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 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) 65 static const char transfertypes[][12] = { 66 "pure inline", /* no chunks */ 67 " read chunk", /* some argument via rdma read */ 68 "*read chunk", /* entire request via rdma read */ 69 "write chunk", /* some result via rdma write */ 70 "reply chunk" /* entire reply via rdma write */ 71 }; 72 #endif 73 74 /* The client can send a request inline as long as the RPCRDMA header 75 * plus the RPC call fit under the transport's inline limit. If the 76 * combined call message size exceeds that limit, the client must use 77 * the read chunk list for this operation. 78 */ 79 static bool rpcrdma_args_inline(struct rpc_rqst *rqst) 80 { 81 unsigned int callsize = RPCRDMA_HDRLEN_MIN + rqst->rq_snd_buf.len; 82 83 return callsize <= RPCRDMA_INLINE_WRITE_THRESHOLD(rqst); 84 } 85 86 /* The client can't know how large the actual reply will be. Thus it 87 * plans for the largest possible reply for that particular ULP 88 * operation. If the maximum combined reply message size exceeds that 89 * limit, the client must provide a write list or a reply chunk for 90 * this request. 91 */ 92 static bool rpcrdma_results_inline(struct rpc_rqst *rqst) 93 { 94 unsigned int repsize = RPCRDMA_HDRLEN_MIN + rqst->rq_rcv_buf.buflen; 95 96 return repsize <= RPCRDMA_INLINE_READ_THRESHOLD(rqst); 97 } 98 99 static int 100 rpcrdma_tail_pullup(struct xdr_buf *buf) 101 { 102 size_t tlen = buf->tail[0].iov_len; 103 size_t skip = tlen & 3; 104 105 /* Do not include the tail if it is only an XDR pad */ 106 if (tlen < 4) 107 return 0; 108 109 /* xdr_write_pages() adds a pad at the beginning of the tail 110 * if the content in "buf->pages" is unaligned. Force the 111 * tail's actual content to land at the next XDR position 112 * after the head instead. 113 */ 114 if (skip) { 115 unsigned char *src, *dst; 116 unsigned int count; 117 118 src = buf->tail[0].iov_base; 119 dst = buf->head[0].iov_base; 120 dst += buf->head[0].iov_len; 121 122 src += skip; 123 tlen -= skip; 124 125 dprintk("RPC: %s: skip=%zu, memmove(%p, %p, %zu)\n", 126 __func__, skip, dst, src, tlen); 127 128 for (count = tlen; count; count--) 129 *dst++ = *src++; 130 } 131 132 return tlen; 133 } 134 135 /* 136 * Chunk assembly from upper layer xdr_buf. 137 * 138 * Prepare the passed-in xdr_buf into representation as RPC/RDMA chunk 139 * elements. Segments are then coalesced when registered, if possible 140 * within the selected memreg mode. 141 * 142 * Returns positive number of segments converted, or a negative errno. 143 */ 144 145 static int 146 rpcrdma_convert_iovs(struct xdr_buf *xdrbuf, unsigned int pos, 147 enum rpcrdma_chunktype type, struct rpcrdma_mr_seg *seg, int nsegs) 148 { 149 int len, n = 0, p; 150 int page_base; 151 struct page **ppages; 152 153 if (pos == 0 && xdrbuf->head[0].iov_len) { 154 seg[n].mr_page = NULL; 155 seg[n].mr_offset = xdrbuf->head[0].iov_base; 156 seg[n].mr_len = xdrbuf->head[0].iov_len; 157 ++n; 158 } 159 160 len = xdrbuf->page_len; 161 ppages = xdrbuf->pages + (xdrbuf->page_base >> PAGE_SHIFT); 162 page_base = xdrbuf->page_base & ~PAGE_MASK; 163 p = 0; 164 while (len && n < nsegs) { 165 if (!ppages[p]) { 166 /* alloc the pagelist for receiving buffer */ 167 ppages[p] = alloc_page(GFP_ATOMIC); 168 if (!ppages[p]) 169 return -ENOMEM; 170 } 171 seg[n].mr_page = ppages[p]; 172 seg[n].mr_offset = (void *)(unsigned long) page_base; 173 seg[n].mr_len = min_t(u32, PAGE_SIZE - page_base, len); 174 if (seg[n].mr_len > PAGE_SIZE) 175 return -EIO; 176 len -= seg[n].mr_len; 177 ++n; 178 ++p; 179 page_base = 0; /* page offset only applies to first page */ 180 } 181 182 /* Message overflows the seg array */ 183 if (len && n == nsegs) 184 return -EIO; 185 186 /* When encoding the read list, the tail is always sent inline */ 187 if (type == rpcrdma_readch) 188 return n; 189 190 if (xdrbuf->tail[0].iov_len) { 191 /* the rpcrdma protocol allows us to omit any trailing 192 * xdr pad bytes, saving the server an RDMA operation. */ 193 if (xdrbuf->tail[0].iov_len < 4 && xprt_rdma_pad_optimize) 194 return n; 195 if (n == nsegs) 196 /* Tail remains, but we're out of segments */ 197 return -EIO; 198 seg[n].mr_page = NULL; 199 seg[n].mr_offset = xdrbuf->tail[0].iov_base; 200 seg[n].mr_len = xdrbuf->tail[0].iov_len; 201 ++n; 202 } 203 204 return n; 205 } 206 207 /* 208 * Create read/write chunk lists, and reply chunks, for RDMA 209 * 210 * Assume check against THRESHOLD has been done, and chunks are required. 211 * Assume only encoding one list entry for read|write chunks. The NFSv3 212 * protocol is simple enough to allow this as it only has a single "bulk 213 * result" in each procedure - complicated NFSv4 COMPOUNDs are not. (The 214 * RDMA/Sessions NFSv4 proposal addresses this for future v4 revs.) 215 * 216 * When used for a single reply chunk (which is a special write 217 * chunk used for the entire reply, rather than just the data), it 218 * is used primarily for READDIR and READLINK which would otherwise 219 * be severely size-limited by a small rdma inline read max. The server 220 * response will come back as an RDMA Write, followed by a message 221 * of type RDMA_NOMSG carrying the xid and length. As a result, reply 222 * chunks do not provide data alignment, however they do not require 223 * "fixup" (moving the response to the upper layer buffer) either. 224 * 225 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64): 226 * 227 * Read chunklist (a linked list): 228 * N elements, position P (same P for all chunks of same arg!): 229 * 1 - PHLOO - 1 - PHLOO - ... - 1 - PHLOO - 0 230 * 231 * Write chunklist (a list of (one) counted array): 232 * N elements: 233 * 1 - N - HLOO - HLOO - ... - HLOO - 0 234 * 235 * Reply chunk (a counted array): 236 * N elements: 237 * 1 - N - HLOO - HLOO - ... - HLOO 238 * 239 * Returns positive RPC/RDMA header size, or negative errno. 240 */ 241 242 static ssize_t 243 rpcrdma_create_chunks(struct rpc_rqst *rqst, struct xdr_buf *target, 244 struct rpcrdma_msg *headerp, enum rpcrdma_chunktype type) 245 { 246 struct rpcrdma_req *req = rpcr_to_rdmar(rqst); 247 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt); 248 int n, nsegs, nchunks = 0; 249 unsigned int pos; 250 struct rpcrdma_mr_seg *seg = req->rl_segments; 251 struct rpcrdma_read_chunk *cur_rchunk = NULL; 252 struct rpcrdma_write_array *warray = NULL; 253 struct rpcrdma_write_chunk *cur_wchunk = NULL; 254 __be32 *iptr = headerp->rm_body.rm_chunks; 255 int (*map)(struct rpcrdma_xprt *, struct rpcrdma_mr_seg *, int, bool); 256 257 if (type == rpcrdma_readch || type == rpcrdma_areadch) { 258 /* a read chunk - server will RDMA Read our memory */ 259 cur_rchunk = (struct rpcrdma_read_chunk *) iptr; 260 } else { 261 /* a write or reply chunk - server will RDMA Write our memory */ 262 *iptr++ = xdr_zero; /* encode a NULL read chunk list */ 263 if (type == rpcrdma_replych) 264 *iptr++ = xdr_zero; /* a NULL write chunk list */ 265 warray = (struct rpcrdma_write_array *) iptr; 266 cur_wchunk = (struct rpcrdma_write_chunk *) (warray + 1); 267 } 268 269 if (type == rpcrdma_replych || type == rpcrdma_areadch) 270 pos = 0; 271 else 272 pos = target->head[0].iov_len; 273 274 nsegs = rpcrdma_convert_iovs(target, pos, type, seg, RPCRDMA_MAX_SEGS); 275 if (nsegs < 0) 276 return nsegs; 277 278 map = r_xprt->rx_ia.ri_ops->ro_map; 279 do { 280 n = map(r_xprt, seg, nsegs, cur_wchunk != NULL); 281 if (n <= 0) 282 goto out; 283 if (cur_rchunk) { /* read */ 284 cur_rchunk->rc_discrim = xdr_one; 285 /* all read chunks have the same "position" */ 286 cur_rchunk->rc_position = cpu_to_be32(pos); 287 cur_rchunk->rc_target.rs_handle = 288 cpu_to_be32(seg->mr_rkey); 289 cur_rchunk->rc_target.rs_length = 290 cpu_to_be32(seg->mr_len); 291 xdr_encode_hyper( 292 (__be32 *)&cur_rchunk->rc_target.rs_offset, 293 seg->mr_base); 294 dprintk("RPC: %s: read chunk " 295 "elem %d@0x%llx:0x%x pos %u (%s)\n", __func__, 296 seg->mr_len, (unsigned long long)seg->mr_base, 297 seg->mr_rkey, pos, n < nsegs ? "more" : "last"); 298 cur_rchunk++; 299 r_xprt->rx_stats.read_chunk_count++; 300 } else { /* write/reply */ 301 cur_wchunk->wc_target.rs_handle = 302 cpu_to_be32(seg->mr_rkey); 303 cur_wchunk->wc_target.rs_length = 304 cpu_to_be32(seg->mr_len); 305 xdr_encode_hyper( 306 (__be32 *)&cur_wchunk->wc_target.rs_offset, 307 seg->mr_base); 308 dprintk("RPC: %s: %s chunk " 309 "elem %d@0x%llx:0x%x (%s)\n", __func__, 310 (type == rpcrdma_replych) ? "reply" : "write", 311 seg->mr_len, (unsigned long long)seg->mr_base, 312 seg->mr_rkey, n < nsegs ? "more" : "last"); 313 cur_wchunk++; 314 if (type == rpcrdma_replych) 315 r_xprt->rx_stats.reply_chunk_count++; 316 else 317 r_xprt->rx_stats.write_chunk_count++; 318 r_xprt->rx_stats.total_rdma_request += seg->mr_len; 319 } 320 nchunks++; 321 seg += n; 322 nsegs -= n; 323 } while (nsegs); 324 325 /* success. all failures return above */ 326 req->rl_nchunks = nchunks; 327 328 /* 329 * finish off header. If write, marshal discrim and nchunks. 330 */ 331 if (cur_rchunk) { 332 iptr = (__be32 *) cur_rchunk; 333 *iptr++ = xdr_zero; /* finish the read chunk list */ 334 *iptr++ = xdr_zero; /* encode a NULL write chunk list */ 335 *iptr++ = xdr_zero; /* encode a NULL reply chunk */ 336 } else { 337 warray->wc_discrim = xdr_one; 338 warray->wc_nchunks = cpu_to_be32(nchunks); 339 iptr = (__be32 *) cur_wchunk; 340 if (type == rpcrdma_writech) { 341 *iptr++ = xdr_zero; /* finish the write chunk list */ 342 *iptr++ = xdr_zero; /* encode a NULL reply chunk */ 343 } 344 } 345 346 /* 347 * Return header size. 348 */ 349 return (unsigned char *)iptr - (unsigned char *)headerp; 350 351 out: 352 for (pos = 0; nchunks--;) 353 pos += r_xprt->rx_ia.ri_ops->ro_unmap(r_xprt, 354 &req->rl_segments[pos]); 355 return n; 356 } 357 358 /* 359 * Copy write data inline. 360 * This function is used for "small" requests. Data which is passed 361 * to RPC via iovecs (or page list) is copied directly into the 362 * pre-registered memory buffer for this request. For small amounts 363 * of data, this is efficient. The cutoff value is tunable. 364 */ 365 static void rpcrdma_inline_pullup(struct rpc_rqst *rqst) 366 { 367 int i, npages, curlen; 368 int copy_len; 369 unsigned char *srcp, *destp; 370 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt); 371 int page_base; 372 struct page **ppages; 373 374 destp = rqst->rq_svec[0].iov_base; 375 curlen = rqst->rq_svec[0].iov_len; 376 destp += curlen; 377 378 dprintk("RPC: %s: destp 0x%p len %d hdrlen %d\n", 379 __func__, destp, rqst->rq_slen, curlen); 380 381 copy_len = rqst->rq_snd_buf.page_len; 382 383 if (rqst->rq_snd_buf.tail[0].iov_len) { 384 curlen = rqst->rq_snd_buf.tail[0].iov_len; 