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_recvfrom. This is called from 48 * svc_recv when the transport indicates there is incoming data to 49 * be read. "Data Ready" is signaled when an RDMA Receive completes, 50 * or when a set of RDMA Reads complete. 51 * 52 * An svc_rqst is passed in. This structure contains an array of 53 * free pages (rq_pages) that will contain the incoming RPC message. 54 * 55 * Short messages are moved directly into svc_rqst::rq_arg, and 56 * the RPC Call is ready to be processed by the Upper Layer. 57 * svc_rdma_recvfrom returns the length of the RPC Call message, 58 * completing the reception of the RPC Call. 59 * 60 * However, when an incoming message has Read chunks, 61 * svc_rdma_recvfrom must post RDMA Reads to pull the RPC Call's 62 * data payload from the client. svc_rdma_recvfrom sets up the 63 * RDMA Reads using pages in svc_rqst::rq_pages, which are 64 * transferred to an svc_rdma_recv_ctxt for the duration of the 65 * I/O. svc_rdma_recvfrom then returns zero, since the RPC message 66 * is still not yet ready. 67 * 68 * When the Read chunk payloads have become available on the 69 * server, "Data Ready" is raised again, and svc_recv calls 70 * svc_rdma_recvfrom again. This second call may use a different 71 * svc_rqst than the first one, thus any information that needs 72 * to be preserved across these two calls is kept in an 73 * svc_rdma_recv_ctxt. 74 * 75 * The second call to svc_rdma_recvfrom performs final assembly 76 * of the RPC Call message, using the RDMA Read sink pages kept in 77 * the svc_rdma_recv_ctxt. The xdr_buf is copied from the 78 * svc_rdma_recv_ctxt to the second svc_rqst. The second call returns 79 * the length of the completed RPC Call message. 80 * 81 * Page Management 82 * 83 * Pages under I/O must be transferred from the first svc_rqst to an 84 * svc_rdma_recv_ctxt before the first svc_rdma_recvfrom call returns. 85 * 86 * The first svc_rqst supplies pages for RDMA Reads. These are moved 87 * from rqstp::rq_pages into ctxt::pages. The consumed elements of 88 * the rq_pages array are set to NULL and refilled with the first 89 * svc_rdma_recvfrom call returns. 90 * 91 * During the second svc_rdma_recvfrom call, RDMA Read sink pages 92 * are transferred from the svc_rdma_recv_ctxt to the second svc_rqst 93 * (see rdma_read_complete() below). 94 */ 95 96 #include <linux/spinlock.h> 97 #include <asm/unaligned.h> 98 #include <rdma/ib_verbs.h> 99 #include <rdma/rdma_cm.h> 100 101 #include <linux/sunrpc/xdr.h> 102 #include <linux/sunrpc/debug.h> 103 #include <linux/sunrpc/rpc_rdma.h> 104 #include <linux/sunrpc/svc_rdma.h> 105 106 #include "xprt_rdma.h" 107 #include <trace/events/rpcrdma.h> 108 109 #define RPCDBG_FACILITY RPCDBG_SVCXPRT 110 111 static void svc_rdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc); 112 113 static inline struct svc_rdma_recv_ctxt * 114 svc_rdma_next_recv_ctxt(struct list_head *list) 115 { 116 return list_first_entry_or_null(list, struct svc_rdma_recv_ctxt, 117 rc_list); 118 } 119 120 static void svc_rdma_recv_cid_init(struct