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 #ifdef RPC_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 #ifdef RPC_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 /* 75 * Chunk assembly from upper layer xdr_buf. 76 * 77 * Prepare the passed-in xdr_buf into representation as RPC/RDMA chunk 78 * elements. Segments are then coalesced when registered, if possible 79 * within the selected memreg mode. 80 * 81 * Note, this routine is never called if the connection's memory 82 * registration strategy is 0 (bounce buffers). 83 */ 84 85 static int 86 rpcrdma_convert_iovs(struct xdr_buf *xdrbuf, unsigned int pos, 87 enum rpcrdma_chunktype type, struct rpcrdma_mr_seg *seg, int nsegs) 88 { 89 int len, n = 0, p; 90 int page_base; 91 struct page **ppages; 92 93 if (pos == 0 && xdrbuf->head[0].iov_len) { 94 seg[n].mr_page = NULL; 95 seg[n].mr_offset = xdrbuf->head[0].iov_base; 96 seg[n].mr_len = xdrbuf->head[0].iov_len; 97 ++n; 98 } 99 100 len = xdrbuf->page_len; 101 ppages = xdrbuf->pages + (xdrbuf->page_base >> PAGE_SHIFT); 102 page_base = xdrbuf->page_base & ~PAGE_MASK; 103 p = 0; 104 while (len && n < nsegs) { 105 seg[n].mr_page = ppages[p]; 106 seg[n].mr_offset = (void *)(unsigned long) page_base; 107 seg[n].mr_len = min_t(u32, PAGE_SIZE - page_base, len); 108 BUG_ON(seg[n].mr_len > PAGE_SIZE); 109 len -= seg[n].mr_len; 110 ++n; 111 ++p; 112 page_base = 0; /* page offset only applies to first page */ 113 } 114 115 /* Message overflows the seg array */ 116 if (len && n == nsegs) 117 return 0; 118 119 if (xdrbuf->tail[0].iov_len) { 120 /* the rpcrdma protocol allows us to omit any trailing 121 * xdr pad bytes, saving the server an RDMA operation. */ 122 if (xdrbuf->tail[0].iov_len < 4 && xprt_rdma_pad_optimize) 123 return n; 124 if (n == nsegs) 125 /* Tail remains, but we're out of segments */ 126 return 0; 127 seg[n].mr_page = NULL; 128 seg[n].mr_offset = xdrbuf->tail[0].iov_base; 129 seg[n].mr_len = xdrbuf->tail[0].iov_len; 130 ++n; 131 } 132 133 return n; 134 } 135 136 /* 137 * Create read/write chunk lists, and reply chunks, for RDMA 138 * 139 * Assume check against THRESHOLD has been done, and chunks are required. 140 * Assume only encoding one list entry for read|write chunks. The NFSv3 141 * protocol is simple enough to allow this as it only has a single "bulk 142 * result" in each procedure - complicated NFSv4 COMPOUNDs are not. (The 143 * RDMA/Sessions NFSv4 proposal addresses this for future v4 revs.) 144 * 145 * When used for a single reply chunk (which is a special write 146 * chunk used for the entire reply, rather than just the data), it 147 * is used primarily for READDIR and READLINK which would otherwise 148 * be severely size-limited by a small rdma inline read max. The server 149 * response will come back as an RDMA Write, followed by a message 150 * of type RDMA_NOMSG carrying the xid and length. As a result, reply 151 * chunks do not provide data alignment, however they do not require 152 * "fixup" (moving the response to the upper layer buffer) either. 