1 /* 2 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved. 3 * 4 * This software is available to you under a choice of one of two 5 * licenses. You may choose to be licensed under the terms of the GNU 6 * General Public License (GPL) Version 2, available from the file 7 * COPYING in the main directory of this source tree, or the BSD-type 8 * license below: 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 14 * Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 17 * Redistributions in binary form must reproduce the above 18 * copyright notice, this list of conditions and the following 19 * disclaimer in the documentation and/or other materials provided 20 * with the distribution. 21 * 22 * Neither the name of the Network Appliance, Inc. nor the names of 23 * its contributors may be used to endorse or promote products 24 * derived from this software without specific prior written 25 * permission. 26 * 27 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 28 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 29 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 30 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 31 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 32 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 33 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 34 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 35 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 36 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 37 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 38 */ 39 40 /* 41 * rpc_rdma.c 42 * 43 * This file contains the guts of the RPC RDMA protocol, and 44 * does marshaling/unmarshaling, etc. It is also where interfacing 45 * to the Linux RPC framework lives. 46 */ 47 48 #include "xprt_rdma.h" 49 50 #include <linux/highmem.h> 51 52 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) 53 # define RPCDBG_FACILITY RPCDBG_TRANS 54 #endif 55 56 enum rpcrdma_chunktype { 57 rpcrdma_noch = 0, 58 rpcrdma_readch, 59 rpcrdma_areadch, 60 rpcrdma_writech, 61 rpcrdma_replych 62 }; 63 64 static const char transfertypes[][12] = { 65 "inline", /* no chunks */ 66 "read list", /* some argument via rdma read */ 67 "*read list", /* entire request via rdma read */ 68 "write list", /* some result via rdma write */ 69 "reply chunk" /* entire reply via rdma write */ 70 }; 71 72 /* Returns size of largest RPC-over-RDMA header in a Call message 73 * 74 * The largest Call header contains a full-size Read list and a 75 * minimal Reply chunk. 76 */ 77 static unsigned int rpcrdma_max_call_header_size(unsigned int maxsegs) 78 { 79 unsigned int size; 80 81 /* Fixed header fields and list discriminators */ 82 size = RPCRDMA_HDRLEN_MIN; 83 84 /* Maximum Read list size */ 85 maxsegs += 2; /* segment for head and tail buffers */ 86 size = maxsegs * sizeof(struct rpcrdma_read_chunk); 87 88 /* Minimal Read chunk size */ 89 size += sizeof(__be32); /* segment count */ 90 size += sizeof(struct rpcrdma_segment); 91 size += sizeof(__be32); /* list discriminator */ 92 93 dprintk("RPC: %s: max call header size = %u\n", 94 __func__, size); 95 return size; 96 } 97 98 /* Returns size of largest RPC-over-RDMA header in a Reply message 99 * 100 * There is only one Write list or one Reply chunk per Reply 101 * message. The larger list is the Write list. 102 */ 103 static unsigned int rpcrdma_max_reply_header_size(unsigned int maxsegs) 104 { 105 unsigned int size; 106 107 /* Fixed header fields and list discriminators */ 108 size = RPCRDMA_HDRLEN_MIN; 109 110 /* Maximum Write list size */ 111 maxsegs += 2; /* segment for head and tail buffers */ 112 size = sizeof(__be32); /* segment count */ 113 size += maxsegs * sizeof(struct rpcrdma_segment); 114 size += sizeof(__be32); /* list discriminator */ 115 116 dprintk("RPC: %s: max reply header size = %u\n", 117 __func__, size); 118 return size; 119 } 120 121 void rpcrdma_set_max_header_sizes(struct rpcrdma_ia *ia, 122 struct rpcrdma_create_data_internal *cdata, 123 unsigned int maxsegs) 124 { 125 ia->ri_max_inline_write = cdata->inline_wsize - 126 rpcrdma_max_call_header_size(maxsegs); 127 ia->ri_max_inline_read = cdata->inline_rsize - 128 rpcrdma_max_reply_header_size(maxsegs); 129 } 130 131 /* The client can send a request inline as long as the RPCRDMA header 132 * plus the RPC call fit under the transport's inline limit. If the 133 * combined call message size exceeds that limit, the client must use 134 * the read chunk list for this operation. 