svcxprt_rdma *rdma, 121 struct rpc_rdma_cid *cid) 122 { 123 cid->ci_queue_id = rdma->sc_rq_cq->res.id; 124 cid->ci_completion_id = atomic_inc_return(&rdma->sc_completion_ids); 125 } 126 127 static struct svc_rdma_recv_ctxt * 128 svc_rdma_recv_ctxt_alloc(struct svcxprt_rdma *rdma) 129 { 130 struct svc_rdma_recv_ctxt *ctxt; 131 dma_addr_t addr; 132 void *buffer; 133 134 ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL); 135 if (!ctxt) 136 goto fail0; 137 buffer = kmalloc(rdma->sc_max_req_size, GFP_KERNEL); 138 if (!buffer) 139 goto fail1; 140 addr = ib_dma_map_single(rdma->sc_pd->device, buffer, 141 rdma->sc_max_req_size, DMA_FROM_DEVICE); 142 if (ib_dma_mapping_error(rdma->sc_pd->device, addr)) 143 goto fail2; 144 145 svc_rdma_recv_cid_init(rdma, &ctxt->rc_cid); 146 147 ctxt->rc_recv_wr.next = NULL; 148 ctxt->rc_recv_wr.wr_cqe = &ctxt->rc_cqe; 149 ctxt->rc_recv_wr.sg_list = &ctxt->rc_recv_sge; 150 ctxt->rc_recv_wr.num_sge = 1; 151 ctxt->rc_cqe.done = svc_rdma_wc_receive; 152 ctxt->rc_recv_sge.addr = addr; 153 ctxt->rc_recv_sge.length = rdma->sc_max_req_size; 154 ctxt->rc_recv_sge.lkey = rdma->sc_pd->local_dma_lkey; 155 ctxt->rc_recv_buf = buffer; 156 ctxt->rc_temp = false; 157 return ctxt; 158 159 fail2: 160 kfree(buffer); 161 fail1: 162 kfree(ctxt); 163 fail0: 164 return NULL; 165 } 166 167 static void svc_rdma_recv_ctxt_destroy(struct svcxprt_rdma *rdma, 168 struct svc_rdma_recv_ctxt *ctxt) 169 { 170 ib_dma_unmap_single(rdma->sc_pd->device, ctxt->rc_recv_sge.addr, 171 ctxt->rc_recv_sge.length, DMA_FROM_DEVICE); 172 kfree(ctxt->rc_recv_buf); 173 kfree(ctxt); 174 } 175 176 /** 177 * svc_rdma_recv_ctxts_destroy - Release all recv_ctxt's for an xprt 178 * @rdma: svcxprt_rdma being torn down 179 * 180 */ 181 void svc_rdma_recv_ctxts_destroy(struct svcxprt_rdma *rdma) 182 { 183 struct svc_rdma_recv_ctxt *ctxt; 184 struct llist_node *node; 185 186 while ((node = llist_del_first(&rdma->sc_recv_ctxts))) { 187 ctxt = llist_entry(node, struct svc_rdma_recv_ctxt, rc_node); 188 svc_rdma_recv_ctxt_destroy(rdma, ctxt); 189 } 190 } 191 192 static struct svc_rdma_recv_ctxt * 193 svc_rdma_recv_ctxt_get(struct svcxprt_rdma *rdma) 194 { 195 struct svc_rdma_recv_ctxt *ctxt; 196 struct llist_node *node; 197 198 node = llist_del_first(&rdma->sc_recv_ctxts); 199 if (!node) 200 goto out_empty; 201 ctxt = llist_entry(node, struct svc_rdma_recv_ctxt, rc_node); 202 203 out: 204 ctxt->rc_page_count = 0; 205 ctxt->rc_read_payload_length = 0; 206 return ctxt; 207 208 out_empty: 209 ctxt = svc_rdma_recv_ctxt_alloc(rdma); 210 if (!ctxt) 211 return NULL; 212 goto out; 213 } 214 215 /** 216 * svc_rdma_recv_ctxt_put - Return recv_ctxt to free list 217 * @rdma: controlling svcxprt_rdma 218 * @ctxt: object to return to the free list 219 * 220 */ 221 void svc_rdma_recv_ctxt_put(struct svcxprt_rdma *rdma, 222 struct svc_rdma_recv_ctxt *ctxt) 223 { 224 unsigned int i; 225 226 for (i = 0; i < ctxt->rc_page_count; i++) 227 put_page(ctxt->rc_pages[i]); 228 229 if (!ctxt->rc_temp) 230 llist_add(&ctxt->rc_node, &rdma->sc_recv_ctxts); 231 