153 * 154 * Encoding key for single-list chunks (HLOO = Handle32 Length32 Offset64): 155 * 156 * Read chunklist (a linked list): 157 * N elements, position P (same P for all chunks of same arg!): 158 * 1 - PHLOO - 1 - PHLOO - ... - 1 - PHLOO - 0 159 * 160 * Write chunklist (a list of (one) counted array): 161 * N elements: 162 * 1 - N - HLOO - HLOO - ... - HLOO - 0 163 * 164 * Reply chunk (a counted array): 165 * N elements: 166 * 1 - N - HLOO - HLOO - ... - HLOO 167 */ 168 169 static unsigned int 170 rpcrdma_create_chunks(struct rpc_rqst *rqst, struct xdr_buf *target, 171 struct rpcrdma_msg *headerp, enum rpcrdma_chunktype type) 172 { 173 struct rpcrdma_req *req = rpcr_to_rdmar(rqst); 174 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt); 175 int nsegs, nchunks = 0; 176 unsigned int pos; 177 struct rpcrdma_mr_seg *seg = req->rl_segments; 178 struct rpcrdma_read_chunk *cur_rchunk = NULL; 179 struct rpcrdma_write_array *warray = NULL; 180 struct rpcrdma_write_chunk *cur_wchunk = NULL; 181 __be32 *iptr = headerp->rm_body.rm_chunks; 182 183 if (type == rpcrdma_readch || type == rpcrdma_areadch) { 184 /* a read chunk - server will RDMA Read our memory */ 185 cur_rchunk = (struct rpcrdma_read_chunk *) iptr; 186 } else { 187 /* a write or reply chunk - server will RDMA Write our memory */ 188 *iptr++ = xdr_zero; /* encode a NULL read chunk list */ 189 if (type == rpcrdma_replych) 190 *iptr++ = xdr_zero; /* a NULL write chunk list */ 191 warray = (struct rpcrdma_write_array *) iptr; 192 cur_wchunk = (struct rpcrdma_write_chunk *) (warray + 1); 193 } 194 195 if (type == rpcrdma_replych || type == rpcrdma_areadch) 196 pos = 0; 197 else 198 pos = target->head[0].iov_len; 199 200 nsegs = rpcrdma_convert_iovs(target, pos, type, seg, RPCRDMA_MAX_SEGS); 201 if (nsegs == 0) 202 return 0; 203 204 do { 205 /* bind/register the memory, then build chunk from result. */ 206 int n = rpcrdma_register_external(seg, nsegs, 207 cur_wchunk != NULL, r_xprt); 208 if (n <= 0) 209 goto out; 210 if (cur_rchunk) { /* read */ 211 cur_rchunk->rc_discrim = xdr_one; 212 /* all read chunks have the same "position" */ 213 cur_rchunk->rc_position = htonl(pos); 214 cur_rchunk->rc_target.rs_handle = htonl(seg->mr_rkey); 215 cur_rchunk->rc_target.rs_length = htonl(seg->mr_len); 216 xdr_encode_hyper( 217 (__be32 *)&cur_rchunk->rc_target.rs_offset, 218 seg->mr_base); 219 dprintk("RPC: %s: read chunk " 220 "elem %d@0x%llx:0x%x pos %u (%s)\n", __func__, 221 seg->mr_len, (unsigned long long)seg->mr_base, 222 seg->mr_rkey, pos, n < nsegs ? "more" : "last"); 223 cur_rchunk++; 224 r_xprt->rx_stats.read_chunk_count++; 225 } else { /* write/reply */ 226 cur_wchunk->wc_target.rs_handle = htonl(seg->mr_rkey); 227 cur_wchunk->wc_target.rs_length = htonl(seg->mr_len); 228 xdr_encode_hyper( 229 (__be32 *)&cur_wchunk->wc_target.rs_offset, 230 seg->mr_base); 231 dprintk("RPC: %s: %s chunk " 232 "elem %d@0x%llx:0x%x (%s)\n", __func__, 233 (type == rpcrdma_replych) ? "reply" : "write", 234 seg->mr_len, (unsigned long long)seg->mr_base, 235 seg->mr_rkey, n < nsegs ? "more" : "last"); 236 cur_wchunk++; 237 if (type == rpcrdma_replych) 238 r_xprt->rx_stats.reply_chunk_count++; 239 else 240 r_xprt->rx_stats.write_chunk_count++; 241 r_xprt->rx_stats.total_rdma_request += seg->mr_len; 242 } 243 nchunks++; 244 seg += n; 245 nsegs -= n; 246 } while (nsegs); 247 248 /* success. all failures return above */ 249 req->rl_nchunks = nchunks; 250 251 BUG_ON(nchunks == 0); 252 BUG_ON((r_xprt->rx_ia.ri_memreg_strategy == RPCRDMA_FRMR) 253 && (nchunks > 3)); 254 255 /* 256 * finish off header. If write, marshal discrim and nchunks. 