ctxt) 202 goto out_empty; 203 list_del(&ctxt->sc_list); 204 spin_unlock(&rdma->sc_send_lock); 205 206 out: 207 ctxt->sc_send_wr.num_sge = 0; 208 ctxt->sc_cur_sge_no = 0; 209 ctxt->sc_page_count = 0; 210 return ctxt; 211 212 out_empty: 213 spin_unlock(&rdma->sc_send_lock); 214 ctxt = svc_rdma_send_ctxt_alloc(rdma); 215 if (!ctxt) 216 return NULL; 217 goto out; 218 } 219 220 /** 221 * svc_rdma_send_ctxt_put - Return send_ctxt to free list 222 * @rdma: controlling svcxprt_rdma 223 * @ctxt: object to return to the free list 224 * 225 * Pages left in sc_pages are DMA unmapped and released. 226 */ 227 void svc_rdma_send_ctxt_put(struct svcxprt_rdma *rdma, 228 struct svc_rdma_send_ctxt *ctxt) 229 { 230 struct ib_device *device = rdma->sc_cm_id->device; 231 unsigned int i; 232 233 /* The first SGE contains the transport header, which 234 * remains mapped until @ctxt is destroyed. 235 */ 236 for (i = 1; i < ctxt->sc_send_wr.num_sge; i++) { 237 ib_dma_unmap_page(device, 238 ctxt->sc_sges[i].addr, 239 ctxt->sc_sges[i].length, 240 DMA_TO_DEVICE); 241 trace_svcrdma_dma_unmap_page(rdma, 242 ctxt->sc_sges[i].addr, 243 ctxt->sc_sges[i].length); 244 } 245 246 for (i = 0; i < ctxt->sc_page_count; ++i) 247 put_page(ctxt->sc_pages[i]); 248 249 spin_lock(&rdma->sc_send_lock); 250 list_add(&ctxt->sc_list, &rdma->sc_send_ctxts); 251 spin_unlock(&rdma->sc_send_lock); 252 } 253 254 /** 255 * svc_rdma_wc_send - Invoked by RDMA provider for each polled Send WC 256 * @cq: Completion Queue context 257 * @wc: Work Completion object 258 * 259 * NB: The svc_xprt/svcxprt_rdma is pinned whenever it's possible that 260 * the Send completion handler could be running. 261 */ 262 static void svc_rdma_wc_send(struct ib_cq *cq, struct ib_wc *wc) 263 { 264 struct svcxprt_rdma *rdma = cq->cq_context; 265 struct ib_cqe *cqe = wc->wr_cqe; 266 struct svc_rdma_send_ctxt *ctxt; 267 268 trace_svcrdma_wc_send(wc); 269 270 atomic_inc(&rdma->sc_sq_avail); 271 wake_up(&rdma->sc_send_wait); 272 273 ctxt = container_of(cqe, struct svc_rdma_send_ctxt, sc_cqe); 274 svc_rdma_send_ctxt_put(rdma, ctxt); 275 276 if (unlikely(wc->status != IB_WC_SUCCESS)) { 277 set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags); 278 svc_xprt_enqueue(&rdma->sc_xprt); 279 } 280 281 svc_xprt_put(&rdma->sc_xprt); 282 } 283 284 /** 285 * svc_rdma_send - Post a single Send WR 286 * @rdma: transport on which to post the WR 287 * @wr: prepared Send WR to post 288 * 289 * Returns zero the Send WR was posted successfully. Otherwise, a 290 * negative errno is returned. 291 */ 292 int svc_rdma_send(struct svcxprt_rdma *rdma, struct ib_send_wr *wr) 293 { 294 int ret; 295 296 might_sleep(); 297 298 /* If the SQ is full, wait until an SQ entry is available */ 299 while (1) { 300 if ((atomic_dec_return(&rdma->sc_sq_avail) < 0)) { 301 atomic_inc(&rdma_stat_sq_starve); 302 trace_svcrdma_sq_full(rdma); 303 atomic_inc(&rdma->sc_sq_avail); 304 wait_event(rdma->sc_send_wait, 305 atomic_read(&rdma->sc_sq_avail) > 1); 306 if (test_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags)) 307 return -ENOTCONN; 308 trace_svcrdma_sq_retry(rdma); 309 continue; 310 } 311 312 svc_xprt_get(&rdma->sc_xprt); 313 ret = ib_post_send(rdma->sc_qp, wr, NULL); 314 trace_svcrdma_post_send(wr, ret); 315 if (ret) { 316 set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags); 317 svc_xprt_put(&rdma->sc_xprt); 318 wake_up(&rdma->sc_send_wait); 319 } 320 break; 321 } 322 return ret; 323 } 324 325 static u32 xdr_padsize(u32 len) 326 { 327 return (len & 3) ? (4 - (len & 3)) : 0; 328 } 329 330 /* Returns length of transport header, in bytes. 