else 232 svc_rdma_recv_ctxt_destroy(rdma, ctxt); 233 } 234 235 /** 236 * svc_rdma_release_rqst - Release transport-specific per-rqst resources 237 * @rqstp: svc_rqst being released 238 * 239 * Ensure that the recv_ctxt is released whether or not a Reply 240 * was sent. For example, the client could close the connection, 241 * or svc_process could drop an RPC, before the Reply is sent. 242 */ 243 void svc_rdma_release_rqst(struct svc_rqst *rqstp) 244 { 245 struct svc_rdma_recv_ctxt *ctxt = rqstp->rq_xprt_ctxt; 246 struct svc_xprt *xprt = rqstp->rq_xprt; 247 struct svcxprt_rdma *rdma = 248 container_of(xprt, struct svcxprt_rdma, sc_xprt); 249 250 rqstp->rq_xprt_ctxt = NULL; 251 if (ctxt) 252 svc_rdma_recv_ctxt_put(rdma, ctxt); 253 } 254 255 static int __svc_rdma_post_recv(struct svcxprt_rdma *rdma, 256 struct svc_rdma_recv_ctxt *ctxt) 257 { 258 int ret; 259 260 trace_svcrdma_post_recv(ctxt); 261 ret = ib_post_recv(rdma->sc_qp, &ctxt->rc_recv_wr, NULL); 262 if (ret) 263 goto err_post; 264 return 0; 265 266 err_post: 267 trace_svcrdma_rq_post_err(rdma, ret); 268 svc_rdma_recv_ctxt_put(rdma, ctxt); 269 return ret; 270 } 271 272 static int svc_rdma_post_recv(struct svcxprt_rdma *rdma) 273 { 274 struct svc_rdma_recv_ctxt *ctxt; 275 276 if (test_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags)) 277 return 0; 278 ctxt = svc_rdma_recv_ctxt_get(rdma); 279 if (!ctxt) 280 return -ENOMEM; 281 return __svc_rdma_post_recv(rdma, ctxt); 282 } 283 284 /** 285 * svc_rdma_post_recvs - Post initial set of Recv WRs 286 * @rdma: fresh svcxprt_rdma 287 * 288 * Returns true if successful, otherwise false. 289 */ 290 bool svc_rdma_post_recvs(struct svcxprt_rdma *rdma) 291 { 292 struct svc_rdma_recv_ctxt *ctxt; 293 unsigned int i; 294 int ret; 295 296 for (i = 0; i < rdma->sc_max_requests; i++) { 297 ctxt = svc_rdma_recv_ctxt_get(rdma); 298 if (!ctxt) 299 return false; 300 ctxt->rc_temp = true; 301 ret = __svc_rdma_post_recv(rdma, ctxt); 302 if (ret) 303 return false; 304 } 305 return true; 306 } 307 308 /** 309 * svc_rdma_wc_receive - Invoked by RDMA provider for each polled Receive WC 310 * @cq: Completion Queue context 311 * @wc: Work Completion object 312 * 313 * NB: The svc_xprt/svcxprt_rdma is pinned whenever it's possible that 314 * the Receive completion handler could be running. 315 */ 316 static void svc_rdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc) 317 { 318 struct svcxprt_rdma *rdma = cq->cq_context; 319 struct ib_cqe *cqe = wc->wr_cqe; 320 struct svc_rdma_recv_ctxt *ctxt; 321 322 /* WARNING: Only wc->wr_cqe and wc->status are reliable */ 323 ctxt = container_of(cqe, struct svc_rdma_recv_ctxt, rc_cqe); 324 325 trace_svcrdma_wc_receive(wc, &ctxt->rc_cid); 326 if (wc->status != IB_WC_SUCCESS) 327 goto flushed; 328 329 if (svc_rdma_post_recv(rdma)) 330 goto post_err; 331 332 /* All wc fields are now known to be valid */ 333 ctxt->rc_byte_len = wc->byte_len; 334 ib_dma_sync_single_for_cpu(rdma->sc_pd->device, 335 ctxt->rc_recv_sge.addr, 336 wc->byte_len, DMA_FROM_DEVICE); 337 338 spin_lock(&rdma->sc_rq_dto_lock); 