257 */ 258 if (cur_rchunk) { 259 iptr = (__be32 *) cur_rchunk; 260 *iptr++ = xdr_zero; /* finish the read chunk list */ 261 *iptr++ = xdr_zero; /* encode a NULL write chunk list */ 262 *iptr++ = xdr_zero; /* encode a NULL reply chunk */ 263 } else { 264 warray->wc_discrim = xdr_one; 265 warray->wc_nchunks = htonl(nchunks); 266 iptr = (__be32 *) cur_wchunk; 267 if (type == rpcrdma_writech) { 268 *iptr++ = xdr_zero; /* finish the write chunk list */ 269 *iptr++ = xdr_zero; /* encode a NULL reply chunk */ 270 } 271 } 272 273 /* 274 * Return header size. 275 */ 276 return (unsigned char *)iptr - (unsigned char *)headerp; 277 278 out: 279 for (pos = 0; nchunks--;) 280 pos += rpcrdma_deregister_external( 281 &req->rl_segments[pos], r_xprt, NULL); 282 return 0; 283 } 284 285 /* 286 * Copy write data inline. 287 * This function is used for "small" requests. Data which is passed 288 * to RPC via iovecs (or page list) is copied directly into the 289 * pre-registered memory buffer for this request. For small amounts 290 * of data, this is efficient. The cutoff value is tunable. 291 */ 292 static int 293 rpcrdma_inline_pullup(struct rpc_rqst *rqst, int pad) 294 { 295 int i, npages, curlen; 296 int copy_len; 297 unsigned char *srcp, *destp; 298 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt); 299 int page_base; 300 struct page **ppages; 301 302 destp = rqst->rq_svec[0].iov_base; 303 curlen = rqst->rq_svec[0].iov_len; 304 destp += curlen; 305 /* 306 * Do optional padding where it makes sense. Alignment of write 307 * payload can help the server, if our setting is accurate. 308 */ 309 pad -= (curlen + 36/*sizeof(struct rpcrdma_msg_padded)*/); 310 if (pad < 0 || rqst->rq_slen - curlen < RPCRDMA_INLINE_PAD_THRESH) 311 pad = 0; /* don't pad this request */ 312 313 dprintk("RPC: %s: pad %d destp 0x%p len %d hdrlen %d\n", 314 __func__, pad, destp, rqst->rq_slen, curlen); 315 316 copy_len = rqst->rq_snd_buf.page_len; 317 318 if (rqst->rq_snd_buf.tail[0].iov_len) { 319 curlen = rqst->rq_snd_buf.tail[0].iov_len; 320 if (destp + copy_len != rqst->rq_snd_buf.tail[0].iov_base) { 321 memmove(destp + copy_len, 322 rqst->rq_snd_buf.tail[0].iov_base, curlen); 323 r_xprt->rx_stats.pullup_copy_count += curlen; 324 } 325 dprintk("RPC: %s: tail destp 0x%p len %d\n", 326 __func__, destp + copy_len, curlen); 327 rqst->rq_svec[0].iov_len += curlen; 328 } 329 r_xprt->rx_stats.pullup_copy_count += copy_len; 330 331 page_base = rqst->rq_snd_buf.page_base; 332 ppages = rqst->rq_snd_buf.pages + (page_base >> PAGE_SHIFT); 333 page_base &= ~PAGE_MASK; 334 npages = PAGE_ALIGN(page_base+copy_len) >> PAGE_SHIFT; 335 for (i = 0; copy_len && i < npages; i++) { 336 curlen = PAGE_SIZE - page_base; 337 if (curlen > copy_len) 338 curlen = copy_len; 339 dprintk("RPC: %s: page %d destp 0x%p len %d curlen %d\n", 340 __func__, i, destp, copy_len, curlen); 341 srcp = kmap_atomic(ppages[i]); 342 memcpy(destp, srcp+page_base, curlen); 343 kunmap_atomic(srcp); 344 rqst->rq_svec[0].iov_len += curlen; 345 destp += curlen; 346 copy_len -= curlen; 347 page_base = 0; 348 } 349 /* header now contains entire send message */ 350 return pad; 351 } 352 353 /* 354 * Marshal a request: the primary job of this routine is to choose 355 * the transfer modes. See comments below. 356 * 357 * Uses multiple RDMA IOVs for a request: 358 * [0] -- RPC RDMA header, which uses memory from the *start* of the 359 * preregistered buffer that already holds the RPC data in 360 * its middle. 361 * [1] -- the RPC header/data, marshaled by RPC and the NFS protocol. 362 * [2] -- optional padding. 