1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 /* 22 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. 23 */ 24 25 /* 26 * Copyright (c) 2007, The Ohio State University. All rights reserved. 27 * 28 * Portions of this source code is developed by the team members of 29 * The Ohio State University's Network-Based Computing Laboratory (NBCL), 30 * headed by Professor Dhabaleswar K. (DK) Panda. 31 * 32 * Acknowledgements to contributions from developors: 33 * Ranjit Noronha: noronha@cse.ohio-state.edu 34 * Lei Chai : chail@cse.ohio-state.edu 35 * Weikuan Yu : yuw@cse.ohio-state.edu 36 * 37 */ 38 39 /* 40 * xdr_rdma.c, XDR implementation using RDMA to move large chunks 41 */ 42 43 #include <sys/param.h> 44 #include <sys/types.h> 45 #include <sys/systm.h> 46 #include <sys/kmem.h> 47 #include <sys/sdt.h> 48 #include <sys/debug.h> 49 50 #include <rpc/types.h> 51 #include <rpc/xdr.h> 52 #include <sys/cmn_err.h> 53 #include <rpc/rpc_sztypes.h> 54 #include <rpc/rpc_rdma.h> 55 #include <sys/sysmacros.h> 56 57 /* 58 * RCP header and xdr encoding overhead. The number was determined by 59 * tracing the msglen in svc_rdma_ksend for sec=sys,krb5,krb5i and krb5p. 60 * If the XDR_RDMA_BUF_OVERHEAD is not large enough the result is the trigger 61 * of the dtrace probe on the server "krpc-e-svcrdma-ksend-noreplycl" from 62 * svc_rdma_ksend. 63 */ 64 #define XDR_RDMA_BUF_OVERHEAD 300 65 66 static bool_t xdrrdma_getint32(XDR *, int32_t *); 67 static bool_t xdrrdma_putint32(XDR *, int32_t *); 68 static bool_t xdrrdma_getbytes(XDR *, caddr_t, int); 69 static bool_t xdrrdma_putbytes(XDR *, caddr_t, int); 70 uint_t xdrrdma_getpos(XDR *); 71 bool_t xdrrdma_setpos(XDR *, uint_t); 72 static rpc_inline_t *xdrrdma_inline(XDR *, int); 73 void xdrrdma_destroy(XDR *); 74 static bool_t xdrrdma_control(XDR *, int, void *); 75 static bool_t xdrrdma_read_a_chunk(XDR *, CONN **); 76 static void xdrrdma_free_xdr_chunks(CONN *, struct clist *); 77 78 struct xdr_ops xdrrdmablk_ops = { 79 xdrrdma_getbytes, 80 xdrrdma_putbytes, 81 xdrrdma_getpos, 82 xdrrdma_setpos, 83 xdrrdma_inline, 84 xdrrdma_destroy, 85 xdrrdma_control, 86 xdrrdma_getint32, 87 xdrrdma_putint32 88 }; 89 90 struct xdr_ops xdrrdma_ops = { 91 xdrrdma_getbytes, 92 xdrrdma_putbytes, 93 xdrrdma_getpos, 94 xdrrdma_setpos, 95 xdrrdma_inline, 96 xdrrdma_destroy, 97 xdrrdma_control, 98 xdrrdma_getint32, 99 xdrrdma_putint32 100 }; 101 102 /* 103 * A chunk list entry identifies a chunk of opaque data to be moved 104 * separately from the rest of the RPC message. xp_min_chunk = 0, is a 105 * special case for ENCODING, which means do not chunk the incoming stream of 106 * data. 107 * 108 * A read chunk can contain part of the RPC message in addition to the 109 * inline message. In such a case, (xp_offp - x_base) will not provide 110 * the correct xdr offset of the entire message. xp_off is used in such 111 * a case to denote the offset or current position in the overall message 112 * covering both the inline and the chunk. This is used only in the case 113 * of decoding and useful to compare read chunk 'c_xdroff' offsets. 114 * 115 * An example for a read chunk containing an XDR message: 116 * An NFSv4 compound as following: 117 * 118 * PUTFH 119 * WRITE [4109 bytes] 120 * GETATTR 121 * 122 * Solaris Encoding is: 123 * ------------------- 124 * 125 * <Inline message>: [PUTFH WRITE4args GETATTR] 126 * | 127 * v 128 * [RDMA_READ chunks]: [write data] 129 * 130 * 131 * Linux encoding is: 132 * ----------------- 133 * 134 * <Inline message>: [PUTFH WRITE4args] 135 * | 136 * v 137 * [RDMA_READ chunks]: [Write data] [Write data2] [Getattr chunk] 138 * chunk1 chunk2 chunk3 139 * 140 * where the READ chunks are as: 141 * 142 * - chunk1 - 4k 143 * write data | 144 * - chunk2 - 13 bytes(4109 - 4k) 145 * getattr op - chunk3 - 19 bytes 146 * (getattr op starts at byte 4 after 3 bytes of roundup) 147 * 148 */ 149 150 typedef struct { 151 caddr_t xp_offp; 152 int xp_min_chunk; 153 uint_t xp_flags; /* Controls setting for rdma xdr */ 154 int xp_buf_size; /* size of xdr buffer */ 155 int xp_off; /* overall offset */ 156 struct clist *xp_rcl; /* head of chunk list */ 157 struct clist **xp_rcl_next; /* location to place/find next chunk */ 158 struct clist *xp_rcl_xdr; /* copy of rcl containing RPC message */ 159 struct clist *xp_wcl; /* head of write chunk list */ 160 CONN *xp_conn; /* connection for chunk data xfer */ 161 uint_t xp_reply_chunk_len; 162 /* used to track length for security modes: integrity/privacy */ 163 uint_t xp_reply_chunk_len_alt; 164 } xrdma_private_t; 165 166 extern kmem_cache_t *clist_cache; 167 168 bool_t 169 xdrrdma_getrdmablk(XDR *xdrs, struct clist **rlist, uint_t *sizep, 170 CONN **conn, const uint_t maxsize) 171 { 172 xrdma_private_t *xdrp = (xrdma_private_t *)(xdrs->x_private); 173 struct clist *cle = *(xdrp->xp_rcl_next); 174 struct clist *rdclist = NULL, *prev = NULL; 175 bool_t retval = TRUE; 176 uint32_t cur_offset = 0; 177 uint32_t total_segments = 0; 178 uint32_t actual_segments = 0; 179 uint32_t alen; 180 uint_t total_len; 181 182 ASSERT(xdrs->x_op != XDR_FREE); 183 184 /* 185 * first deal with the length since xdr bytes are counted 186 */ 187 if (!xdr_u_int(xdrs, sizep)) { 188 DTRACE_PROBE(xdr__e__getrdmablk_sizep_fail); 189 return (FALSE); 190 } 191 total_len = *sizep; 192 if (total_len > maxsize) { 193 DTRACE_PROBE2(xdr__e__getrdmablk_bad_size, 194 int, total_len, int, maxsize); 195 return (FALSE); 196 } 197 (*conn) = xdrp->xp_conn; 198 199 /* 200 * if no data we are done 201 */ 202 if (total_len == 0) 203 return (TRUE); 204 205 while (cle) { 206 total_segments++; 207 cle = cle->c_next; 208 } 209 210 cle = *(xdrp->xp_rcl_next); 211 212 /* 213 * If there was a chunk at the current offset, then setup a read 214 * chunk list which records the destination address and length 215 * and will RDMA READ the data in later. 216 */ 217 if (cle == NULL) 218 return (FALSE); 219 220 if (cle->c_xdroff != (xdrp->xp_offp - xdrs->x_base)) 221 return (FALSE); 222 223 /* 224 * Setup the chunk list with appropriate 225 * address (offset) and length 226 */ 227 for (actual_segments = 0; 228 actual_segments < total_segments; actual_segments++) { 229 230 DTRACE_PROBE3(krpc__i__xdrrdma_getrdmablk, uint32_t, cle->c_len, 231 uint32_t, total_len, uint32_t, cle->c_xdroff); 232 233 if (total_len <= 0) 234 break; 235 236 /* 237 * not the first time in the loop 238 */ 239 if (actual_segments > 0) 240 cle = cle->c_next; 241 242 cle->u.c_daddr = (uint64) cur_offset; 243 alen = 0; 244 if (cle->c_len > total_len) { 245 alen = cle->c_len; 246 cle->c_len = total_len; 247 } 248 if (!alen) 249 xdrp->xp_rcl_next = &cle->c_next; 250 251 cur_offset += cle->c_len; 252 total_len -= cle->c_len; 253 254 if ((total_segments - actual_segments - 1) == 0 && 255 total_len > 0) { 256 DTRACE_PROBE(krpc__e__xdrrdma_getblk_chunktooshort); 257 retval = FALSE; 258 } 259 260 if ((total_segments - actual_segments - 1) > 0 && 261 total_len == 0) { 262 DTRACE_PROBE2(krpc__e__xdrrdma_getblk_toobig, 263 int, total_segments, int, actual_segments); 264 } 265 266 rdclist = clist_alloc(); 267 (*rdclist) = (*cle); 268 if ((*rlist) == NULL) 269 (*rlist) = rdclist; 270 if (prev == NULL) 271 prev = rdclist; 272 else { 273 prev->c_next = rdclist; 274 prev = rdclist; 275 } 276 277 } 278 279 out: 280 if (prev != NULL) 281 prev->c_next = NULL; 282 283 /* 284 * Adjust the chunk length, if we read only a part of 285 * a chunk. 286 */ 287 288 if (alen) { 289 cle->w.c_saddr = 290 (uint64)(uintptr_t)cle->w.c_saddr + cle->c_len; 291 cle->c_len = alen - cle->c_len; 292 } 293 294 return (retval); 295 } 296 297 /* 298 * The procedure xdrrdma_create initializes a stream descriptor for a memory 299 * buffer. 300 */ 301 void 302 xdrrdma_create(XDR *xdrs, caddr_t addr, uint_t size, 303 int min_chunk, struct clist *cl, enum xdr_op op, CONN *conn) 304 { 305 xrdma_private_t *xdrp; 306 struct clist *cle; 307 308 xdrs->x_op = op; 309 xdrs->x_ops = &xdrrdma_ops; 310 xdrs->x_base = addr; 311 xdrs->x_handy = size; 312 xdrs->x_public = NULL; 313 314 xdrp = (xrdma_private_t *)kmem_zalloc(sizeof (xrdma_private_t), 315 KM_SLEEP); 316 xdrs->x_private = (caddr_t)xdrp; 317 xdrp->xp_offp = addr; 318 xdrp->xp_min_chunk = min_chunk; 319 xdrp->xp_flags = 0; 320 xdrp->xp_buf_size = size; 321 xdrp->xp_rcl = cl; 322 xdrp->xp_reply_chunk_len = 0; 323 xdrp->xp_reply_chunk_len_alt = 0; 324 325 if (op == XDR_ENCODE && cl != NULL) { 326 /* Find last element in chunk list and set xp_rcl_next */ 327 for (cle = cl; cle->c_next != NULL; cle = cle->c_next) 328 continue; 329 330 xdrp->xp_rcl_next = &(cle->c_next); 331 } else { 332 xdrp->xp_rcl_next = &(xdrp->xp_rcl); 333 } 334 335 xdrp->xp_wcl = NULL; 336 337 xdrp->xp_conn = conn; 338 if (xdrp->xp_min_chunk != 0) 339 xdrp->xp_flags |= XDR_RDMA_CHUNK; 340 } 341 342 /* ARGSUSED */ 343 void 344 xdrrdma_destroy(XDR * xdrs) 345 { 346 xrdma_private_t *xdrp = (xrdma_private_t *)(xdrs->x_private); 347 348 if (xdrp == NULL) 349 return; 350 351 if (xdrp->xp_wcl) { 352 if (xdrp->xp_flags & XDR_RDMA_WLIST_REG) { 353 (void) clist_deregister(xdrp->xp_conn, xdrp->xp_wcl); 354 rdma_buf_free(xdrp->xp_conn, 355 &xdrp->xp_wcl->rb_longbuf); 356 } 357 clist_free(xdrp->xp_wcl); 358 } 359 360 if (xdrp->xp_rcl) { 361 if (xdrp->xp_flags & XDR_RDMA_RLIST_REG) { 362 (void) clist_deregister(xdrp->xp_conn, xdrp->xp_rcl); 363 rdma_buf_free(xdrp->xp_conn, 364 &xdrp->xp_rcl->rb_longbuf); 365 } 366 clist_free(xdrp->xp_rcl); 367 } 368 369 if (xdrp->xp_rcl_xdr) 370 xdrrdma_free_xdr_chunks(xdrp->xp_conn, xdrp->xp_rcl_xdr); 371 372 (void) kmem_free(xdrs->x_private, sizeof (xrdma_private_t)); 373 xdrs->x_private = NULL; 374 } 375 376 static bool_t 377 xdrrdma_getint32(XDR *xdrs, int32_t *int32p) 378 { 379 