339 list_add_tail(&ctxt->rc_list, &rdma->sc_rq_dto_q); 340 /* Note the unlock pairs with the smp_rmb in svc_xprt_ready: */ 341 set_bit(XPT_DATA, &rdma->sc_xprt.xpt_flags); 342 spin_unlock(&rdma->sc_rq_dto_lock); 343 if (!test_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags)) 344 svc_xprt_enqueue(&rdma->sc_xprt); 345 return; 346 347 flushed: 348 post_err: 349 svc_rdma_recv_ctxt_put(rdma, ctxt); 350 set_bit(XPT_CLOSE, &rdma->sc_xprt.xpt_flags); 351 svc_xprt_enqueue(&rdma->sc_xprt); 352 } 353 354 /** 355 * svc_rdma_flush_recv_queues - Drain pending Receive work 356 * @rdma: svcxprt_rdma being shut down 357 * 358 */ 359 void svc_rdma_flush_recv_queues(struct svcxprt_rdma *rdma) 360 { 361 struct svc_rdma_recv_ctxt *ctxt; 362 363 while ((ctxt = svc_rdma_next_recv_ctxt(&rdma->sc_read_complete_q))) { 364 list_del(&ctxt->rc_list); 365 svc_rdma_recv_ctxt_put(rdma, ctxt); 366 } 367 while ((ctxt = svc_rdma_next_recv_ctxt(&rdma->sc_rq_dto_q))) { 368 list_del(&ctxt->rc_list); 369 svc_rdma_recv_ctxt_put(rdma, ctxt); 370 } 371 } 372 373 static void svc_rdma_build_arg_xdr(struct svc_rqst *rqstp, 374 struct svc_rdma_recv_ctxt *ctxt) 375 { 376 struct xdr_buf *arg = &rqstp->rq_arg; 377 378 arg->head[0].iov_base = ctxt->rc_recv_buf; 379 arg->head[0].iov_len = ctxt->rc_byte_len; 380 arg->tail[0].iov_base = NULL; 381 arg->tail[0].iov_len = 0; 382 arg->page_len = 0; 383 arg->page_base = 0; 384 arg->buflen = ctxt->rc_byte_len; 385 arg->len = ctxt->rc_byte_len; 386 } 387 388 /* This accommodates the largest possible Write chunk. 389 */ 390 #define MAX_BYTES_WRITE_CHUNK ((u32)(RPCSVC_MAXPAGES << PAGE_SHIFT)) 391 392 /* This accommodates the largest possible Position-Zero 393 * Read chunk or Reply chunk. 394 */ 395 #define MAX_BYTES_SPECIAL_CHUNK ((u32)((RPCSVC_MAXPAGES + 2) << PAGE_SHIFT)) 396 397 /* Sanity check the Read list. 398 * 399 * Implementation limits: 400 * - This implementation supports only one Read chunk. 401 * 402 * Sanity checks: 403 * - Read list does not overflow Receive buffer. 404 * - Segment size limited by largest NFS data payload. 405 * 406 * The segment count is limited to how many segments can 407 * fit in the transport header without overflowing the 408 * buffer. That's about 40 Read segments for a 1KB inline 409 * threshold. 410 * 411 * Return values: 412 * %true: Read list is valid. @rctxt's xdr_stream is updated 413 * to point to the first byte past the Read list. 414 * %false: Read list is corrupt. @rctxt's xdr_stream is left 415 * in an unknown state. 416 */ 417 static bool xdr_check_read_list(struct svc_rdma_recv_ctxt *rctxt) 418 { 419 u32 position, len; 420 bool first; 421 __be32 *p; 422 423 p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p)); 424 if (!p) 425 return false; 426 427 len = 0; 428 first = true; 429 while (xdr_item_is_present(p)) { 430 p = xdr_inline_decode(&rctxt->rc_stream, 431 rpcrdma_readseg_maxsz * sizeof(*p)); 432 if (!p) 433 return false; 434 435 if (first) { 436 position = be32_to_cpup(p); 437 first = false; 438 } else if (be32_to_cpup(p) != position) { 439 return false; 440 } 441 p += 2; 442 len += be32_to_cpup(p); 443 444 p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p)); 445 if (!p) 446 return false; 447 } 448 return len <= MAX_BYTES_SPECIAL_CHUNK; 449 } 450 451 /* The segment count is limited to how many segments can 452 * fit in the transport header without overflowing the 453 * buffer. That's about 60 Write segments for a 1KB inline 454 * threshold. 