135 */ 136 static bool rpcrdma_args_inline(struct rpcrdma_xprt *r_xprt, 137 struct rpc_rqst *rqst) 138 { 139 struct rpcrdma_ia *ia = &r_xprt->rx_ia; 140 141 return rqst->rq_snd_buf.len <= ia->ri_max_inline_write; 142 } 143 144 /* The client can't know how large the actual reply will be. Thus it 145 * plans for the largest possible reply for that particular ULP 146 * operation. If the maximum combined reply message size exceeds that 147 * limit, the client must provide a write list or a reply chunk for 148 * this request. 149 */ 150 static bool rpcrdma_results_inline(struct rpcrdma_xprt *r_xprt, 151 struct rpc_rqst *rqst) 152 { 153 struct rpcrdma_ia *ia = &r_xprt->rx_ia; 154 155 return rqst->rq_rcv_buf.buflen <= ia->ri_max_inline_read; 156 } 157 158 static int 159 rpcrdma_tail_pullup(struct xdr_buf *buf) 160 { 161 size_t tlen = buf->tail[0].iov_len; 162 size_t skip = tlen & 3; 163 164 /* Do not include the tail if it is only an XDR pad */ 165 if (tlen < 4) 166 return 0; 167 168 /* xdr_write_pages() adds a pad at the beginning of the tail 169 * if the content in "buf->pages" is unaligned. Force the 170 * tail's actual content to land at the next XDR position 171 * after the head instead. 172 */ 173 if (skip) { 174 unsigned char *src, *dst; 175 unsigned int count; 176 177 src = buf->tail[0].iov_base; 178 dst = buf->head[0].iov_base; 179 dst += buf->head[0].iov_len; 180 181 src += skip; 182 tlen -= skip; 183 184 dprintk("RPC: %s: skip=%zu, memmove(%p, %p, %zu)\n", 185 __func__, skip, dst, src, tlen); 186 187 for (count = tlen; count; count--) 188 *dst++ = *src++; 189 } 190 191 return tlen; 192 } 193 194 /* Split "vec" on page boundaries into segments. FMR registers pages, 195 * not a byte range. Other modes coalesce these segments into a single 196 * MR when they can. 197 */ 198 static int 199 rpcrdma_convert_kvec(struct kvec *vec, struct rpcrdma_mr_seg *seg, int n) 200 { 201 size_t page_offset; 202 u32 remaining; 203 char *base; 204 205 base = vec->iov_base; 206 page_offset = offset_in_page(base); 207 remaining = vec->iov_len; 208 while (remaining && n < RPCRDMA_MAX_SEGS) { 209 seg[n].mr_page = NULL; 210 seg[n].mr_offset = base; 211 seg[n].mr_len = min_t(u32, PAGE_SIZE - page_offset, remaining); 212 remaining -= seg[n].mr_len; 213 base += seg[n].mr_len; 214 ++n; 215 page_offset = 0; 216 } 217 return n; 218 } 219 220 /* 221 * Chunk assembly from upper layer xdr_buf. 222 * 223 * Prepare the passed-in xdr_buf into representation as RPC/RDMA chunk 224 * elements. Segments are then coalesced when registered, if possible 225 * within the selected memreg mode. 226 * 227 * Returns positive number of segments converted, or a negative errno. 228 */ 229 230 static int 231 rpcrdma_convert_iovs(struct xdr_buf *xdrbuf, unsigned int pos, 232 enum rpcrdma_chunktype type, struct rpcrdma_mr_seg *seg) 233 { 234 int len, n, p, page_base; 235 struct page **ppages; 236 237 n = 0; 238 if (pos == 0) { 239 n = rpcrdma_convert_kvec(&xdrbuf->head[0], seg, n); 240 if (n == RPCRDMA_MAX_SEGS) 241 goto out_overflow; 242 } 243 244 len = xdrbuf->page_len; 245 ppages = xdrbuf->pages + (xdrbuf->page_base >> PAGE_SHIFT); 246 page_base = xdrbuf->page_base & ~PAGE_MASK; 247 p = 0; 248 while (len && n < RPCRDMA_MAX_SEGS) { 249 if (!ppages[p]) { 250 /* alloc the pagelist for receiving buffer */ 251 ppages[p] = alloc_page(GFP_ATOMIC); 252 if (!ppages[p]) 253 return -EAGAIN; 254 } 255 seg[n].mr_page = ppages[p]; 256 seg[n].mr_offset = (void *)(unsigned long) page_base; 257 seg[n].mr_len = min_t(u32, PAGE_SIZE - page_base, len); 258 if (seg[n].mr_len > PAGE_SIZE) 259 goto out_overflow; 260 len -= seg[n].mr_len; 261 ++n; 262 ++p; 263 page_base = 0; /* page offset only applies to first page */ 264 } 265 266 /* Message overflows the seg array */ 267 if (len && n == RPCRDMA_MAX_SEGS) 268 goto out_overflow; 269 270 /* When encoding the read list, the tail is always sent inline */ 271 if (type == rpcrdma_readch) 272 return n; 273 274 if (xdrbuf->tail[0].iov_len) { 275 /* the rpcrdma protocol allows us to omit any trailing 276 * xdr pad bytes, saving the server an RDMA operation. */ 277 if (xdrbuf->tail[0].iov_len < 4 && xprt_rdma_pad_optimize) 278 return n; 279 n = rpcrdma_convert_kvec(&xdrbuf->tail[0], seg, n); 280 if (n == RPCRDMA_MAX_SEGS) 281 goto out_overflow; 282 } 283 284 return n; 285 286 out_overflow: 287 pr_err("rpcrdma: segment array overflow\n"); 288 return -EIO; 289 } 290 291 static inline __be32 * 292 xdr_encode_rdma_segment(__be32 *iptr, struct rpcrdma_mw *mw) 293 { 294 *iptr++ = cpu_to_be32(mw->mw_handle); 295 *iptr++ = cpu_to_be32(mw->mw_length); 296 return xdr_encode_hyper(iptr, mw->mw_offset); 297 } 298 299 /* XDR-encode the Read list. Supports encoding a list of read 300 * segments that belong to a single read chunk. 