xrdma_private_t *xdrp = (xrdma_private_t *)(xdrs->x_private); 380 int chunked = 0; 381 382 if ((xdrs->x_handy -= (int)sizeof (int32_t)) < 0) { 383 /* 384 * check if rest of the rpc message is in a chunk 385 */ 386 if (!xdrrdma_read_a_chunk(xdrs, &xdrp->xp_conn)) { 387 return (FALSE); 388 } 389 chunked = 1; 390 } 391 392 /* LINTED pointer alignment */ 393 *int32p = (int32_t)ntohl((uint32_t)(*((int32_t *)(xdrp->xp_offp)))); 394 395 DTRACE_PROBE1(krpc__i__xdrrdma_getint32, int32_t, *int32p); 396 397 xdrp->xp_offp += sizeof (int32_t); 398 399 if (chunked) 400 xdrs->x_handy -= (int)sizeof (int32_t); 401 402 if (xdrp->xp_off != 0) { 403 xdrp->xp_off += sizeof (int32_t); 404 } 405 406 return (TRUE); 407 } 408 409 static bool_t 410 xdrrdma_putint32(XDR *xdrs, int32_t *int32p) 411 { 412 xrdma_private_t *xdrp = (xrdma_private_t *)(xdrs->x_private); 413 414 if ((xdrs->x_handy -= (int)sizeof (int32_t)) < 0) 415 return (FALSE); 416 417 /* LINTED pointer alignment */ 418 *(int32_t *)xdrp->xp_offp = (int32_t)htonl((uint32_t)(*int32p)); 419 xdrp->xp_offp += sizeof (int32_t); 420 421 return (TRUE); 422 } 423 424 /* 425 * DECODE bytes from XDR stream for rdma. 426 * If the XDR stream contains a read chunk list, 427 * it will go through xdrrdma_getrdmablk instead. 428 */ 429 static bool_t 430 xdrrdma_getbytes(XDR *xdrs, caddr_t addr, int len) 431 { 432 xrdma_private_t *xdrp = (xrdma_private_t *)(xdrs->x_private); 433 struct clist *cle = *(xdrp->xp_rcl_next); 434 struct clist *cls = *(xdrp->xp_rcl_next); 435 struct clist cl; 436 bool_t retval = TRUE; 437 uint32_t total_len = len; 438 uint32_t cur_offset = 0; 439 uint32_t total_segments = 0; 440 uint32_t actual_segments = 0; 441 uint32_t status = RDMA_SUCCESS; 442 uint32_t alen = 0; 443 uint32_t xpoff; 444 445 while (cle) { 446 total_segments++; 447 cle = cle->c_next; 448 } 449 450 cle = *(xdrp->xp_rcl_next); 451 452 if (xdrp->xp_off) { 453 xpoff = xdrp->xp_off; 454 } else { 455 xpoff = (xdrp->xp_offp - xdrs->x_base); 456 } 457 458 /* 459 * If there was a chunk at the current offset, then setup a read 460 * chunk list which records the destination address and length 461 * and will RDMA READ the data in later. 462 */ 463 464 if (cle != NULL && cle->c_xdroff == xpoff) { 465 for (actual_segments = 0; 466 actual_segments < total_segments; actual_segments++) { 467 468 if (total_len <= 0) 469 break; 470 471 if (status != RDMA_SUCCESS) 472 goto out; 473 474 cle->u.c_daddr = (uint64)(uintptr_t)addr + cur_offset; 475 alen = 0; 476 if (cle->c_len > total_len) { 477 alen = cle->c_len; 478 cle->c_len = total_len; 479 } 480 if (!alen) 481 xdrp->xp_rcl_next = &cle->c_next; 482 483 cur_offset += cle->c_len; 484 total_len -= cle->c_len; 485 486 if ((total_segments - actual_segments - 1) == 0 && 487 total_len > 0) { 488 DTRACE_PROBE( 489 krpc__e__xdrrdma_getbytes_chunktooshort); 490 retval = FALSE; 491 } 492 493 if ((total_segments - actual_segments - 1) > 0 && 494 total_len == 0) { 495 DTRACE_PROBE2(krpc__e__xdrrdma_getbytes_toobig, 496 int, total_segments, int, actual_segments); 497 } 498 499 /* 500 * RDMA READ the chunk data from the remote end. 501 * First prep the destination buffer by registering 502 * it, then RDMA READ the chunk data. Since we are 503 * doing streaming memory, sync the destination 504 * buffer to CPU and deregister the buffer. 505 */ 506 if (xdrp->xp_conn == NULL) { 507 return (FALSE); 508 } 509 cl = *cle; 510 cl.c_next = NULL; 511 status = clist_register(xdrp->xp_conn, &cl, 512 CLIST_REG_DST); 513 if (status != RDMA_SUCCESS) { 514 retval = FALSE; 515 /* 516 * Deregister the previous chunks 517 * before return 518 */ 519 goto out; 520 } 521 522 cle->c_dmemhandle = cl.c_dmemhandle; 523 cle->c_dsynchandle = cl.c_dsynchandle; 524 525 /* 526 * Now read the chunk in 527 */ 528 if ((total_segments - actual_segments - 1) == 0 || 529 total_len == 0) { 530 status = RDMA_READ(xdrp->xp_conn, &cl, WAIT); 531 } else { 532 status = RDMA_READ(xdrp->xp_conn, &cl, NOWAIT); 533 } 534 if (status != RDMA_SUCCESS) { 535 DTRACE_PROBE1( 536 krpc__i__xdrrdma_getblk_readfailed, 537 int, status); 538 retval = FALSE; 539 } 540 541 cle = cle->c_next; 542 543 } 544 545 /* 546 * sync the memory for cpu 547 */ 548 cl = *cls; 549 cl.c_next = NULL; 550 cl.c_len = cur_offset; 551 if (clist_syncmem( 552 xdrp->xp_conn, &cl, CLIST_REG_DST) != RDMA_SUCCESS) { 553 retval = FALSE; 554 } 555 out: 556 557 /* 558 * Deregister the chunks 559 */ 560 cle = cls; 561 while (actual_segments != 0) { 562 cl = *cle; 563 cl.c_next = NULL; 564 565 cl.c_regtype = CLIST_REG_DST; 566 (void) clist_deregister(xdrp->xp_conn, &cl); 567 568 cle = cle->c_next; 569 actual_segments--; 570 } 571 572 if (alen) { 573 cle = *(xdrp->xp_rcl_next); 574 cle->w.c_saddr = 575 (uint64)(uintptr_t)cle->w.c_saddr + cle->c_len; 576 cle->c_len = alen - cle->c_len; 577 } 578 579 return (retval); 580 } 581 582 if ((xdrs->x_handy -= len) < 0) 583 return (FALSE); 584 585 bcopy(xdrp->xp_offp, addr, len); 586 587 xdrp->xp_offp += len; 588 589 if (xdrp->xp_off != 0) 590 xdrp->xp_off += len; 591 592 return (TRUE); 593 } 594 595 /* 596 * ENCODE some bytes into an XDR stream xp_min_chunk = 0, means the stream of 597 * bytes contain no chunks to seperate out, and if the bytes do not fit in 598 * the supplied buffer, grow the buffer and free the old buffer. 