331 */ 332 static unsigned int svc_rdma_reply_hdr_len(__be32 *rdma_resp) 333 { 334 unsigned int nsegs; 335 __be32 *p; 336 337 p = rdma_resp; 338 339 /* RPC-over-RDMA V1 replies never have a Read list. */ 340 p += rpcrdma_fixed_maxsz + 1; 341 342 /* Skip Write list. */ 343 while (*p++ != xdr_zero) { 344 nsegs = be32_to_cpup(p++); 345 p += nsegs * rpcrdma_segment_maxsz; 346 } 347 348 /* Skip Reply chunk. */ 349 if (*p++ != xdr_zero) { 350 nsegs = be32_to_cpup(p++); 351 p += nsegs * rpcrdma_segment_maxsz; 352 } 353 354 return (unsigned long)p - (unsigned long)rdma_resp; 355 } 356 357 /* One Write chunk is copied from Call transport header to Reply 358 * transport header. Each segment's length field is updated to 359 * reflect number of bytes consumed in the segment. 360 * 361 * Returns number of segments in this chunk. 362 */ 363 static unsigned int xdr_encode_write_chunk(__be32 *dst, __be32 *src, 364 unsigned int remaining) 365 { 366 unsigned int i, nsegs; 367 u32 seg_len; 368 369 /* Write list discriminator */ 370 *dst++ = *src++; 371 372 /* number of segments in this chunk */ 373 nsegs = be32_to_cpup(src); 374 *dst++ = *src++; 375 376 for (i = nsegs; i; i--) { 377 /* segment's RDMA handle */ 378 *dst++ = *src++; 379 380 /* bytes returned in this segment */ 381 seg_len = be32_to_cpu(*src); 382 if (remaining >= seg_len) { 383 /* entire segment was consumed */ 384 *dst = *src; 385 remaining -= seg_len; 386 } else { 387 /* segment only partly filled */ 388 *dst = cpu_to_be32(remaining); 389 remaining = 0; 390 } 391 dst++; src++; 392 393 /* segment's RDMA offset */ 394 *dst++ = *src++; 395 *dst++ = *src++; 396 } 397 398 return nsegs; 399 } 400 401 /* The client provided a Write list in the Call message. Fill in 402 * the segments in the first Write chunk in the Reply's transport 403 * header with the number of bytes consumed in each segment. 404 * Remaining chunks are returned unused. 405 * 406 * Assumptions: 407 * - Client has provided only one Write chunk 408 */ 409 static void svc_rdma_xdr_encode_write_list(__be32 *rdma_resp, __be32 *wr_ch, 410 unsigned int consumed) 411 { 412 unsigned int nsegs; 413 __be32 *p, *q; 414 415 /* RPC-over-RDMA V1 replies never have a Read list. */ 416 p = rdma_resp + rpcrdma_fixed_maxsz + 1; 417 418 q = wr_ch; 419 while (*q != xdr_zero) { 420 nsegs = xdr_encode_write_chunk(p, q, consumed); 421 q += 2 + nsegs * rpcrdma_segment_maxsz; 422 p += 2 + nsegs * rpcrdma_segment_maxsz; 423 consumed = 0; 424 } 425 426 /* Terminate Write list */ 427 *p++ = xdr_zero; 428 429 /* Reply chunk discriminator; may be replaced later */ 430 *p = xdr_zero; 431 } 432 433 /* The client provided a Reply chunk in the Call message. Fill in 434 * the segments in the Reply chunk in the Reply message with the 435 * number of bytes consumed in each segment. 