363 * [3] -- if padded, header only in [1] and data here. 364 */ 365 366 int 367 rpcrdma_marshal_req(struct rpc_rqst *rqst) 368 { 369 struct rpc_xprt *xprt = rqst->rq_xprt; 370 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 371 struct rpcrdma_req *req = rpcr_to_rdmar(rqst); 372 char *base; 373 size_t hdrlen, rpclen, padlen; 374 enum rpcrdma_chunktype rtype, wtype; 375 struct rpcrdma_msg *headerp; 376 377 /* 378 * rpclen gets amount of data in first buffer, which is the 379 * pre-registered buffer. 380 */ 381 base = rqst->rq_svec[0].iov_base; 382 rpclen = rqst->rq_svec[0].iov_len; 383 384 /* build RDMA header in private area at front */ 385 headerp = (struct rpcrdma_msg *) req->rl_base; 386 /* don't htonl XID, it's already done in request */ 387 headerp->rm_xid = rqst->rq_xid; 388 headerp->rm_vers = xdr_one; 389 headerp->rm_credit = htonl(r_xprt->rx_buf.rb_max_requests); 390 headerp->rm_type = htonl(RDMA_MSG); 391 392 /* 393 * Chunks needed for results? 394 * 395 * o If the expected result is under the inline threshold, all ops 396 * return as inline (but see later). 397 * o Large non-read ops return as a single reply chunk. 398 * o Large read ops return data as write chunk(s), header as inline. 399 * 400 * Note: the NFS code sending down multiple result segments implies 401 * the op is one of read, readdir[plus], readlink or NFSv4 getacl. 402 */ 403 404 /* 405 * This code can handle read chunks, write chunks OR reply 406 * chunks -- only one type. If the request is too big to fit 407 * inline, then we will choose read chunks. If the request is 408 * a READ, then use write chunks to separate the file data 409 * into pages; otherwise use reply chunks. 410 */ 411 if (rqst->rq_rcv_buf.buflen <= RPCRDMA_INLINE_READ_THRESHOLD(rqst)) 412 wtype = rpcrdma_noch; 413 else if (rqst->rq_rcv_buf.page_len == 0) 414 wtype = rpcrdma_replych; 415 else if (rqst->rq_rcv_buf.flags & XDRBUF_READ) 416 wtype = rpcrdma_writech; 417 else 418 wtype = rpcrdma_replych; 419 420 /* 421 * Chunks needed for arguments? 422 * 423 * o If the total request is under the inline threshold, all ops 424 * are sent as inline. 425 * o Large non-write ops are sent with the entire message as a 426 * single read chunk (protocol 0-position special case). 427 * o Large write ops transmit data as read chunk(s), header as 428 * inline. 429 * 430 * Note: the NFS code sending down multiple argument segments 431 * implies the op is a write. 432 * TBD check NFSv4 setacl 433 */ 434 if (rqst->rq_snd_buf.len <= RPCRDMA_INLINE_WRITE_THRESHOLD(rqst)) 435 rtype = rpcrdma_noch; 436 else if (rqst->rq_snd_buf.page_len == 0) 437 rtype = rpcrdma_areadch; 438 else 439 rtype = rpcrdma_readch; 440 441 /* The following simplification is not true forever */ 442 if (rtype != rpcrdma_noch && wtype == rpcrdma_replych) 443 wtype = rpcrdma_noch; 444 BUG_ON(rtype != rpcrdma_noch && wtype != rpcrdma_noch); 445 446 if (r_xprt->rx_ia.ri_memreg_strategy == RPCRDMA_BOUNCEBUFFERS && 447 (rtype != rpcrdma_noch || wtype != rpcrdma_noch)) { 448 /* forced to "pure inline"? */ 449 dprintk("RPC: %s: too much data (%d/%d) for inline\n", 450 __func__, rqst->rq_rcv_buf.len, rqst->rq_snd_buf.len); 451 return -1; 452 } 453 454 hdrlen = 28; /*sizeof *headerp;*/ 455 padlen = 0; 456 457 /* 458 * Pull up any extra send data into the preregistered buffer. 