599 */ 600 static bool_t 601 xdrrdma_putbytes(XDR *xdrs, caddr_t addr, int len) 602 { 603 xrdma_private_t *xdrp = (xrdma_private_t *)(xdrs->x_private); 604 /* 605 * Is this stream accepting chunks? 606 * If so, does the either of the two following conditions exist? 607 * - length of bytes to encode is greater than the min chunk size? 608 * - remaining space in this stream is shorter than length of 609 * bytes to encode? 610 * 611 * If the above exists, then create a chunk for this encoding 612 * and save the addresses, etc. 613 */ 614 if (xdrp->xp_flags & XDR_RDMA_CHUNK && 615 ((xdrp->xp_min_chunk != 0 && 616 len >= xdrp->xp_min_chunk) || 617 (xdrs->x_handy - len < 0))) { 618 struct clist *cle; 619 int offset = xdrp->xp_offp - xdrs->x_base; 620 621 cle = clist_alloc(); 622 cle->c_xdroff = offset; 623 cle->c_len = len; 624 cle->w.c_saddr = (uint64)(uintptr_t)addr; 625 cle->c_next = NULL; 626 627 *(xdrp->xp_rcl_next) = cle; 628 xdrp->xp_rcl_next = &(cle->c_next); 629 630 return (TRUE); 631 } 632 /* Is there enough space to encode what is left? */ 633 if ((xdrs->x_handy -= len) < 0) { 634 return (FALSE); 635 } 636 bcopy(addr, xdrp->xp_offp, len); 637 xdrp->xp_offp += len; 638 639 return (TRUE); 640 } 641 642 uint_t 643 xdrrdma_getpos(XDR *xdrs) 644 { 645 xrdma_private_t *xdrp = (xrdma_private_t *)(xdrs->x_private); 646 647 return ((uint_t)((uintptr_t)xdrp->xp_offp - (uintptr_t)xdrs->x_base)); 648 } 649 650 bool_t 651 xdrrdma_setpos(XDR *xdrs, uint_t pos) 652 { 653 xrdma_private_t *xdrp = (xrdma_private_t *)(xdrs->x_private); 654 655 caddr_t newaddr = xdrs->x_base + pos; 656 caddr_t lastaddr = xdrp->xp_offp + xdrs->x_handy; 657 ptrdiff_t diff; 658 659 if (newaddr > lastaddr) 660 return (FALSE); 661 662 xdrp->xp_offp = newaddr; 663 diff = lastaddr - newaddr; 664 xdrs->x_handy = (int)diff; 665 666 return (TRUE); 667 } 668 669 /* ARGSUSED */ 670 static rpc_inline_t * 671 xdrrdma_inline(XDR *xdrs, int len) 672 { 673 rpc_inline_t *buf = NULL; 674 xrdma_private_t *xdrp = (xrdma_private_t *)(xdrs->x_private); 675 struct clist *cle = *(xdrp->xp_rcl_next); 676 677 if (xdrs->x_op == XDR_DECODE) { 678 /* 679 * Since chunks aren't in-line, check to see whether there is 680 * a chunk in the inline range. 681 */ 682 if (cle != NULL && 683 cle->c_xdroff <= (xdrp->xp_offp - xdrs->x_base + len)) 684 return (NULL); 685 } 686 687 /* LINTED pointer alignment */ 688 buf = (rpc_inline_t *)xdrp->xp_offp; 689 if (!IS_P2ALIGNED(buf, sizeof (int32_t))) 690 return (NULL); 691 692 if ((xdrs->x_handy < len) || (xdrp->xp_min_chunk != 0 && 693 len >= xdrp->xp_min_chunk)) { 694 return (NULL); 695 } else { 696 xdrs->x_handy -= len; 697 xdrp->xp_offp += len; 698 return (buf); 699 } 700 } 701 702 static bool_t 703 xdrrdma_control(XDR *xdrs, int request, void *info) 704 { 705 int32_t *int32p; 706 int len, i; 707 uint_t in_flags; 708 xrdma_private_t *xdrp = (xrdma_private_t *)(xdrs->x_private); 709 rdma_chunkinfo_t *rcip = NULL; 710 rdma_wlist_conn_info_t *rwcip = NULL; 711 rdma_chunkinfo_lengths_t *rcilp = NULL; 712 struct uio *uiop; 713 struct clist *rwl = NULL; 714 struct clist *prev = NULL; 715 716 switch (request) { 717 case XDR_PEEK: 718 /* 719 * Return the next 4 byte unit in the XDR stream. 720 */ 721 if (xdrs->x_handy < sizeof (int32_t)) 722 return (FALSE); 723 724 int32p = (int32_t *)info; 725 *int32p = (int32_t)ntohl((uint32_t) 726 (*((int32_t *)(xdrp->xp_offp)))); 727 728 return (TRUE); 729 730 case XDR_SKIPBYTES: 731 /* 732 * Skip the next N bytes in the XDR stream. 733 */ 734 int32p = (int32_t *)info; 735 len = RNDUP((int)(*int32p)); 736 if ((xdrs->x_handy -= len) < 0) 737 return (FALSE); 738 xdrp->xp_offp += len; 739 740 return (TRUE); 741 742 case XDR_RDMA_SET_FLAGS: 743 /* 744 * Set the flags provided in the *info in xp_flags for rdma 745 * xdr stream control. 