436 * 437 * Assumptions: 438 * - Reply can always fit in the provided Reply chunk 439 */ 440 static void svc_rdma_xdr_encode_reply_chunk(__be32 *rdma_resp, __be32 *rp_ch, 441 unsigned int consumed) 442 { 443 __be32 *p; 444 445 /* Find the Reply chunk in the Reply's xprt header. 446 * RPC-over-RDMA V1 replies never have a Read list. 447 */ 448 p = rdma_resp + rpcrdma_fixed_maxsz + 1; 449 450 /* Skip past Write list */ 451 while (*p++ != xdr_zero) 452 p += 1 + be32_to_cpup(p) * rpcrdma_segment_maxsz; 453 454 xdr_encode_write_chunk(p, rp_ch, consumed); 455 } 456 457 /* Parse the RPC Call's transport header. 458 */ 459 static void svc_rdma_get_write_arrays(__be32 *rdma_argp, 460 __be32 **write, __be32 **reply) 461 { 462 __be32 *p; 463 464 p = rdma_argp + rpcrdma_fixed_maxsz; 465 466 /* Read list */ 467 while (*p++ != xdr_zero) 468 p += 5; 469 470 /* Write list */ 471 if (*p != xdr_zero) { 472 *write = p; 473 while (*p++ != xdr_zero) 474 p += 1 + be32_to_cpu(*p) * 4; 475 } else { 476 *write = NULL; 477 p++; 478 } 479 480 /* Reply chunk */ 481 if (*p != xdr_zero) 482 *reply = p; 483 else 484 *reply = NULL; 485 } 486 487 static int svc_rdma_dma_map_page(struct svcxprt_rdma *rdma, 488 struct svc_rdma_send_ctxt *ctxt, 489 struct page *page, 490 unsigned long offset, 491 unsigned int len) 492 { 493 struct ib_device *dev = rdma->sc_cm_id->device; 494 dma_addr_t dma_addr; 495 496 dma_addr = ib_dma_map_page(dev, page, offset, len, DMA_TO_DEVICE); 497 trace_svcrdma_dma_map_page(rdma, dma_addr, len); 498 if (ib_dma_mapping_error(dev, dma_addr)) 499 goto out_maperr; 500 501 ctxt->sc_sges[ctxt->sc_cur_sge_no].addr = dma_addr; 502 ctxt->sc_sges[ctxt->sc_cur_sge_no].length = len; 503 ctxt->sc_send_wr.num_sge++; 504 return 0; 505 506 out_maperr: 507 return -EIO; 508 } 509 510 /* ib_dma_map_page() is used here because svc_rdma_dma_unmap() 511 * handles DMA-unmap and it uses ib_dma_unmap_page() exclusively. 512 */ 513 static int svc_rdma_dma_map_buf(struct svcxprt_rdma *rdma, 514 struct svc_rdma_send_ctxt *ctxt, 515 unsigned char *base, 516 unsigned int len) 517 { 518 return svc_rdma_dma_map_page(rdma, ctxt, virt_to_page(base), 519 offset_in_page(base), len); 520 } 521 522 /** 523 * svc_rdma_sync_reply_hdr - DMA sync the transport header buffer 524 * @rdma: controlling transport 525 * @ctxt: send_ctxt for the Send WR 526 * @len: length of transport header 527 * 528 */ 529 void svc_rdma_sync_reply_hdr(struct svcxprt_rdma *rdma, 530 struct svc_rdma_send_ctxt *ctxt, 531 unsigned int len) 532 { 533 ctxt->sc_sges[0].length = len; 534 ctxt->sc_send_wr.num_sge++; 535 ib_dma_sync_single_for_device(rdma->sc_pd->device, 536 ctxt->sc_sges[0].addr, len, 537 DMA_TO_DEVICE); 538 } 539 540 /* If the xdr_buf has more elements than the device can 541 * transmit in a single RDMA Send, then the reply will 542 * have to be copied into a bounce buffer. 543 */ 544 static bool svc_rdma_pull_up_needed(struct svcxprt_rdma *rdma, 545 struct xdr_buf *xdr, 546 __be32 *wr_lst) 547 { 548 int elements; 549 550 /* xdr->head */ 551 elements = 1; 552 553 /* xdr->pages */ 554 if (!wr_lst) { 555 unsigned int remaining; 556 unsigned long pageoff; 557 558 pageoff = xdr->page_base & ~PAGE_MASK; 559 remaining = xdr->page_len; 560 while (remaining) { 561 ++elements; 562 remaining -= min_t(u32, PAGE_SIZE - pageoff, 563 remaining); 564 pageoff = 0; 565 } 566 } 567 568 /* xdr->tail */ 569 if (xdr->tail[0].iov_len) 570 ++elements; 571 572 /* assume 1 SGE is needed for the transport header */ 573 return elements >= rdma->sc_max_send_sges; 574 } 575 576 /* The device is not capable of sending the reply directly. 