459 * When padding is in use and applies to the transfer, insert 460 * it and change the message type. 461 */ 462 if (rtype == rpcrdma_noch) { 463 464 padlen = rpcrdma_inline_pullup(rqst, 465 RPCRDMA_INLINE_PAD_VALUE(rqst)); 466 467 if (padlen) { 468 headerp->rm_type = htonl(RDMA_MSGP); 469 headerp->rm_body.rm_padded.rm_align = 470 htonl(RPCRDMA_INLINE_PAD_VALUE(rqst)); 471 headerp->rm_body.rm_padded.rm_thresh = 472 htonl(RPCRDMA_INLINE_PAD_THRESH); 473 headerp->rm_body.rm_padded.rm_pempty[0] = xdr_zero; 474 headerp->rm_body.rm_padded.rm_pempty[1] = xdr_zero; 475 headerp->rm_body.rm_padded.rm_pempty[2] = xdr_zero; 476 hdrlen += 2 * sizeof(u32); /* extra words in padhdr */ 477 BUG_ON(wtype != rpcrdma_noch); 478 479 } else { 480 headerp->rm_body.rm_nochunks.rm_empty[0] = xdr_zero; 481 headerp->rm_body.rm_nochunks.rm_empty[1] = xdr_zero; 482 headerp->rm_body.rm_nochunks.rm_empty[2] = xdr_zero; 483 /* new length after pullup */ 484 rpclen = rqst->rq_svec[0].iov_len; 485 /* 486 * Currently we try to not actually use read inline. 487 * Reply chunks have the desirable property that 488 * they land, packed, directly in the target buffers 489 * without headers, so they require no fixup. The 490 * additional RDMA Write op sends the same amount 491 * of data, streams on-the-wire and adds no overhead 492 * on receive. Therefore, we request a reply chunk 493 * for non-writes wherever feasible and efficient. 494 */ 495 if (wtype == rpcrdma_noch && 496 r_xprt->rx_ia.ri_memreg_strategy > RPCRDMA_REGISTER) 497 wtype = rpcrdma_replych; 498 } 499 } 500 501 /* 502 * Marshal chunks. This routine will return the header length 503 * consumed by marshaling. 504 */ 505 if (rtype != rpcrdma_noch) { 506 hdrlen = rpcrdma_create_chunks(rqst, 507 &rqst->rq_snd_buf, headerp, rtype); 508 wtype = rtype; /* simplify dprintk */ 509 510 } else if (wtype != rpcrdma_noch) { 511 hdrlen = rpcrdma_create_chunks(rqst, 512 &rqst->rq_rcv_buf, headerp, wtype); 513 } 514 515 if (hdrlen == 0) 516 return -1; 517 518 dprintk("RPC: %s: %s: hdrlen %zd rpclen %zd padlen %zd" 519 " headerp 0x%p base 0x%p lkey 0x%x\n", 520 __func__, transfertypes[wtype], hdrlen, rpclen, padlen, 521 headerp, base, req->rl_iov.lkey); 522 523 /* 524 * initialize send_iov's - normally only two: rdma chunk header and 525 * single preregistered RPC header buffer, but if padding is present, 526 * then use a preregistered (and zeroed) pad buffer between the RPC 527 * header and any write data. In all non-rdma cases, any following 528 * data has been copied into the RPC header buffer. 529 */ 530 req->rl_send_iov[0].addr = req->rl_iov.addr; 531 req->rl_send_iov[0].length = hdrlen; 532 req->rl_send_iov[0].lkey = req->rl_iov.lkey; 533 534 req->rl_send_iov[1].addr = req->rl_iov.addr + (base - req->rl_base); 535 req->rl_send_iov[1].length = rpclen; 536 req->rl_send_iov[1].lkey = req->rl_iov.lkey; 537 538 req->rl_niovs = 2; 539 540 if (padlen) { 541 struct rpcrdma_ep *ep = &r_xprt->rx_ep; 542 543 req->rl_send_iov[2].addr = ep->rep_pad.addr; 544 req->rl_send_iov[2].length = padlen; 545 req->rl_send_iov[2].lkey = ep->rep_pad.lkey; 546 547 req->rl_send_iov[3].addr = req->rl_send_iov[1].addr + rpclen; 548 req->rl_send_iov[3].length = rqst->rq_slen - rpclen; 549 req->rl_send_iov[3].lkey = req->rl_iov.lkey; 550 551 req->rl_niovs = 4; 552 } 553 554 return 0; 555 } 556 557 /* 558 * Chase down a received write or reply chunklist to get length 559 * RDMA'd by server. See map at rpcrdma_create_chunks()! :-) 560 */ 561 static int 562 rpcrdma_count_chunks(struct rpcrdma_rep *rep, unsigned int max, int wrchunk, __be32 **iptrp) 563 { 