746 */ 747 int32p = (int32_t *)info; 748 in_flags = (uint_t)(*int32p); 749 750 xdrp->xp_flags |= in_flags; 751 return (TRUE); 752 753 case XDR_RDMA_GET_FLAGS: 754 /* 755 * Get the flags provided in xp_flags return through *info 756 */ 757 int32p = (int32_t *)info; 758 759 *int32p = (int32_t)xdrp->xp_flags; 760 return (TRUE); 761 762 case XDR_RDMA_GET_CHUNK_LEN: 763 rcilp = (rdma_chunkinfo_lengths_t *)info; 764 rcilp->rcil_len = xdrp->xp_reply_chunk_len; 765 rcilp->rcil_len_alt = xdrp->xp_reply_chunk_len_alt; 766 767 return (TRUE); 768 769 case XDR_RDMA_ADD_CHUNK: 770 /* 771 * Store wlist information 772 */ 773 774 rcip = (rdma_chunkinfo_t *)info; 775 776 DTRACE_PROBE2(krpc__i__xdrrdma__control__add__chunk, 777 rci_type_t, rcip->rci_type, uint32, rcip->rci_len); 778 switch (rcip->rci_type) { 779 case RCI_WRITE_UIO_CHUNK: 780 xdrp->xp_reply_chunk_len_alt += rcip->rci_len; 781 782 if ((rcip->rci_len + XDR_RDMA_BUF_OVERHEAD) < 783 xdrp->xp_min_chunk) { 784 xdrp->xp_wcl = NULL; 785 *(rcip->rci_clpp) = NULL; 786 return (TRUE); 787 } 788 uiop = rcip->rci_a.rci_uiop; 789 790 for (i = 0; i < uiop->uio_iovcnt; i++) { 791 rwl = clist_alloc(); 792 rwl->c_len = uiop->uio_iov[i].iov_len; 793 rwl->u.c_daddr = 794 (uint64)(uintptr_t) 795 (uiop->uio_iov[i].iov_base); 796 /* 797 * if userspace address, put adspace ptr in 798 * clist. If not, then do nothing since it's 799 * already set to NULL (from kmem_zalloc) 800 */ 801 if (uiop->uio_segflg == UIO_USERSPACE) { 802 rwl->c_adspc = ttoproc(curthread)->p_as; 803 } 804 805 if (prev == NULL) 806 prev = rwl; 807 else { 808 prev->c_next = rwl; 809 prev = rwl; 810 } 811 } 812 813 rwl->c_next = NULL; 814 xdrp->xp_wcl = rwl; 815 *(rcip->rci_clpp) = rwl; 816 817 break; 818 819 case RCI_WRITE_ADDR_CHUNK: 820 rwl = clist_alloc(); 821 822 rwl->c_len = rcip->rci_len; 823 rwl->u.c_daddr3 = rcip->rci_a.rci_addr; 824 rwl->c_next = NULL; 825 xdrp->xp_reply_chunk_len_alt += rcip->rci_len; 826 827 xdrp->xp_wcl = rwl; 828 *(rcip->rci_clpp) = rwl; 829 830 break; 831 832 case RCI_REPLY_CHUNK: 833 xdrp->xp_reply_chunk_len += rcip->rci_len; 834 break; 835 } 836 return (TRUE); 837 838 case XDR_RDMA_GET_WLIST: 839 *((struct clist **)info) = xdrp->xp_wcl; 840 return (TRUE); 841 842 case XDR_RDMA_SET_WLIST: 843 xdrp->xp_wcl = (struct clist *)info; 844 return (TRUE); 845 846 case XDR_RDMA_GET_RLIST: 847 *((struct clist **)info) = xdrp->xp_rcl; 848 return (TRUE); 849 850 case XDR_RDMA_GET_WCINFO: 851 rwcip = (rdma_wlist_conn_info_t *)info; 852 853 rwcip->rwci_wlist = xdrp->xp_wcl; 854 rwcip->rwci_conn = xdrp->xp_conn; 855 856 return (TRUE); 857 858 default: 859 return (FALSE); 860 } 861 } 862 863 bool_t xdr_do_clist(XDR *, clist **); 864 865 /* 866 * Not all fields in struct clist are interesting to the RPC over RDMA 867 * protocol. Only XDR the interesting fields. 868 */ 869 bool_t 870 xdr_clist(XDR *xdrs, clist *objp) 871 { 872 if (!xdr_uint32(xdrs, &objp->c_xdroff)) 873 return (FALSE); 874 if (!xdr_uint32(xdrs, &objp->c_smemhandle.mrc_rmr)) 875 return (FALSE); 876 if (!xdr_uint32(xdrs, &objp->c_len)) 877 return (FALSE); 878 if (!xdr_uint64(xdrs, &objp->w.c_saddr)) 879 return (FALSE); 880 if (!xdr_do_clist(xdrs, &objp->c_next)) 881 return (FALSE); 882 return (TRUE); 883 } 884 885 /* 886 * The following two functions are forms of xdr_pointer() 887 * and xdr_reference(). Since the generic versions just 888 * kmem_alloc() a new clist, we actually want to use the 889 * rdma_clist kmem_cache. 890 */ 891 892 /* 893 * Generate or free a clist structure from the 894 * kmem_cache "rdma_clist" 895 */ 896 bool_t 897 xdr_ref_clist(XDR *xdrs, caddr_t *pp) 898 { 899 caddr_t loc = *pp; 900 bool_t stat; 901 902 if (loc == NULL) { 903 switch (xdrs->x_op) { 904 case XDR_FREE: 905 return (TRUE); 906 907 case XDR_DECODE: 908 *pp = loc = (caddr_t)clist_alloc(); 909 break; 910 911 case XDR_ENCODE: 912 ASSERT(loc); 913 break; 914 } 915 } 916 917 stat = xdr_clist(xdrs, (struct clist *)loc); 918 919 if (xdrs->x_op == XDR_FREE) { 920 kmem_cache_free(clist_cache, loc); 921 *pp = NULL; 922 } 923 return (stat); 924 } 925 926 /* 927 * XDR a pointer to a possibly recursive clist. This differs 928 * with xdr_reference in that it can serialize/deserialiaze 929 * trees correctly. 930 * 931 * What is sent is actually a union: 932 * 933 * union object_pointer switch (boolean b) { 934 * case TRUE: object_data data; 935 * case FALSE: void nothing; 936 * } 937 * 938 * > objpp: Pointer to the pointer to the object. 939 * 940 */ 941 942 bool_t 943 xdr_do_clist(XDR *xdrs, clist **objpp) 944 { 945 bool_t more_data; 946 947 more_data = (*objpp != NULL); 948 if (!xdr_bool(xdrs, &more_data)) 949 return (FALSE); 950 if (!more_data) { 951 *objpp = NULL; 952 return (TRUE); 953 } 954 return (xdr_ref_clist(xdrs, (caddr_t *)objpp)); 955 } 956 957 uint_t 958 xdr_getbufsize(XDR *xdrs) 959 { 960 xrdma_private_t *xdrp = (xrdma_private_t *)(xdrs->x_private); 961 962 return ((uint_t)xdrp->xp_buf_size); 963 } 964 965 /* ARGSUSED */ 966 bool_t 967 xdr_encode_rlist_svc(XDR *xdrs, clist *rlist) 968 { 969 bool_t vfalse = FALSE; 970 971 ASSERT(rlist == NULL); 972 return (xdr_bool(xdrs, &vfalse)); 973 } 974 975 bool_t 976 xdr_encode_wlist(XDR *xdrs, clist *w) 977 { 978 bool_t vfalse = FALSE, vtrue = TRUE; 979 int i; 980 uint_t num_segment = 0; 981 struct clist *cl; 982 983 /* does a wlist exist? */ 984 if (w == NULL) { 985 return (xdr_bool(xdrs, &vfalse)); 986 } 987 /* Encode N consecutive segments, 1, N, HLOO, ..., HLOO, 0 */ 988 if (!xdr_bool(xdrs, &vtrue)) 989 return (FALSE); 990 991 for (cl = w; cl != NULL; cl = cl->c_next) { 992 num_segment++; 993 } 994 995 if (!xdr_uint32(xdrs, &num_segment)) 996 return (FALSE); 997 for (i = 0; i < num_segment; i++) { 998 999 DTRACE_PROBE1(krpc__i__xdr_encode_wlist_len, uint_t, w->c_len); 1000 1001 if (!xdr_uint32(xdrs, &w->c_dmemhandle.mrc_rmr)) 1002 return (FALSE); 1003 1004 if (!xdr_uint32(xdrs, &w->c_len)) 1005 return (FALSE); 1006 1007 if (!xdr_uint64(xdrs, &w->u.c_daddr)) 1008 return (FALSE); 1009 1010 w = w->c_next; 1011 } 1012 1013 if (!xdr_bool(xdrs, &vfalse)) 1014 return (FALSE); 1015 1016 return (TRUE); 1017 } 1018 1019 1020 /* 1021 * Conditionally decode a RDMA WRITE chunk list from XDR stream. 