301 * 302 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64): 303 * 304 * Read chunklist (a linked list): 305 * N elements, position P (same P for all chunks of same arg!): 306 * 1 - PHLOO - 1 - PHLOO - ... - 1 - PHLOO - 0 307 * 308 * Returns a pointer to the XDR word in the RDMA header following 309 * the end of the Read list, or an error pointer. 310 */ 311 static __be32 * 312 rpcrdma_encode_read_list(struct rpcrdma_xprt *r_xprt, 313 struct rpcrdma_req *req, struct rpc_rqst *rqst, 314 __be32 *iptr, enum rpcrdma_chunktype rtype) 315 { 316 struct rpcrdma_mr_seg *seg; 317 struct rpcrdma_mw *mw; 318 unsigned int pos; 319 int n, nsegs; 320 321 if (rtype == rpcrdma_noch) { 322 *iptr++ = xdr_zero; /* item not present */ 323 return iptr; 324 } 325 326 pos = rqst->rq_snd_buf.head[0].iov_len; 327 if (rtype == rpcrdma_areadch) 328 pos = 0; 329 seg = req->rl_segments; 330 nsegs = rpcrdma_convert_iovs(&rqst->rq_snd_buf, pos, rtype, seg); 331 if (nsegs < 0) 332 return ERR_PTR(nsegs); 333 334 do { 335 n = r_xprt->rx_ia.ri_ops->ro_map(r_xprt, seg, nsegs, 336 false, &mw); 337 if (n < 0) 338 return ERR_PTR(n); 339 list_add(&mw->mw_list, &req->rl_registered); 340 341 *iptr++ = xdr_one; /* item present */ 342 343 /* All read segments in this chunk 344 * have the same "position". 345 */ 346 *iptr++ = cpu_to_be32(pos); 347 iptr = xdr_encode_rdma_segment(iptr, mw); 348 349 dprintk("RPC: %5u %s: pos %u %u@0x%016llx:0x%08x (%s)\n", 350 rqst->rq_task->tk_pid, __func__, pos, 351 mw->mw_length, (unsigned long long)mw->mw_offset, 352 mw->mw_handle, n < nsegs ? "more" : "last"); 353 354 r_xprt->rx_stats.read_chunk_count++; 355 seg += n; 356 nsegs -= n; 357 } while (nsegs); 358 359 /* Finish Read list */ 360 *iptr++ = xdr_zero; /* Next item not present */ 361 return iptr; 362 } 363 364 /* XDR-encode the Write list. Supports encoding a list containing 365 * one array of plain segments that belong to a single write chunk. 366 * 367 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64): 368 * 369 * Write chunklist (a list of (one) counted array): 370 * N elements: 371 * 1 - N - HLOO - HLOO - ... - HLOO - 0 372 * 373 * Returns a pointer to the XDR word in the RDMA header following 374 * the end of the Write list, or an error pointer. 375 */ 376 static __be32 * 377 rpcrdma_encode_write_list(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req, 378 struct rpc_rqst *rqst, __be32 *iptr, 379 enum rpcrdma_chunktype wtype) 380 { 381 struct rpcrdma_mr_seg *seg; 382 struct rpcrdma_mw *mw; 383 int n, nsegs, nchunks; 384 __be32 *segcount; 385 386 if (wtype != rpcrdma_writech) { 387 *iptr++ = xdr_zero; /* no Write list present */ 388 return iptr; 389 } 390 391 seg = req->rl_segments; 392 nsegs = rpcrdma_convert_iovs(&rqst->rq_rcv_buf, 393 rqst->rq_rcv_buf.head[0].iov_len, 394 wtype, seg); 395 if (nsegs < 0) 396 return ERR_PTR(nsegs); 397 398 *iptr++ = xdr_one; /* Write list present */ 399 segcount = iptr++; /* save location of segment count */ 400 401 nchunks = 0; 402 do { 403 n = r_xprt->rx_ia.ri_ops->ro_map(r_xprt, seg, nsegs, 404 true, &mw); 405 if (n < 0) 406 return ERR_PTR(n); 407 list_add(&mw->mw_list, &req->rl_registered); 408 409 iptr = xdr_encode_rdma_segment(iptr, mw); 410 411 dprintk("RPC: %5u %s: %u@0x016%llx:0x%08x (%s)\n", 412 rqst->rq_task->tk_pid, __func__, 413 mw->mw_length, (unsigned long long)mw->mw_offset, 414 mw->mw_handle, n < nsegs ? "more" : "last"); 415 416 r_xprt->rx_stats.write_chunk_count++; 417 r_xprt->rx_stats.total_rdma_request += seg->mr_len; 418 nchunks++; 419 seg += n; 420 nsegs -= n; 421 } while (nsegs); 422 423 /* Update count of segments in this Write chunk */ 424 *segcount = cpu_to_be32(nchunks); 425 426 /* Finish Write list */ 427 *iptr++ = xdr_zero; /* Next item not present */ 428 return iptr; 429 } 430 431 /* XDR-encode the Reply chunk. Supports encoding an array of plain 432 * segments that belong to a single write (reply) chunk. 433 * 434 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64): 435 * 436 * Reply chunk (a counted array): 437 * N elements: 438 * 1 - N - HLOO - HLOO - ... - HLOO 439 * 440 * Returns a pointer to the XDR word in the RDMA header following 441 * the end of the Reply chunk, or an error pointer. 