455 */ 456 static bool xdr_check_write_chunk(struct svc_rdma_recv_ctxt *rctxt, u32 maxlen) 457 { 458 u32 i, segcount, total; 459 __be32 *p; 460 461 p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p)); 462 if (!p) 463 return false; 464 segcount = be32_to_cpup(p); 465 466 total = 0; 467 for (i = 0; i < segcount; i++) { 468 u32 handle, length; 469 u64 offset; 470 471 p = xdr_inline_decode(&rctxt->rc_stream, 472 rpcrdma_segment_maxsz * sizeof(*p)); 473 if (!p) 474 return false; 475 476 xdr_decode_rdma_segment(p, &handle, &length, &offset); 477 trace_svcrdma_decode_wseg(handle, length, offset); 478 479 total += length; 480 } 481 return total <= maxlen; 482 } 483 484 /* Sanity check the Write list. 485 * 486 * Implementation limits: 487 * - This implementation currently supports only one Write chunk. 488 * 489 * Sanity checks: 490 * - Write list does not overflow Receive buffer. 491 * - Chunk size limited by largest NFS data payload. 492 * 493 * Return values: 494 * %true: Write list is valid. @rctxt's xdr_stream is updated 495 * to point to the first byte past the Write list. 496 * %false: Write list is corrupt. @rctxt's xdr_stream is left 497 * in an unknown state. 498 */ 499 static bool xdr_check_write_list(struct svc_rdma_recv_ctxt *rctxt) 500 { 501 u32 chcount = 0; 502 __be32 *p; 503 504 p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p)); 505 if (!p) 506 return false; 507 rctxt->rc_write_list = p; 508 while (xdr_item_is_present(p)) { 509 if (!xdr_check_write_chunk(rctxt, MAX_BYTES_WRITE_CHUNK)) 510 return false; 511 ++chcount; 512 p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p)); 513 if (!p) 514 return false; 515 } 516 if (!chcount) 517 rctxt->rc_write_list = NULL; 518 return chcount < 2; 519 } 520 521 /* Sanity check the Reply chunk. 522 * 523 * Sanity checks: 524 * - Reply chunk does not overflow Receive buffer. 525 * - Chunk size limited by largest NFS data payload. 526 * 527 * Return values: 528 * %true: Reply chunk is valid. @rctxt's xdr_stream is updated 529 * to point to the first byte past the Reply chunk. 530 * %false: Reply chunk is corrupt. @rctxt's xdr_stream is left 531 * in an unknown state. 532 */ 533 static bool xdr_check_reply_chunk(struct svc_rdma_recv_ctxt *rctxt) 534 { 535 __be32 *p; 536 537 p = xdr_inline_decode(&rctxt->rc_stream, sizeof(*p)); 538 if (!p) 539 return false; 540 rctxt->rc_reply_chunk = NULL; 541 if (xdr_item_is_present(p)) { 542 if (!xdr_check_write_chunk(rctxt, MAX_BYTES_SPECIAL_CHUNK)) 543 return false; 544 rctxt->rc_reply_chunk = p; 545 } 546 return true; 547 } 548 549 /* RPC-over-RDMA Version One private extension: Remote Invalidation. 