577 * Assemble the elements of @xdr into the transport header 578 * buffer. 579 */ 580 static int svc_rdma_pull_up_reply_msg(struct svcxprt_rdma *rdma, 581 struct svc_rdma_send_ctxt *ctxt, 582 struct xdr_buf *xdr, __be32 *wr_lst) 583 { 584 unsigned char *dst, *tailbase; 585 unsigned int taillen; 586 587 dst = ctxt->sc_xprt_buf; 588 dst += ctxt->sc_sges[0].length; 589 590 memcpy(dst, xdr->head[0].iov_base, xdr->head[0].iov_len); 591 dst += xdr->head[0].iov_len; 592 593 tailbase = xdr->tail[0].iov_base; 594 taillen = xdr->tail[0].iov_len; 595 if (wr_lst) { 596 u32 xdrpad; 597 598 xdrpad = xdr_padsize(xdr->page_len); 599 if (taillen && xdrpad) { 600 tailbase += xdrpad; 601 taillen -= xdrpad; 602 } 603 } else { 604 unsigned int len, remaining; 605 unsigned long pageoff; 606 struct page **ppages; 607 608 ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT); 609 pageoff = xdr->page_base & ~PAGE_MASK; 610 remaining = xdr->page_len; 611 while (remaining) { 612 len = min_t(u32, PAGE_SIZE - pageoff, remaining); 613 614 memcpy(dst, page_address(*ppages), len); 615 remaining -= len; 616 dst += len; 617 pageoff = 0; 618 } 619 } 620 621 if (taillen) 622 memcpy(dst, tailbase, taillen); 623 624 ctxt->sc_sges[0].length += xdr->len; 625 ib_dma_sync_single_for_device(rdma->sc_pd->device, 626 ctxt->sc_sges[0].addr, 627 ctxt->sc_sges[0].length, 628 DMA_TO_DEVICE); 629 630 return 0; 631 } 632 633 /* svc_rdma_map_reply_msg - Map the buffer holding RPC message 634 * @rdma: controlling transport 635 * @ctxt: send_ctxt for the Send WR 636 * @xdr: prepared xdr_buf containing RPC message 637 * @wr_lst: pointer to Call header's Write list, or NULL 638 * 639 * Load the xdr_buf into the ctxt's sge array, and DMA map each 640 * element as it is added. 641 * 642 * Returns zero on success, or a negative errno on failure. 643 */ 644 int svc_rdma_map_reply_msg(struct svcxprt_rdma *rdma, 645 struct svc_rdma_send_ctxt *ctxt, 646 struct xdr_buf *xdr, __be32 *wr_lst) 647 { 648 unsigned int len, remaining; 649 unsigned long page_off; 650 struct page **ppages; 651 unsigned char *base; 652 u32 xdr_pad; 653 int ret; 654 655 if (svc_rdma_pull_up_needed(rdma, xdr, wr_lst)) 656 return svc_rdma_pull_up_reply_msg(rdma, ctxt, xdr, wr_lst); 657 658 ++ctxt->sc_cur_sge_no; 659 ret = svc_rdma_dma_map_buf(rdma, ctxt, 660 xdr->head[0].iov_base, 661 xdr->head[0].iov_len); 662 if (ret < 0) 663 return ret; 664 665 /* If a Write chunk is present, the xdr_buf's page list 666 * is not included inline. However the Upper Layer may 667 * have added XDR padding in the tail buffer, and that 668 * should not be included inline. 