442 */ 443 static __be32 * 444 rpcrdma_encode_reply_chunk(struct rpcrdma_xprt *r_xprt, 445 struct rpcrdma_req *req, struct rpc_rqst *rqst, 446 __be32 *iptr, enum rpcrdma_chunktype wtype) 447 { 448 struct rpcrdma_mr_seg *seg; 449 struct rpcrdma_mw *mw; 450 int n, nsegs, nchunks; 451 __be32 *segcount; 452 453 if (wtype != rpcrdma_replych) { 454 *iptr++ = xdr_zero; /* no Reply chunk present */ 455 return iptr; 456 } 457 458 seg = req->rl_segments; 459 nsegs = rpcrdma_convert_iovs(&rqst->rq_rcv_buf, 0, wtype, seg); 460 if (nsegs < 0) 461 return ERR_PTR(nsegs); 462 463 *iptr++ = xdr_one; /* Reply chunk present */ 464 segcount = iptr++; /* save location of segment count */ 465 466 nchunks = 0; 467 do { 468 n = r_xprt->rx_ia.ri_ops->ro_map(r_xprt, seg, nsegs, 469 true, &mw); 470 if (n < 0) 471 return ERR_PTR(n); 472 list_add(&mw->mw_list, &req->rl_registered); 473 474 iptr = xdr_encode_rdma_segment(iptr, mw); 475 476 dprintk("RPC: %5u %s: %u@0x%016llx:0x%08x (%s)\n", 477 rqst->rq_task->tk_pid, __func__, 478 mw->mw_length, (unsigned long long)mw->mw_offset, 479 mw->mw_handle, n < nsegs ? "more" : "last"); 480 481 r_xprt->rx_stats.reply_chunk_count++; 482 r_xprt->rx_stats.total_rdma_request += seg->mr_len; 483 nchunks++; 484 seg += n; 485 nsegs -= n; 486 } while (nsegs); 487 488 /* Update count of segments in the Reply chunk */ 489 *segcount = cpu_to_be32(nchunks); 490 491 return iptr; 492 } 493 494 /* 495 * Copy write data inline. 496 * This function is used for "small" requests. Data which is passed 497 * to RPC via iovecs (or page list) is copied directly into the 498 * pre-registered memory buffer for this request. For small amounts 499 * of data, this is efficient. The cutoff value is tunable. 500 */ 501 static void rpcrdma_inline_pullup(struct rpc_rqst *rqst) 502 { 503 int i, npages, curlen; 504 int copy_len; 505 unsigned char *srcp, *destp; 506 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt); 507 int page_base; 508 struct page **ppages; 509 510 destp = rqst->rq_svec[0].iov_base; 511 curlen = rqst->rq_svec[0].iov_len; 512 destp += curlen; 513 514 dprintk("RPC: %s: destp 0x%p len %d hdrlen %d\n", 515 __func__, destp, rqst->rq_slen, curlen); 516 517 copy_len = rqst->rq_snd_buf.page_len; 518 519 if (rqst->rq_snd_buf.tail[0].iov_len) { 520 curlen = rqst->rq_snd_buf.tail[0].iov_len; 521 if (destp + copy_len != rqst->rq_snd_buf.tail[0].iov_base) { 522 memmove(destp + copy_len, 523 rqst->rq_snd_buf.tail[0].iov_base, curlen); 524 r_xprt->rx_stats.pullup_copy_count += curlen; 525 } 526 dprintk("RPC: %s: tail destp 0x%p len %d\n", 527 __func__, destp + copy_len, curlen); 528 rqst->rq_svec[0].iov_len += curlen; 529 } 530 r_xprt->rx_stats.pullup_copy_count += copy_len; 531 532 page_base = rqst->rq_snd_buf.page_base; 533 ppages = rqst->rq_snd_buf.pages + (page_base >> PAGE_SHIFT); 534 page_base &= ~PAGE_MASK; 535 npages = PAGE_ALIGN(page_base+copy_len) >> PAGE_SHIFT; 536 for (i = 0; copy_len && i < npages; i++) { 537 curlen = PAGE_SIZE - page_base; 538 if (curlen > copy_len) 539 curlen = copy_len; 540 dprintk("RPC: %s: page %d destp 0x%p len %d curlen %d\n", 541 __func__, i, destp, copy_len, curlen); 542 srcp = kmap_atomic(ppages[i]); 543 memcpy(destp, srcp+page_base, curlen); 544 kunmap_atomic(srcp); 545 rqst->rq_svec[0].iov_len += curlen; 546 destp += curlen; 547 copy_len -= curlen; 548 page_base = 0; 549 } 550 /* header now contains entire send message */ 551 } 552 553 /* 554 * Marshal a request: the primary job of this routine is to choose 555 * the transfer modes. See comments below. 556 * 557 * Prepares up to two IOVs per Call message: 558 * 559 * [0] -- RPC RDMA header 560 * [1] -- the RPC header/data 561 * 562 * Returns zero on success, otherwise a negative errno. 563 */ 564 565 int 566 rpcrdma_marshal_req(struct rpc_rqst *rqst) 567 { 568 struct rpc_xprt *xprt = rqst->rq_xprt; 569 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 570 struct rpcrdma_req *req = rpcr_to_rdmar(rqst); 571 enum rpcrdma_chunktype rtype, wtype; 572 struct rpcrdma_msg *headerp; 573 bool ddp_allowed; 574 ssize_t hdrlen; 575 size_t rpclen; 576 __be32 *iptr; 577 578 #if defined(CONFIG_SUNRPC_BACKCHANNEL) 579 if (test_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state)) 580 return rpcrdma_bc_marshal_reply(rqst); 581 #endif 582 583 headerp = rdmab_to_msg(req->rl_rdmabuf); 584 /* don't byte-swap XID, it's already done in request */ 585 headerp->rm_xid = rqst->rq_xid; 586 headerp->rm_vers = rpcrdma_version; 587 headerp->rm_credit = cpu_to_be32(r_xprt->rx_buf.rb_max_requests); 588 headerp->rm_type = rdma_msg; 589 590 /* When the ULP employs a GSS flavor that guarantees integrity 591 * or privacy, direct data placement of individual data items 592 * is not allowed. 