385 if (destp + copy_len != rqst->rq_snd_buf.tail[0].iov_base) { 386 memmove(destp + copy_len, 387 rqst->rq_snd_buf.tail[0].iov_base, curlen); 388 r_xprt->rx_stats.pullup_copy_count += curlen; 389 } 390 dprintk("RPC: %s: tail destp 0x%p len %d\n", 391 __func__, destp + copy_len, curlen); 392 rqst->rq_svec[0].iov_len += curlen; 393 } 394 r_xprt->rx_stats.pullup_copy_count += copy_len; 395 396 page_base = rqst->rq_snd_buf.page_base; 397 ppages = rqst->rq_snd_buf.pages + (page_base >> PAGE_SHIFT); 398 page_base &= ~PAGE_MASK; 399 npages = PAGE_ALIGN(page_base+copy_len) >> PAGE_SHIFT; 400 for (i = 0; copy_len && i < npages; i++) { 401 curlen = PAGE_SIZE - page_base; 402 if (curlen > copy_len) 403 curlen = copy_len; 404 dprintk("RPC: %s: page %d destp 0x%p len %d curlen %d\n", 405 __func__, i, destp, copy_len, curlen); 406 srcp = kmap_atomic(ppages[i]); 407 memcpy(destp, srcp+page_base, curlen); 408 kunmap_atomic(srcp); 409 rqst->rq_svec[0].iov_len += curlen; 410 destp += curlen; 411 copy_len -= curlen; 412 page_base = 0; 413 } 414 /* header now contains entire send message */ 415 } 416 417 /* 418 * Marshal a request: the primary job of this routine is to choose 419 * the transfer modes. See comments below. 420 * 421 * Uses multiple RDMA IOVs for a request: 422 * [0] -- RPC RDMA header, which uses memory from the *start* of the 423 * preregistered buffer that already holds the RPC data in 424 * its middle. 425 * [1] -- the RPC header/data, marshaled by RPC and the NFS protocol. 426 * [2] -- optional padding. 427 * [3] -- if padded, header only in [1] and data here. 428 * 429 * Returns zero on success, otherwise a negative errno. 430 */ 431 432 int 433 rpcrdma_marshal_req(struct rpc_rqst *rqst) 434 { 435 struct rpc_xprt *xprt = rqst->rq_xprt; 436 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 437 struct rpcrdma_req *req = rpcr_to_rdmar(rqst); 438 char *base; 439 size_t rpclen; 440 ssize_t hdrlen; 441 enum rpcrdma_chunktype rtype, wtype; 442 struct rpcrdma_msg *headerp; 443 444 #if defined(CONFIG_SUNRPC_BACKCHANNEL) 445 if (test_bit(RPC_BC_PA_IN_USE, &rqst->rq_bc_pa_state)) 446 return rpcrdma_bc_marshal_reply(rqst); 447 #endif 448 449 /* 450 * rpclen gets amount of data in first buffer, which is the 451 * pre-registered buffer. 452 */ 453 base = rqst->rq_svec[0].iov_base; 454 rpclen = rqst->rq_svec[0].iov_len; 455 456 headerp = rdmab_to_msg(req->rl_rdmabuf); 457 /* don't byte-swap XID, it's already done in request */ 458 headerp->rm_xid = rqst->rq_xid; 459 headerp->rm_vers = rpcrdma_version; 460 headerp->rm_credit = cpu_to_be32(r_xprt->rx_buf.rb_max_requests); 461 headerp->rm_type = rdma_msg; 462 463 /* 464 * Chunks needed for results? 465 * 466 * o Read ops return data as write chunk(s), header as inline. 467 * o If the expected result is under the inline threshold, all ops 468 * return as inline. 469 * o Large non-read ops return as a single reply chunk. 470 */ 471 if (rqst->rq_rcv_buf.flags & XDRBUF_READ) 472 wtype = rpcrdma_writech; 473 else if (rpcrdma_results_inline(rqst)) 474 wtype = rpcrdma_noch; 475 else 476 wtype = rpcrdma_replych; 477 478 /* 479 * Chunks needed for arguments? 480 * 481 * o If the total request is under the inline threshold, all ops 482 * are sent as inline. 483 * o Large write ops transmit data as read chunk(s), header as 484 * inline. 485 * o Large non-write ops are sent with the entire message as a 486 * single read chunk (protocol 0-position special case). 487 * 488 * This assumes that the upper layer does not present a request 489 * that both has a data payload, and whose non-data arguments 490 * by themselves are larger than the inline threshold. 