669 */ 670 if (wr_lst) { 671 base = xdr->tail[0].iov_base; 672 len = xdr->tail[0].iov_len; 673 xdr_pad = xdr_padsize(xdr->page_len); 674 675 if (len && xdr_pad) { 676 base += xdr_pad; 677 len -= xdr_pad; 678 } 679 680 goto tail; 681 } 682 683 ppages = xdr->pages + (xdr->page_base >> PAGE_SHIFT); 684 page_off = xdr->page_base & ~PAGE_MASK; 685 remaining = xdr->page_len; 686 while (remaining) { 687 len = min_t(u32, PAGE_SIZE - page_off, remaining); 688 689 ++ctxt->sc_cur_sge_no; 690 ret = svc_rdma_dma_map_page(rdma, ctxt, *ppages++, 691 page_off, len); 692 if (ret < 0) 693 return ret; 694 695 remaining -= len; 696 page_off = 0; 697 } 698 699 base = xdr->tail[0].iov_base; 700 len = xdr->tail[0].iov_len; 701 tail: 702 if (len) { 703 ++ctxt->sc_cur_sge_no; 704 ret = svc_rdma_dma_map_buf(rdma, ctxt, base, len); 705 if (ret < 0) 706 return ret; 707 } 708 709 return 0; 710 } 711 712 /* The svc_rqst and all resources it owns are released as soon as 713 * svc_rdma_sendto returns. Transfer pages under I/O to the ctxt 714 * so they are released by the Send completion handler. 715 */ 716 static void svc_rdma_save_io_pages(struct svc_rqst *rqstp, 717 struct svc_rdma_send_ctxt *ctxt) 718 { 719 int i, pages = rqstp->rq_next_page - rqstp->rq_respages; 720 721 ctxt->sc_page_count += pages; 722 for (i = 0; i < pages; i++) { 723 ctxt->sc_pages[i] = rqstp->rq_respages[i]; 724 rqstp->rq_respages[i] = NULL; 725 } 726 727 /* Prevent svc_xprt_release from releasing pages in rq_pages */ 728 rqstp->rq_next_page = rqstp->rq_respages; 729 } 730 731 /* Prepare the portion of the RPC Reply that will be transmitted 732 * via RDMA Send. The RPC-over-RDMA transport header is prepared 733 * in sc_sges[0], and the RPC xdr_buf is prepared in following sges. 734 * 735 * Depending on whether a Write list or Reply chunk is present, 736 * the server may send all, a portion of, or none of the xdr_buf. 737 * In the latter case, only the transport header (sc_sges[0]) is 738 * transmitted. 739 * 740 * RDMA Send is the last step of transmitting an RPC reply. Pages 741 * involved in the earlier RDMA Writes are here transferred out 742 * of the rqstp and into the sctxt's page array. These pages are 743 * DMA unmapped by each Write completion, but the subsequent Send 744 * completion finally releases these pages. 745 * 746 * Assumptions: 747 * - The Reply's transport header will never be larger than a page. 748 */ 749 static int svc_rdma_send_reply_msg(struct svcxprt_rdma *rdma, 750 struct svc_rdma_send_ctxt *sctxt, 751 struct svc_rdma_recv_ctxt *rctxt, 752 struct svc_rqst *rqstp, 753 __be32 *wr_lst, __be32 *rp_ch) 754 { 755 int ret; 756 757 if (!rp_ch) { 758 ret = svc_rdma_map_reply_msg(rdma, sctxt, 759 &rqstp->rq_res, wr_lst); 760 if (ret < 0) 761 return ret; 762 } 763 764 svc_rdma_save_io_pages(rqstp, sctxt); 765 766 if (rctxt->rc_inv_rkey) { 767 sctxt->sc_send_wr.opcode = IB_WR_SEND_WITH_INV; 768 sctxt->sc_send_wr.ex.invalidate_rkey = rctxt->rc_inv_rkey; 769 } else { 770 sctxt->sc_send_wr.opcode = IB_WR_SEND; 771 } 772 dprintk("svcrdma: posting Send WR with %u sge(s)\n", 773 sctxt->sc_send_wr.num_sge); 774 return svc_rdma_send(rdma, &sctxt->sc_send_wr); 775 } 776 777 /* Given the client-provided Write and Reply chunks, the server was not 778 * able to form a complete reply. Return an RDMA_ERROR message so the 779 * client can retire this RPC transaction. As above, the Send completion 780 * routine releases payload pages that were part of a previous RDMA Write. 781 * 782 * Remote Invalidation is skipped for simplicity. 