593 */ 594 ddp_allowed = !(rqst->rq_cred->cr_auth->au_flags & 595 RPCAUTH_AUTH_DATATOUCH); 596 597 /* 598 * Chunks needed for results? 599 * 600 * o If the expected result is under the inline threshold, all ops 601 * return as inline. 602 * o Large read ops return data as write chunk(s), header as 603 * inline. 604 * o Large non-read ops return as a single reply chunk. 605 */ 606 if (rpcrdma_results_inline(r_xprt, rqst)) 607 wtype = rpcrdma_noch; 608 else if (ddp_allowed && rqst->rq_rcv_buf.flags & XDRBUF_READ) 609 wtype = rpcrdma_writech; 610 else 611 wtype = rpcrdma_replych; 612 613 /* 614 * Chunks needed for arguments? 615 * 616 * o If the total request is under the inline threshold, all ops 617 * are sent as inline. 618 * o Large write ops transmit data as read chunk(s), header as 619 * inline. 620 * o Large non-write ops are sent with the entire message as a 621 * single read chunk (protocol 0-position special case). 622 * 623 * This assumes that the upper layer does not present a request 624 * that both has a data payload, and whose non-data arguments 625 * by themselves are larger than the inline threshold. 626 */ 627 if (rpcrdma_args_inline(r_xprt, rqst)) { 628 rtype = rpcrdma_noch; 629 rpcrdma_inline_pullup(rqst); 630 rpclen = rqst->rq_svec[0].iov_len; 631 } else if (ddp_allowed && rqst->rq_snd_buf.flags & XDRBUF_WRITE) { 632 rtype = rpcrdma_readch; 633 rpclen = rqst->rq_svec[0].iov_len; 634 rpclen += rpcrdma_tail_pullup(&rqst->rq_snd_buf); 635 } else { 636 r_xprt->rx_stats.nomsg_call_count++; 637 headerp->rm_type = htonl(RDMA_NOMSG); 638 rtype = rpcrdma_areadch; 639 rpclen = 0; 640 } 641 642 /* This implementation supports the following combinations 643 * of chunk lists in one RPC-over-RDMA Call message: 644 * 645 * - Read list 646 * - Write list 647 * - Reply chunk 648 * - Read list + Reply chunk 649 * 650 * It might not yet support the following combinations: 651 * 652 * - Read list + Write list 653 * 654 * It does not support the following combinations: 655 * 656 * - Write list + Reply chunk 657 * - Read list + Write list + Reply chunk 658 * 659 * This implementation supports only a single chunk in each 660 * Read or Write list. Thus for example the client cannot 661 * send a Call message with a Position Zero Read chunk and a 662 * regular Read chunk at the same time. 663 */ 664 iptr = headerp->rm_body.rm_chunks; 665 iptr = rpcrdma_encode_read_list(r_xprt, req, rqst, iptr, rtype); 666 if (IS_ERR(iptr)) 667 goto out_unmap; 668 iptr = rpcrdma_encode_write_list(r_xprt, req, rqst, iptr, wtype); 669 if (IS_ERR(iptr)) 670 goto out_unmap; 671 iptr = rpcrdma_encode_reply_chunk(r_xprt, req, rqst, iptr, wtype); 672 if (IS_ERR(iptr)) 673 goto out_unmap; 674 hdrlen = (unsigned char *)iptr - (unsigned char *)headerp; 675 676 if (hdrlen + rpclen > RPCRDMA_INLINE_WRITE_THRESHOLD(rqst)) 677 goto out_overflow; 678 679 dprintk("RPC: %5u %s: %s/%s: hdrlen %zd rpclen %zd\n", 680 rqst->rq_task->tk_pid, __func__, 681 transfertypes[rtype], transfertypes[wtype], 682 hdrlen, rpclen); 683 684 req->rl_send_iov[0].addr = rdmab_addr(req->rl_rdmabuf); 685 req->rl_send_iov[0].length = hdrlen; 686 req->rl_send_iov[0].lkey = rdmab_lkey(req->rl_rdmabuf); 687 688 req->rl_niovs = 1; 689 if (rtype == rpcrdma_areadch) 690 return 0; 691 692 req->rl_send_iov[1].addr = rdmab_addr(req->rl_sendbuf); 693 req->rl_send_iov[1].length = rpclen; 694 req->rl_send_iov[1].lkey = rdmab_lkey(req->rl_sendbuf); 695 696 req->rl_niovs = 2; 697 return 0; 698 699 out_overflow: 700 pr_err("rpcrdma: send overflow: hdrlen %zd rpclen %zu %s/%s\n", 701 hdrlen, rpclen, transfertypes[rtype], transfertypes[wtype]); 702 iptr = ERR_PTR(-EIO); 703 704 out_unmap: 705 r_xprt->rx_ia.ri_ops->ro_unmap_safe(r_xprt, req, false); 706 return PTR_ERR(iptr); 707 } 708 709 /* 710 * Chase down a received write or reply chunklist to get length 711 * RDMA'd by server. See map at rpcrdma_create_chunks()! :-) 712 */ 713 static int 714 rpcrdma_count_chunks(struct rpcrdma_rep *rep, int wrchunk, __be32 **iptrp) 715 { 716 unsigned int i, total_len; 717 struct rpcrdma_write_chunk *cur_wchunk; 718 char *base = (char *)rdmab_to_msg(rep->rr_rdmabuf); 719 720 i = be32_to_cpu(**iptrp); 721 cur_wchunk = (struct rpcrdma_write_chunk *) (*iptrp + 1); 722 total_len = 0; 723 while (i--) { 724 struct rpcrdma_segment *seg = &cur_wchunk->wc_target; 725 ifdebug(FACILITY) { 726 u64 off; 727 xdr_decode_hyper((__be32 *)&seg->rs_offset, &off); 728 dprintk("RPC: %s: chunk %d@0x%llx:0x%x\n", 729 __func__, 730 be32_to_cpu(seg->rs_length), 731 (unsigned long long)off, 732 be32_to_cpu(seg->rs_handle)); 733 } 734 total_len += be32_to_cpu(seg->rs_length); 735 ++cur_wchunk; 736 } 737 /* check and adjust for properly terminated write chunk */ 738 if (wrchunk) { 739 __be32 *w = (__be32 *) cur_wchunk; 740 if (*w++ != xdr_zero) 741 return -1; 742 cur_wchunk = (struct rpcrdma_write_chunk *) w; 743 } 744 if ((char *)cur_wchunk > base + rep->rr_len) 745 return -1; 746 747 *iptrp = (__be32 *) cur_wchunk; 748 return total_len; 749 } 750 751 /** 752 * rpcrdma_inline_fixup - Scatter inline received data into rqst's iovecs 753 * @rqst: controlling RPC request 754 * @srcp: points to RPC message payload in receive buffer 755 * @copy_len: remaining length of receive buffer content 756 * @pad: Write chunk pad bytes needed (zero for pure inline) 757 * 758 * The upper layer has set the maximum number of bytes it can 759 * receive in each component of rq_rcv_buf. These values are set in 760 * the head.iov_len, page_len, tail.iov_len, and buflen fields. 761 * 762 * Unlike the TCP equivalent (xdr_partial_copy_from_skb), in 763 * many cases this function simply updates iov_base pointers in 764 * rq_rcv_buf to point directly to the received reply data, to 765 * avoid copying reply data. 766 * 767 * Returns the count of bytes which had to be memcopied. 768 */ 769 static unsigned long 770 rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad) 771 { 772 unsigned long fixup_copy_count; 773 int i, npages, curlen; 774 char *destp; 775 struct page **ppages; 776 int page_base; 777 778 /* The head iovec is redirected to the RPC reply message 779 * in the receive buffer, to avoid a memcopy. 780 */ 781 rqst->rq_rcv_buf.head[0].iov_base = srcp; 782 rqst->rq_private_buf.head[0].iov_base = srcp; 783 784 /* The contents of the receive buffer that follow 785 * head.iov_len bytes are copied into the page list. 786 */ 787 curlen = rqst->rq_rcv_buf.head[0].iov_len; 788 if (curlen > copy_len) 789 curlen = copy_len; 790 dprintk("RPC: %s: srcp 0x%p len %d hdrlen %d\n", 791 __func__, srcp, copy_len, curlen); 792 srcp += curlen; 793 copy_len -= curlen; 794 795 page_base = rqst->rq_rcv_buf.page_base; 796 ppages = rqst->rq_rcv_buf.pages + (page_base >> PAGE_SHIFT); 797 page_base &= ~PAGE_MASK; 798 fixup_copy_count = 0; 799 if (copy_len && rqst->rq_rcv_buf.page_len) { 800 int pagelist_len; 801 802 pagelist_len = rqst->rq_rcv_buf.page_len; 803 if (pagelist_len > copy_len) 804 pagelist_len = copy_len; 805 npages = PAGE_ALIGN(page_base + pagelist_len) >> PAGE_SHIFT; 806 for (i = 0; i < npages; i++) { 807 curlen = PAGE_SIZE - page_base; 808 if (curlen > pagelist_len) 809 curlen = pagelist_len; 810 811 dprintk("RPC: %s: page %d" 812 " srcp 0x%p len %d curlen %d\n", 813 __func__, i, srcp, copy_len, curlen); 814 destp = kmap_atomic(ppages[i]); 815 memcpy(destp + page_base, srcp, curlen); 816 flush_dcache_page(ppages[i]); 817 kunmap_atomic(destp); 818 srcp += curlen; 819 copy_len -= curlen; 820 fixup_copy_count += curlen; 821 pagelist_len -= curlen; 822 if (!pagelist_len) 823 break; 824 page_base = 0; 825 } 826 827 /* Implicit padding for the last segment in a Write 828 * chunk is inserted inline at the front of the tail 829 * iovec. The upper layer ignores the content of 830 * the pad. Simply ensure inline content in the tail 831 * that follows the Write chunk is properly aligned. 832 */ 833 if (pad) 834 srcp -= pad; 835 } 836 837 /* The tail iovec is redirected to the remaining data 838 * in the receive buffer, to avoid a memcopy. 