491 */ 492 if (rpcrdma_args_inline(rqst)) { 493 rtype = rpcrdma_noch; 494 } else if (rqst->rq_snd_buf.flags & XDRBUF_WRITE) { 495 rtype = rpcrdma_readch; 496 } else { 497 r_xprt->rx_stats.nomsg_call_count++; 498 headerp->rm_type = htonl(RDMA_NOMSG); 499 rtype = rpcrdma_areadch; 500 rpclen = 0; 501 } 502 503 /* The following simplification is not true forever */ 504 if (rtype != rpcrdma_noch && wtype == rpcrdma_replych) 505 wtype = rpcrdma_noch; 506 if (rtype != rpcrdma_noch && wtype != rpcrdma_noch) { 507 dprintk("RPC: %s: cannot marshal multiple chunk lists\n", 508 __func__); 509 return -EIO; 510 } 511 512 hdrlen = RPCRDMA_HDRLEN_MIN; 513 514 /* 515 * Pull up any extra send data into the preregistered buffer. 516 * When padding is in use and applies to the transfer, insert 517 * it and change the message type. 518 */ 519 if (rtype == rpcrdma_noch) { 520 521 rpcrdma_inline_pullup(rqst); 522 523 headerp->rm_body.rm_nochunks.rm_empty[0] = xdr_zero; 524 headerp->rm_body.rm_nochunks.rm_empty[1] = xdr_zero; 525 headerp->rm_body.rm_nochunks.rm_empty[2] = xdr_zero; 526 /* new length after pullup */ 527 rpclen = rqst->rq_svec[0].iov_len; 528 } else if (rtype == rpcrdma_readch) 529 rpclen += rpcrdma_tail_pullup(&rqst->rq_snd_buf); 530 if (rtype != rpcrdma_noch) { 531 hdrlen = rpcrdma_create_chunks(rqst, &rqst->rq_snd_buf, 532 headerp, rtype); 533 wtype = rtype; /* simplify dprintk */ 534 535 } else if (wtype != rpcrdma_noch) { 536 hdrlen = rpcrdma_create_chunks(rqst, &rqst->rq_rcv_buf, 537 headerp, wtype); 538 } 539 if (hdrlen < 0) 540 return hdrlen; 541 542 dprintk("RPC: %s: %s: hdrlen %zd rpclen %zd" 543 " headerp 0x%p base 0x%p lkey 0x%x\n", 544 __func__, transfertypes[wtype], hdrlen, rpclen, 545 headerp, base, rdmab_lkey(req->rl_rdmabuf)); 546 547 /* 548 * initialize send_iov's - normally only two: rdma chunk header and 549 * single preregistered RPC header buffer, but if padding is present, 550 * then use a preregistered (and zeroed) pad buffer between the RPC 551 * header and any write data. In all non-rdma cases, any following 552 * data has been copied into the RPC header buffer. 553 */ 554 req->rl_send_iov[0].addr = rdmab_addr(req->rl_rdmabuf); 555 req->rl_send_iov[0].length = hdrlen; 556 req->rl_send_iov[0].lkey = rdmab_lkey(req->rl_rdmabuf); 557 558 req->rl_niovs = 1; 559 if (rtype == rpcrdma_areadch) 560 return 0; 561 562 req->rl_send_iov[1].addr = rdmab_addr(req->rl_sendbuf); 563 req->rl_send_iov[1].length = rpclen; 564 req->rl_send_iov[1].lkey = rdmab_lkey(req->rl_sendbuf); 565 566 req->rl_niovs = 2; 567 return 0; 568 } 569 570 /* 571 * Chase down a received write or reply chunklist to get length 572 * RDMA'd by server. See map at rpcrdma_create_chunks()! :-) 573 */ 574 static int 575 rpcrdma_count_chunks(struct rpcrdma_rep *rep, unsigned int max, int wrchunk, __be32 **iptrp) 576 { 577 unsigned int i, total_len; 578 struct rpcrdma_write_chunk *cur_wchunk; 579 char *base = (char *)rdmab_to_msg(rep->rr_rdmabuf); 580 581 i = be32_to_cpu(**iptrp); 582 if (i > max) 583 return -1; 584 cur_wchunk = (struct rpcrdma_write_chunk *) (*iptrp + 1); 585 total_len = 0; 586 while (i--) { 587 struct rpcrdma_segment *seg = &cur_wchunk->wc_target; 588 ifdebug(FACILITY) { 589 u64 off; 590 xdr_decode_hyper((__be32 *)&seg->rs_offset, &off); 591 dprintk("RPC: %s: chunk %d@0x%llx:0x%x\n", 592 __func__, 593 be32_to_cpu(seg->rs_length), 594 (unsigned long long)off, 595 be32_to_cpu(seg->rs_handle)); 596 } 597 total_len += be32_to_cpu(seg->rs_length); 