783 */ 784 static int svc_rdma_send_error_msg(struct svcxprt_rdma *rdma, 785 struct svc_rdma_send_ctxt *ctxt, 786 struct svc_rqst *rqstp) 787 { 788 __be32 *p; 789 int ret; 790 791 p = ctxt->sc_xprt_buf; 792 trace_svcrdma_err_chunk(*p); 793 p += 3; 794 *p++ = rdma_error; 795 *p = err_chunk; 796 svc_rdma_sync_reply_hdr(rdma, ctxt, RPCRDMA_HDRLEN_ERR); 797 798 svc_rdma_save_io_pages(rqstp, ctxt); 799 800 ctxt->sc_send_wr.opcode = IB_WR_SEND; 801 ret = svc_rdma_send(rdma, &ctxt->sc_send_wr); 802 if (ret) { 803 svc_rdma_send_ctxt_put(rdma, ctxt); 804 return ret; 805 } 806 807 return 0; 808 } 809 810 /** 811 * svc_rdma_sendto - Transmit an RPC reply 812 * @rqstp: processed RPC request, reply XDR already in ::rq_res 813 * 814 * Any resources still associated with @rqstp are released upon return. 815 * If no reply message was possible, the connection is closed. 816 * 817 * Returns: 818 * %0 if an RPC reply has been successfully posted, 819 * %-ENOMEM if a resource shortage occurred (connection is lost), 820 * %-ENOTCONN if posting failed (connection is lost). 821 */ 822 int svc_rdma_sendto(struct svc_rqst *rqstp) 823 { 824 struct svc_xprt *xprt = rqstp->rq_xprt; 825 struct svcxprt_rdma *rdma = 826 container_of(xprt, struct svcxprt_rdma, sc_xprt); 827 struct svc_rdma_recv_ctxt *rctxt = rqstp->rq_xprt_ctxt; 828 __be32 *p, *rdma_argp, *rdma_resp, *wr_lst, *rp_ch; 829 struct xdr_buf *xdr = &rqstp->rq_res; 830 struct svc_rdma_send_ctxt *sctxt; 831 int ret; 832 833 rdma_argp = rctxt->rc_recv_buf; 834 svc_rdma_get_write_arrays(rdma_argp, &wr_lst, &rp_ch); 835 836 /* Create the RDMA response header. xprt->xpt_mutex, 837 * acquired in svc_send(), serializes RPC replies. The 838 * code path below that inserts the credit grant value 839 * into each transport header runs only inside this 840 * critical section. 841 */ 842 ret = -ENOMEM; 843 sctxt = svc_rdma_send_ctxt_get(rdma); 844 if (!sctxt) 845 goto err0; 846 rdma_resp = sctxt->sc_xprt_buf; 847 848 p = rdma_resp; 849 *p++ = *rdma_argp; 850 *p++ = *(rdma_argp + 1); 851 *p++ = rdma->sc_fc_credits; 852 *p++ = rp_ch ? rdma_nomsg : rdma_msg; 853 854 /* Start with empty chunks */ 855 *p++ = xdr_zero; 856 *p++ = xdr_zero; 857 *p = xdr_zero; 858 859 if (wr_lst) { 860 /* XXX: Presume the client sent only one Write chunk */ 861 ret = svc_rdma_send_write_chunk(rdma, wr_lst, xdr); 862 if (ret < 0) 863 goto err2; 864 svc_rdma_xdr_encode_write_list(rdma_resp, wr_lst, ret); 865 } 866 if (rp_ch) { 867 ret = svc_rdma_send_reply_chunk(rdma, rp_ch, wr_lst, xdr); 868 if (ret < 0) 869 goto err2; 870 svc_rdma_xdr_encode_reply_chunk(rdma_resp, rp_ch, ret); 871 } 872 873 svc_rdma_sync_reply_hdr(rdma, sctxt, svc_rdma_reply_hdr_len(rdma_resp)); 874 ret = svc_rdma_send_reply_msg(rdma, sctxt, rctxt, rqstp, 875 wr_lst, rp_ch); 876 if (ret < 0) 877 goto err1; 878 ret = 0; 879 880 out: 881 rqstp->rq_xprt_ctxt = NULL; 882 svc_rdma_recv_ctxt_put(rdma, rctxt); 883 return ret; 884 885 err2: 886 if (ret != -E2BIG && ret != -EINVAL) 887 goto err1; 888 889 ret = svc_rdma_send_error_msg(rdma, sctxt, rqstp); 890 if (ret < 0) 891 goto err1; 892 ret = 0; 893 goto out; 894 895 err1: 896 svc_rdma_send_ctxt_put(rdma, sctxt); 897 err0: 898 trace_svcrdma_send_failed(rqstp, ret); 899 set_bit(XPT_CLOSE, &xprt->xpt_flags); 900 ret = -ENOTCONN; 901 goto out; 902 } 903