839 */ 840 if (copy_len || pad) { 841 rqst->rq_rcv_buf.tail[0].iov_base = srcp; 842 rqst->rq_private_buf.tail[0].iov_base = srcp; 843 } 844 845 return fixup_copy_count; 846 } 847 848 void 849 rpcrdma_connect_worker(struct work_struct *work) 850 { 851 struct rpcrdma_ep *ep = 852 container_of(work, struct rpcrdma_ep, rep_connect_worker.work); 853 struct rpcrdma_xprt *r_xprt = 854 container_of(ep, struct rpcrdma_xprt, rx_ep); 855 struct rpc_xprt *xprt = &r_xprt->rx_xprt; 856 857 spin_lock_bh(&xprt->transport_lock); 858 if (++xprt->connect_cookie == 0) /* maintain a reserved value */ 859 ++xprt->connect_cookie; 860 if (ep->rep_connected > 0) { 861 if (!xprt_test_and_set_connected(xprt)) 862 xprt_wake_pending_tasks(xprt, 0); 863 } else { 864 if (xprt_test_and_clear_connected(xprt)) 865 xprt_wake_pending_tasks(xprt, -ENOTCONN); 866 } 867 spin_unlock_bh(&xprt->transport_lock); 868 } 869 870 #if defined(CONFIG_SUNRPC_BACKCHANNEL) 871 /* By convention, backchannel calls arrive via rdma_msg type 872 * messages, and never populate the chunk lists. This makes 873 * the RPC/RDMA header small and fixed in size, so it is 874 * straightforward to check the RPC header's direction field. 875 */ 876 static bool 877 rpcrdma_is_bcall(struct rpcrdma_msg *headerp) 878 { 879 __be32 *p = (__be32 *)headerp; 880 881 if (headerp->rm_type != rdma_msg) 882 return false; 883 if (headerp->rm_body.rm_chunks[0] != xdr_zero) 884 return false; 885 if (headerp->rm_body.rm_chunks[1] != xdr_zero) 886 return false; 887 if (headerp->rm_body.rm_chunks[2] != xdr_zero) 888 return false; 889 890 /* sanity */ 891 if (p[7] != headerp->rm_xid) 892 return false; 893 /* call direction */ 894 if (p[8] != cpu_to_be32(RPC_CALL)) 895 return false; 896 897 return true; 898 } 899 #endif /* CONFIG_SUNRPC_BACKCHANNEL */ 900 901 /* 902 * This function is called when an async event is posted to 903 * the connection which changes the connection state. All it 904 * does at this point is mark the connection up/down, the rpc 905 * timers do the rest. 906 */ 907 void 908 rpcrdma_conn_func(struct rpcrdma_ep *ep) 909 { 910 schedule_delayed_work(&ep->rep_connect_worker, 0); 911 } 912 913 /* Process received RPC/RDMA messages. 914 * 915 * Errors must result in the RPC task either being awakened, or 916 * allowed to timeout, to discover the errors at that time. 917 */ 918 void 919 rpcrdma_reply_handler(struct rpcrdma_rep *rep) 920 { 921 struct rpcrdma_msg *headerp; 922 struct rpcrdma_req *req; 923 struct rpc_rqst *rqst; 924 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt; 925 struct rpc_xprt *xprt = &r_xprt->rx_xprt; 926 __be32 *iptr; 927 int rdmalen, status, rmerr; 928 unsigned long cwnd; 929 930 dprintk("RPC: %s: incoming rep %p\n", __func__, rep); 931 932 if (rep->rr_len == RPCRDMA_BAD_LEN) 933 goto out_badstatus; 934 if (rep->rr_len < RPCRDMA_HDRLEN_ERR) 935 goto out_shortreply; 936 937 headerp = rdmab_to_msg(rep->rr_rdmabuf); 938 #if defined(CONFIG_SUNRPC_BACKCHANNEL) 939 if (rpcrdma_is_bcall(headerp)) 940 goto out_bcall; 941 #endif 942 943 /* Match incoming rpcrdma_rep to an rpcrdma_req to 944 * get context for handling any incoming chunks. 945 */ 946 spin_lock_bh(&xprt->transport_lock); 947 rqst = xprt_lookup_rqst(xprt, headerp->rm_xid); 948 if (!rqst) 949 goto out_nomatch; 950 951 req = rpcr_to_rdmar(rqst); 952 if (req->rl_reply) 953 goto out_duplicate; 954 955 /* Sanity checking has passed. We are now committed 956 * to complete this transaction. 957 */ 958 list_del_init(&rqst->rq_list); 959 spin_unlock_bh(&xprt->transport_lock); 960 dprintk("RPC: %s: reply %p completes request %p (xid 0x%08x)\n", 961 __func__, rep, req, be32_to_cpu(headerp->rm_xid)); 962 963 /* from here on, the reply is no longer an orphan */ 964 req->rl_reply = rep; 965 xprt->reestablish_timeout = 0; 966 967 if (headerp->rm_vers != rpcrdma_version) 968 goto out_badversion; 969 970 /* check for expected message types */ 971 /* The order of some of these tests is important. */ 972 switch (headerp->rm_type) { 973 case rdma_msg: 974 /* never expect read chunks */ 975 /* never expect reply chunks (two ways to check) */ 976 /* never expect write chunks without having offered RDMA */ 977 if (headerp->rm_body.rm_chunks[0] != xdr_zero || 978 (headerp->rm_body.rm_chunks[1] == xdr_zero && 979 headerp->rm_body.rm_chunks[2] != xdr_zero) || 980 (headerp->rm_body.rm_chunks[1] != xdr_zero && 981 list_empty(&req->rl_registered))) 982 goto badheader; 983 if (headerp->rm_body.rm_chunks[1] != xdr_zero) { 984 /* count any expected write chunks in read reply */ 985 /* start at write chunk array count */ 986 iptr = &headerp->rm_body.rm_chunks[2]; 987 rdmalen = rpcrdma_count_chunks(rep, 1, &iptr); 988 /* check for validity, and no reply chunk after */ 989 if (rdmalen < 0 || *iptr++ != xdr_zero) 990 goto badheader; 991 rep->rr_len -= 992 ((unsigned char *)iptr - (unsigned char *)headerp); 993 status = rep->rr_len + rdmalen; 994 r_xprt->rx_stats.total_rdma_reply += rdmalen; 995 /* special