564 unsigned int i, total_len; 565 struct rpcrdma_write_chunk *cur_wchunk; 566 567 i = ntohl(**iptrp); /* get array count */ 568 if (i > max) 569 return -1; 570 cur_wchunk = (struct rpcrdma_write_chunk *) (*iptrp + 1); 571 total_len = 0; 572 while (i--) { 573 struct rpcrdma_segment *seg = &cur_wchunk->wc_target; 574 ifdebug(FACILITY) { 575 u64 off; 576 xdr_decode_hyper((__be32 *)&seg->rs_offset, &off); 577 dprintk("RPC: %s: chunk %d@0x%llx:0x%x\n", 578 __func__, 579 ntohl(seg->rs_length), 580 (unsigned long long)off, 581 ntohl(seg->rs_handle)); 582 } 583 total_len += ntohl(seg->rs_length); 584 ++cur_wchunk; 585 } 586 /* check and adjust for properly terminated write chunk */ 587 if (wrchunk) { 588 __be32 *w = (__be32 *) cur_wchunk; 589 if (*w++ != xdr_zero) 590 return -1; 591 cur_wchunk = (struct rpcrdma_write_chunk *) w; 592 } 593 if ((char *) cur_wchunk > rep->rr_base + rep->rr_len) 594 return -1; 595 596 *iptrp = (__be32 *) cur_wchunk; 597 return total_len; 598 } 599 600 /* 601 * Scatter inline received data back into provided iov's. 602 */ 603 static void 604 rpcrdma_inline_fixup(struct rpc_rqst *rqst, char *srcp, int copy_len, int pad) 605 { 606 int i, npages, curlen, olen; 607 char *destp; 608 struct page **ppages; 609 int page_base; 610 611 curlen = rqst->rq_rcv_buf.head[0].iov_len; 612 if (curlen > copy_len) { /* write chunk header fixup */ 613 curlen = copy_len; 614 rqst->rq_rcv_buf.head[0].iov_len = curlen; 615 } 616 617 dprintk("RPC: %s: srcp 0x%p len %d hdrlen %d\n", 618 __func__, srcp, copy_len, curlen); 619 620 /* Shift pointer for first receive segment only */ 621 rqst->rq_rcv_buf.head[0].iov_base = srcp; 622 srcp += curlen; 623 copy_len -= curlen; 624 625 olen = copy_len; 626 i = 0; 627 rpcx_to_rdmax(rqst->rq_xprt)->rx_stats.fixup_copy_count += olen; 628 page_base = rqst->rq_rcv_buf.page_base; 629 ppages = rqst->rq_rcv_buf.pages + (page_base >> PAGE_SHIFT); 630 page_base &= ~PAGE_MASK; 631 632 if (copy_len && rqst->rq_rcv_buf.page_len) { 633 npages = PAGE_ALIGN(page_base + 634 rqst->rq_rcv_buf.page_len) >> PAGE_SHIFT; 635 for (; i < npages; i++) { 636 curlen = PAGE_SIZE - page_base; 637 if (curlen > copy_len) 638 curlen = copy_len; 639 dprintk("RPC: %s: page %d" 640 " srcp 0x%p len %d curlen %d\n", 641 __func__, i, srcp, copy_len, curlen); 642 destp = kmap_atomic(ppages[i]); 643 memcpy(destp + page_base, srcp, curlen); 644 flush_dcache_page(ppages[i]); 645 kunmap_atomic(destp); 646 srcp += curlen; 647 copy_len -= curlen; 648 if (copy_len == 0) 649 break; 650 page_base = 0; 651 } 652 rqst->rq_rcv_buf.page_len = olen - copy_len; 653 } else 654 rqst->rq_rcv_buf.page_len = 0; 655 656 if (copy_len && rqst->rq_rcv_buf.tail[0].iov_len) { 657 curlen = copy_len; 658 if (curlen > rqst->rq_rcv_buf.tail[0].iov_len) 659 curlen = rqst->rq_rcv_buf.tail[0].iov_len; 660 if (rqst->rq_rcv_buf.tail[0].iov_base != srcp) 661 memmove(rqst->rq_rcv_buf.tail[0].iov_base, srcp, curlen); 662 dprintk("RPC: %s: tail srcp 0x%p len %d curlen %d\n", 663 __func__, srcp, copy_len, curlen); 664 rqst->rq_rcv_buf.tail[0].iov_len = curlen; 665 copy_len -= curlen; ++i; 666 } else 667 rqst->rq_rcv_buf.tail[0].iov_len = 0; 668 669 if (pad) { 670 /* implicit padding on terminal chunk */ 671 unsigned char *p = rqst->rq_rcv_buf.tail[0].iov_base; 672 while (pad--) 673 p[rqst->rq_rcv_buf.tail[0].iov_len++] = 0; 674 } 675 676 if (copy_len) 677 dprintk("RPC: %s: %d bytes in" 678 " %d extra segments (%d