550 * Responder's choice: requester signals it can handle Send With 551 * Invalidate, and responder chooses one R_key to invalidate. 552 * 553 * If there is exactly one distinct R_key in the received transport 554 * header, set rc_inv_rkey to that R_key. Otherwise, set it to zero. 555 * 556 * Perform this operation while the received transport header is 557 * still in the CPU cache. 558 */ 559 static void svc_rdma_get_inv_rkey(struct svcxprt_rdma *rdma, 560 struct svc_rdma_recv_ctxt *ctxt) 561 { 562 __be32 inv_rkey, *p; 563 u32 i, segcount; 564 565 ctxt->rc_inv_rkey = 0; 566 567 if (!rdma->sc_snd_w_inv) 568 return; 569 570 inv_rkey = xdr_zero; 571 p = ctxt->rc_recv_buf; 572 p += rpcrdma_fixed_maxsz; 573 574 /* Read list */ 575 while (xdr_item_is_present(p++)) { 576 p++; /* position */ 577 if (inv_rkey == xdr_zero) 578 inv_rkey = *p; 579 else if (inv_rkey != *p) 580 return; 581 p += 4; 582 } 583 584 /* Write list */ 585 while (xdr_item_is_present(p++)) { 586 segcount = be32_to_cpup(p++); 587 for (i = 0; i < segcount; i++) { 588 if (inv_rkey == xdr_zero) 589 inv_rkey = *p; 590 else if (inv_rkey != *p) 591 return; 592 p += 4; 593 } 594 } 595 596 /* Reply chunk */ 597 if (xdr_item_is_present(p++)) { 598 segcount = be32_to_cpup(p++); 599 for (i = 0; i < segcount; i++) { 600 if (inv_rkey == xdr_zero) 601 inv_rkey = *p; 602 else if (inv_rkey != *p) 603 return; 604 p += 4; 605 } 606 } 607 608 ctxt->rc_inv_rkey = be32_to_cpu(inv_rkey); 609 } 610 611 /** 612 * svc_rdma_xdr_decode_req - Decode the transport header 613 * @rq_arg: xdr_buf containing ingress RPC/RDMA message 614 * @rctxt: state of decoding 615 * 616 * On entry, xdr->head[0].iov_base points to first byte of the 617 * RPC-over-RDMA transport header. 618 * 619 * On successful exit, head[0] points to first byte past the 620 * RPC-over-RDMA header. For RDMA_MSG, this is the RPC message. 621 * 622 * The length of the RPC-over-RDMA header is returned. 623 * 624 * Assumptions: 625 * - The transport header is entirely contained in the head iovec. 626 */ 627 static int svc_rdma_xdr_decode_req(struct xdr_buf *rq_arg, 628 struct svc_rdma_recv_ctxt *rctxt) 629 { 630 __be32 *p, *rdma_argp; 631 unsigned int hdr_len; 632 633 rdma_argp = rq_arg->head[0].iov_base; 634 xdr_init_decode(&rctxt->rc_stream, rq_arg, rdma_argp, NULL); 635 636 p = xdr_inline_decode(&rctxt->rc_stream, 637 rpcrdma_fixed_maxsz * sizeof(*p)); 638 if (unlikely(!p)) 639 goto out_short; 640 p++; 641 if (*p != rpcrdma_version) 642 goto out_version; 643 p += 2; 644 switch (*p) { 645 case rdma_msg: 646 break; 647 case rdma_nomsg: 648 break; 649 case rdma_done: 650 goto out_drop; 651 case rdma_error: 652 goto out_drop; 653 default: 654 goto out_proc; 655 } 656 657 if (!xdr_check_read_list(rctxt)) 658 goto out_inval; 659 if (!xdr_check_write_list(rctxt)) 660 goto out_inval; 661 if (!xdr_check_reply_chunk(rctxt)) 662 goto out_inval; 663 664 rq_arg->head[0].iov_base = rctxt->rc_stream.p; 665 hdr_len = xdr_stream_pos(&rctxt->rc_stream); 666 rq_arg->head[0].iov_len -= hdr_len; 667 rq_arg->len -= hdr_len; 668 trace_svcrdma_decode_rqst(rctxt, rdma_argp, hdr_len); 