598 ++cur_wchunk; 599 } 600 /* check and adjust for properly terminated write chunk */ 601 if (wrchunk) { 602 __be32 *w = (__be32 *) cur_wchunk; 603 if (*w++ != xdr_zero) 604 return -1; 605 cur_wchunk = (struct rpcrdma_write_chunk *) w; 606 } 607 if ((char *)cur_wchunk > base + rep->rr_len) 608 return -1; 609 610 *iptrp = (__be32 *) cur_wchunk; 611 return total_len; 612 } 613 614 /* 615 * Scatter inline received data back into provided iov's. 616 */ 617 static void 618 rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad) 619 { 620 int i, npages, curlen, olen; 621 char *destp; 622 struct page **ppages; 623 int page_base; 624 625 curlen = rqst->rq_rcv_buf.head[0].iov_len; 626 if (curlen > copy_len) { /* write chunk header fixup */ 627 curlen = copy_len; 628 rqst->rq_rcv_buf.head[0].iov_len = curlen; 629 } 630 631 dprintk("RPC: %s: srcp 0x%p len %d hdrlen %d\n", 632 __func__, srcp, copy_len, curlen); 633 634 /* Shift pointer for first receive segment only */ 635 rqst->rq_rcv_buf.head[0].iov_base = srcp; 636 srcp += curlen; 637 copy_len -= curlen; 638 639 olen = copy_len; 640 i = 0; 641 rpcx_to_rdmax(rqst->rq_xprt)->rx_stats.fixup_copy_count += olen; 642 page_base = rqst->rq_rcv_buf.page_base; 643 ppages = rqst->rq_rcv_buf.pages + (page_base >> PAGE_SHIFT); 644 page_base &= ~PAGE_MASK; 645 646 if (copy_len && rqst->rq_rcv_buf.page_len) { 647 npages = PAGE_ALIGN(page_base + 648 rqst->rq_rcv_buf.page_len) >> PAGE_SHIFT; 649 for (; i < npages; i++) { 650 curlen = PAGE_SIZE - page_base; 651 if (curlen > copy_len) 652 curlen = copy_len; 653 dprintk("RPC: %s: page %d" 654 " srcp 0x%p len %d curlen %d\n", 655 __func__, i, srcp, copy_len, curlen); 656 destp = kmap_atomic(ppages[i]); 657 memcpy(destp + page_base, srcp, curlen); 658 flush_dcache_page(ppages[i]); 659 kunmap_atomic(destp); 660 srcp += curlen; 661 copy_len -= curlen; 662 if (copy_len == 0) 663 break; 664 page_base = 0; 665 } 666 } 667 668 if (copy_len && rqst->rq_rcv_buf.tail[0].iov_len) { 669 curlen = copy_len; 670 if (curlen > rqst->rq_rcv_buf.tail[0].iov_len) 671 curlen = rqst->rq_rcv_buf.tail[0].iov_len; 672 if (rqst->rq_rcv_buf.tail[0].iov_base != srcp) 673 memmove(rqst->rq_rcv_buf.tail[0].iov_base, srcp, curlen); 674 dprintk("RPC: %s: tail srcp 0x%p len %d curlen %d\n", 675 __func__, srcp, copy_len, curlen); 676 rqst->rq_rcv_buf.tail[0].iov_len = curlen; 677 copy_len -= curlen; ++i; 678 } else 679 rqst->rq_rcv_buf.tail[0].iov_len = 0; 680 681 if (pad) { 682 /* implicit padding on terminal chunk */ 683 unsigned char *p = rqst->rq_rcv_buf.tail[0].iov_base; 684 while (pad--) 685 p[rqst->rq_rcv_buf.tail[0].iov_len++] = 0; 686 } 687 688 if (copy_len) 689 dprintk("RPC: %s: %d bytes in" 690 " %d extra segments (%d lost)\n", 691 __func__, olen, i, copy_len); 692 693 /* TBD avoid a warning from call_decode() */ 694 rqst->rq_private_buf = rqst->rq_rcv_buf; 695 } 696 697 void 698 rpcrdma_connect_worker(struct work_struct *work) 699 { 700 struct rpcrdma_ep *ep = 701 container_of(work, struct rpcrdma_ep, rep_connect_worker.work); 702 struct rpcrdma_xprt *r_xprt = 703 container_of(ep, struct rpcrdma_xprt, rx_ep); 704 struct rpc_xprt *xprt = &r_xprt->rx_xprt; 705 706 spin_lock_bh(&xprt->transport_lock); 707 if (++xprt->connect_cookie == 0) /* maintain a reserved value */ 708 ++xprt->connect_cookie; 709 if (ep->rep_connected > 0) { 710 if (!xprt_test_and_set_connected(xprt)) 711 xprt_wake_pending_tasks(xprt, 0); 712 } else { 713 if (xprt_test_and_clear_connected(xprt)) 714 xprt_wake_pending_tasks(xprt, -ENOTCONN); 715 } 716 spin_unlock_bh(&xprt->transport_lock); 717 } 718 719 #if defined(CONFIG_SUNRPC_BACKCHANNEL) 720 /* By convention, backchannel calls arrive via rdma_msg type 721 * messages, and never populate the chunk lists. This makes 722 * the RPC/RDMA header small and fixed in size, so it is 723 * straightforward to check the RPC header's direction field. 