lost)\n", 679 __func__, olen, i, copy_len); 680 681 /* TBD avoid a warning from call_decode() */ 682 rqst->rq_private_buf = rqst->rq_rcv_buf; 683 } 684 685 /* 686 * This function is called when an async event is posted to 687 * the connection which changes the connection state. All it 688 * does at this point is mark the connection up/down, the rpc 689 * timers do the rest. 690 */ 691 void 692 rpcrdma_conn_func(struct rpcrdma_ep *ep) 693 { 694 struct rpc_xprt *xprt = ep->rep_xprt; 695 696 spin_lock_bh(&xprt->transport_lock); 697 if (++xprt->connect_cookie == 0) /* maintain a reserved value */ 698 ++xprt->connect_cookie; 699 if (ep->rep_connected > 0) { 700 if (!xprt_test_and_set_connected(xprt)) 701 xprt_wake_pending_tasks(xprt, 0); 702 } else { 703 if (xprt_test_and_clear_connected(xprt)) 704 xprt_wake_pending_tasks(xprt, -ENOTCONN); 705 } 706 spin_unlock_bh(&xprt->transport_lock); 707 } 708 709 /* 710 * This function is called when memory window unbind which we are waiting 711 * for completes. Just use rr_func (zeroed by upcall) to signal completion. 712 */ 713 static void 714 rpcrdma_unbind_func(struct rpcrdma_rep *rep) 715 { 716 wake_up(&rep->rr_unbind); 717 } 718 719 /* 720 * Called as a tasklet to do req/reply match and complete a request 721 * Errors must result in the RPC task either being awakened, or 722 * allowed to timeout, to discover the errors at that time. 723 */ 724 void 725 rpcrdma_reply_handler(struct rpcrdma_rep *rep) 726 { 727 struct rpcrdma_msg *headerp; 728 struct rpcrdma_req *req; 729 struct rpc_rqst *rqst; 730 struct rpc_xprt *xprt = rep->rr_xprt; 731 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 732 __be32 *iptr; 733 int i, rdmalen, status; 734 735 /* Check status. If bad, signal disconnect and return rep to pool */ 736 if (rep->rr_len == ~0U) { 737 rpcrdma_recv_buffer_put(rep); 738 if (r_xprt->rx_ep.rep_connected == 1) { 739 r_xprt->rx_ep.rep_connected = -EIO; 740 rpcrdma_conn_func(&r_xprt->rx_ep); 741 } 742 return; 743 } 744 if (rep->rr_len < 28) { 745 dprintk("RPC: %s: short/invalid reply\n", __func__); 746 goto repost; 747 } 748 headerp = (struct rpcrdma_msg *) rep->rr_base; 749 if (headerp->rm_vers != xdr_one) { 750 dprintk("RPC: %s: invalid version %d\n", 751 __func__, ntohl(headerp->rm_vers)); 752 goto repost; 753 } 754 755 /* Get XID and try for a match. */ 756 spin_lock(&xprt->transport_lock); 757 rqst = xprt_lookup_rqst(xprt, headerp->rm_xid); 758 if (rqst == NULL) { 759 spin_unlock(&xprt->transport_lock); 760 dprintk("RPC: %s: reply 0x%p failed " 761 "to match any request xid 0x%08x len %d\n", 762 __func__, rep, headerp->rm_xid, rep->rr_len); 763 repost: 764 r_xprt->rx_stats.bad_reply_count++; 765 rep->rr_func = rpcrdma_reply_handler; 766 if (rpcrdma_ep_post_recv(&r_xprt->rx_ia, &r_xprt->rx_ep, rep)) 767 rpcrdma_recv_buffer_put(rep); 768 769 return; 770 } 771 772 /* get request object */ 773 req = rpcr_to_rdmar(rqst); 774 if (req->rl_reply) { 775 spin_unlock(&xprt->transport_lock); 776 dprintk("RPC: %s: duplicate reply 0x%p to RPC " 777 "request 0x%p: xid 0x%08x\n", __func__, rep, req, 778 headerp->rm_xid); 779 goto repost; 780 } 781 782 dprintk("RPC: %s: reply 0x%p completes request 0x%p\n" 783 " RPC request 0x%p xid 0x%08x\n", 784 __func__, rep, req, rqst, headerp->rm_xid); 785 786 /* from here on, the reply is no longer an orphan */ 787 req->rl_reply = rep; 788 789 /* check for expected message types */ 790 /* The order of some