1022 * 1023 * If the next boolean in the XDR stream is false there is no 1024 * RDMA WRITE chunk list present. Otherwise iterate over the 1025 * array and for each entry: allocate a struct clist and decode. 1026 * Pass back an indication via wlist_exists if we have seen a 1027 * RDMA WRITE chunk list. 1028 */ 1029 bool_t 1030 xdr_decode_wlist(XDR *xdrs, struct clist **w, bool_t *wlist_exists) 1031 { 1032 struct clist *tmp; 1033 bool_t more = FALSE; 1034 uint32_t seg_array_len; 1035 uint32_t i; 1036 1037 if (!xdr_bool(xdrs, &more)) 1038 return (FALSE); 1039 1040 /* is there a wlist? */ 1041 if (more == FALSE) { 1042 *wlist_exists = FALSE; 1043 return (TRUE); 1044 } 1045 *wlist_exists = TRUE; 1046 1047 if (!xdr_uint32(xdrs, &seg_array_len)) 1048 return (FALSE); 1049 1050 tmp = *w = clist_alloc(); 1051 for (i = 0; i < seg_array_len; i++) { 1052 1053 if (!xdr_uint32(xdrs, &tmp->c_dmemhandle.mrc_rmr)) 1054 return (FALSE); 1055 if (!xdr_uint32(xdrs, &tmp->c_len)) 1056 return (FALSE); 1057 1058 DTRACE_PROBE1(krpc__i__xdr_decode_wlist_len, 1059 uint_t, tmp->c_len); 1060 1061 if (!xdr_uint64(xdrs, &tmp->u.c_daddr)) 1062 return (FALSE); 1063 if (i < seg_array_len - 1) { 1064 tmp->c_next = clist_alloc(); 1065 tmp = tmp->c_next; 1066 } else { 1067 tmp->c_next = NULL; 1068 } 1069 } 1070 1071 more = FALSE; 1072 if (!xdr_bool(xdrs, &more)) 1073 return (FALSE); 1074 1075 return (TRUE); 1076 } 1077 1078 /* 1079 * Server side RDMA WRITE list decode. 1080 * XDR context is memory ops 1081 */ 1082 bool_t 1083 xdr_decode_wlist_svc(XDR *xdrs, struct clist **wclp, bool_t *wwl, 1084 uint32_t *total_length, CONN *conn) 1085 { 1086 struct clist *first, *ncl; 1087 char *memp; 1088 uint32_t num_wclist; 1089 uint32_t wcl_length = 0; 1090 uint32_t i; 1091 bool_t more = FALSE; 1092 1093 *wclp = NULL; 1094 *wwl = FALSE; 1095 *total_length = 0; 1096 1097 if (!xdr_bool(xdrs, &more)) { 1098 return (FALSE); 1099 } 1100 1101 if (more == FALSE) { 1102 return (TRUE); 1103 } 1104 1105 *wwl = TRUE; 1106 1107 if (!xdr_uint32(xdrs, &num_wclist)) { 1108 DTRACE_PROBE(krpc__e__xdrrdma__wlistsvc__listlength); 1109 return (FALSE); 1110 } 1111 1112 first = ncl = clist_alloc(); 1113 1114 for (i = 0; i < num_wclist; i++) { 1115 1116 if (!xdr_uint32(xdrs, &ncl->c_dmemhandle.mrc_rmr)) 1117 goto err_out; 1118 if (!xdr_uint32(xdrs, &ncl->c_len)) 1119 goto err_out; 1120 if (!xdr_uint64(xdrs, &ncl->u.c_daddr)) 1121 goto err_out; 1122 1123 if (ncl->c_len > MAX_SVC_XFER_SIZE) { 1124 DTRACE_PROBE( 1125 krpc__e__xdrrdma__wlistsvc__chunklist_toobig); 1126 ncl->c_len = MAX_SVC_XFER_SIZE; 1127 } 1128 1129 DTRACE_PROBE1(krpc__i__xdr_decode_wlist_svc_len, 1130 uint_t, ncl->c_len); 1131 1132 wcl_length += ncl->c_len; 1133 1134 if (i < num_wclist - 1) { 1135 ncl->c_next = clist_alloc(); 1136 ncl = ncl->c_next; 1137 } 1138 } 1139 1140 if (!xdr_bool(xdrs, &more)) 1141 goto err_out; 1142 1143 first->rb_longbuf.type = RDMA_LONG_BUFFER; 1144 first->rb_longbuf.len = 1145 wcl_length > WCL_BUF_LEN ? wcl_length : WCL_BUF_LEN; 1146 1147 if (rdma_buf_alloc(conn, &first->rb_longbuf)) { 1148 clist_free(first); 1149 return (FALSE); 1150 } 1151 1152 memp = first->rb_longbuf.addr; 1153 1154 ncl = first; 1155 for (i = 0; i < num_wclist; i++) { 1156 ncl->w.c_saddr3 = (caddr_t)memp; 1157 memp += ncl->c_len; 1158 ncl = ncl->c_next; 1159 } 1160 1161 *wclp = first; 1162 *total_length = wcl_length; 1163 return (TRUE); 1164 1165 err_out: 1166 clist_free(first); 1167 return (FALSE); 1168 } 1169 1170 /* 1171 * XDR decode the long reply write chunk. 1172 */ 1173 bool_t 1174 xdr_decode_reply_wchunk(XDR *xdrs, struct clist **clist) 1175 { 1176 bool_t have_rchunk = FALSE; 1177 struct clist *first = NULL, *ncl = NULL; 1178 uint32_t num_wclist; 1179 uint32_t i; 1180 1181 if (!xdr_bool(xdrs, &have_rchunk)) 1182 return (FALSE); 1183 1184 if (have_rchunk == FALSE) 1185 return (TRUE); 1186 1187 if (!xdr_uint32(xdrs, &num_wclist)) { 1188 DTRACE_PROBE(krpc__e__xdrrdma__replywchunk__listlength); 1189 return (FALSE); 1190 } 1191 1192 if (num_wclist == 0) { 1193 return (FALSE); 1194 } 1195 1196 first = ncl = clist_alloc(); 1197 1198 for (i = 0; i < num_wclist; i++) { 1199 1200 if (i > 0) { 1201 ncl->c_next = clist_alloc(); 1202 ncl = ncl->c_next; 1203 } 1204 1205 if (!xdr_uint32(xdrs, &ncl->c_dmemhandle.mrc_rmr)) 1206 goto err_out; 1207 if (!xdr_uint32(xdrs, &ncl->c_len)) 1208 goto err_out; 1209 if (!xdr_uint64(xdrs, &ncl->u.c_daddr)) 1210 goto err_out; 1211 1212 if (ncl->c_len > MAX_SVC_XFER_SIZE) { 1213 DTRACE_PROBE( 1214 krpc__e__xdrrdma__replywchunk__chunklist_toobig); 1215 ncl->c_len = MAX_SVC_XFER_SIZE; 1216 } 1217 if (!