669 return hdr_len; 670 671 out_short: 672 trace_svcrdma_decode_short_err(rctxt, rq_arg->len); 673 return -EINVAL; 674 675 out_version: 676 trace_svcrdma_decode_badvers_err(rctxt, rdma_argp); 677 return -EPROTONOSUPPORT; 678 679 out_drop: 680 trace_svcrdma_decode_drop_err(rctxt, rdma_argp); 681 return 0; 682 683 out_proc: 684 trace_svcrdma_decode_badproc_err(rctxt, rdma_argp); 685 return -EINVAL; 686 687 out_inval: 688 trace_svcrdma_decode_parse_err(rctxt, rdma_argp); 689 return -EINVAL; 690 } 691 692 static void rdma_read_complete(struct svc_rqst *rqstp, 693 struct svc_rdma_recv_ctxt *head) 694 { 695 int page_no; 696 697 /* Move Read chunk pages to rqstp so that they will be released 698 * when svc_process is done with them. 699 */ 700 for (page_no = 0; page_no < head->rc_page_count; page_no++) { 701 put_page(rqstp->rq_pages[page_no]); 702 rqstp->rq_pages[page_no] = head->rc_pages[page_no]; 703 } 704 head->rc_page_count = 0; 705 706 /* Point rq_arg.pages past header */ 707 rqstp->rq_arg.pages = &rqstp->rq_pages[head->rc_hdr_count]; 708 rqstp->rq_arg.page_len = head->rc_arg.page_len; 709 710 /* rq_respages starts after the last arg page */ 711 rqstp->rq_respages = &rqstp->rq_pages[page_no]; 712 rqstp->rq_next_page = rqstp->rq_respages + 1; 713 714 /* Rebuild rq_arg head and tail. */ 715 rqstp->rq_arg.head[0] = head->rc_arg.head[0]; 716 rqstp->rq_arg.tail[0] = head->rc_arg.tail[0]; 717 rqstp->rq_arg.len = head->rc_arg.len; 718 rqstp->rq_arg.buflen = head->rc_arg.buflen; 719 } 720 721 static void svc_rdma_send_error(struct svcxprt_rdma *rdma, 722 struct svc_rdma_recv_ctxt *rctxt, 723 int status) 724 { 725 struct svc_rdma_send_ctxt *sctxt; 726 727 sctxt = svc_rdma_send_ctxt_get(rdma); 728 if (!sctxt) 729 return; 730 svc_rdma_send_error_msg(rdma, sctxt, rctxt, status); 731 } 732 733 /* By convention, backchannel calls arrive via rdma_msg type 734 * messages, and never populate the chunk lists. This makes 735 * the RPC/RDMA header small and fixed in size, so it is 736 * straightforward to check the RPC header's direction field. 737 */ 738 static bool svc_rdma_is_backchannel_reply(struct svc_xprt *xprt, 739 __be32 *rdma_resp) 740 { 741 __be32 *p; 742 743 if (!xprt->xpt_bc_xprt) 744 return false; 745 746 p = rdma_resp + 3; 747 if (*p++ != rdma_msg) 748 return false; 749 750 if (*p++ != xdr_zero) 751 return false; 752 if (*p++ != xdr_zero) 753 return false; 754 if (*p++ != xdr_zero) 755 return false; 756 757 /* XID sanity */ 758 if (*p++ != *rdma_resp) 759 return false; 760 /* call direction */ 761 if (*p == cpu_to_be32(RPC_CALL)) 762 return false; 763 764 return true; 765 } 766 767 /** 768 * svc_rdma_recvfrom - Receive an RPC call 769 * @rqstp: request structure into which to receive an RPC Call 770 * 771 * Returns: 772 * The positive number of bytes in the RPC Call message, 773 * %0 if there were no Calls ready to return, 774 * %-EINVAL if the Read chunk data is too large, 775 * %-ENOMEM if rdma_rw context pool was exhausted, 776 * %-ENOTCONN if posting failed (connection is lost), 777 * %-EIO if rdma_rw initialization failed (DMA mapping, etc). 