case - last chunk may omit padding */ 996 if (rdmalen &= 3) { 997 rdmalen = 4 - rdmalen; 998 status += rdmalen; 999 } 1000 } else { 1001 /* else ordinary inline */ 1002 rdmalen = 0; 1003 iptr = (__be32 *)((unsigned char *)headerp + 1004 RPCRDMA_HDRLEN_MIN); 1005 rep->rr_len -= RPCRDMA_HDRLEN_MIN; 1006 status = rep->rr_len; 1007 } 1008 1009 r_xprt->rx_stats.fixup_copy_count += 1010 rpcrdma_inline_fixup(rqst, (char *)iptr, rep->rr_len, 1011 rdmalen); 1012 break; 1013 1014 case rdma_nomsg: 1015 /* never expect read or write chunks, always reply chunks */ 1016 if (headerp->rm_body.rm_chunks[0] != xdr_zero || 1017 headerp->rm_body.rm_chunks[1] != xdr_zero || 1018 headerp->rm_body.rm_chunks[2] != xdr_one || 1019 list_empty(&req->rl_registered)) 1020 goto badheader; 1021 iptr = (__be32 *)((unsigned char *)headerp + 1022 RPCRDMA_HDRLEN_MIN); 1023 rdmalen = rpcrdma_count_chunks(rep, 0, &iptr); 1024 if (rdmalen < 0) 1025 goto badheader; 1026 r_xprt->rx_stats.total_rdma_reply += rdmalen; 1027 /* Reply chunk buffer already is the reply vector - no fixup. */ 1028 status = rdmalen; 1029 break; 1030 1031 case rdma_error: 1032 goto out_rdmaerr; 1033 1034 badheader: 1035 default: 1036 dprintk("RPC: %5u %s: invalid rpcrdma reply (type %u)\n", 1037 rqst->rq_task->tk_pid, __func__, 1038 be32_to_cpu(headerp->rm_type)); 1039 status = -EIO; 1040 r_xprt->rx_stats.bad_reply_count++; 1041 break; 1042 } 1043 1044 out: 1045 /* Invalidate and flush the data payloads before waking the 1046 * waiting application. This guarantees the memory region is 1047 * properly fenced from the server before the application 1048 * accesses the data. It also ensures proper send flow 1049 * control: waking the next RPC waits until this RPC has 1050 * relinquished all its Send Queue entries. 1051 */ 1052 if (!list_empty(&req->rl_registered)) 1053 r_xprt->rx_ia.ri_ops->ro_unmap_sync(r_xprt, req); 1054 1055 spin_lock_bh(&xprt->transport_lock); 1056 cwnd = xprt->cwnd; 1057 xprt->cwnd = atomic_read(&r_xprt->rx_buf.rb_credits) << RPC_CWNDSHIFT; 1058 if (xprt->cwnd > cwnd) 1059 xprt_release_rqst_cong(rqst->rq_task); 1060 1061 xprt_complete_rqst(rqst->rq_task, status); 1062 spin_unlock_bh(&xprt->transport_lock); 1063 dprintk("RPC: %s: xprt_complete_rqst(0x%p, 0x%p, %d)\n", 1064 __func__, xprt, rqst, status); 1065 return; 1066 1067 out_badstatus: 1068 rpcrdma_recv_buffer_put(rep); 1069 if (r_xprt->rx_ep.rep_connected == 1) { 1070 r_xprt->rx_ep.rep_connected = -EIO; 1071 rpcrdma_conn_func(&r_xprt->rx_ep); 1072 } 1073 return; 1074 1075 #if defined(CONFIG_SUNRPC_BACKCHANNEL) 1076 out_bcall: 1077 rpcrdma_bc_receive_call(r_xprt, rep); 1078 return; 1079 #endif 1080 1081 /* If the incoming reply terminated a pending RPC, the next 1082 * RPC call will post a replacement receive buffer as it is 1083 * being marshaled. 1084 */ 1085 out_badversion: 1086 dprintk("RPC: %s: invalid version %d\n", 1087 __func__, be32_to_cpu(headerp->rm_vers)); 1088 status = -EIO; 1089 r_xprt->rx_stats.bad_reply_count++; 1090 goto out; 1091 1092 out_rdmaerr: 1093 rmerr = be32_to_cpu(headerp->rm_body.rm_error.rm_err); 1094 switch (rmerr) { 1095 case ERR_VERS: 1096 pr_err("%s: server reports header version error (%u-%u)\n", 1097 __func__, 1098 be32_to_cpu(headerp->rm_body.rm_error.rm_vers_low), 1099 be32_to_cpu(headerp->rm_body.rm_error.rm_vers_high)); 1100 break; 1101 case ERR_CHUNK: 1102 pr_err("%s: server reports header decoding error\n", 1103 __func__); 1104 break; 1105 default: 1106 pr_err("%s: server reports unknown error %d\n", 1107 __func__, rmerr); 1108 } 1109 status = -EREMOTEIO; 1110 r_xprt->rx_stats.bad_reply_count++; 1111 goto out; 1112 1113 /* If no pending RPC transaction was matched, post a replacement 1114 * receive buffer before returning. 1115 */ 1116 out_shortreply: 1117 dprintk("RPC: %s: short/invalid reply\n", __func__); 1118 goto repost; 1119 1120 out_nomatch: 1121 spin_unlock_bh(&xprt->transport_lock); 1122 dprintk("RPC: %s: no match for incoming xid 0x%08x len %d\n", 1123 __func__, be32_to_cpu(headerp->rm_xid), 1124 rep->rr_len); 1125 goto repost; 1126 1127 out_duplicate: 1128 spin_unlock_bh(&xprt->transport_lock); 1129 dprintk("RPC: %s: " 1130 "duplicate reply %p to RPC request %p: xid 0x%08x\n", 1131 __func__, rep, req, be32_to_cpu(headerp->rm_xid)); 1132 1133 repost: 1134 r_xprt->rx_stats.bad_reply_count++; 1135 if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, &r_xprt->rx_ep, rep)) 1136 rpcrdma_recv_buffer_put(rep); 1137 } 1138