(ncl->c_dmemhandle.mrc_rmr && 1218 (ncl->c_len > 0) && ncl->u.c_daddr)) 1219 DTRACE_PROBE( 1220 krpc__e__xdrrdma__replywchunk__invalid_segaddr); 1221 1222 DTRACE_PROBE1(krpc__i__xdr_decode_reply_wchunk_c_len, 1223 uint32_t, ncl->c_len); 1224 1225 } 1226 *clist = first; 1227 return (TRUE); 1228 1229 err_out: 1230 clist_free(first); 1231 return (FALSE); 1232 } 1233 1234 1235 bool_t 1236 xdr_encode_reply_wchunk(XDR *xdrs, 1237 struct clist *cl_longreply, uint32_t seg_array_len) 1238 { 1239 int i; 1240 bool_t long_reply_exists = TRUE; 1241 uint32_t length; 1242 uint64 offset; 1243 1244 if (seg_array_len > 0) { 1245 if (!xdr_bool(xdrs, &long_reply_exists)) 1246 return (FALSE); 1247 if (!xdr_uint32(xdrs, &seg_array_len)) 1248 return (FALSE); 1249 1250 for (i = 0; i < seg_array_len; i++) { 1251 if (!cl_longreply) 1252 return (FALSE); 1253 length = cl_longreply->c_len; 1254 offset = (uint64) cl_longreply->u.c_daddr; 1255 1256 DTRACE_PROBE1( 1257 krpc__i__xdr_encode_reply_wchunk_c_len, 1258 uint32_t, length); 1259 1260 if (!xdr_uint32(xdrs, 1261 &cl_longreply->c_dmemhandle.mrc_rmr)) 1262 return (FALSE); 1263 if (!xdr_uint32(xdrs, &length)) 1264 return (FALSE); 1265 if (!xdr_uint64(xdrs, &offset)) 1266 return (FALSE); 1267 cl_longreply = cl_longreply->c_next; 1268 } 1269 } else { 1270 long_reply_exists = FALSE; 1271 if (!xdr_bool(xdrs, &long_reply_exists)) 1272 return (FALSE); 1273 } 1274 return (TRUE); 1275 } 1276 bool_t 1277 xdrrdma_read_from_client(struct clist *rlist, CONN **conn, uint_t count) 1278 { 1279 struct clist *rdclist; 1280 struct clist cl; 1281 uint_t total_len = 0; 1282 uint32_t status; 1283 bool_t retval = TRUE; 1284 1285 rlist->rb_longbuf.type = RDMA_LONG_BUFFER; 1286 rlist->rb_longbuf.len = 1287 count > RCL_BUF_LEN ? count : RCL_BUF_LEN; 1288 1289 if (rdma_buf_alloc(*conn, &rlist->rb_longbuf)) { 1290 return (FALSE); 1291 } 1292 1293 /* 1294 * The entire buffer is registered with the first chunk. 1295 * Later chunks will use the same registered memory handle. 1296 */ 1297 1298 cl = *rlist; 1299 cl.c_next = NULL; 1300 if (clist_register(*conn, &cl, CLIST_REG_DST) != RDMA_SUCCESS) { 1301 rdma_buf_free(*conn, &rlist->rb_longbuf); 1302 DTRACE_PROBE( 1303 krpc__e__xdrrdma__readfromclient__clist__reg); 1304 return (FALSE); 1305 } 1306 1307 rlist->c_regtype = CLIST_REG_DST; 1308 rlist->c_dmemhandle = cl.c_dmemhandle; 1309 rlist->c_dsynchandle = cl.c_dsynchandle; 1310 1311 for (rdclist = rlist; 1312 rdclist != NULL; rdclist = rdclist->c_next) { 1313 total_len += rdclist->c_len; 1314 #if (defined(OBJ32)||defined(DEBUG32)) 1315 rdclist->u.c_daddr3 = 1316 (caddr_t)((char *)rlist->rb_longbuf.addr + 1317 (uint32) rdclist->u.c_daddr3); 1318 #else 1319 rdclist->u.c_daddr3 = 1320 (caddr_t)((char *)rlist->rb_longbuf.addr + 1321 (uint64) rdclist->u.c_daddr); 1322 1323 #endif 1324 cl = (*rdclist); 1325 cl.c_next = NULL; 1326 1327 /* 1328 * Use the same memory handle for all the chunks 1329 */ 1330 cl.c_dmemhandle = rlist->c_dmemhandle; 1331 cl.c_dsynchandle = rlist->c_dsynchandle; 1332 1333 1334 DTRACE_PROBE1(krpc__i__xdrrdma__readfromclient__buflen, 1335 int, rdclist->c_len); 1336 1337 /* 1338 * Now read the chunk in 1339 */ 1340 if (rdclist->c_next == NULL) { 1341 status = RDMA_READ(*conn, &cl, WAIT); 1342 } else { 1343 status = RDMA_READ(*conn, &cl, NOWAIT); 1344 } 1345 if (status != RDMA_SUCCESS) { 1346 DTRACE_PROBE( 1347 krpc__e__xdrrdma__readfromclient__readfailed); 1348 rdma_buf_free(*conn, &rlist->rb_longbuf); 1349 return (FALSE); 1350 } 1351 } 1352 1353 cl = (*rlist); 1354 cl.c_next = NULL; 1355 cl.c_len = total_len; 1356 if (clist_syncmem(*conn, &cl, CLIST_REG_DST) != RDMA_SUCCESS) { 1357 retval = FALSE; 1358 } 1359 return (retval); 1360 } 1361 1362 bool_t 1363 xdrrdma_free_clist(CONN *conn, struct clist *clp) 1364 { 1365 rdma_buf_free(conn, &clp->rb_longbuf); 1366 clist_free(clp); 1367 return (TRUE); 1368 } 1369 1370 bool_t 1371 xdrrdma_send_read_data(XDR *xdrs, uint_t data_len, struct clist *wcl) 1372 { 1373 int status; 1374 xrdma_private_t *xdrp = (xrdma_private_t *)(xdrs->x_private); 1375 struct xdr_ops *xops = xdrrdma_xops(); 1376 struct clist *tcl, *wrcl, *cl; 1377 struct clist fcl; 1378 int rndup_present, rnduplen; 1379 1380 rndup_present = 0; 1381 wrcl = NULL; 1382 1383 /* caller is doing a sizeof */ 1384 if (xdrs->x_ops != &xdrrdma_ops || xdrs->x_ops == xops) 1385 return (TRUE); 1386 1387 /* copy of the first chunk */ 1388 fcl = *wcl; 1389 fcl.c_next = NULL; 1390 1391 /* 1392 * The entire buffer is registered with the first chunk. 