778 * 779 * Called in a loop when XPT_DATA is set. XPT_DATA is cleared only 780 * when there are no remaining ctxt's to process. 781 * 782 * The next ctxt is removed from the "receive" lists. 783 * 784 * - If the ctxt completes a Read, then finish assembling the Call 785 * message and return the number of bytes in the message. 786 * 787 * - If the ctxt completes a Receive, then construct the Call 788 * message from the contents of the Receive buffer. 789 * 790 * - If there are no Read chunks in this message, then finish 791 * assembling the Call message and return the number of bytes 792 * in the message. 793 * 794 * - If there are Read chunks in this message, post Read WRs to 795 * pull that payload and return 0. 796 */ 797 int svc_rdma_recvfrom(struct svc_rqst *rqstp) 798 { 799 struct svc_xprt *xprt = rqstp->rq_xprt; 800 struct svcxprt_rdma *rdma_xprt = 801 container_of(xprt, struct svcxprt_rdma, sc_xprt); 802 struct svc_rdma_recv_ctxt *ctxt; 803 __be32 *p; 804 int ret; 805 806 rqstp->rq_xprt_ctxt = NULL; 807 808 spin_lock(&rdma_xprt->sc_rq_dto_lock); 809 ctxt = svc_rdma_next_recv_ctxt(&rdma_xprt->sc_read_complete_q); 810 if (ctxt) { 811 list_del(&ctxt->rc_list); 812 spin_unlock(&rdma_xprt->sc_rq_dto_lock); 813 rdma_read_complete(rqstp, ctxt); 814 goto complete; 815 } 816 ctxt = svc_rdma_next_recv_ctxt(&rdma_xprt->sc_rq_dto_q); 817 if (!ctxt) { 818 /* No new incoming requests, terminate the loop */ 819 clear_bit(XPT_DATA, &xprt->xpt_flags); 820 spin_unlock(&rdma_xprt->sc_rq_dto_lock); 821 return 0; 822 } 823 list_del(&ctxt->rc_list); 824 spin_unlock(&rdma_xprt->sc_rq_dto_lock); 825 826 atomic_inc(&rdma_stat_recv); 827 828 svc_rdma_build_arg_xdr(rqstp, ctxt); 829 830 /* Prevent svc_xprt_release from releasing pages in rq_pages 831 * if we return 0 or an error. 832 */ 833 rqstp->rq_respages = rqstp->rq_pages; 834 rqstp->rq_next_page = rqstp->rq_respages; 835 836 p = (__be32 *)rqstp->rq_arg.head[0].iov_base; 837 ret = svc_rdma_xdr_decode_req(&rqstp->rq_arg, ctxt); 838 if (ret < 0) 839 goto out_err; 840 if (ret == 0) 841 goto out_drop; 842 rqstp->rq_xprt_hlen = ret; 843 844 if (svc_rdma_is_backchannel_reply(xprt, p)) 845 goto out_backchannel; 846 847 svc_rdma_get_inv_rkey(rdma_xprt, ctxt); 848 849 p += rpcrdma_fixed_maxsz; 850 if (*p != xdr_zero) 851 goto out_readchunk; 852 853 complete: 854 rqstp->rq_xprt_ctxt = ctxt; 855 rqstp->rq_prot = IPPROTO_MAX; 856 svc_xprt_copy_addrs(rqstp, xprt); 857 return rqstp->rq_arg.len; 858 859 out_readchunk: 860 ret = svc_rdma_recv_read_chunk(rdma_xprt, rqstp, ctxt, p); 861 if (ret < 0) 862 goto out_postfail; 863 return 0; 864 865 out_err: 866 svc_rdma_send_error(rdma_xprt, ctxt, ret); 867 svc_rdma_recv_ctxt_put(rdma_xprt, ctxt); 868 return 0; 869 870 out_postfail: 871 if (ret == -EINVAL) 872 svc_rdma_send_error(rdma_xprt, ctxt, ret); 873 svc_rdma_recv_ctxt_put(rdma_xprt, ctxt); 874 return ret; 875 876 out_backchannel: 877 svc_rdma_handle_bc_reply(rqstp, ctxt); 878 out_drop: 879 svc_rdma_recv_ctxt_put(rdma_xprt, ctxt); 880 return 0; 881 } 882