of these tests is important. */ 791 switch (headerp->rm_type) { 792 case htonl(RDMA_MSG): 793 /* never expect read chunks */ 794 /* never expect reply chunks (two ways to check) */ 795 /* never expect write chunks without having offered RDMA */ 796 if (headerp->rm_body.rm_chunks[0] != xdr_zero || 797 (headerp->rm_body.rm_chunks[1] == xdr_zero && 798 headerp->rm_body.rm_chunks[2] != xdr_zero) || 799 (headerp->rm_body.rm_chunks[1] != xdr_zero && 800 req->rl_nchunks == 0)) 801 goto badheader; 802 if (headerp->rm_body.rm_chunks[1] != xdr_zero) { 803 /* count any expected write chunks in read reply */ 804 /* start at write chunk array count */ 805 iptr = &headerp->rm_body.rm_chunks[2]; 806 rdmalen = rpcrdma_count_chunks(rep, 807 req->rl_nchunks, 1, &iptr); 808 /* check for validity, and no reply chunk after */ 809 if (rdmalen < 0 || *iptr++ != xdr_zero) 810 goto badheader; 811 rep->rr_len -= 812 ((unsigned char *)iptr - (unsigned char *)headerp); 813 status = rep->rr_len + rdmalen; 814 r_xprt->rx_stats.total_rdma_reply += rdmalen; 815 /* special case - last chunk may omit padding */ 816 if (rdmalen &= 3) { 817 rdmalen = 4 - rdmalen; 818 status += rdmalen; 819 } 820 } else { 821 /* else ordinary inline */ 822 rdmalen = 0; 823 iptr = (__be32 *)((unsigned char *)headerp + 28); 824 rep->rr_len -= 28; /*sizeof *headerp;*/ 825 status = rep->rr_len; 826 } 827 /* Fix up the rpc results for upper layer */ 828 rpcrdma_inline_fixup(rqst, (char *)iptr, rep->rr_len, rdmalen); 829 break; 830 831 case htonl(RDMA_NOMSG): 832 /* never expect read or write chunks, always reply chunks */ 833 if (headerp->rm_body.rm_chunks[0] != xdr_zero || 834 headerp->rm_body.rm_chunks[1] != xdr_zero || 835 headerp->rm_body.rm_chunks[2] != xdr_one || 836 req->rl_nchunks == 0) 837 goto badheader; 838 iptr = (__be32 *)((unsigned char *)headerp + 28); 839 rdmalen = rpcrdma_count_chunks(rep, req->rl_nchunks, 0, &iptr); 840 if (rdmalen < 0) 841 goto badheader; 842 r_xprt->rx_stats.total_rdma_reply += rdmalen; 843 /* Reply chunk buffer already is the reply vector - no fixup. */ 844 status = rdmalen; 845 break; 846 847 badheader: 848 default: 849 dprintk("%s: invalid rpcrdma reply header (type %d):" 850 " chunks[012] == %d %d %d" 851 " expected chunks <= %d\n", 852 __func__, ntohl(headerp->rm_type), 853 headerp->rm_body.rm_chunks[0], 854 headerp->rm_body.rm_chunks[1], 855 headerp->rm_body.rm_chunks[2], 856 req->rl_nchunks); 857 status = -EIO; 858 r_xprt->rx_stats.bad_reply_count++; 859 break; 860 } 861 862 /* If using mw bind, start the deregister process now. */ 863 /* (Note: if mr_free(), cannot perform it here, in tasklet context) */ 864 if (req->rl_nchunks) switch (r_xprt->rx_ia.ri_memreg_strategy) { 865 case RPCRDMA_MEMWINDOWS: 866 for (i = 0; req->rl_nchunks-- > 1;) 867 i += rpcrdma_deregister_external( 868 &req->rl_segments[i], r_xprt, NULL); 869 /* Optionally wait (not here) for unbinds to complete */ 870 rep->rr_func = rpcrdma_unbind_func; 871 (void) rpcrdma_deregister_external(&req->rl_segments[i], 872 r_xprt, rep); 873 break; 874 case RPCRDMA_MEMWINDOWS_ASYNC: 875 for (i = 0; req->rl_nchunks--;) 876 i += rpcrdma_deregister_external(&req->rl_segments[i], 877 r_xprt, NULL); 878 break; 879 default: 880 break; 881 } 882 883 dprintk("RPC: %s: xprt_complete_rqst(0x%p, 0x%p, %d)\n", 884 __func__, xprt, rqst, status); 885 xprt_complete_rqst(rqst->rq_task, status); 886 spin_unlock(&xprt->transport_lock); 887 } 888