1393 * Later chunks will use the same registered memory handle. 1394 */ 1395 1396 status = clist_register(xdrp->xp_conn, &fcl, CLIST_REG_SOURCE); 1397 if (status != RDMA_SUCCESS) { 1398 return (FALSE); 1399 } 1400 1401 wcl->c_regtype = CLIST_REG_SOURCE; 1402 wcl->c_smemhandle = fcl.c_smemhandle; 1403 wcl->c_ssynchandle = fcl.c_ssynchandle; 1404 1405 /* 1406 * Only transfer the read data ignoring any trailing 1407 * roundup chunks. A bit of work, but it saves an 1408 * unnecessary extra RDMA_WRITE containing only 1409 * roundup bytes. 1410 */ 1411 1412 rnduplen = clist_len(wcl) - data_len; 1413 1414 if (rnduplen) { 1415 1416 tcl = wcl->c_next; 1417 1418 /* 1419 * Check if there is a trailing roundup chunk 1420 */ 1421 while (tcl) { 1422 if ((tcl->c_next == NULL) && (tcl->c_len == rnduplen)) { 1423 rndup_present = 1; 1424 break; 1425 } 1426 tcl = tcl->c_next; 1427 } 1428 1429 /* 1430 * Make a copy chunk list skipping the last chunk 1431 */ 1432 if (rndup_present) { 1433 cl = wcl; 1434 tcl = NULL; 1435 while (cl) { 1436 if (tcl == NULL) { 1437 tcl = clist_alloc(); 1438 wrcl = tcl; 1439 } else { 1440 tcl->c_next = clist_alloc(); 1441 tcl = tcl->c_next; 1442 } 1443 1444 *tcl = *cl; 1445 cl = cl->c_next; 1446 /* last chunk */ 1447 if (cl->c_next == NULL) 1448 break; 1449 } 1450 tcl->c_next = NULL; 1451 } 1452 } 1453 1454 if (wrcl == NULL) { 1455 /* No roundup chunks */ 1456 wrcl = wcl; 1457 } 1458 1459 /* 1460 * Set the registered memory handles for the 1461 * rest of the chunks same as the first chunk. 1462 */ 1463 tcl = wrcl->c_next; 1464 while (tcl) { 1465 tcl->c_smemhandle = fcl.c_smemhandle; 1466 tcl->c_ssynchandle = fcl.c_ssynchandle; 1467 tcl = tcl->c_next; 1468 } 1469 1470 /* 1471 * Sync the total len beginning from the first chunk. 1472 */ 1473 fcl.c_len = clist_len(wrcl); 1474 status = clist_syncmem(xdrp->xp_conn, &fcl, CLIST_REG_SOURCE); 1475 if (status != RDMA_SUCCESS) { 1476 return (FALSE); 1477 } 1478 1479 status = RDMA_WRITE(xdrp->xp_conn, wrcl, WAIT); 1480 1481 if (rndup_present) 1482 clist_free(wrcl); 1483 1484 if (status != RDMA_SUCCESS) { 1485 return (FALSE); 1486 } 1487 1488 return (TRUE); 1489 } 1490 1491 1492 /* 1493 * Reads one chunk at a time 1494 */ 1495 1496 static bool_t 1497 xdrrdma_read_a_chunk(XDR *xdrs, CONN **conn) 1498 { 1499 int status; 1500 int32_t len = 0; 1501 xrdma_private_t *xdrp = (xrdma_private_t *)(xdrs->x_private); 1502 struct clist *cle = *(xdrp->xp_rcl_next); 1503 struct clist *rclp = xdrp->xp_rcl; 1504 struct clist *clp; 1505 1506 /* 1507 * len is used later to decide xdr offset in 1508 * the chunk factoring any 4-byte XDR alignment 1509 * (See read chunk example top of this file) 1510 */ 1511 while (rclp != cle) { 1512 len += rclp->c_len; 1513 rclp = rclp->c_next; 1514 } 1515 1516 len = RNDUP(len) - len; 1517 1518 ASSERT(xdrs->x_handy <= 0); 1519 1520 /* 1521 * If this is the first chunk to contain the RPC 1522 * message set xp_off to the xdr offset of the 1523 * inline message. 1524 */ 1525 if (xdrp->xp_off == 0) 1526 xdrp->xp_off = (xdrp->xp_offp - xdrs->x_base); 1527 1528 if (cle == NULL || (cle->c_xdroff != xdrp->xp_off)) 1529 return (FALSE); 1530 1531 /* 1532 * Make a copy of the chunk to read from client. 1533 * Chunks are read on demand, so read only one 1534 * for now. 1535 */ 1536 1537 rclp = clist_alloc(); 1538 *rclp = *cle; 1539 rclp->c_next = NULL; 1540 1541 xdrp->xp_rcl_next = &cle->c_next; 1542 1543 /* 1544 * If there is a roundup present, then skip those 1545 * bytes when reading. 1546 */ 1547 if (len) { 1548 rclp->w.c_saddr = 1549 (uint64)(uintptr_t)rclp->w.c_saddr + len; 1550 rclp->c_len = rclp->c_len - len; 1551 } 1552 1553 status = xdrrdma_read_from_client(rclp, conn, rclp->c_len); 1554 1555 if (status == FALSE) { 1556 clist_free(rclp); 1557 return (status); 1558 } 1559 1560 xdrp->xp_offp = rclp->rb_longbuf.addr; 1561 xdrs->x_base = xdrp->xp_offp; 1562 xdrs->x_handy = rclp->c_len; 1563 1564 /* 1565 * This copy of read chunks containing the XDR 1566 * message is freed later in xdrrdma_destroy() 1567 */ 1568 1569 if (xdrp->xp_rcl_xdr) { 1570 /* Add the chunk to end of the list */ 1571 clp = xdrp->xp_rcl_xdr; 1572 while (clp->c_next != NULL) 1573 clp = clp->c_next; 1574 clp->c_next = rclp; 1575 } else { 1576 xdrp->xp_rcl_xdr = rclp; 1577 } 1578 return (TRUE); 1579 } 1580 1581 static void 1582 xdrrdma_free_xdr_chunks(CONN *conn, struct clist *xdr_rcl) 1583 { 1584 struct clist *cl; 1585 1586 (void) clist_deregister(conn, xdr_rcl); 1587 1588 /* 1589 * Read chunks containing parts XDR message are 1590 * special: in case of multiple chunks each has 1591 * its own buffer. 1592 */ 1593 1594 cl = xdr_rcl; 1595 while (cl) { 1596 rdma_buf_free(conn, &cl->rb_longbuf); 1597 cl = cl->c_next; 1598 } 1599 1600 clist_free(xdr_rcl); 1601 } 1602