724 */ 725 static bool 726 rpcrdma_is_bcall(struct rpcrdma_msg *headerp) 727 { 728 __be32 *p = (__be32 *)headerp; 729 730 if (headerp->rm_type != rdma_msg) 731 return false; 732 if (headerp->rm_body.rm_chunks[0] != xdr_zero) 733 return false; 734 if (headerp->rm_body.rm_chunks[1] != xdr_zero) 735 return false; 736 if (headerp->rm_body.rm_chunks[2] != xdr_zero) 737 return false; 738 739 /* sanity */ 740 if (p[7] != headerp->rm_xid) 741 return false; 742 /* call direction */ 743 if (p[8] != cpu_to_be32(RPC_CALL)) 744 return false; 745 746 return true; 747 } 748 #endif /* CONFIG_SUNRPC_BACKCHANNEL */ 749 750 /* 751 * This function is called when an async event is posted to 752 * the connection which changes the connection state. All it 753 * does at this point is mark the connection up/down, the rpc 754 * timers do the rest. 755 */ 756 void 757 rpcrdma_conn_func(struct rpcrdma_ep *ep) 758 { 759 schedule_delayed_work(&ep->rep_connect_worker, 0); 760 } 761 762 /* Process received RPC/RDMA messages. 763 * 764 * Errors must result in the RPC task either being awakened, or 765 * allowed to timeout, to discover the errors at that time. 766 */ 767 void 768 rpcrdma_reply_handler(struct rpcrdma_rep *rep) 769 { 770 struct rpcrdma_msg *headerp; 771 struct rpcrdma_req *req; 772 struct rpc_rqst *rqst; 773 struct rpcrdma_xprt *r_xprt = rep->rr_rxprt; 774 struct rpc_xprt *xprt = &r_xprt->rx_xprt; 775 __be32 *iptr; 776 int rdmalen, status; 777 unsigned long cwnd; 778 u32 credits; 779 780 dprintk("RPC: %s: incoming rep %p\n", __func__, rep); 781 782 if (rep->rr_len == RPCRDMA_BAD_LEN) 783 goto out_badstatus; 784 if (rep->rr_len < RPCRDMA_HDRLEN_MIN) 785 goto out_shortreply; 786 787 headerp = rdmab_to_msg(rep->rr_rdmabuf); 788 if (headerp->rm_vers != rpcrdma_version) 789 goto out_badversion; 790 #if defined(CONFIG_SUNRPC_BACKCHANNEL) 791 if (rpcrdma_is_bcall(headerp)) 792 goto out_bcall; 793 #endif 794 795 /* Match incoming rpcrdma_rep to an rpcrdma_req to 796 * get context for handling any incoming chunks. 797 */ 798 spin_lock_bh(&xprt->transport_lock); 799 rqst = xprt_lookup_rqst(xprt, headerp->rm_xid); 800 if (!rqst) 801 goto out_nomatch; 802 803 req = rpcr_to_rdmar(rqst); 804 if (req->rl_reply) 805 goto out_duplicate; 806 807 dprintk("RPC: %s: reply 0x%p completes request 0x%p\n" 808 " RPC request 0x%p xid 0x%08x\n", 809 __func__, rep, req, rqst, 810 be32_to_cpu(headerp->rm_xid)); 811 812 /* from here on, the reply is no longer an orphan */ 813 req->rl_reply = rep; 814 xprt->reestablish_timeout = 0; 815 816 /* check for expected message types */ 817 /* The order of some of these tests is important. */ 818 switch (headerp->rm_type) { 819 case rdma_msg: 820 /* never expect read chunks */ 821 /* never expect reply chunks (two ways to check) */ 822 /* never expect write chunks without having offered RDMA */ 823 if (headerp->rm_body.rm_chunks[0] != xdr_zero || 824 (headerp->rm_body.rm_chunks[1] == xdr_zero && 825 headerp->rm_body.rm_chunks[2] != xdr_zero) || 826 (headerp->rm_body.rm_chunks[1] != xdr_zero && 827 req->rl_nchunks == 0)) 828 goto badheader; 829 if (headerp->rm_body.rm_chunks[1] != xdr_zero) { 830 /* count any expected write chunks in read reply */ 831 /* start at write chunk array count */ 832 iptr = &headerp->rm_body.rm_chunks[2]; 833 rdmalen = rpcrdma_count_chunks(rep, 834 req->rl_nchunks, 1, &iptr); 835 /* check for validity, and no reply chunk after */ 836 if (rdmalen < 0 || *iptr++ != xdr_zero) 837 goto badheader; 838 rep->rr_len -= 839 ((unsigned char *)iptr - (unsigned char *)headerp); 840 status = rep->rr_len + rdmalen; 841 r_xprt->rx_stats.total_rdma_reply += rdmalen; 842 /* special case - last chunk may omit padding */ 843 if (rdmalen &= 3) { 844 rdmalen = 4 - rdmalen; 845 status += rdmalen; 846 } 847 } else { 848 /* else ordinary inline */ 849 rdmalen = 0; 850 iptr = (__be32 *)((unsigned char *)headerp + 851 RPCRDMA_HDRLEN_MIN); 852 rep->rr_len -= RPCRDMA_HDRLEN_MIN; 853 status = rep->rr_len; 854 } 855 /* Fix up the rpc results for upper layer */ 856 rpcrdma_inline_fixup(rqst, (char *)iptr, rep->rr_len, rdmalen); 857 break; 858 859 case rdma_nomsg: 860 /* never expect read or write chunks, always reply chunks */ 861 if (headerp->rm_body.rm_chunks[0] != xdr_zero || 862 headerp->rm_body.rm_chunks[1] != xdr_zero || 863 headerp->rm_body.rm_chunks[2] != xdr_one || 864 req->rl_nchunks == 0) 865 goto badheader; 866 iptr = (__be32 *)((unsigned char *)headerp + 867 RPCRDMA_HDRLEN_MIN); 868 rdmalen = rpcrdma_count_chunks(rep, req->rl_nchunks, 0, &iptr); 869 if (rdmalen < 0) 870 goto badheader; 871 r_xprt->rx_stats.total_rdma_reply += rdmalen; 872 /* Reply chunk buffer already is the reply vector - no fixup. */ 873 status = rdmalen; 874 break; 875 876 badheader: 877 default: 878 dprintk("%s: invalid rpcrdma reply header (type %d):" 879 " chunks[012] == %d %d %d" 880 " expected chunks <= %d\n", 881 __func__, be32_to_cpu(headerp->rm_type), 882 headerp->rm_body.rm_chunks[0], 883 headerp->rm_body.rm_chunks[1], 884 headerp->rm_body.rm_chunks[2], 885 req->rl_nchunks); 886 status = -EIO; 887 r_xprt->rx_stats.bad_reply_count++; 888 break; 889 } 890 891 credits = be32_to_cpu(headerp->rm_credit); 892 if (credits == 0) 893 credits = 1; /* don't deadlock */ 894 else if (credits > r_xprt->rx_buf.rb_max_requests) 895 credits = r_xprt->rx_buf.rb_max_requests; 896 897 cwnd = xprt->cwnd; 898 xprt->cwnd = credits << RPC_CWNDSHIFT; 899 if (xprt->cwnd > cwnd) 900 xprt_release_rqst_cong(rqst->rq_task); 901 902 xprt_complete_rqst(rqst->rq_task, status); 903 spin_unlock_bh(&xprt->transport_lock); 904 dprintk("RPC: %s: xprt_complete_rqst(0x%p, 0x%p, %d)\n", 905 __func__, xprt, rqst, status); 906 return; 907 908 out_badstatus: 909 rpcrdma_recv_buffer_put(rep); 910 if (r_xprt->rx_ep.rep_connected == 1) { 911 r_xprt->rx_ep.rep_connected = -EIO; 912 rpcrdma_conn_func(&r_xprt->rx_ep); 913 } 914 return; 915 916 #if defined(CONFIG_SUNRPC_BACKCHANNEL) 917 out_bcall: 918 rpcrdma_bc_receive_call(r_xprt, rep); 919 return; 920 #endif 921 922 out_shortreply: 923 dprintk("RPC: %s: short/invalid reply\n", __func__); 924 goto repost; 925 926 out_badversion: 927 dprintk("RPC: %s: invalid version %d\n", 928 __func__, be32_to_cpu(headerp->rm_vers)); 929 goto repost; 930 931 out_nomatch: 932 spin_unlock_bh(&xprt->transport_lock); 933 dprintk("RPC: %s: no match for incoming xid 0x%08x len %d\n", 934 __func__, be32_to_cpu(headerp->rm_xid), 935 rep->rr_len); 936 goto repost; 937 938 out_duplicate: 939 spin_unlock_bh(&xprt->transport_lock); 940 dprintk("RPC: %s: " 941 "duplicate reply %p to RPC request %p: xid 0x%08x\n", 942 __func__, rep, req, be32_to_cpu(headerp->rm_xid)); 943 944 repost: 945 r_xprt->rx_stats.bad_reply_count++; 946 if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, &r_xprt->rx_ep, rep)) 947 rpcrdma_recv_buffer_put(rep); 948 } 949